Split (1)

train · 7.63k rows

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
Locating What You Need: Towards Adapting Diffusion Models to OOD Concepts In-the-Wild	Accept (poster)	Summary: The paper introduces CATOD, a framework designed to enhance the adaptation of text-to-image models for out-of-distribution (OOD) concepts. It addresses the issue of low-quality training data by employing an active learning approach that iteratively improves the training set. The framework utilizes a scoring system comprising aesthetic and concept-matching scores to guide data selection and training. This scoring system dynamically determines the relative weight of the aesthetic score and content matching score to prioritize samples for selecting the training set. The authors demonstrate significant performance improvements using CATOD, achieving notable improvements on standard metrics. Strengths: 1. The paper is well-written and easy to follow. 2. The research problem of out-of-distribution (OOD) on text-to-image models is essential, and the motivation is quite clear. 3. The approach is innovative, which combines active learning with a novel scoring system to enhance model adaptation. 4. The paper has very adequate quantitative experimental results and show clear improvements over baseline methods. Weaknesses: 1. The paper does not provide enough detail to clarify how the CMMD metric differs from other commonly-used metrics for image generation, such as Inception Score and FID. The authors should further explain this. 2. Since your proposed CATOD is trained on few samples (100 for single-concept and 200 for multi-concept), the generation model may be biased toward specific samples. Reporting diversity metrics of the generated samples can help clarify this issue. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. I wonder if your proposed CATOD also outperforms other methods on previous metrics such as FID and the Inception Score. 2. In the paper, random sampling and CLIP-score-based sampling are used as the comparison method, and finally 100-200 images are selected for training. Would it be better to fine-tune the model directly using all 1000 images? 3. In addition to the three methods mentioned in the paper, are there other custom content generation methods that your proposed CATOD framework is still applicable to, such as LyCORIS[1] and ELITE[2]? [1] YEH, SHIH-YING, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. "Navigating text-to-image customization: From lycoris fine-tuning to model evaluation." In The Twelfth International Conference on Learning Representations. 2023. [2] Wei, Yuxiang, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. "Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation." In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 15943-15953. 2023. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank for the good words! We are happy that you enjoyed the paper! Part 1: Why CMMD is a good metric (W1, Q1) As reported in many recent works [1,2,3], the most popular image-matching metrics like Inception Score, Precision/Recall, and FID (Frechet Inception Distance) may disagree with human raters, thus ill-suited for evaluating recent text-to-image models. In comparison, CMMD uses CLIP embeddings and Maximum Mean Discrepancy (MMD) that correctly ranks the image sets based on the severity of the distortion. The limitations of FID and previous metrics can be listed as follows: \| \| FID and other previous metrics \| MMD distance \| \| -------------------- \| ------------------------------------------------------------ \| ------------------------------------------------------------ \| \| Inception embeddings \| × Weak image embeddings<br />× Normality assumption<br />× Sample inefficient<br />× Biased estimator \| × Weak image embeddings<br />√ Distribution-free<br />√ Sample inefficient<br />√ Unbiased estimator \| \| CLIP embeddings \| √ Rich image embeddings<br />× Normality assumption<br />× Sample inefficient<br />× Biased estimator \| √ Rich image embeddings<br />√ Distribution-free<br />√ Sample inefficient<br />√ Unbiased estimator \| Also, we are glad to report our experimental results with FID, Precision, and Recall as follows based on CATOD adapted through LoRA: \| Insect \| FID \| Precision \| Recall \| \| ------------ \| ----- \| --------- \| ------ \| \| LoRA + RAND \| 35.02 \| 0.55 \| 0.83 \| \| LoRA + CLIP \| 28.46 \| 0.85 \| 0.91 \| \| LoRA + CATOD \| 26.74 \| 0.92 \| 0.94 \| \| Penguin \| FID \| Precision \| Recall \| \| ------------ \| ----- \| --------- \| ------ \| \| LoRA + RAND \| 11.17 \| 0.63 \| 0.79 \| \| LoRA + CLIP \| 5.22 \| 0.80 \| 0.87 \| \| LoRA + CATOD \| 4.03 \| 0.84 \| 0.88 \| [1] Jayasumana, Sadeep, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. "Rethinking fid: Towards a better evaluation metric for image generation." arXiv preprint arXiv:2401.09603 (2023). [2] Grimal, Paul, Hervé Le Borgne, Olivier Ferret, and Julien Tourille. "TIAM-A Metric for Evaluating Alignment in Text-to-Image Generation." In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 2890-2899. 2024. [3] Lee, Tony, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, et al. "Holistic evaluation of text-to-image models." Advances in Neural Information Processing Systems 36 (2024). Part 2: About the diversity evaluation (W2) CATOD maintains the diversity to produce OOD concepts with different angles or poses in generative results. To validate the diversity of our generative results, we further provide a quantitative evaluation with LPIPS~[1,2]. The results are shown in the Table below: \| Comparison Methods \| Insect \| Lizard \| Penguin \| Seafish \| Snake \| \| ------------------ \| --------- \| --------- \| --------- \| --------- \| --------- \| \| DreamBooth + CLIP \| 0.254 \| 0.156 \| 0.149 \| 0.305 \| 0.265 \| \| DreamBooth + CATOD \| 0.362 \| 0.198 \| 0.197 \| 0.349 \| 0.286 \| \| TI + CLIP \| 0.198 \| 0.257 \| 0.174 \| 0.108 \| 0.208 \| \| TI + CATOD \| 0.267 \| 0.278 \| 0.152 \| 0.278 \| 0.133 \| \| LoRA + CLIP \| 0.178 \| 0.184 \| 0.155 \| 0.203 \| 0.094 \| \| LoRA + CATOD \| 0.245 \| 0.203 \| 0.217 \| 0.244 \| 0.178 \| In this table, we can see that CATOD gives out a diversity improvement up to 0.17 in the LPIPS score compared to CLIP-based sampling, and outperforms CLIP over most categories. From this, we conclude that CATOD also preserves diversity. Since the training set contains samples with different angles, it is also easy to produce objects with different angles. We will surely add a visualization of our selected samples and generative results with different poses/angles in our final revision. --- Rebuttal 2: Comment: Part 3: Using 1000 data is not feasible (Q2) Using all the 1000 data for training is not feasible for adapting OOD concepts. The experimental results of using all data are as follows: \| Combinations \| Axolotl (CLIP$\uparrow$) \| Axolotl (CMMD$\downarrow$) \| Emperor Penguin Chick (CLIP$\uparrow$) \| Emperor Penguin Chick (CMMD$\downarrow$) \| \| --------------------- \| ------------------------ \| -------------------------- \| -------------------------------------- \| ---------------------------------------- \| \| DreamBooth + RAND \| 66.19 \| 1.13 \| 67.34 \| 1.54 \| \| DreamBooth + ALL DATA \| 68.24 \| 1.01 \| 70.12 \| 1.23 \| \| DreamBooth + CATOD \| 72.25 \| 0.88 \| 74.38 \| 0.83 \| \| TI + RAND \| 59.67 \| 2.11 \| 62.89 \| 2.10 \| \| TI + ALL DATA \| 52.79 \| 3.04 \| 55.63 \| 2.67 \| \| TI + CATOD \| 69.95 \| 1.49 \| 68.14 \| 1.15 \| \| LoRA + RAND \| 65.35 \| 1.41 \| 68.23 \| 1.58 \| \| LoRA + ALL DATA \| 67.24 \| 1.47 \| 71.17 \| 1.33 \| \| LoRA + CATOD \| 72.87 \| 1.37 \| 73.19 \| 1.13 \| As we see in this table, performing adaptation over the full dataset can lead to severe performance loss compared to using carefully selected ones. On Textual Inversion (TI), it even becomes worse than random sampling. This is due to that the full dataset contains lots of "bad" samples, which exhibit too many disruptive elements -- in turn -- introducing wrong details that mislead the concept adaptation. Part 4: Generalizing CATOD to other customization methods. (Q3) Thanks for the advice! CATOD can be easily generalized to ELITE and LyCORIS since it is designed as a general framework for the adapters. We show the results for applying CATOD to ELITE/LyCORIS on concepts "axolotl" and "emperor penguin chick" as follows: \| Combinations \| Axolotl (CLIP) \| Axolotl (CMMD) \| Emperor Penguin Chick (CLIP) \| Emperor Penguin Chick (CMMD) \| \| --------------- \| -------------- \| -------------- \| ---------------------------- \| ---------------------------- \| \| LoRA + CLIP \| 72.87 \| 1.37 \| 73.19 \| 1.13 \| \| ELITE + CLIP \| 74.14 \| 1.03 \| 74.37 \| 1.01 \| \| LyCORIS + CLIP \| 72.95 \| 1.18 \| 74.08 \| 1.08 \| \| LoRA + CATOD \| 75.35 \| 0.85 \| 74.89 \| 0.79 \| \| ELITE + CATOD \| 75.68 \| 0.83 \| 74.68 \| 0.74 \| \| LyCORIS + CATOD \| 76.04 \| 0.79 \| 75.05 \| 0.73 \| From these results, we can also see that CATOD gives a consistent boost to ELITE and LyCORIS compared to other sampling strategies. Despite that CATOD is compatible with ELITE and LyCORIS, they do not exhibit a notable performance gain compared to LoRA.	Summary: This paper introduces a novel framework called Controllable Adaptor Towards Out-of-Distribution (OOD) Concepts (CATOD). CATOD is designed to adapt text-to-image models to OOD concepts and generate high-quality images of those OOD concepts accordingly. The authors identified the challenge of accurately depicting OOD concepts due to low-quality training data. CATOD employs an active learning approach to iteratively accumulate high-quality training data and update the adaptor. The framework incorporates an aesthetic score and a concept-matching score to guide the accumulation of training data. Extensive experiments demonstrate significant improvements in both the CLIP score and the CMMD metric. Strengths: 1. The presentation of the paper is clear and concise. The method is easy to understand and the delivery of the paper is smooth. 2. The paper addresses a significant challenge in text-to-image generation by focusing on OOD concepts, and serves as the first work to address the OOD challenge from a data-centric perspective. 3. The method is well-motivated and supported by theoretical insights into the importance of aesthetics and concept-matching scores. 4. The experimental results demonstrate significant improvements over existing methods. Weaknesses: 1. To support the claim in Line 189 that your selection "guarantees the sample diversity", the authors should report the diversity metrics of the generated samples. 2. The paper provides ablation results on the design of weighted scoring strategies, which is valuable. I think more extensive ablation studies on the impact of initial data size, the type of text-to-image encoders, etc. can help signify the paper's contribution. 3. To further investigate how the weighting strategy work in your proposed CATOD, it would better to list how the weights of aesthetic/concept-matching scores change as your learning cycle proceeds over different concepts. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Following weakness 1, can you compare the diversity of the fine-tuned model under different strategies? 2. It would be intriguing to investigate how initial data size impact the experiments, especially for the multiple concept scenario as you mentioned in Sec. 5.2 following weakness 2. 3. For different concepts, which of the two scores gets high weightage (in equation 11) as the training progresses? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions! We hope that our rebuttal addresses your concerns. Part 1: About the diversity evaluation (W1, Q1) CATOD maintains the diversity to produce OOD concepts with different angles or poses in generative results. To validate the diversity of our generative results, we further provide a quantitative evaluation with LPIPS~[1,2]. The results are shown in the Table below: \| Comparison Methods \| Insect \| Lizard \| Penguin \| Seafish \| Snake \| \| ------------------ \| --------- \| --------- \| --------- \| --------- \| --------- \| \| DreamBooth + CLIP \| 0.254 \| 0.156 \| 0.149 \| 0.305 \| 0.265 \| \| DreamBooth + CATOD \| 0.362 \| 0.198 \| 0.197 \| 0.349 \| 0.286 \| \| TI + CLIP \| 0.198 \| 0.257 \| 0.174 \| 0.108 \| 0.208 \| \| TI + CATOD \| 0.267 \| 0.278 \| 0.152 \| 0.278 \| 0.133 \| \| LoRA + CLIP \| 0.178 \| 0.184 \| 0.155 \| 0.203 \| 0.094 \| \| LoRA + CATOD \| 0.245 \| 0.203 \| 0.217 \| 0.244 \| 0.178 \| In this table, we can see that CATOD gives out a diversity improvement up to 0.17 in the LPIPS score compared to CLIP-based sampling, and outperforms CLIP over most categories. From this, we conclude that CATOD also preserves diversity. Since the training set contains samples with different angles, it is also easy to produce objects with different angles. We will surely add a visualization of our selected samples and generative results with different poses/angles in our final revision. Part 2: Ablating initial size/quality in CATOD (W2, Q2) It is intriguing to ablate CATOD over the initial training set size and quality! In detail, we retest CATOD with different numbers of initial samples (10,20,50) and different quality (high-quality and random sampling), with 100 images sampled at last and 10 images selected per cycle. These experiments are tested with adaptor LoRA. The results are shown as follows: \| Initial Setting \| Axolotl (CLIP) \| Axolotl (CMMD) \| Emperor Penguin Chick (CLIP) \| Emperor Penguin Chick (CMMD) \| \| ----------------------- \| -------------- \| -------------- \| ---------------------------- \| ---------------------------- \| \| 10 HQ initial samples \| 73.99 \| 0.84 \| 74.97 \| 0.71 \| \| 10 RAND initial samples \| 72.89 \| 0.93 \| 73.95 \| 0.75 \| \| 20 HQ initial samples \| 75.35 \| 0.82 \| 74.89 \| 0.69 \| \| 20 RAND initial samples \| 72.27 \| 1.10 \| 72.38 \| 0.73 \| \| 50 HQ initial samples \| 75.37 \| 0.81 \| 74.56 \| 0.63 \| \| 50 RAND initial samples \| 70.39 \| 1.34 \| 68.85 \| 1.12 \| From these results, we may draw the following conclusions: 1. With HQ initial samples, the initial batch size does not have a significant compact, since a change in this size is not more than 0.3 in the CLIP score and 0.08 in the MMD score. However, this initial size has a significant impact with randomly initialized samples with at most a 5.71 loss in CLIP score and 0.49 in CMMD. We account this phenomenon for that randomly initialized images contain more bad-quality ones that mislead the adaptation. 2. The quality of initial samples does have an impact on generative results since we can see a consistent performance loss when changing HQ initial samples to randomly initialized samples. With the initial size increase, this impact tends to be even more significant. The full results will be added in our revisions. Part 3: Ablating the text-to-image encoder (W2) Thanks for the advice! It is straightforward to extend our experiments to other text-to-image encoders (like SD 1.5 and SDXL). We show some of these results with LoRA in the tables below: \| T2I Model Ablation \| Axolotl (CLIP) \| Axolotl (CMMD) \| Emperor Penguin Chick (CLIP) \| Emperor Penguin Chick (CMMD) \| \| ------------------------- \| -------------- \| --------------- \| ---------------------------- \| ---------------------------- \| \| CATOD + SD 2.0 (original) \| 75.35 \| 0.85 \| 74.89 \| 0.79 \| \| CATOD + SD 1.5 \| 74.79 \| 0.88 \| 73.01 \| 0.92 \| \| CATOD + SDXL \| 79.36 \| 0.72 \| 80.37 \| 0.66 \| This table shows that our proposed CATOD is also compatible with different text-to-image encoders with notable performance. We will add these ablations to our final revisions. Part 4: How the weight of different scores changes (W3, Q3) Following Eq. (11) in the main context, the overall aesthetic score $\gamma_{aes}(A)$, overall concept matching score $\gamma_{con}(A)$ and the corresponding weights on concept "emperor penguin chick" can be listed as follows: \| Cycle \| 1 \| 2 \| 3 \| 4 \| 5 \| \| ----------------------- \| ----- \| ----- \| ----- \| ----- \| ----- \| \| $\gamma_{aes}(A)$ \| 6.26 \| 8.57 \| 9.03 \| 9.11 \| 9.22 \| \| aesthetic weight \| 0.374 \| 0.143 \| 0.097 \| 0.099 \| 0.078 \| \| $\gamma_{con}(A)$ \| 1.86 \| 8.59 \| 8.54 \| 9.35 \| 9.67 \| \| concept-matching weight \| 0.814 \| 0.141 \| 0.146 \| 0.065 \| 0.033 \| During the first three cycles, concept-matching takes precedence over aesthetics because the OOD concepts are not yet fully learned by the adapter. As the cycles progress, concept-matching is quickly achieved. In the last two cycles, the adapter stabilizes and focuses more on enhancing aesthetics.	Summary: This work tackles the challenge of adapting text-to-image diffusion models to out-of-distribution (OOD) concepts. The authors introduce a framework called Controllable Adaptor Towards Out-of-Distribution Concepts (CATOD), which employs an active learning paradigm to iteratively accumulate high-quality training data and optimize adaptor training. The framework features a dual scoring system that balances aesthetics and concept matching, ensuring improved quality and conciseness of the generated images. Strengths: 1. This paper is easy to follow due to its clear writing, concise figures, and well-structured formulation of problem definitions, methods, and theories. 2. The motivation of this paper is innovative, addressing the overlooked problem of adapting recent text-to-image (T2I) models to out-of-distribution concepts. 3. The paper presents sufficient experimental results that effectively support its claims. Weaknesses: 1. The key contribution of this paper is Equation (6) and (7), which are claimed to be use as a signal to reduce the learning rate, stop the training in time, and select samples for the next cycle. It seems the innovation is that the authors identify two relevant factors and utilize the strategy of active learning. Am I correct? 2. Authors use MMSE to reinterpret the loss function of the LDM and decompose it into two terms (Eq. 10). As the description in lines 213 to 220 of the paper, associating these two terms directly with aesthetic preservation and concept-matching score is rather far-fetched and vague. Kindly advise if I am wrong. 3. If the aesthetic and concept-matching scores are really that important, why just use those two scores to pick suitable samples for adaptation from the beginning? This would also avoid the later calculations. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Is the CMMD shown in Figure 1 calculated using Real Images and before adapting images? 2. Are the high-quality samples shown in Figure 2 generated by vanilla Textual Inversion, DreamBooth, and LoRA? Or are they already using the CATOD framework? 3. Is “if r_k is the closet representation for r_k” in line 145 of the paper correct? 4. Eq. 6 and the description in lines 173 to 174 of the paper confuse me. Shouldn’t the maximum value of \gamma (A) be \gamma_aes (A)=\gamma_con (A)=10+40k, k=0...n, and the minimum value be \gamma_aes (A)=30+40k, or \gamma_con (A)=30+40k, k=0...n? 5. Associating the two terms of Eq. 10 directly with aesthetic and concept-matching scores requires further theoretical explanation and empirical analysis. 6. I noticed in the experiment section that the training period is about 20 epochs. Would this somehow be too less? If there were enough training cycles, we would get an overfitted LDM capable of generating images containing specific OOD concepts. 7. In Algorithm 1, lines 9 to 19, it does not look like new training samples are being selected for each epoch. 8. Does the learning rate set R determine the stopping of the training? How would you guarantee that the model has converged at this point? As question 5, a sufficiently overfitted LDM should be able to generate these OOD concepts. 9. This paper tests the proposed CATOD on SD 2.0. Have you conducted experiments to evaluate its performance on other text-to-image models, such as SD 1.5 and SDXL? Including these results would help verify the versatility of your method. 10. I am curious about how CATOD performs with different initial training set sizes (e.g., 10 and 50) and varying quality. Adding this as an ablation experiment would provide valuable insights into the robustness of your method. 11. I am also interested in whether similar categories (e.g., other kinds of penguins) might be distorted after adaptation. Have you conducted any experiments to evaluate the degree of distortion in these cases? 12. Since CMMD is a newly proposed metric by Google in CVPR 2024, it would be helpful to list the advantages of using CMMD compared to traditional metrics like Inception Score, FID, and MMD. This would help readers understand the rationale behind choosing CMMD. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: n/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions. We have carefully addressed your concerns as follows: Part 1: About the key technical contribution (W1) The key technical contributions of this paper are twofold: (1) our method for estimating the impact of each sample on the model without training, corresponding to the aesthetic/concept-matching scorer described in Section 3.2; (2) our approach to effectively utilizing these scores and balancing the trade-off between the two impact factors (Eq. 6 and Eq. 7). This is achieved through Active Learning, a vital paradigm that enables dynamic trade-offs cycle-by-cycle. Part 2: A more detailed explanation for the theories (W2, Q5) We are pleased to provide a detailed explanation of the connection between our proposed method and the underlying theories! As introduced in our background Section (Sec.2), the LDMs consist of two core components: An encoder $\mathcal{E}$ learns to map images $x$ to a latent code $z=\mathcal{E}(x)$, and a diffusion model trained to produce code conditioned on texts. Let $c\_\theta(y)$ be a model mapping the conditioning text $y$ into latent vectors, $A$ the adapter embedded into the diffusion model, the LDM loss is given by: $L\_{LDM}(x,A)= \mathbb{E}\_z\sim\mathcal{E}(x),t,\epsilon\sim\mathcal{N}(0,1)\left[\Vert \epsilon-\epsilon\_{\theta, A}(z\_t,t,c\_{\theta, A}(y)) \Vert\_2\^2\right]$ in Eq.1, where t is the time step, $z\_t$ the latent noised to time $t$, $\epsilon$ the noise sample, $\epsilon\_\theta$ the denoising network. When optimizing the LDM loss, there are two important factors we should consider: the adapter $A$ that adapts the underly model to OOD image $x$, and the conditional text $y$ for the images $x$ to match. As we claimed in the main context that disrupting $x$'s can lead to misunderstanding of concept $y$, our objective is in Eq.2: $A^,X\_{T}^=\mbox{argmin}\_{A,X\_T} L\_{\mbox{LDM}}(x,A,y)$. Since the optimal set for $A, X_T$ is initially unknown and cannot located at once, is common to use an iterative paradigm that alternatively optimizes $A$ and $X_T$. We have the following conjugate forms: $\mathbf{X}\_T\^{(t)} = \mbox{argmin}\_{\mathbf{X}\_T\subset D\_{pool}}\mathbb{E}\_{x\sim \mathbf{X}\_T}L\_{LDM}(x,A\^{(t-1)},y)$ (Eq.3 in main context) $A\^{(t)} = \mbox{argmin}\_{A}\mathbb{E}\_{x\sim \mathbf{X}\_T\^{(t-1)}}L\_{LDM}(x,A,y)$ (Eq.4 in main context) For the first equation (Eq.3), we optimize the training $X_T$ towards both conditional text $y$ (image-text matching) and loss reduction (aesthetic preserving), which in turn leads to the motivation of decoupling aesthetic/concept-matching scores. Theorem 4.3 verifies the necessity of this decomposition while Section 3.2 implements this decomposition. The second equation (Eq.4) just performs an ordinary adaptation on select samples. Since the iterative paradigm in Eq. 3 and Eq. 4 may lead to convergence to local optima, we also need a dynamic weighted scoring system to trade-off between aesthetics and concept-matching, to alleviate the potential bias towards only one of them (leading to the practical implementation in Sec 3.3). To this end, we have explained the motivation of our theory and how we organize our CATOD. Part 3: The pre-calculated scores are not feasible (W3) Since Out-of-Distribution (OOD) concepts are usually unseen by both the underlying text-to-model and the aesthetic/concept-matching scorers, it is likely that the preset scorers provide inaccurate results. To address this, we must use the currently available high-quality samples to correct these inaccuracies. Following the iterative paradigm outlined in Part 2, the scorers and adaptors are alternately adjusted to ensure accurate results. Additionally, we update the scorers after selecting high-quality samples in each cycle. To demonstrate this experimentally, we conduct another ablation experiment with pre-calculated scores, as shown in the following table: \| Method \| Lizard (CLIP$\uparrow$) \| penguin(CLIP$\uparrow$) \| Lizard (CMMD$\downarrow$) \| penguin(CMMD$\downarrow$) \| \| -------------------------------- \| ----------------------- \| ----------------------- \| ------------------------- \| ------------------------- \| \| CATOD \| 77.00 \| 74.11 \| 0.89 \| 0.71 \| \| CATOD with pre-calculated scores \| 75.29 \| 73.57 \| 0.93 \| 0.76 \| We observe a noticeable performance loss when the scores are pre-calculated. Part 4: Other architectures (Q9). Thanks for the advice! It is straightforward to extend our experiments to other text-to-image models (like SD 1.5 and SDXL). We show some of these results with LoRA in the tables below: \| T2I Model Ablation \| Axolotl (CLIP) \| Axolotl (CMMD) \| Emperor Penguin Chick (CLIP) \| Emperor Penguin Chick (CMMD) \| \| ------------------------- \| -------------- \| --------------- \| ---------------------------- \| ---------------------------- \| \| CATOD + SD 2.0 (original) \| 75.35 \| 0.85 \| 74.89 \| 0.79 \| \| CATOD + SD 1.5 \| 74.79 \| 0.88 \| 73.01 \| 0.92 \| \| CATOD + SDXL \| 79.36 \| 0.72 \| 80.37 \| 0.66 \| This table shows that our proposed CATOD is also compatible with different text-to-image encoders with notable performance. We will add these ablations to our final revisions. --- Rebuttal Comment 1.1: Title: Post-rebuttal comments. Comment: The authors have addressed my concerns. I keep positive attitude towards this paper and increase my score by one. --- Rebuttal 2: Comment: Part 5: The impact of initial size/quality in CATOD (Q10) Thanks for the advice! It is intriguing to ablate CATOD over the initial training set size/quality. In detail, we retest CATOD with different numbers of initial samples (10,20,50) and different quality (high-quality and random sampling), with 100 images sampled at last and 10 images selected per cycle. These experiments are tested with adaptor LoRA. The results are shown as follows: \| Initial Setting \| Axolotl (CLIP) \| Axolotl (CMMD) \| Emperor Penguin Chick (CLIP) \| Emperor Penguin Chick (CMMD) \| \| ----------------------- \| -------------- \| -------------- \| ---------------------------- \| ---------------------------- \| \| 10 HQ initial samples \| 73.99 \| 0.84 \| 74.97 \| 0.71 \| \| 10 RAND initial samples \| 72.89 \| 0.93 \| 73.95 \| 0.75 \| \| 20 HQ initial samples \| 75.35 \| 0.82 \| 74.89 \| 0.69 \| \| 20 RAND initial samples \| 72.27 \| 1.10 \| 72.38 \| 0.73 \| \| 50 HQ initial samples \| 75.37 \| 0.81 \| 74.56 \| 0.63 \| \| 50 RAND initial samples \| 70.39 \| 1.34 \| 68.85 \| 1.12 \| From these results, we may draw the following conclusions: 1. With HQ initial samples, the initial batch size does not have a significant compact, since a change in this size is not more than 0.3 in the CLIP score and 0.08 in the MMD score. However, this initial size has a significant impact with randomly initialized samples with at most a 5.71 loss in CLIP score and 0.49 in CMMD. We account this phenomenon for that randomly initialized images contain more bad-quality ones that mislead the adaptation. 2. The quality of initial samples does have an impact on generative results since we can see a consistent performance loss when changing HQ initial samples to randomly initialized samples. With the initial size increase, this impact tends to be even more significant. The full results will be added in our revisions. Part 6: Evaluating the quality of ID concepts after adapting OOD concepts. (Q11) In brief, our method will not cause much degradation in non-OOD concepts. To prove this, we list a comparison of the performance of the following non-OOD concepts (corresponding to Q1) over SD 2.0 when fine-tuning with LoRA. The results are shown in the table below: \| Insect \| Thrips \| Flea Beetle \| Aphids \| Red Spider \| Meadow Moth \| \| ----------------- \| ------ \| ----------- \| ------ \| ---------- \| ----------- \| \| CLIP (before) \| 75.35 \| 72.76 \| 75.59 \| 74.27 \| 73.93 \| \| CLIP (after) \| 75.17 \| 71.89 \| 75.76 \| 74.25 \| 73.78 \| \| CMMD (before) \| 2.05 \| 1.28 \| 1.16 \| 1.85 \| 1.47 \| \| CMMD (after) \| 2.08 \| 1.32 \| 1.30 \| 1.93 \| 1.49 \| \| Penguin \| Emperor Penguin \| King Penguin \| Little penguin \| Magellanic Penguin \| Adelie Penguin \| \| ----------------- \| --------------- \| ------------ \| -------------- \| ------------------ \| -------------- \| \| CLIP (before) \| 76.89 \| 79.23 \| 80.45 \| 75.45 \| 73.68 \| \| CLIP (after) \| 76.56 \| 78.97 \| 79.78 \| 75.42 \| 73.55 \| \| CMMD (before) \| 1.19 \| 0.75 \| 1.46 \| 1.60 \| 0.77 \| \| CMMD (after) \| 1.26 \| 0.80 \| 1.58 \| 1.61 \| 0.86 \| From this table, we can see that our proposed CATOD only leads to a performance loss of at most 0.15 on the CLIP score and at most 0.12 on CMMD on the non-OOD concepts after fine-tuning. This proves that our method does not distort non-OOD concepts when adapting native models to OOD concepts. --- Rebuttal 3: Comment: Part 7: Why CMMD is a good metric (Q12) As reported in many recent works [1,2,3], most popular image matching metrics like Inception Score, Precision/Recall and FID (Frechet Inception Distance) may disagree with human raters, thus ill-suited for evaluating recent text-to-image models. In comparison, CMMD uses CLIP embeddings and Maximum Mean Discerpancy (MMD) that correctly ranks the image sets based on the severity of the distortion. The limitations of FID and previous metrics can be listed as follows: \| \| FID and other previous metrics \| MMD distance \| \| -------------------- \| ------------------------------------------------------------ \| ------------------------------------------------------------ \| \| Inception embeddings \| × Weak image embeddings<br />× Normality assumption<br />× Sample inefficient<br />× Biased estimator \| × Weak image embeddings<br />√ Distribution-free<br />√ Sample inefficient<br />√ Unbiased estimator \| \| CLIP embeddings \| √ Rich image embeddings<br />× Normality assumption<br />× Sample inefficient<br />× Biased estimator \| √ Rich image embeddings<br />√ Distribution-free<br />√ Sample inefficient<br />√ Unbiased estimator \| [1] Jayasumana, Sadeep, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. "Rethinking fid: Towards a better evaluation metric for image generation." arXiv preprint arXiv:2401.09603 (2023). [2] Grimal, Paul, Hervé Le Borgne, Olivier Ferret, and Julien Tourille. "TIAM-A Metric for Evaluating Alignment in Text-to-Image Generation." In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 2890-2899. 2024. [3] Lee, Tony, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang et al. "Holistic evaluation of text-to-image models." Advances in Neural Information Processing Systems 36 (2024). Part 8: Other minor questions Q1 (About Figure 1): Yes, the CMMD in Figure 1 is calculated before adapting. This is used to show how out-of-distribution a concept is before adapting. Q2 (About Figure 2): Yes, the high-quality samples in the third part of Figure 2 are selected by the CATOD framework. Q4 (About Eq. 6): Apologies for the confusion! Since both the aesthetic and concept-matching scores range from 0 to 10, Equation 6 monotonically increases according to both variables, with both sine functions ranging from 0 to 1. We will ensure that the scores $\gamma_{aes}$ and $\gamma_{con}$ are clearly stated to range from 0 to 10 in future revisions. Q6 (About the epochs to use): Training for 20 epochs is common for adaptors such as LoRA, DreamBooth, and other related methods, as they typically require much less data (usually around 100 samples per concept). Training for too many epochs can lead to the underlying model replicating the images as exact copies. Q7 (About algorithm 1): We don't need to select new samples at each training epoch. The selection is performed in lines 2-6, prior to training. The algorithm demonstrates the process of a single Active Learning cycle, while training CATOD involves several cycles. The selection is conducted at the beginning of each cycle. Q8 (About the schedule): R includes 5 learning rates: $5\times 10^{-4},2.5\times 10^{-4},7.5\times10^{-5},5\times 10^{-5},2.5\times 10^{-5}$. Following Q7, we calculate the indicator $\gamma(A)$ after every epoch. When this indicator falls below the previous evaluation, we reduce the learning rate. If the learning rate cannot be lowered further, we conclude that the model has converged and the training is complete. Part 9: Some typo errors (According to Q3) Thanks for pointing out the mistake! In lines 173 and 174, the sentence should be "if $r_k$ is the closet representation for $r_x$".	Summary: This work presents CATOD, a new method for dealing with the challenging scenario of adapting generative models to OOD concepts. In particular, CATOD leverages an active learning setting to update the adaptor to generate better OOD images more broadly. Furthermore, the authors provide both extensive empirical evaluations and theoretical analysis which further support the validity of the CATOD methodology. Strengths: Strength: - This paper provides a timely analysis of the difficult problem of adapting generative models to OOD concepts which is often overlooked in recent literature. - CATOD showcases both strong empirical performance (CLIP score and CMMD score) on a reasonable set of 25 OOD concepts. - Additionally, the paper provides strong theoretical links between the introduced aesthetic/concept-matching score and their necessity to the performance of adaptors. Weaknesses: The primary concern with this paper is the sparse empirical evaluations presented for CATOD. In particular, the 25 OOD concepts chosen seem few with only a small set of 100 samples left out for validation. Technical Quality: 3 Clarity: 2 Questions for Authors: The reviewer would like to have further clarification regarding the experimental setup of the paper. In particular, the reviewer would like to hear the author more directly address whether the small evaluation set used for the experiments is a reasonable concern. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors don't provide any full discussion of the limitations of the proposed methodology. Additionally, no direct comment has been made regarding societal impacts, however, the reviewer does not see any potential negative impacts derived from this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! Part 1: About the number of samples to use for training and validation. In brief, since most recent adaptors require only a small number of training samples (usually around 100), our validation set is also set to be of the same scale as the training samples. As mentioned in the paper, adaptors are designed to reduce training costs when introducing new concepts. Specifically, they require significantly fewer samples compared to direct fine-tuning, typically no more than 1000 samples, and often around 100 samples. The following table provides a comparison of the most commonly used adaptors (including our proposed CATOD): \| Aspect \| Textual Inversion [1] \| DreamBooth [2] \| LoRA [3,4] \| CATOD (Ours) \| \| ---------------------------- \| --------------------------------------------------------- \| ---------------------------------------------------------- \| ------------------------------------------------------------ \| -------------------------------------------------------- \| \| Training Data \| Requires a few images (5-20) \| Requires a large amount of labeled data (100-1000) \| Requires intermediate amount of data (around 100) \| Requires intermediate amount of data (around 100) \| \| Flexibility \| High flexibility, but usually limited to only one concept \| Not flexible, can learn multiple detailed attributes \| High flexibility, can learn multiple detailed attributes \| High flexibility, can learn multiple detailed attributes \| \| Model Size \| No change to model size \| Model size can increase \| Slight increase due to additional adaptation layers \| Depend on the adaptor, usually slight increalowse \| \| Quality for New Concepts \| Usually not good \| Good for ID concepts, but usually deteriorates on OOD ones. \| Good for both ID concepts, but usually deteriorates on OOD ones. \| Good for both ID and OOD concepts. \| Since previous adaptors often fail to accurately depict out-of-distribution (OOD) concepts due to inconsistent image quality, we also ensure that our validation set is of high quality to provide accurate validation. [1] Gal, Rinon, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. "An image is worth one word: Personalizing text-to-image generation using textual inversion." arXiv preprint arXiv:2208.01618 (2022). [2] Ruiz, Nataniel, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. "Dreambooth: Fine-tuning text-to-image diffusion models for subject-driven generation." In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 22500-22510. 2023. [3] Hu, Edward J., Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. "Lora: Low-rank adaptation of large language models." arXiv preprint arXiv:2106.09685 (2021). [4] Yeh, Shih-Ying, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. "Navigating text-to-image customization: From lycoris fine-tuning to model evaluation." In The Twelfth International Conference on Learning Representations. 2023. Part 2: About the number of categories to use for the adaptors. Thank you for reminding me to investigate the capacity of adaptors in learning categories with OOD concepts! Extending CATOD to encompass even more categories is straightforward. Using the category "Insect" as an example, we explored the impact of adding three additional OOD concepts. Specifically, we conducted experiments with the following additional categories: "Chlumetia transversa," "Mango flat beak leafhopper," and "Rhytidodera bowrinii white." The table below shows the average performance in the "Insect" category as we add new concepts: \| Number of Categories \| DreamBooth + CATOD (CLIP $\uparrow$) \| DreamBooth +for CATOD (CMMD$\downarrow$) \| LoRA + CATOD (CLIP$\uparrow$) \| LoRA + CATOD (CMMD$\downarrow$) \| \| ------------------------------- \| ------------------------------------ \| ---------------------------------------- \| ----------------------------- \| ------------------------------- \| \| 5 \| 70.83 \| 1.39 \| 71.19 \| 1.13 \| \| 6 (+Chlumetia transversa) \| 69.23 \| 1.57 \| 68.05 \| 1.52 \| \| 7 (+Mango flat beak leafhopper) \| 66.35 \| 1.96 \| 65.39 \| 2.03 \| \| 8 (+Rhytidodera bowrinii white) \| 64.07 \| 2.65 \| 63.28 \| 2.77 \| From this table, we can see that while CATOD can be easily generalized to more concepts, it experiences a performance decline as the number of categories increases. This decline is because recently proposed adaptors were originally designed to accommodate only a few concepts. However, since CATOD is designed as a general framework, it can be easily adapted to work with any future adaptors that are capable of handling a larger number of concepts.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
SeTAR: Out-of-Distribution Detection with Selective Low-Rank Approximation	Accept (poster)	Summary: This work proposes SeTAR for CLIP-based OOD detection. SeTAR is based on low-rank approximation. It determines the optimal rank for each weight block with a greedy hyperparameter search using validation samples. While SeTAR itself is training-free, the paper also proposes SeTAR+FT which incorporates LoRA as a training extension. Experiments demonstrate that SeTAR outperforms existing methods on CLIP-specific benchmarks. Strengths: The paper is written in good quality, and the proposed method sounds reasonable. Weaknesses: 1. My main concern is that the performance improvement of SeTAR (SeTAR+FT) seems limited compared with MCM and GL-MCM (LoCoOp). In Table 1, for example, all improvements of SeTAR in terms of AUROC are within 1%. The only case where SeTAR leads to noticeable/significant improvements is when the model is the image-only Swin-T (Table 2). Would the authors comment on why this is the case? 2. I'm confused by some of the numbers discussed in the main text. In line 220-222, it says `For example, using Pascal VOC as ID, SeTAR yields an average reduction of 12.84% FPR95 on MCM and 18.95% FPR95 on GL-MCM`. How is the "12.84" and "18.95" computed? According to Table 1, on Pascal VOC, the average FPR95 of MCM and SeTAR is 38.88 and 32.46, respectively. The average FPR95 of GL-MCM and SeTAR is 31.12 and 23.86, respectively. A similar confusion of mine is regarding line 238-240 `when scaled up to CLIP-large, SeTAR+FT outperforms LoCoOp and LoRA by 17.92% and 12.45% FPR95 on the same benchmark`, which doesn't seem to match with the numbers reported in Table 2. Am I missing anything? 3. All considered OOD datasets are far-OOD w.r.t. the ID set. Near-OOD detection has long been recognized as a more challenging and realistic problem [1, 2]. I would be very interested to see how well SeTAR performs on ImageNet v.s. NINCO or SSB [2]. I'm willing to adjust my score if my comments are addressed or clarified. [1] Detecting semantic anomalies [2] OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # 1. Performance Gains 1. Limited Room for AUC Improvement: - The baseline AUC scores are above 90, leaving limited room for significant improvement. Despite this, our method still achieves AUC improvements, demonstrating its effectiveness even in a high-performance context. 2. Significant FPR Improvement: - For the False Positive Rate (FPR), our method shows notable improvements. These reductions in FPR are significant and highlight the practical benefits of our approach in reducing false positives, enhancing the reliability of OOD detection. 3. Swin-Base Performance: - The performance boost on Swin-Base is more pronounced because Swin is trained directly on ImageNet, lacking a text-encoder and thus requiring training solely on IN1K. This can lead to overfitting on ID data and poor recognition of OOD images, providing more room for improvement. Our method helps alleviate this issue, improving Swin's generalization to OOD samples. - In contrast, CLIP models are pretrained on large image-text datasets, which provides robust representation capabilities for both ID and OOD images. Consequently, CLIP’s baseline OOD performance is already strong, leaving less room for further improvement compared to Swin-Base. Therefore, the pronounced performance boost seen with Swin-Base is due to its initial lower performance and higher potential for enhancement through our method. # 2. Numerical Clarifications 1. Clarification on Pascal VOC FPR95 Reductions: - Mean MCM FPR95 reduction: (37.24 - 32.46) / 37.24 = 12.84% - Mean GL-MCM FPR95 reduction: (29.44 - 23.86) / 29.44 = 18.95% 2. Clarification on CLIP-large FPR95 Improvements: - Mean LoCoOp FPR95 improvement: (40.74 + 46.74 - 34.75 - 37.05) / (40.74 + 46.74) = 17.92% - Mean LoRA FPR95 improvement: (38.62 + 43.39 - 34.75 - 37.05) / (38.62 + 43.39) = 12.45% These clarifications demonstrate the calculations behind the reported improvements, providing transparency and accuracy in our results. # 3. Near-OOD Results - We appreciate the suggestion and have added results for the CLIP-base backbone, with ImageNet1K as the ID dataset and SSB-Hard as the OOD dataset. SeTAR and SeTAR+FT show superior performance compared to the baselines. \| CLIP-base \| Category \| MCM Score FPR↓ \| MCM Score AUC↑ \| GL-MCM Score FPR↓ \| GL-MCM Score AUC↑ \| \|:-------------\|:------------------\|-----------------:\|-----------------:\|--------------------:\|--------------------:\| \| Vanilla \| Training-Free \| 89.28 \| 63.88 \| 85.62 \| 67.63 \| \| SeTAR \| Training-Free \| 88.29 \| 64.20 \| 84.03 \| 68.29 \| \| LoCoOp \| Finetuning (3run) \| 89.72 \| 63.45 \| 86.79 \| 65.93 \| \| LoRA \| Finetuning (3run) \| 88.52 \| 65.38 \| 84.39 \| 68.85 \| \| SeTAR+FT \| Finetuning (3run) \| 87.16 \| 68.13 \| 84.72 \| 70.42 \| These results highlight SeTAR's and SeTAR+FT's robust performance in challenging near-OOD scenarios. --- Rebuttal 2: Title: Thanks for the rebuttal Comment: My concerns are addressed well and I'd like to raise my score to 6. For point 2, please make it clear in the manuscript that the performance improvement is relative (by default I think people assume/expect absolute improvement being discussed). For point 3, I strongly recommend including the near-OOD results in the main text so that later works can continue to work the meaningful & challenging near-OOD detection problem (rather than keep working only on far-OOD by following previous works). --- Rebuttal Comment 2.1: Title: Thanks for the Feedback and Revised Score Comment: Thank you very much for your constructive feedback and for raising the score. We are grateful for your detailed review and are glad that our responses addressed your concerns. We will ensure that the manuscript clearly indicates that the performance improvements are relative. Additionally, we will incorporate the near-OOD results into the main text as suggested. Your insights have been invaluable in enhancing our work, and we sincerely appreciate your support and encouragement.	Summary: The paper introduces SeTAR, a novel training-free out-of-distribution (OOD) detection method that leverages selective low-rank approximation of weight matrices in vision-language and vision-only models. SeTAR enhances OOD detection by post-hoc modifying the model’s weight matrices using a greedy search algorithm. The paper also extends this method to SeTAR+FT, a fine-tuning approach to further optimize OOD detection performance. Strengths: 1. The paper proposes a unique training-free method for enhancing OOD detection, which is novel. 2. The paper includes thorough ablation studies and sensitivity analyses, which help in understanding the robustness and generalizability of the proposed approach. 3. The authors provide extensive empirical evaluations on multiple benchmarks, showing the effectiveness of SeTAR and SeTAR+FT. Weaknesses: - The performance gains compared to baselines are relatively small. Given that these results are achieved using a greedy search strategy, the potential of SeTAR may be limited. - There is a lack of theoretical analysis explaining why SeTAR is effective. Understanding the underlying principles is crucial for advancing the method and its applications. - The experiments indicate that the optimal hyperparameters (λ and K) vary significantly across different backbones. In practical OOD detection scenarios, it is challenging to obtain OOD samples for hyperparameter tuning. Therefore, SeTAR needs to be strengthened in this aspect to ensure robust performance without extensive hyperparameter tuning. P.S. The layout of Figure 1 is somewhat cluttered. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # 1. Performance Gains 1. Limited Room for AUC Improvement: - The baseline AUC scores are above 90, leaving limited room for significant improvement. Despite this, our method still achieves AUC improvements, demonstrating its effectiveness even in a high-performance context. 2. Significant FPR Improvement: - For the False Positive Rate (FPR), our method shows notable improvements. These reductions in FPR highlight the practical benefits of our approach in reducing false positives, enhancing the reliability of OOD detection. 3. Pronounced Performance on Swin-Base: - As shown in Table 2 and Table 7, the performance boost on Swin-Base is significant. For example, using SeTAR with the Energy score on ImageNet1K, the reductions in FPR and improvements in AUC are over 20% and 9.8% compared to the baseline. SeTAR+FT further reduces FPR and improves AUC by more than 36% and 20% compared to the baseline. These points emphasize the substantial performance gains achieved by our method, particularly on Swin-Base, highlighting its effectiveness in various contexts. # 2. Theoretical Analysis of SeTAR's Effectiveness To address this, we draw on theoretical principles from recent work on SVD-based weight pruning, particularly from the study titled "[Enhancing In-Context Learning Performance with just SVD-Based Weight Pruning: A Theoretical Perspective](https://arxiv.org/pdf/2406.03768)". SeTAR employs selective low-rank approximation through SVD-based pruning, which aligns with established theoretical frameworks that explain how such techniques can enhance model performance: 1. Gradient Stability and Generalization: - The theoretical analysis of SVD-based weight pruning shows that pruning can stabilize the gradient updates in neural networks. By pruning weights, particularly in layers where noise might be higher, we reduce the sensitivity of the network to small perturbations, leading to more stable and robust performance. This stability is crucial for OOD detection tasks, as it allows the model to maintain its performance across varying inputs. 2. Matrix Condition Number and Noise Reduction: - The effectiveness of weight pruning can also be understood through the concept of matrix condition numbers. High condition numbers indicate ill-conditioned problems prone to significant errors due to small perturbations. Pruning minor singular values reduces the condition number, stabilizing the model and enhancing its robustness to noise. This is crucial for tasks like OOD detection, where stability and robustness are key. 3. Retention of Principal Components: - By retaining the principal components (i.e., the components corresponding to the largest singular values), SeTAR ensures that the most critical information is preserved while reducing noise. This principle, rooted in the Eckart-Young-Mirsky theorem, provides an optimal low-rank approximation that maintains essential features necessary for effective OOD detection. # 3. Hyperparameter Tuning 1. Robustness of Top-K Hyperparameter: - The optimal Top-K parameter is related to the number of ID categories and cannot be directly transferred between different ID datasets. However, as shown in Figure 6, this parameter is quite robust. We generally recommend setting it to 30% of the total number of categories. For example, ImageNet1K (300/1000) and Pascal-VOC (4/14). For the Swin-base model, setting Top-K to 300 also yields good performance. \| Backbone \| Score \| Vanilla Method FPR↓ \| Vanilla Method AUC↑ \| SeTAR (TopK700) FPR↓ \| SeTAR (TopK700) AUC↑ \| SeTAR (TopK300) FPR↓ \| SeTAR (TopK300) AUC↑ \| \|:-----\|:--------\|-------:\|-------:\|--------:\|--------:\|--------:\|--------:\| \| Swin-base\| MSP \| 59.25\| 84.12\| 56.05 \| 85.77 \| 56.82 \| 85.68 \| \| Swin-base\| Energy\| 65.01\| 76.10\| 51.61 \| 84.42 \| 52.56 \| 84.51 \| 2. Transferability of λ Hyperparameter: - The λ parameter shows transferability across different datasets for the same backbone. As shown in Table 9, the λ for CLIP-base ranges between 0.05 and 0.1, while for CLIP-large it ranges between 0.3 and 0.5. 3. Hyperparameter Transferability Across Datasets: - The optimal hyperparameters exhibit transferability across different datasets. For instance, the optimal hyperparameters for CLIP-Base on ImageNet1K and Pascal-VOC can be interchanged with minimal performance degradation, and both outperform the vanilla method. \| ImageNet1K \| Score \| Vanilla Method FPR↓ \| Vanilla Method AUC↑ \| SeTAR Optimal hyperpara FPR↓ \| SeTAR Optimal hyperpara AUC↑ \| SeTAR with VOC optimal hyperpara FPR↓ \| SeTAR with VOC optimal hyperpara AUC↑ \| \|:--\|:-------\|---:\|---:\|---:\|---:\|---:\|---:\| \| CLIP-base \| MCM\| 43.09\| 90.74\| 40.24\| 91.05\| 40.41 \| 91.02 \| \|\| GL-MCM \| 35.29\| 90.86\| 33.12\| 91.32\| 33.55 \| 91.17 \| \| CLIP-large \| MCM\| 37.19\| 91.73\| 36.26\| 91.92\| 36.73 \| 91.81 \| \|\| GL-MCM \| 40.65\| 89.98\| 39.54\| 90.22\| 39.18 \| 90.10 \| \| Pascal-VOC \| Score\| Vanilla Method FPR↓ \| Vanilla Method AUC↑ \| SeTAR Optimal hyperpara FPR↓ \| SeTAR Optimal hyperpara AUC↑ \| SeTAR with ImageNet1K optimal hyp FPR↓ \| SeTAR with ImageNet1K optimal hyp AUC↑ \| \|:-----------\|:-------\|------:\|------:\|--:\|--:\|------------:\|------------:\| \| CLIP-base\| MCM\| 37.24\| 92.98\| 32.46 \| 93.74 \| 33.18\|93.65\| \|\| GL-MCM \| 29.44\| 93.88\| 23.86 \| 94.87\| 23.57\| 94.86\| \| CLIP-large \| MCM\| 52.21\| 91.68\| 42.57 \| 92.91 \| 44.39\| 92.34\| \|\| GL-MCM \| 43.96 \| 92.45 \| 31.12 \| 94.00\| 33.74 \| 93.76\| 4. Figure 1 Layout: - We appreciate the feedback on Figure 1 and have adjusted its layout for better clarity. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed reply. I find the explanation about the "limited room for AUC improvement" somewhat inadequate. Additionally, I noticed that your primary comparisons are with MCM and its variant GL-MCM. I suggest comparing SeTAR with more recent baselines, such as “CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No” (ICCV2023) and “Negative Label Guided OOD Detection with Pretrained Vision-Language Models” (ICLR2024), to more convincingly demonstrate its performance advantages. Your responses to the theoretical analysis and hyperparameter tuning concerns have addressed some of my doubts. Thank you for the additional clarifications. --- Rebuttal 2: Title: Comparisons with CLIPN and NegLabel Comment: We apologize for the late reply and appreciate your patience. # 1. Comparisons with CLIPN - CLIPN [1] is pre-trained on the CC-3M dataset for 10 epochs, involves training an additional NO-encoder for over 64 million parameters - On ViT-B/16, CLIPN-C achieves an FPR95 of 38.59 and an AUROC of 86.35, while CLIPN-A achieves an FPR95 of 31.10 and an AUROC of 93.10. - In contrast, our SeTAR method does not involve any training. When combined with fine-tuning (SeTAR+FT), it operates in a 1-shot setting, utilizing only 1,000 images for training, with just 1.6% of the parameters being trainable over 5 epochs. - For comparison, SeTAR achieves an FPR95 of 33.12 and an AUROC of 91.32 without any training. With minimal fine-tuning (SeTAR+FT), it achieves an FPR95 of 32.19 and an AUROC of 92.31 with only 1,000 training samples. Given the significant differences in the amount of training data, the number of trainable parameters, and the computational resources required, we believe that a direct comparison between CLIPN and our method is not entirely fair. Despite the minimal computational resources involved, SeTAR and SeTAR+FT achieve performance levels close to those of CLIPN, demonstrating the efficiency of our approach. # 2. Comparisons with NegLabel - Difference from NegLabel: Our method differs from NegLabel [2] in its focus. SeTAR primarily aims to enhance the model's intrinsic performance through SVD pruning, without incorporating any additional knowledge or inputs. In contrast, NegLabel improves OOD detection by constructing large scale virtual negative labels from the data perspective. - Compatibility with NegLabel: Since SeTAR and NegLabel have different focuses, they are not mutually exclusive. As mentioned in our paper, SeTAR is highly compatible with various score functions and can also work alongside data augmentation methods like NegLabel. For instance, by simply merging negative labels into the label space for searching and testing, SeTAR can further surpass NegLabel's performance. \|ViT-B/16 \|iNaturalist FPR↓\|iNaturalist AUC↑\|SUN FPR↓\|SUN AUC↑\|Places FPR↓\|Places AUC↑\|Texture FPR↓\|Texture AUC↑\|Average FPR↓\|Average AUC↑\| \|:---\|-:\|-:\|-:\|----:\|-:\|-:\|-:\|-:\|-:\|-:\| \|CLIPN\\|23.94 \|95.27\|26.17\|93.93\|33.45\|92.28\|40.83\|90.93\|31.10\|93.10\| \|NegLabel\ \|1.91 \|99.49\|20.53\|95.49\|35.59\|91.64\|43.56\|90.22\|25.40\|94.21\| \|SeTAR\|0.15 \|99.54\|19.06\|95.84\|30.63\|92.22\|42.54\|90.30\|23.09\|94.48\| > \* cited from [2] In summary, SeTAR demonstrates strong performance through its structural enhancements, particularly with SVD pruning, which significantly improves model robustness and OOD detection without requiring additional training. Furthermore, when combined with data augmentation methods like NegLabel, SeTAR's effectiveness is further amplified, showing even greater improvements in OOD detection metrics. --- References: - [1] CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No - [2] Negative Label Guided OOD Detection with Pretrained Vision-Language Models --- Rebuttal Comment 2.1: Title: Anticipating Further Remarks Comment: Dear Reviewer 6fvA As the discussion period is coming to a close, we would appreciate it if you could let us know whether our recent rebuttal has addressed some of your concerns or questions. We are more than happy to address any further issues you may have. Engaging in this discussion will greatly help us in refining and improving our paper. We look forward to your response or acknowledgment once you have read our message, as your support is very important to us. Best regards, The Authors	Summary: This paper proposes an algorithm for OOD detection along with CLIP models. It observes that pruning based on SVD decomposition on CLIP models can improve the OOD detection performance. A greedy search algorithm is developed for searching pruning rations of each weight in CLIP models. Experiments on regular settings demonstrate obvious performance when compared with the Vanilla GL_MCM method. Strengths: (1) The paper is written clearly and easy to follow. (2) The method is simple but effective. (3) The ablation studies are sound. Weaknesses: (1) It seems that the proposed method is general enough to apply to regular models not just CLIP models. Will it also work on CNN-based ResNets and ViT models trained on datasets, like ImageNet and CIFAR? The experimental results focus on CLIP-based models in Table 1. More comparisons with previous methods using CNN-based models should be included. (2) SVD decomposition pruning has strong connections with sparsification-based methods [1,2,3]. Comparisons with this kind of method should be included. [1] Dice: leveraging sparsification for out-of-distribution detection. [2] Extremely simple activation shaping for out-of-distribution detection. [3] ReAct: out-of-distribution detection with rectified activations. (3) The proposed greedy search algorithm is efficient. However, it can not be guaranteed to be converged. What's the performance variance for 2 independent runs of the search algorithm? Technical Quality: 3 Clarity: 3 Questions for Authors: (1) The main results in the paper concentrate on CLIP models. Results on other CNN-based models and comparisons with previous methods should be included in the paper. (2) The proposed greedy search algorithm can not be guaranteed to be converged. Thus, a discussion on the stability of the proposed algorithm is required. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: A section of Impact Statements including limitations is included in the paper Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # 1. Applicability to CNN-Models 1. Will it work on CNN-based ResNets? - No. Our method is not applicable to pure CNN models like traditional ResNet50. The loss function (Eq. 12) in our approach includes an OOD loss (Eq. 11), which relies on local features from the attention layer. Since pure CNN models lack self-attention layers, our method cannot be directly used or compared with CNN-based models. - However, in CLIP-ResNet models, a self-attention layer is added to the last layer of the ResNet tower. Therefore, we can conduct experiments on CLIP-ResNet models. - Specifically, we used the CLIP-ResNet50x4 model as the backbone with ImageNet1K as the ID dataset. By applying SVD pruning on the conv1 layer of each vision layer and W_{up} on each text layer, we obtained the following results, demonstrating that our method is also applicable to CLIP-ResNet models (\* stands for our re-run). \| CLIP-ResNet50x4 \| iNaturalist FPR↓ \| iNaturalist AUC↑ \| SUN FPR↓ \| SUN AUC↑ \| Places FPR↓ \| Places AUC↑ \| Texture FPR↓ \| Texture AUC↑ \| Average FPR↓ \| Average AUC↑ \| \|:----------------\|-------------------:\|-------------------:\|-----------:\|-----------:\|--------------:\|--------------:\|---------------:\|---------------:\|---------------:\|---------------:\| \| Vanilla MCM\* \| 44.03 \| 91.58 \| 35.18 \| 92.83 \| 44.38 \| 89.38 \| 57.29 \| 85.99 \| 45.22 \| 89.95 \| \| SeTAR+MCM \| 41.29 \| 92.12 \| 35.44 \| 92.76 \| 43.01 \| 89.85 \| 54.82 \| 86.89 \| 43.64 \| 90.40 \| \| Vanilla GL-MCM\* \| 32.17 \| 93.09 \| 46.64 \| 89.27 \| 51.85 \| 85.86 \| 44.47 \| 86.49 \| 43.78 \| 88.68 \| \| SeTAR+GL-MCM \| 30.15 \| 93.73 \| 45.01 \| 89.58 \| 49.82 \| 86.68 \| 42.32 \| 87.30 \| 41.83 \| 89.32 \| 2. Will it work on ViT models trained on datasets like ImageNet and CIFAR? - Yes. The Swin-Transformer is a ViT model that does not include a text-encoder but only an image-encoder. As shown in Table2 and Table 7, we used the Swin-base model trained on ImageNet1K as the backbone and observed significant improvements over the original model. 3. More Comparisons with Previous Methods Using CNN-based Models: - As mentioned, our method is not applicable to CNN-based models. More relevant comparisons are addressed in the response to question 2. These points clarify that while our method is not suitable for pure CNN models, it is effective for models incorporating attention mechanisms, such as CLIP-ResNet and ViT models like Swin-Transformer. # 2. Comparisons with Sparsification-Based CNN Methods We appreciate the reviewer's suggestion and have carefully reviewed the referenced papers. These methods are based on CNN-ResNet models and the ImageNet1K dataset. However, there is a significant difference between these models and the CLIP-ResNet: 1. Training Differences: - CNN-ResNet models, due to the lack of a text-encoder, are fine-tuned directly on ImageNet1K. - In contrast, CLIP-ResNet models are not trained directly on ImageNet1K. Therefore, directly comparing the results of CLIP-ResNet with CNN-ResNet models is not meaningful, since ID-domain training would largely boost the model performance. 2. Lack of Suitable CLIP-ResNet Models: - We attempted to find CLIP-ResNet models that were fine-tuned on ImageNet1K for a fair comparison but were unable to locate such models. 3. Truly Training-Free Nature of Our Method: - Our method is literally training-free, as it does not require any training on the ID dataset. In contrast, the backbones used in the sparsification-based methods require training on the ID dataset. These points highlight the challenges in making direct comparisons with sparsification-based methods and underscore the unique, training-free advantage of our approach. # 3. Convergence and Performance Variance Concerns The SVD algorithm used in our method is quite stable. We have compared the results using different random seeds (3, 4, and 5), and the SVD results are consistent across these different seeds. Therefore, SeTAR does not involve performance variance due to the deterministic nature of the SVD algorithm used in our approach. --- Rebuttal Comment 1.1: Title: Further questions Comment: Hi, thanks for the responses from the authors. I still have some confusion on the paper. (1) Is there any alternative loss function for Eq. 11? Although the proposed method is general, its capability is heavily limited by this loss. (2) Could the authors provide comparisons with sparsification-based methods using Swin or ViT backbones? --- Rebuttal 2: Title: Reply to further questions Comment: We sincerely appreciate the thoughtful feedback provided by the reviewer. Here are our responses: # 1. Alternative Loss Function for Eq. 11 - Our loss function (Eq. 11) leverages local features because it requires pseudo-OOD features, which are inherently present in ViTs and can also be constructed in CNNs. In CNNs, alternative methods for OOD feature construction are available, such as the approach proposed in [NPOS](https://arxiv.org/pdf/2303.02966)[4], where boundary ID embeddings are selected based on non-parametric k-NN distances, and outliers are synthesized by sampling from a multivariate Gaussian distribution centered around these boundary embeddings. Additionally, [CLIP-OS](https://arxiv.org/pdf/2404.00323)[5] suggests using CLIP for outlier synthesis, which can similarly be adapted for constructing OOD features in CNNs. - However, due to the complexity of implementation and time constraints, we were unable to conduct the corresponding experiments. Nevertheless, our method is not limited to CLIP models; by utilizing different OOD feature construction methods, our approach can be readily adapted to CNN models as well. # 2. Comparisons on Swin Backbone - We conducted experiments using the Swin backbone with ReAct [3], DICE [1], and ASH [2] methods. Specifically: 1. Codebase: For ASH and ReAct, we implemented Swin-transformer based on the [official ASH repository](https://github.com/andrijazz/ash/blob/main/config/vit_config.yml), strictly following the original settings and applying sparsification before the final linear layer. Since ASH does not provide an implementation for DICE, we based our implementation on the [DICE official repository](https://github.com/deeplearning-wisc/dice/tree/master/models). 2. Hyperparameter Search: Following ASH, we experimented with various parameter combinations and reported the best results: DICE pruning thresholds included {10%, 15%, 70%}, ReAct clipping thresholds {1.0, 1.5, 1.33}, and ASH-S pruning thresholds included {60%, 65%, 90%, 95%}. 3. Datasets: We used ImageNet1K as the ID dataset. - The results are shown below. It is evident that compared to sparsification-based methods, SeTAR exhibits the best performance. ReAct shows improvements over the baseline in both scoring functions, while DICE shows improvement only with MSP; ASH-S performs poorly across the board. \| SwinV2-Base\| Score\| iNaturalist FPR↓ \| iNaturalist AUC↑ \| SUN FPR↓ \| SUN AUC↑ \| Places FPR↓ \| Places AUC↑ \| Texture FPR↓ \| Texture AUC↑ \| Average FPR↓ \| Average AUC↑ \| \|:-\|:-\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\| \| Vanilla* \| MSP\|44.78 \|89.89 \|63.12 \|82.81 \| 67.07 \| 81.45 \|62.04 \|82.33 \|59.25 \|84.12 \| \| React* \| MSP\|42.98 \|90.39 \|61.34 \|83.89 \| 65.11 \| 82.64 \|61.22 \|81.37 \|57.66 \|84.57 \| \| DICE\| MSP\|43.02 \|89.03 \|62.22 \|78.31 \| 65.82 \| 79.35 \|57.75 \|81.48 \|57.20 \|82.04 \| \| ASH-S \| MSP\|53.21 \|78.72 \|73.71 \|66.56 \| 79.75 \| 60.75 \|60.27 \|75.52 \|66.73 \|70.39 \| \| SeTAR (Ours) \| MSP\| 41.44 \|91.08 \|60.05 \|85.04 \| 64.31 \| 83.70 \|58.39 \|83.26 \|56.05 \|85.77 \| \| Vanilla* \| Energy \|57.52 \|81.60 \|71.98 \|72.93 \| 76.90 \| 68.90 \|53.65 \|80.96 \|65.01 \|76.10 \| \| React* \| Energy \|41.78 \|88.34 \|60.98 \|79.19 \| 68.07 \| 75.56 \|53.72\|80.39\|56.14\|80.87\| \| DICE\| Energy \|64.45 \|74.91\|83.04\|62.67\|95.18\|47.05\|94.65\|33.45\|84.33\|58.18\| \| ASH-S \| Energy \|99.67 \|19.26 \|99.29 \|22.91 \| 99.51 \| 21.49 \|98.51 \|33.45 \|99.25 \|24.28 \| \| SeTAR (Ours) \| Energy \|41.71 \|89.42 \|56.53 \|83.29 \| 62.84 \| 80.20 \|45.37 \|84.76 \|51.61 \|84.42 \| > \* denotes our rerun These results demonstrate that SeTAR significantly outperforms sparsification-based methods, highlighting the effectiveness of our approach. # 3. Verification of ASH-S Implementation - Due to the poor performance of ASH-S, we verified the results using the official code. Although the paper does not report ViT results, related scripts are available in their codebase. We tested the official [ViT script](https://github.com/andrijazz/ash/blob/main/config/vit_config.yml), and the results are shown below, confirming that ASH-S continues to perform poorly on ViT. \| ViT-B/16 \| Score\| iNaturalist FPR↓ \| iNaturalist AUC↑ \| SUN FPR↓ \| SUN AUC↑ \| Places FPR↓ \| Places AUC↑ \| Texture FPR↓ \| Texture AUC↑ \| Average FPR↓ \| Average AUC↑ \| \|:-\|:-\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\|-:\| \|Vanilla\| Energy \|64.08 \|79.24\|72.77\|70.25\|74.30\|68.44\|58.46\|79.30\|67.40\|74.31\| \|ASH-S\| Energy \|99.98 \|7.28 \|99.64\|17.82\|99.59\|19.72\|98.09\|27.31\|99.32\|18.03\| > \* denotes our rerun References: - [1] Dice: leveraging sparsification for out-of-distribution detection. - [2] Extremely simple activation shaping for out-of-distribution detection. - [3] ReAct: out-of-distribution detection with rectified activations. - [4] Non-Parametric Outlier Synthesis. - [5] CLIP-driven Outliers Synthesis for few-shot OOD detection. --- Rebuttal Comment 2.1: Title: Thanks for the responses from the authors Comment: Thanks for the responses from the authors. Q1. I think it is a crucial weakness that the method heavily depends on the loss (Eq. 11). Most previous research conducts experiments on CNNs. Although the authors provide comparisons with previous methods using the Swin backbone, the results seem to be a little bit wired, especially for ASH-S. I still recommend the authors could compare their method and previous works on CNNs for fair comparisons. --- Rebuttal 3: Title: Results on CNNs Comment: We sincerely appreciate the reviewer’s insightful comments. Here are our detailed responses: # 1. Results on ResNet50 1. Setup: We conducted experiments using only the ID loss, applying low-rank approximation on the in-feature and out-feature dimensions of the convolutional layers, combined with ASH for search. The results are as follows: \| ResNet50 \|iNaturalist FPR↓ \|iNaturalist AUC↑ \|SUN FPR↓ \|SUN AUC↑ \|Places FPR↓ \|Places AUC↑ \|Texture FPR↓ \|Texture AUC↑ \|Average FPR↓ \|Average AUC↑ \| \|:--\|---:\|---:\|---:\|---:\|-:\|-:\|--:\|--:\|--:\|--:\| \| Softmax \* \|54.99\|87.74 \|70.83 \|80.86 \|73.99 \|79.76 \| 68.00 \| 79.61 \| 66.95 \| 81.99 \| \| Energy \* \|55.72\|89.95 \|59.26 \|85.89 \|64.92 \|82.86 \| 53.72 \| 85.99 \| 58.41 \| 86.17 \| \| ReAct \* \|20.38\|96.22 \|24.20 \|94.20 \|33.85 \|91.58 \| 47.30 \| 89.80 \| 31.43 \| 92.95 \| \| DICE \\|25.63\|94.49 \|35.15 \|90.83 \|46.49 \|87.48 \| 31.72 \| 90.30 \| 34.75 \| 90.77 \| \| ASH-P \ \|44.57\|92.51 \|52.88 \|88.35 \|61.79 \|61.79 \| 42.06 \| 89.70 \| 50.32 \| 89.04 \| \| ASH-B \* \|14.21\|97.32 \|22.08 \|95.10 \|33.45 \|92.31 \| 21.17 \| 95.50 \| 22.73 \| 95.06 \| \| ASH-S \* \|11.49\|97.87 \|27.98 \|94.02 \|39.78 \|90.98 \| 11.93 \|97.60\| 22.80 \| 95.12 \| \| SeTAR\|10.08\|98.11\|27.68\|94.15\|39.22\|91.24\| 12.54 \| 97.51 \| 22.38 \| 95.25\| > \* cite from "Extremely simple activation shaping for out-of-distribution detection." 2. Not Highly Dependent on OOD Loss (Eq. 11): Even under this simple setup, SeTAR outperforms ASH's best performance. This demonstrates that our method can effectively enhance the model's ability to distinguish between ID and OOD samples, even when relying solely on the ID loss. It’s important to note that this result was obtained under a basic setup, and due to time constraints, we did not further tune or explore other detailed configurations specific to CNNs, such as: - Low-Rank Settings for Convolutional Layers: The optimal low-rank structure and dimensions for convolutional layers have not been thoroughly researched. For example, ELRT [1] proposes low-rank approximation directly in the high-order tensor format, while other methods [2][3][4] conduct and maintain the 4-D convolutional layer in the format of a low-rank 2-D matrix. - Pseudo-OOD Feature Extraction in CNNs: As mentioned in our previous response, there are methods to construct OOD features within CNNs as well, which could further improve our model's performance. --- Rebuttal Comment 3.1: Comment: Thanks for your responses to address my concerns. I strongly recommend including the comparisons with CNNs in the main paper. Although the improvements seem not obvious with CNNs, I still think the paper deserves to be accepted for its applications on CLIP models. --- Reply to Comment 3.1.1: Title: Grateful for Your Continued Feedback and Consideration Comment: We sincerely appreciate the reviewer’s continued engagement and thoughtful feedback. We will include the CNN comparison results in the main paper as suggested. We hope that the additional comparisons enhance the overall evaluation of our work and would be grateful if the reviewer could kindly reconsider the rating in light of these updates. --- Reply to Comment 3.1.2: Title: Appreciation for Feedback and Request for Score Review Comment: Dear Reviewer SyYP, Thank you for your recognition of our paper. We appreciate your comment that “the paper deserves to be accepted for its applications on CLIP models.” Your feedback has been incredibly valuable to us. As the rebuttal period is nearing its end, we have provided a quick summary of our responses and updates. We kindly ask you to consider these in your final scoring. Thank you once again for your valuable review. Best regards, The Authors --- Rebuttal 4: Title: Results on CNNs (Continued) Comment: # 2. Effectiveness of Our Method Across CLIP, Swin, and CNN Architectures 1. Broad Effectiveness Across Architectures: Our method has consistently proven effective across a range of architectures. As detailed in Table 1 and Table 7 of our paper, SeTAR outperforms vanilla methods on both CLIP and Swin models. Furthermore, even with a basic setup on ResNet, SeTAR surpasses current state-of-the-art sparsification-based methods. The comparison with Swin in our previous response underscores the limitations of previous sparsification-based approaches, which struggle with models like Swin. In contrast, SeTAR achieves state-of-the-art performance across all major vision architectures, demonstrating its versatility and generalizability. 2. Significance for ViT-based Models: ViT-based models, like CLIP, are receiving increasing attention in research due to their scalability and strong performance [5][6][7][8][9]. Our method’s superior results on ViT models highlight its potential for advancing OOD detection in these architectures, making it particularly relevant as ViT models become more widely adopted. 3. Limitations of Sparsification-Based CNN Methods: Sparsification-based CNN methods like ReAct and ASH cannot be applied as a post-hoc method to CLIP-based zero-shot OOD detection model. Both methods rely on the assumption that ID and OOD images produce distinct activations in models trained specifically on ID data, such as ResNet50. However, in large-scale pretrained models like CLIP, the activations for ID and OOD images are not significantly different. Consequently, methods like ReAct and ASH, which are limited to models trained on downstream ID-domain tasks, constrain their effectiveness in enhancing CLIP’s zero-shot OOD detection capabilities. In contrast, our method can be applied as a post-hoc method to enhance CLIP's zero-shot OOD detection capabilities. 4. Contribution of Specialized Methods: Despite their limitations, specialized methods play a crucial role in their respective domains. For instance, sparsification-based CNN methods significantly enhance OOD detection in CNN models, even though they may not perform well on CLIP models. Similarly, methods like GL-MCM [10] and LoCoOp [11], which utilize CLIP’s local features, substantially improve MCM scores and the performance of fine-tuned models. Although these methods are specialized, they contribute meaningfully to the ongoing development and advancement of the field. --- References: - [1] ELRT: Towards Efficient Low-Rank Training for Compact Neural Networks - [2] Learning Low-rank Deep-Neural Networks via Singular Vector Orthogonality Regularization and Singular Value Sparsification - [3] Training CNNs with Low-Rank Filters for Efficient Image Classification - [4] Convolutional Neural Networks with Low-Rank Regularization - [5] Multimodal Learning with Transformers: A Survey - [6] Self-Supervised Multimodal Learning: A Survey - [7] The Llama 3 Herd of Models - [8] Scaling Vision Transformers to 22 Billion Parameters - [9] How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites - [10] Zero-Shot In-Distribution Detection in Multi-Object Settings Using Vision-Language Foundation Models - [11] LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning	Summary: The paper presents SETAR, a novel method designed to enhance out-of-distribution (OOD) detection without requiring additional training. The proposed method leverages rank reduction techniques applied to the model weights, specifically targeting the minor singular components, while retaining the principal components that significantly contribute to the model’s performance. SETAR is evaluated across various model backbones, including CLIP-base, CLIP-large, and Swin-base, and demonstrates notable improvements in OOD detection tasks. The paper also provides comprehensive experiments and ablation studies to validate the effectiveness and efficiency of SETAR. Strengths: 1. Novelty and Innovation: The introduction of a training-free method for improving OOD detection is a significant contribution. By focusing on rank reduction of model weights, the method offers a fresh perspective compared to traditional training-intensive approaches. 2. Comprehensive Evaluation: The paper provides extensive experimental results across multiple datasets and model backbones, showcasing the robustness and generalizability of SETAR. 3. Effective Performance: The method achieves substantial improvements in OOD detection metrics, such as FPR95 and AUROC, demonstrating its practical utility. The significant performance boost on Swin-base compared to CLIP-base and CLIP-large is particularly notable, likely due to the inherent design differences and stronger zero-shot performance of CLIP models. 4. Detailed Analysis: The inclusion of ablation studies and sensitivity analyses helps in understanding the impact of various components and hyperparameters, offering valuable insights into the method’s functioning. Weaknesses: I hope the authors could address my following questions. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Table 2: Could you provide more insights on why the performance boost on Swin-Base is significantly more pronounced compared to CLIP-Base and CLIP-Large? Is this disparity related to CLIP’s stronger zero-shot performance? 2. Figure 2: For ImageNet-1K, it appears that decomposing the vision encoder alone is sufficient for OOD detection, with the combination of both vision and text encoders yielding only minor improvements. Could you explain this in more detail? 3. The proposed method is training-free, but the greedy search algorithm used for determining the rank reduction ratio list and performing rank reduction does require computation time. Can you provide details on the time required for the greedy search and rank reduction for SeTAR, as well as the overall computation time of SeTAR+FT, compared to previous fine-tuning methods? 4. The proposed method seems to be an application of LASER, the only contribution is the adaptation of LASER to CLIP model, and benefits OOD detection. Can the authors justify this? The paper is technically solid with comprehensive experiments, but I think the paper is not novel enough given the similarities compared to LASER. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Limitation discussed in Appendix Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # 1. Performance Disparity on Swin-Base - The performance boost on Swin-Base is more pronounced because Swin is trained directly on ImageNet, lacking a text-encoder and thus requiring training solely on IN1K. This can lead to overfitting on ID data and poor recognition of OOD images, providing more room for improvement. Our method helps alleviate this issue, improving Swin's generalization to OOD samples. - In contrast, CLIP models are pretrained on large image-text datasets, which provides robust representation capabilities for both ID and OOD images. Consequently, CLIP’s baseline OOD performance is already strong, leaving less room for further improvement compared to Swin-Base. Therefore, the pronounced performance boost seen with Swin-Base is due to its initial lower performance and higher potential for enhancement through our method. # 2. Modality Difference for Performance 1. Significance of Vision Modality: As shown in Figure 2, the experiments demonstrate that the vision encoder is more critical for OOD detection tasks compared to the text encoder. This is intuitive since, in image-based OOD detection, the vision encoder is essential for extracting features and identifying OOD patches. The text encoder, on the other hand, shows limited improvement in performance for this specific task, indicating that the vision modality holds greater importance. 2. Combined Modality for Optimal Performance: While decomposing the vision encoder alone provides substantial improvements, incorporating the text encoder, albeit with minor gains, ensures we maximize the model’s capabilities. By leveraging both vision and text modalities, we achieve better overall performance, which is why we opted for the Vision+Text modality approach in our method. # 3. Computation Time Concerns for SeTAR and SeTAR+FT We appreciate the reviewer's concern regarding the computation time required for our proposed SeTAR method. Here is the comparison for CLIP-base on ImageNet1K: 1. SeTAR and Fine-Tuning Times: - SeTAR requires approximately 14 minutes to complete the Vision+Text greedy search and rank reduction for 1K images. If we only apply the search to the Vision modality, the total time is about 7 minutes, which also achieves competitive performance as noted in the previous point. - SeTAR+FT takes a total of 14 minutes and 11 seconds, consisting of two stages. The first stage is SeTAR for low-rank searching, which takes about 14 minutes. The second stage is LoRA tuning, which takes around 11 seconds for 5 epochs on the same 1,000 images, primarily due to the small model size and limited development set. - For comparison, LoCoOp fine-tuning takes about 16 minutes for 50 epochs. 2. Detailed Time Analysis for SeTAR: - Greedy Search Loss Calculation: This accounts for 47.67% of the total time. - Dataloader: Takes up 11.24% of the time. This delay occurs because the CPU cannot match the GPU speed for smaller models and few samples; for larger models, this delay can be negligible. - SVD Pruning and Reloading: To maintain code clarity and compatibility, each step involves reloading and applying SVD pruning, which takes 20% of the total time. We plan to optimize this by loading the model once at the beginning and pre-computing SVD in parallel to avoid redundant calculations. This optimization could reduce the time taken by these steps to about 1/24 of its current value. - With these optimizations, we estimate the overall time for SeTAR could be reduced to about half of the current duration, approximately 7 minutes for Vision+Text and 3.5 minutes for Vision-Only searching. These points highlight the efficiency of SeTAR and SeTAR+FT in both low-rank searching and fine-tuning compared to existing methods. # 4. Novelty Concerns Compared to LASER We appreciate the reviewer's feedback and the opportunity to clarify the unique contributions of SeTAR compared to LASER: 1. Distinct Approach of SeTAR: - Beyond Simple Application: SeTAR is not just an adaptation of LASER to the CLIP model. LASER primarily focuses on pruning individual layers to enhance factual answering capabilities and does not extensively explore different greedy pruning strategies. Additionally, LASER relies on a validation set for selection, which is not suitable for OOD detection. - Greedy Pruning Algorithm and OOD Detection: SeTAR focuses on designing a greedy pruning algorithm tailored for OOD detection. To address the challenge of unavailable OOD validation sets, we extract OOD information from ID images to guide the algorithm. We also conducted a sensitivity analysis of different parameters. Furthermore, SeTAR includes a comprehensive analysis comparing various modalities, search algorithms, pruning strategies, and backbones, enhancing our understanding of low-rank pruning beyond what LASER provides. 2. Innovations in SeTAR+FT: - Dynamic Rank Adjustment: In addition to the training-free search algorithm, we explored the potential of using SeTAR for fine-tuning. Traditional LoRA distributes the rank evenly across all layers, leading to inefficiencies and performance losses. - Effective and Efficient Fine-Tuning: By combining SeTAR with LoRA, SeTAR first identifies the impact of different layers on performance and dynamically adjusts the rank accordingly. This approach initializes different LoRA weights more effectively, tailored to the specific ID dataset, resulting in a more effective and efficient fine-tuning process. These points illustrate that SeTAR offers significant advancements over LASER, both in methodology and application, particularly for OOD detection. --- Rebuttal Comment 1.1: Comment: Thank you to the authors for providing comprehensive clarifications. Most of my concerns have been addressed. I would like to note that the contribution of greedy rank search across multiple layers has already been proposed in the LASER paper (Sec. 5.1, “Composing reductions across layers”). Given the technical soundness, extensive experiments, and the innovative application of rank reduction in the domain of OOD detection using a Vision-Language model, I will maintain my original score. --- Rebuttal 2: Title: Clarification and Innovations in SeTAR for OOD Detection Comment: We sincerely appreciate the reviewer’s thoughtful feedback and the opportunity to clarify and expand on certain aspects of our work. # 1. Clarification on LASER We appreciate the reviewer’s reminder regarding Section 5.1 of the LASER paper. We also took note of this section during our review. What we intended to highlight in our original rebuttal is that, while LASER employs a single greedy search strategy, our work delves deeper into the nuances of conducting greedy search effectively for OOD detection (as detailed in Section 4.4). Specifically, LASER’s approach focuses on composing reductions across layers without exploring the broader landscape of greedy search strategies. In contrast, we systematically analyze and compare different greedy search techniques, evaluating their effectiveness across various layers and backbones. This detailed exploration allows our method to be more finely tuned for the specific challenges of OOD detection, thereby providing a more robust and versatile solution. # 2. Innovation in Post-Hoc Sparsification Methods Post-hoc sparsification methods are widely utilized to enhance CNN-based OOD detection, with well-known examples including ReAct[1], ASH[2], and Dice[3]. These methods typically operate in the weight or activation space, aiming to improve OOD detection by modifying the model’s internal structures. However, a significant limitation of methods like ASH and ReAct is that they are designed for models that have been trained on in-domain (ID) data, such as ResNet50, where distinct activations for ID and OOD samples are expected. These methods fail to generalize to zero-shot OOD detection models like CLIP, where the model has not been fine-tuned on any ID data, and the activations for ID and OOD samples share similar distributions. Our approach addresses these limitations by introducing two key innovations: first, we operate within the space derived from SVD decomposition, which allows us to capture the most informative components of the model while discarding noise. Second, our method is specifically designed to function as a post-hoc approach compatible with CLIP’s zero-shot OOD detection capabilities. This dual innovation not only enables our method to bypass the limitations of traditional sparsification techniques but also allows it to enhance OOD detection in models that are not specifically trained on ID data. We will ensure that these distinctions and innovations are clearly articulated in the revised version of our paper. --- References: - [1] ReAct: out-of-distribution detection with rectified activations. - [2] Extremely simple activation shaping for out-of-distribution detection. - [3] Dice: leveraging sparsification for out-of-distribution detection. --- Rebuttal 3: Comment: Thank you for your additional clarifications. Based on your responses to all the reviewers, I think the paper’s contributions are sufficient for acceptance. I recommend that the authors expand the discussion of related works on OOD Detection in Computer Vision [1,2,3] in the revised version to further strengthen the paper. I will raise my score to 6. [1] How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection? [2] Learning with Mixture of Prototypes for Out-of-Distribution Detection [3] Energy-based Hopfield Boosting for Out-of-Distribution Detection --- Rebuttal Comment 3.1: Title: Gratitude for Your Valuable Feedback and Support Comment: We sincerely appreciate the reviewer’s thoughtful comments and the decision to raise the score. We will ensure that the discussion on related works in OOD detection, particularly the references provided, is expanded and integrated into the revised version of our paper. This will further strengthen the context and positioning of our contributions. Thank you again for your valuable feedback and support throughout the review process.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Exploring Molecular Pretraining Model at Scale	Accept (poster)	Summary: # Summary This paper presents Uni - Mol2, a molecular pretraining model, and systematically investigates the scaling law within molecular pretraining models. # Contributions 1. The largest dataset for molecular pretraining: Curated a dataset of approximately 884 million 3D conformations for pretraining. 2. Scaling law: It is the first time to demonstrate the scaling law of molecular pretraining and its impact on downstream task performance. 3. Significant improvement: Uni - Mol2 is the SOTA model and demonstrates consistent improvement in downstream task performance with increasing model parameters. Strengths: 1. Presents Uni - Mol2, a novel molecular pretraining model that leverages a two - track transformer to integrate features at multiple levels. 2. Conducts the first exploration of the scaling law in molecular pretraining models. 3. Curates a large dataset of approximately 884 million 3D conformations, providing a solid foundation for training large - scale molecular models. Weaknesses: Limited details of pretraining hyper-parameters: The paper primarily focuses on analyzing the power-law relationships, but does not give details of the pretraining hyper-parameters. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. What is the computational cost and time complexity of training the Uni - Mol2 model with different parameters on the large dataset? 2. Could you please provide more details on how the temperature - based sampling method affects the balance of the training dataset and the performance of the model? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: See the weakness and questions plz. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We extend our sincere thanks to Reviewer ER2p for the positive evaluation and the thoughtful time invested in reviewing our manuscript. Your encouraging feedback is greatly appreciated, and we have carefully addressed each of your comments in the detailed responses provided below. Response to Weaknesses 1. Limited Details of Pretraining Hyper-Parameters Thank you for pointing out the lack of detailed information regarding the pretraining hyper-parameters in the current manuscript. We have outlined the details of the pre-training hyperparameters in Section 3.3 of the paper, which includes comprehensive details on the hyper-parameters used, including learning rates, batch sizes, number of epochs, optimizer settings, and regularization techniques. If there are specific details or additional aspects you would like us to clarify, please let us know, and we would be happy to discuss them further. Response to Questions 1. Computational Cost and Time Complexity The time complexity of training the Uni-Mol2 model primarily depends on the number of parameters, dataset size, and computational resources. We provide the details of the training process, including hardware specifications, training duration, and computational resources used in the general rebuttal. We will subsequently integrate this information into the manuscript. 2. Impact of Temperature-Based Sampling In our current study, we faced a significant imbalance in the skeletal clustering results of the data. Specifically, the first two categories had a disproportionately high proportion of molecules, while the subsequent categories had much lower proportions. Our dataset contains 73 million scaffolds, with the top two scaffolds alone accounting for 2.78% of the total number of molecules. In contrast, the last 25 million scaffolds together represent only 2.8% of the molecules. To address this issue and ensure that rare scaffolds can be sampled, we referred to some sampling techniques [1][2] used in language models. We adopted a temperature-based sampling approach that reduces the sampling probability of selecting the top two scaffolds by a factor of 19,000, which in turn increases the chances of sampling rarer scaffolds. However, due to current computational resource constraints, we have not been able to conduct extensive ablation studies on different sampling strategies. Nevertheless, based on recent advancements in Large Language Models (LLMs) [3][4][5], it is evident that various data sampling methods and the mixture of data types present a promising direction for future research. We appreciate your suggestion and consider it a valuable area for further investigation. References [1] Wang X, Tsvetkov Y, Neubig G. Balancing training for multilingual neural machine translation[J]. arXiv preprint arXiv:2004.06748, 2020. [2] Fan A, Bhosale S, Schwenk H, et al. Beyond english-centric multilingual machine translation[J]. Journal of Machine Learning Research, 2021, 22(107): 1-48. [3] Shao Y, Li L, Fei Z, et al. Balanced Data Sampling for Language Model Training with Clustering[J]. arXiv preprint arXiv:2402.14526, 2024. [4] Ye J, Liu P, Sun T, et al. Data mixing laws: Optimizing data mixtures by predicting language modeling performance[J]. arXiv preprint arXiv:2403.16952, 2024. [5] Gu J, Yang Z, Ding C, et al. CMR Scaling Law: Predicting Critical Mixture Ratios for Continual Pre-training of Language Models[J]. arXiv preprint arXiv:2407.17467, 2024. --- Rebuttal Comment 1.1: Title: Nice Rebuttal! Comment: I really like your rebuttal and the paper. You've added some training details, and I'm curious about how you ultimately chose the learning rate scheduler? For example, I've noticed some interesting recent papers on training dynamics, such as: MiniCPM: https://arxiv.org/abs/2404.06395 Qwen technical report: https://arxiv.org/abs/2309.16609 Could you discuss or possibly explore the impact of adopting these new learning rate schedulers and include this part in your paper? --- Reply to Comment 1.1.1: Comment: Thank you for your kind words and your interest in our work. We appreciate your attention to the training details and your insightful question regarding the choice of the learning rate scheduler. Based on our observations, selecting an appropriate learning rate scheduler often lacks a straightforward or universally applicable solution. This process predominantly relies on experiential knowledge. We chose the polynomial decay scheduler based on prior work [1], as we found that, with certain hyperparameter adjustments, it effectively optimized Uni-Mol2 to convergence. We also noticed that the cosine scheduler has been widely adopted in the training of many large language models [2, 3, 4]. In fact, we conducted some preliminary experiments using Uni-Mol2 84M to compare the cosine scheduler with the polynomial decay scheduler employed in this paper. Our preliminary experimental results indicate that in the pretraining task, the performance of these two schedulers is consistent. We will supplement some additional results in this part as required in the future. Thank you again for your valuable suggestions. References [1] Zhou G, Gao Z, Ding Q, et al. Uni-mol: A universal 3d molecular representation learning framework[J]. 2023. [2] Bai J, Bai S, Chu Y, et al. Qwen technical report[J]. arXiv preprint arXiv:2309.16609, 2023. [3] Dubey A, Jauhri A, Pandey A, et al. The Llama 3 Herd of Models[J]. arXiv preprint arXiv:2407.21783, 2024. [4] Kaplan J, McCandlish S, Henighan T, et al. Scaling laws for neural language models[J]. arXiv preprint arXiv:2001.08361, 2020.	Summary: In this work, the authors propose Uni-Mol2 , a molecular pretraining model that leverages a two-track transformer to integrate features at the atomic level, graph level, and geometry structure level. The authors also investigate the scaling law within molecular pretraining models, characterizing the power-law correlations between validation loss and model size, dataset size, and computational resources. Strengths: The work investigates an important problem in chemistry and machine learning, and proposes a useful LLM model. The figures are very informative, especially Fig. 2 on the architectural pipeline. Weaknesses: In the Related Work, it would also be relevant if the authors can discuss graph neural network (GNN) models that have recently been used to aid in improved representation learning for molecular graphs e.g., YieldGNN. Furthermore, GNNs are significantly more computationally efficient as compared to LLM-based models. This paper provides a survey on molecular representation learning, and seems to be a relevant recent reference: Zhichun Guo, Kehan Guo, Bozhao Nan, Yijun Tian, Roshni G. Iyer, Yihong Ma, Olaf Wiest, Xiangliang Zhang, Wei Wang, Chuxu Zhang, and Nitesh V. Chawla. 2023. Graph-based molecular representation learning. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI '23). Article 744, 6638–6646. https://doi.org/10.24963/ijcai.2023/744 Technical Quality: 3 Clarity: 2 Questions for Authors: Can the authors more clearly describe (1) the contributions of their work, and (2) their design choice behind considering an LLM model as opposed to GNN model. Why not also consider a combination of both? What is the time complexity of training time of the LLM model. Also, how many GPU resources were used? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The work sufficiently describes the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank Reviewer XhCf for the detailed and constructive review. Below, we respond to all your comments. We appreciate the opportunity to clarify a key point at the outset: typically, when referring to LLM models, the term denotes large language models that primarily aim at next-token prediction within the natural language processing domain [1][2]. However, our manuscript builds upon a pre-trained model specifically tailored for the small molecule domain. Fundamentally, our model is a two-track transformer aimed at masked token prediction and coordinate denoising. Our research focuses on examining the scaling laws relevant to this category of molecular pre-training models. Hence, we believe that a more appropriate comparison would be between transformer-based models and GNN-based models within the small molecule domain. Response to Questions 1. Contributions, Design Choice of Uni-Mol2 and Comparision with GNN Our research's main contribution is to investigate the scaling laws relevant to this category of molecular pre-training models. Our results reveal power-law relationships between validation loss and model size, dataset size, and computational resources. Additionally, we observe consistent improvements in downstream tasks as the model size increases. We recognize the advantages of GNNs, particularly their efficiency in processing molecular graph structures and their strong capability to capture local relationships within a molecule. However, the locally connected graph fails to adequately represent long-range interactions between atoms. These long-range interactions are crucial in molecular representation learning (MRL). In contrast, transformer-based models have shown exceptional performance in various tasks within the molecular domain, demonstrating remarkable representation capabilities. Some researchers even view transformers as a form of fully-connected GNN [3]. Additionally, recent advancements in transformer engineering optimization have further enhanced their effectiveness [7]. 2. Time Complexity and GPU Resources The time complexity of training the Uni-Mol2 model primarily depends on the number of parameters, dataset size, and computational resources. We provide the details of the training process, including hardware specifications, training duration, and computational resources used in the general rebuttal. We will subsequently integrate this information into the manuscript. Response to Weaknesses 1. Discussion of Graph Neural Network (GNN) Models We appreciate the suggestion to include a discussion on Graph Neural Network (GNN) models, such as YieldGNN [9], in the Related Work section. Specifically, we will reference the survey on graph-based molecular representation learning by Guo et al. (2023) [8] and highlight the advantages of GNNs, particularly in terms of representation power and computational efficiency. We believe that including this related work will significantly enhance the completeness and context of our manuscript. References [1] Achiam J, Adler S, Agarwal S, et al. Gpt-4 technical report[J]. arXiv preprint arXiv:2303.08774, 2023. [2] Reid M, Savinov N, Teplyashin D, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context[J]. arXiv preprint arXiv:2403.05530, 2024. [3] Min E, Chen R, Bian Y, et al. Transformer for graphs: An overview from architecture perspective[J]. arXiv preprint arXiv:2202.08455, 2022. [4] Zhou, Gengmo, et al. "Uni-mol: A universal 3d molecular representation learning framework." (2023). [5] Yu, Q., Zhang, Y., Ni, Y., Feng, S., Lan, Y., Zhou, H., and Liu, J. Unified molecular modeling via modality blending. arXiv preprint arXiv:2307.06235, 2023. [6] Luo S, Chen T, Xu Y, et al. One transformer can understand both 2d & 3d molecular data[C]//The Eleventh International Conference on Learning Representations. 2022. [7] Dao T, Fu D, Ermon S, et al. Flashattention: Fast and memory-efficient exact attention with io-awareness[J]. Advances in Neural Information Processing Systems, 2022, 35: 16344-16359. [8] Zhichun Guo, Kehan Guo, Bozhao Nan, Yijun Tian, Roshni G. Iyer, Yihong Ma, Olaf Wiest, Xiangliang Zhang, Wei Wang, Chuxu Zhang, and Nitesh V. Chawla. 2023. Graph-based molecular representation learning. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI '23). Article 744, 6638–6646. https://doi.org/10.24963/ijcai.2023/744 [9] Shi R, Yu G, Huo X, et al. Prediction of chemical reaction yields with large-scale multi-view pre-training[J]. Journal of Cheminformatics, 2024, 16(1): 22. --- Rebuttal Comment 1.1: Comment: Thanks for your response and clarifications. I also look forward to seeing the new revisions in the future version of the paper, which will improve quality of this work. I will increase my score by one point. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your feedback, and will incorporate the revisions in the next version of the manuscript. Thank you for your consideration and for raising the score, your support is greatly valued.	Summary: This paper studies large scale pretraining for molecule. The authors compose the largest molecular pretraining dataset, Uni-Mol2, and train the largest molecular pretraining model with 1.1B parameters. This paper also fits a scaling law in this domain that can accurately predict losses on a validation set. Downstream performance on QM9 and COMPAS-1D datasets demonstrate the proposed model can outperform previous approaches. Strengths: 1. This paper creates the largest molecular pretraining dataset and the authors indicate the intent to open-source the dataset in the checklist. 2. The authors pretrain various scales of molecular models up to 1.1B parameters on the new dataset, a scale that has not been studied previously. 3. The results on downstream tasks surpass previous works. Weaknesses: 1. On the downstream numbers reported in the paper, the scores do not scale well with the model size for most of them. For example, on most of the properties Uni-Mol2 1.1B is not apparently better than much smaller models like Uni-Mol2 310M, and on COMPAS-1D the Uni-Mol2 1.1B is even comparable to Uni-Mol2 84M in some cases – it is likely that I am not familiar with these datasets and do not perceive the score difference well, yet I am wondering whether a gain like 0.0001 is really meaningful and how can we know it is statistically significant? Is this because the tasks are too simple or the large 1.1B model is not trained properly? If the reason is the former, I think more difficult tasks should be included to demonstrate the benefit of larger models – otherwise the impact of this paper on scaling up model sizes is limited. 2. The paper’s presentation should be improved. For example, in the results tables it is better to indicate the parameter size of the baselines (like Uni-Mol) to understand the effect of model sizes on the results. There are also some grammar errors such as Line 39 “Most” -> “most”. Line 65 “To” -> “to”. Technical Quality: 3 Clarity: 2 Questions for Authors: Related questions have been asked above. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have included a limitation section in Appendix C, on the prediction accuracy of the scaling law. However, I would like to see a discussion on scaling up models with the proposed approach when the experimental results have shown saturation with only 1.1B parameters – which may indicate that the proposed training losses cannot be used to train powerful larger and larger models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We deeply appreciate Reviewer buoE's careful review and thoughtful feedback. Your suggestions have greatly contributed to improving our manuscript. We respond to your comments and questions in detail below. Response to Weaknesses 1. Performance Improvements We acknowledge your concern regarding the insufficient improvement of Uni-Mol2 in certain downstream tasks. We have provided additional clarification for our downstream experimental results and added some new downstream experiments. Overall, Uni-Mol2 1.1B demonstrates a considerably greater performance improvement compared to Uni-Mol2 84M and Uni-Mol2 310M in downstream tasks. For the aEA and aIP tasks in the COMPAS-1D dataset, Uni-Mol2 1.1B with graph features demonstrates a more noticeable MAE reduction compared to Uni-Mol2 84M (0.96% -> 10.58% for aEA and 1.23% -> 3.90% for aIP). On the 10 downstream tasks of the QM9 dataset, Uni-Mol2 1.1B achieved an average improvement of 8.472% compared to Uni-Mol2 310M. Detailed explanations and results supporting these clarifications can be found in the general rebuttal section. 2. Presentation Improvements We appreciate the suggestions for improving the presentation of our manuscripts. We will revise the results tables to include the parameter sizes of baseline models, such as Uni-Mol, to better illustrate the impact of model size on performance. Additionally, we will correct the identified grammar errors and thoroughly review the manuscript to ensure clarity and accuracy. Response to Limitations We appreciate the reviewer's observation regarding the limitations of our proposed approach, particularly in relation to the observed saturation of experimental results with models containing 1.1 billion parameters. We agree that this finding suggests a potential limitation in the scalability of the proposed training losses. In response, we would like to clarify that the saturation observed may be due to the specific experimental setup and dataset limitations. While some experiments of the current results show a plateau at 1.1B parameters, it does not necessarily imply an inherent ceiling of the proposed method. We added some new extra experiments in the supplementary appendix; the increase in Uni-Mol2 1B compared with 310M is still significant and does not indicate that the results have reached saturation. Returning to the starting point of this paper, we primarily investigate the relation of validation loss as the model, data, and computation size increase, which is consistent with scaling laws. In the field of molecular pretraining, establishing appropriate pretraining loss to ensure the scaling effects of the model is indeed a compelling direction[1,2,3]. Given our current result, we have ensured stable training, and this has led to consistently superior results in downstream tasks. We will discuss these considerations in the manuscript to provide a more comprehensive understanding of the limitations and to further explore this direction in future work. References [1] Yang J, Zheng K, Long S, et al. MOL-AE: Auto-Encoder Based Molecular Representation Learning With 3D Cloze Test Objective[J]. bioRxiv, 2024: 2024.04. 13.589331. [2] Ni Y, Feng S, Ma W Y, et al. Sliced Denoising: A Physics-Informed Molecular Pre-Training Method[J]. arXiv preprint arXiv:2311.02124, 2023. [3] Feng, Shikun, et al. "UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning." arXiv preprint arXiv:2405.10343 (2024). --- Rebuttal Comment 1.1: Comment: Thanks for your response! I appreciate the authors for the added the results and it is encouraged to include those results in the next revision. These new results mitigate my concern on the improvement, though in many cases the improvement still seems small and such scaling does not seem to be universally successful. Thus I would like to raise by score a bit to 5. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your positive feedback and the increase in the score. We are glad that the additional results we provided have helped to mitigate some of your concerns regarding the improvements in our approach. We fully concur with your suggestion to incorporate these results in the next revision.	Summary: This paper studies the pretraining task on molecular domain. The main contribution includes extending the size of pretraining dataset, scaling the model to 1.1B size, investigating pretraining scaling law behavior. The evaluation of the pretrained model is conducted for some molecular property prediction task. Strengths: 1. Extend the pretraining dataset size, including both strings, graphs, and 3D conformation. 2. While the model architecture is not new and is based on existing work, the effort to scale it up to 1B model is great for the current literature. 3. The scaling law behavior is interesting, while it is different from language model. 4. While the downstreaming tasks are limited, consistent performance improvement is shown with increasing model size. Weaknesses: 1. The evaluation of the pretrained model is limited, it would be great if the author can show more diverse and complicated chemistry tasks. Better include some domain-intense tasks. 2. The model training time is not reported, the author should include enough detail for reproducing the experiment. 3. Pretraining tasks/losses have some hyperparameter, the author should include ablation study result for these choices. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For these downstreaming evaluation tasks, the improvement of performance from small model to large model is not that high, (for example the number in Table 5 and Table 6) can the author provide some justification about these numbers? 1.1B model is significantly costly than 84M, so we would expect a larger improvement. 2. Can the author include more downstreaming tasks? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful to Reviewer V7dw for the thorough review and insightful comments. Below, we have carefully considered your feedback and provided detailed responses to each point. Response to Weaknesses: 1. Limited Evaluation of the Pretrained Model We begin by highlighting the promising results achieved in the fields of small molecules and photoelectrics, while also acknowledging the limitations of our evaluation tasks. Due to constraints on computational resources, we focused on a specific set of molecular property prediction tasks. We agree that including more diverse and complex chemistry tasks could provide a broader evaluation of the model's capabilities. To address this, we have added new Biogen ADME benchmark results to demonstrate the model's capabilities. In future work, we plan to incorporate additional domain-intensive tasks to better showcase the utility of our pre-trained model. 2. Model Training Time and Reproducibility Thank you for pointing out the omission of detailed information regarding the model training time. We have outlined the details of the pre-training hyperparameters in Section 3.3 of the paper. Additionally, we provide comprehensive details about the training process, including hardware specifications, training duration, and computational resources used in the general rebuttal. We will subsequently integrate this information into the manuscript. If there are specific details or additional aspects you would like us to clarify, please let us know, and we would be happy to discuss them further. 3. Ablation Study on Pretraining Tasks/Losses Hyperparameters Thank you for your valuable suggestion. We have indeed adjusted the hyperparameters of the loss function during model training to ensure the stability of the training process. In the end, we set the weight for each loss component to 1 across all model sizes. However, systematically conducting ablation studies on the loss function at expanded model and dataset scales would require substantial computational resources, which are currently beyond our capacity for comprehensive exploration in the short term. Nevertheless, we recognize the importance of this research direction and will consider it a significant area for future investigation. Response to Additional Concerns: 1. Justification for Performance Improvement We fully understand your concern regarding the relatively modest performance improvement observed when transitioning from the small model (310M) to the large model (1.1B) in the downstream evaluation tasks, as illustrated in Table 5 and Table 6. To address this, we have provided further clarification of the results and supplemented our study with additional experiments. The detailed findings and explanations have been included in the general rebuttal section. 2. Inclusion of More Downstream Tasks Yes, expanding the range of tasks will allow for a more comprehensive evaluation of the model's performance across different aspects. We will explore more applications in this area. --- Rebuttal Comment 1.1: Comment: Thank you for answering my questions. I'm looking forward to see your model achieve some substantial influence for downstreaming tasks in future. I will keep the score. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your thoughtful questions and your engagement with our work. We are encouraged by your interest in seeing our model have a substantial impact on downstream tasks in the future. Thank you for your valuable feedback and for taking the time to review our paper.	Rebuttal 1: Rebuttal: ## General Rebuttal (R IDs: R1=V7dw, R2=buoE, R3=XhCf, R4=ER2p) We thank the reviewers for the detailed and helpful reviews. Next, we address the main concerns from reviewers. 1. Time Complexity and GPU Resources We utilized a computational cluster comprising 64 NVIDIA A100 GPUs, each equipped with 80GB of HBM2 memory. The GPUs were interconnected via a high-speed Nvidia Infiniband fabric, offering 400 Gbps bandwidth for inter-GPU communication. The details of each model size are listed as follows. \| Params \| Compute Resouce(GPUs) \| Training Time(GPU hours) \| \|-----------------\|---------------\|-----------\| \| 84M \| 32 \| 2585.6 \| \| 164M \| 32 \| 5120 \| \| 310M \| 32 \| 7680 \| \| 510M \| 64 \| 13824 \| \| 1.1 B \| 64 \| 30720 \| 2. Performance Improvement Concern Similar to other works on scaling laws [1][2], our paper explores the power-law relationship between the validation loss of pre-trained models in the molecular domain and the scales of data, model, and computation. This represents the first validation at the scale of billions of data points and pre-trained model parameters. On the other hand, while our proposed Uni-Mol2 consistently outperforms other baseline models across various downstream tasks, the performance on diversified downstream tasks is influenced by several factors such as data partitioning, data quality, and label noise levels [3][4][5]. Consequently, in some downstream tasks, there isn't always a substantial performance improvement as the model scale increases because the data quality and quantity may exert a more significant impact on the final results compared to the model scale. A typical example is the aEA and aIP tasks in the COMPAS-1D dataset, where adding features to Uni-Mol2 1.1B leads to significant performance improvements, achieving an average 14% improvement compared with Uni-Mol. Regarding the differences in model performance on the QM9 dataset that you mentioned, we have supplemented four additional property prediction tasks (U0, U, G, and H) in the supplementary material. Overall, when comparing Uni-Mol2 310M with Uni-Mol2 1.1B on the QM9 dataset, eight of ten tasks showed improvements exceeding 2.7% with Uni-Mol2 1.1B, while two tasks achieved improvements above 26%. It is worth noting that in the fine-tuning process for QM9, we only used atomic features and conformational features. We believe this is one of the potential reasons contributing to the observed convergence in model performance during scaling up Uni-Mol2. In summary, our experiments affirm that scaling models is effective and leads to noticeable improvements. The results clearly demonstrate that increasing model size yields significant performance gains in downstream tasks. References [1] Kaplan J, McCandlish S, Henighan T, et al. Scaling laws for neural language models[J]. arXiv preprint arXiv:2001.08361, 2020. [2] Hoffmann J, Borgeaud S, Mensch A, et al. Training compute-optimal large language models[J]. arXiv preprint arXiv:2203.15556, 2022. [3] Sultan A, Sieg J, Mathea M, et al. Transformers for molecular property prediction: Lessons learned from the past five years[J]. arXiv preprint arXiv:2404.03969, 2024. [4] Deng J, Yang Z, Wang H, et al. Unraveling Key Elements Underlying Molecular Property Prediction: A Systematic Study[J]. arXiv preprint arXiv:2209.13492, 2022. [5] Martinez-Mayorga K, Rosas-Jiménez J G, Gonzalez-Ponce K, et al. The pursuit of accurate predictive models of the bioactivity of small molecules[J]. Chemical Science, 2024, 15(6): 1938-1952. Pdf: /pdf/8ee4297181bb06c9544fec540adf775811d8e922.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models	Accept (poster)	Summary: This work aims to tackle the image-to-LiDAR contrastive learning problem for LiDAR-based point cloud segmentation. Previous approaches designed the cross-modal contrastive learning objective for model pretraining, using superpixels and superpoints as guidance. In this work, the authors observe that the superpixel-driven contrastive loss tends to involve ‘’self-conflict’’ issues during representation learning. A weakly-supervised contrastive distillation method is proposed, which generates semantic superpixels/superpoints using the Segment Anything Model (SAM). Additionally, to balance the imbalanced class distributions of LiDAR scene categories during representation, a density and category-aware sampling strategy is proposed to adjust the sampling probabilities of different anchor points using the weak semantic labels. The overall framework is named OLIVINE, which adopts three optimization objectives: - Weakly-supervised contrastive distillation using coarse semantic labels to identify positive pairs by category. - Self-supervised contrastive distillation applied to randomly sampled point-pixel pairs. - A regularization framework based on the von Mises-Fisher (vMF) distribution to ensure semantic consistency. The proposed OLIVINE method is evaluated on the nuScenes, SemanticKITTI, and KITTI object detection datasets. The results exhibit a consistent improvement of the proposed method compared to existing approaches. Strengths: (+) This work aims to improve the image-to-LiDAR self-supervised representation learning problem on LiDAR-based point cloud datasets, which is one of the current research hotspots, especially for applications related to autonomous driving and robotics. (+) The proposed method has exhibited promising performance on mainstream benchmarks, including nuScenes linear probing, nuScenes fine-tuning, SemanticKITTI fine-tuning, and KITTI object detection. Weaknesses: (-) The weakly-supervised contrastive distillation method has been used in previous literature, such as [R1] and [R2]. Adding semantic categories seems not to cause a major improvement over class-agnostic masks, as the Segment Anything Model is able to segment rather complete and semantically consistent objects and backgrounds. Additionally, using weak labels (which might be erroneous) could introduce additional errors during pretraining. (-) The motivation for using the von Mises-Fisher (vMF) distribution to enforce consistency regularization for image-to-LiDAR representation learning is not clear enough to demonstrate its superiority. A more detailed explanation and theoretical justification would strengthen this aspect of the work. (-) Compared to some of the most related works, for example, [R1] and [R3], the scale and depth regarding the experiments (for example, downstream fine-tuning on other datasets than SemanticKITTI) could be further enhanced. --- ### References: - [R1] Youquan Liu, et al. “Segment Any Point Cloud Sequences by Distilling Vision Foundation Models,” NeurIPS, 2023. - [R2] Ayça Takmaz, et al. “OpenMask3D: Open-Vocabulary 3D Instance Segmentation,” NeurIPS, 2023. - [R3] Gilles Puy, et al. “Revisiting the Distillation of Image Representations into Point Clouds for Autonomous Driving,” arXiv, 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: - Q1: As mentioned in Weakness 1, the semantic masks generated by the Segment Anything Model could inevitably involve errors (e.g., wrong segmentation results). How do the authors handle the propagated errors during image-to-LiDAR representation learning? - Q2: As mentioned in Weakness 2, could the authors provide more details on the hyperparameter settings for the vMF distribution and the reasoning behind their chosen values? Adding a more detailed explanation and theoretical justification would be even better. - Q3: As mentioned in Weakness 3, having more thorough experimental analyses on other LiDAR-based point cloud datasets, such as SemanticPOSS, Waymo, SynLiDAR, etc., could further consolidate the findings and conclusions drawn in the manuscript. - Q4: As most 2D and 3D representation learning approaches (MoCo, SimCLR, Seal, etc.) do, having empirical analyses of models under out-of-distribution datasets is recommended. - [Minor]: The computational cost of the proposed multi-modal contrastive distillation approach is not thoroughly analyzed, which is crucial for real-time applications in autonomous driving. - [Minor]: The generalizability of OLIVINE to other types of sensors (for example, hybrid-solid LiDARs) or environments (for example, off-board environments) beyond the evaluated datasets is not discussed. - [Minor]: “NuScenes” should be revised to “nuScenes”. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors mentioned "Semantic Label Accuracy" as one of their limitations. As also discussed in Weakness 1, more analyses are needed to address the impact of inaccuracies in the weak labels generated by the Segment Anything Model. These inaccuracies could propagate errors during the image-to-LiDAR representation learning process, potentially affecting the overall performance of the proposed method. Additionally, while the von Mises-Fisher distribution is used for consistency regularization, the motivation and theoretical foundation for its use are not fully elaborated. A deeper exploration of its advantages and potential drawbacks in this context would be beneficial. The computational cost associated with the multi-modal contrastive distillation approach is another important aspect that is not thoroughly analyzed. For practical applications, especially in real-time scenarios such as autonomous driving, it is crucial to understand the resource requirements and efficiency of the proposed method. Lastly, the scalability and generalizability of OLIVINE to other sensor types and different environments have not been extensively discussed. Exploring its applicability in diverse settings and with various sensor configurations would provide a more comprehensive evaluation of its robustness and versatility. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your time and effort in reviewing our submission and valuable comments. In the following, we will address your concerns and correct the potential misunderstandings. Q: The weakly-supervised contrastive distillation method has been used in previous literature [R1, R2]. A: We believe there may be a misunderstanding regarding the mentioned methods. Seal [R1] generates semantically coherent superpixels for distinct objects and backgrounds in the 3D scene. However, it does not infer semantic labels or use them to supervise contrastive distillation. Consequently, superpoints and superpixels within the same category may still be mistakenly considered negative pairs during pretraining. [R2] is NOT relevant to contrastive distillation or weakly-supervised learning. Q: Adding semantic categories seems not to cause a major improvement over class-agnostic masks, as the SAM is able to segment rather complete and semantically consistent objects and backgrounds. A: Although Seal [R1] uses VFMs to generate semantically coherent superpixels, it can still mistakenly treat superpoints and superpixels of the same category as negative pairs during contrastive distillation. Our method explicitly defines the points and pixels with the same semantic label as positive pairs during weakly-supervised contrastive learning. Besides, our method can achieve better performance with weak labels generated by stronger VFMs. We refer you to Table M1 of the uploaded PDF file for the additional results. Q: Using weak labels could introduce additional errors during pretraining. A: We acknowledge there is a trade-off. While the weak labels might be erroneous, they enable semantic-guided image-to-LiDAR contrastive distillation and indeed yield state-of-the-art performance on downstream tasks. Similarly, the superpixels widely used in previous image-to-LiDAR knowledge transfer methods can also be inaccurate but effective. Q: The motivation for using the vMF distribution is not clear enough. A more detailed explanation and theoretical justification would strengthen this aspect. A: Thank you for your valuable suggestion. We provide more explanations as follows. - The representation of samples in the same class can vary significantly across different batches during the contrastive distillation, so the model will struggle to learn stable semantic features. By making point features of the same class closely aligned, our method aims to create a more consistent and structured feature space. - The vMF distribution is defined on a hypersphere, making it well suited for directional data in feature space. The concentration parameter can be dynamically adapted during training to refine the feature alignment process. Early in training, a lower $\kappa$ might allow for more exploration, while later stages can benefit from higher $\kappa$ to solidify the learned representations. - Due to the space limitation of each response, we will provide more theoretical justification in another window. Q: The experiments could be enhanced on more datasets. A: Following your valuable suggestion, we conducted experiments on more datasets. We refer you to the Table M2 in the uploaded PDF file. Q: How do the authors handle the propagated errors during image-to-LiDAR representation learning? A: Thank you for raising this insightful point. We acknowledge that the inaccuracy of the labels generated by the SAM is a limitation of the current pipeline. We are actively developing a new label disambiguation module for future work. This module will utilize the learned feature similarities to refine the coarse labels. We believe that weak labels can help mitigate self-conflict, while structured semantic representations can assist in refining these labels. Together, they mutually reinforce each other, ultimately leading to more robust and accurate representations. Q: Could authors provide details on the hyperparameter settings for the vMF distribution and the reason behind chosen values? A: Regarding the vMF distribution, we did not set many hyperparameters. The mean direction and the concentration parameter are learned from the statistical values of features via the EMA (Exponential Moving Average) algorithm. The smoothness coefficient for the moving average is empirically set to 0.0001. Q: As most 2D and 3D representation learning approaches do, having empirical analyses of models under out-of-distribution datasets is recommended. A: Following your valuable suggestion, we added experiments on nuScenes-C datasets. We refer you to the Table M3 in the uploaded PDF file. Q: The computational cost of the proposed approach is not analyzed, which is crucial for real-time applications. A: Thanks for your comments. Our approach only provides pre-trained weights, which do not affect the inference speed of the model on downstream tasks. As shown in the table below, our OLIVINE does not require obviously more GPU memory or training time compared to other pre-training methods. \|Method\|GPUMemory (GB)\|TrainingTime (Hour)\| \|:-\|:-:\|:-:\| \|PPKT\|7\.6\|35\.7\| \|SLidR\|10\.7\|38\.9\| \|OLIVINE\|8\.1\|36\.5\| Q: The generalizability of OLIVINE to other types of sensors or environments is not discussed. A: Following your suggestion, we have supplemented experiments on another six datasets, which demonstrate the generalizability of OLIVINE to some extent (see Table M2 of the uploaded PDF file). Since the computational resources are limited, we will further enhance the experiments part after the rebuttal. Q: "NuScenes" should be "nuScenes". A: Thanks for your careful review. We will correct the typo in the final version. References:\ [R1] Liu et al. Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. NeurIPS2023.\ [R2] Takmaz et al. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. NeurIPS2023. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for putting tremendous effort into addressing the raised concerns. I have read the authors' rebuttal, as well as other reviewers' comments, I believe key issues of this work regarding the following aspects have been addressed or partially addressed: - The motivation has been re-stated and is now more straightforward. - The scale of experiments has been largely improved; a substantial amount of downstream tasks on a diverse set of datasets were added, which provide a more comprehensive and convincing evaluation of the proposed method against previous methods. - Several clarifications regarding technical details were provided, which resolved the related issues. In addition to the above modifications, the authors also attempted to provide some theoretical analyses. However, since I am not an expert in machine learning theory, I leave more room to the ACs and other reviewers to validate the correctness of these theoretical analyses. One key concern remaining is that the authors may over-claim the contribution of using "semantic superpixels" over the "class-agnostic superpixels". As stated in the previous review: the use of semantic categories seems not to cause a major improvement over class-agnostic masks. Therefore, the authors are suggested to re-elaborate the claim on this aspect to avoid possible "over-claim" issues. Taking into consideration the authors' rebuttal and other reviewers' comments, I would like to upgrade the rating from Borderline Reject to Borderline Accept. Meanwhile, I am looking forward to more discussions with the authors and other reviewers during the discussion period. --- Rebuttal 2: Title: Theoretical Perspectives (Part1) Comment: Proposition 1: The features of each class $k$ can be modeled as a von Mises-Fisher (vMF) distribution. This means that for class $k$, the feature vectors $g_i$ lie on a unit hypersphere and are centered around a mean direction $\mu_k$ with a concentration parameter $\kappa_k$. Justification: To show that the features of each class can be effectively modeled by a vMF distribution, we use maximum likelihood estimation (MLE) to determine that the parameters $\mu_k$ and $\kappa_k$ are optimal for the given set of feature vectors. For a set of $M_k$ feature vectors $\\{g_i\\}_{i=1}^{M_k}$ from class $k$, the likelihood function for the vMF distribution is: $L(\mu_k, \kappa_k) = \prod_{i=1}^{M_k} f(g_i; \mu_k, \kappa_k) = \prod_{i=1}^{M_k} \mathcal{K}_{C}(\kappa_k) \exp(\kappa_k \mu_k^T g_i)$ Taking the natural logarithm of the likelihood function, we get the log-likelihood: $\log L(\mu_k, \kappa_k) = \sum_{i=1}^{M_k} \log f(g_i; \mu_k, \kappa_k) = M_k \log \mathcal{K}_{C}(\kappa_k) + \kappa_k \sum _{i=1}^{M_k} \mu_k^T g_i$ Substituting the expression for $\mathcal{K}_{C}(\kappa_k)$, we get: $\log L(\mu_k, \kappa_k) = M_k \left[ \log \left( \frac{\kappa_k^{C/2-1}}{(2\pi)^{C/2} I_{C/2-1}(\kappa_k)} \right) + \frac{\kappa_k}{M_k} \sum_{i=1}^{M_k} \mu_k^T g_i \right]$ $\log L(\mu_k, \kappa_k) = M_k \left[ (C/2-1) \log \kappa_k - \log I_{C/2-1}(\kappa_k) - \frac{C}{2} \log(2\pi) + \frac{\kappa_k}{M_k} \sum_{i=1}^{M_k} \mu_k^T g_i \right]$ To maximize the log-likelihood, we normalize $\mu_k$ by setting it to the normalized sum of the feature vectors: $\mu_k = \frac{\sum_{i=1}^{M_k} g_i}{\\|\sum_{i=1}^{M_k} g_i\\|}$ The derivative of the log-likelihood with respect to $\kappa_k$ is: $\frac{\partial \log L(\mu_k, \kappa_k)}{\partial \kappa_k} = M_k \left[ \frac{C/2-1}{\kappa_k} - \frac{I_{C/2}(\kappa_k)}{I_{C/2-1}(\kappa_k)} + \frac{1}{M_k} \sum_{i=1}^{M_k} \mu_k^T g_i \right]$ Setting this derivative to zero, we get: $\frac{C/2-1}{\kappa_k} - \frac{I_{C/2}(\kappa_k)}{I_{C/2-1}(\kappa_k)} + \frac{1}{M_k} \sum_{i=1}^{M_k} \mu_k^T g_i = 0$ Solving for $\kappa_k$, we obtain: $\kappa_k = \frac{\\|\sum_{i=1}^{M_k} g_i\\| (C - \\|\sum_{i=1}^{M_k} g_i\\|^2)}{1 - \\|\sum_{i=1}^{M_k} g_i\\|^2}$ This equation allows us to compute the concentration parameter $\kappa_k$ based on the alignment of the feature vectors. The concentration parameter $\kappa_k$ is larger when the distribution is more tightly clustered around the mean direction, and smaller when the features are more uniformly spread across the hypersphere. By maximizing the likelihood function for the vMF distribution, we have shown that the parameters $\mu_k$ and $\kappa_k$ can be estimated to model the distribution of feature vectors for each class. The mean direction $\mu_k$ denotes the central direction of the feature cluster, and the concentration parameter $\kappa_k$ controls the tightness of this clustering. Moreover, the way we estimate the parameters of vMF distribution in EMA is also consistent with the results of the above theoretical derivation. [Known issues] If the equations do not display correctly, please refresh the page or try using a different browser. --- Rebuttal 3: Title: Theoretical Perspectives (Part2) Comment: Proposition 2: The representation of samples in the same class can vary significantly across different batches during contrastive distillation, and semantic-guided consistency regularization helps to learn structured features. Justification: Without regularization, the representation of samples within the same class can vary significantly across different batches during contrastive distillation. This variance arises due to random sampling and the influence of negative samples in different batches. The weakly-supervised contrastive loss is defined as: $\mathcal{L}_{\mathrm{sup}} = - \frac{1}{M_s} \sum _{i=1}^{M_s} \log \left[ \frac{1}{\|A(i)\|} \sum _{a\in A(i)} \frac{\mathrm{exp}{(\langle\mathbf{G}^{\mathrm{3D}}_i,\mathbf{G}^{\mathrm{2D}}_a \rangle/\tau)}}{\sum _{j=1}^{M_s} \mathrm{exp}{(\langle\mathbf{G}^{\mathrm{3D}}_i,\mathbf{G}^{\mathrm{2D}}_j \rangle /\tau)}}\right]$ The features of negative samples $\mathbf{G}^{\mathrm{2D}}_j$ vary across batches, leading to different optimization paths for each mini-batch. This introduces variability in the learned representations $\mathbf{G}^{\mathrm{3D}}_i$ for samples of the same class $k$. When we do not use semantic-guided consistency regularization, the within-class variance for class $k$ across different batches is: $\sigma_W^2 = \frac{1}{\|B\|} \sum_{B} \frac{1}{M_k} \sum_{i=1}^{M_k^B} \\|g_i^k - \mu_k^B\\|^2$ For ease of reading, we use $g_i$ to refer to point feature $\mathbf{G}^{\mathrm{3D}}_i$. And $\mu_k^B$ is the mean feature vector for class $k$ in batch $B$. Due to the batch-wise variability in negative samples, $\mu_k^B$ can differ significantly across batches, leading to high within-class variance. By minimizing the KL divergence, we align feature vectors $g_i$ of class $k$ with the mean direction $\mu_k$, reducing the spread of feature vectors within the same class. The within-class variance with regularization is: $\sigma_W^2 = \frac{1}{K} \sum_{k=1}^K \frac{1}{M_k} \sum_{i=1}^{M_k} \\|g_i^k - \mu_k\\|^2$ Since $\mu_k$ is consistent across batches due to the regularization, the within-class variance is significantly reduced. This results in structured feature representations, enhancing class separability and improving performance in downstream tasks. --- Proposition 3: Learning structural representation during pretraining can benefit downstream tasks. Justification: Structured features are those well-aligned within the same class (low within-class variance $\sigma_W^2$) and well-separated between different classes (high between-class variance $\sigma_B^2$). With semantic-guided consistency regularization, feature vectors $g_i^k$ for class $k$ are closely aligned with the mean direction $\mu_k$. This alignment reduces the within-class variance $\sigma_W^2$. Weakly-supervised contrastive learning pushes apart feature vectors of different classes, increasing the separation between class means $\mu_k$. This increases the between-class variance $\sigma_B^2$. Take the linear classifier as an example, the decision boundary is determined by the separation between class means. Higher $\sigma_B^2$ and lower $\sigma_W^2$ result in clearer decision boundaries, reducing classification errors. Consider a simple linear classifier with weight vector $w$ and bias $b$. The decision function is: $f(x) = w^T x + b$ The decision boundary is given by: $w^T x + b = 0$ For well-structured features, the margin (distance between decision boundary and nearest samples) is maximized. The margin $ \gamma $ for class $ k $ can be expressed as: $\gamma = \frac{w^T (\mu_k - \mu)}{\\|w\\|}$ Higher between-class variance ($\sigma_B^2$) and lower within-class variance ($\sigma_W^2$) increase this margin, leading to better classification performance. [Known issues] If the equations do not display correctly, please refresh the page or try using a different browser. --- Rebuttal 4: Title: Authors' Response to Reviewer svRz Comment: Thank you for the positive feedback provided and the time devoted to this review. We are glad that our efforts have addressed your concerns. Next, we will address your remaining concerns. --- Comments: The authors may over-claim the contribution of using "semantic superpixels" over the "class-agnostic superpixels". Response: We believe there may be a misunderstanding regarding our proposed methods. We would like to clarify the following points to address your concerns: - We have not claimed that using semantic superpixels is our contribution. In fact, our method does not rely on superpixels at all, which is different from previous methods [R1-R4]. - Previous methods [R1-R4] use superpixels to pool 3D point features and 2D pixel features, learning with a superpixel-to-superpoint contrastive loss. In contrast, our method directly uses the features of individual points and pixels for contrastive distillation. - The semantic labels can be flexibly utilized in multiple aspects of the proposed method, such as weakly-supervised contrastive distillation, semantic-guided consistency regularization, and category-aware anchor point sampling. These aspects cannot be effectively addressed using only the class-agnostic superpixels. --- Comments: The use of semantic categories seems not to cause a major improvement over class-agnostic masks. Response: Extensive experiments demonstrate that this approach substantially outperforms superpixels-based (class-agnostic mask-based) methods [R1-R4] in various downstream tasks. - Our method achieves a significant improvement over superpixels-based pretraining methods on nuScenes and SemanticKITTI datasets. As shown in the table below, our method outperforms Seal [R1] by a significant margin, achieving an improvement of 5.14\% under the setting of linear probing. The full results are available in Table M1 of the uploaded PDF file. - Following your suggestions, we have added experiments on six additional LiDAR-based point cloud datasets and one out-of-distribution dataset. And our proposed OLIVINE consistently outperforms the superpixels-based methods on all datasets. For the full results, please refer to Tables M2 and M3 of the uploaded PDF file. \| Method \| LP \| 1% \| 5% \| 10% \| 25% \| 100% \| \| :-------- \| :----- \| :----- \| :----- \| :----- \| :----- \| :----- \| \| Random \| 8\.10 \| 30\.30 \| 47\.84 \| 56\.15 \| 65\.48 \| 74\.66 \| \| PPKT \| 35\.90 \| 37\.80 \| 53\.74 \| 60\.25 \| 67\.14 \| 74\.52 \| \| SLidR \| 38\.80 \| 38\.30 \| 52\.49 \| 59\.84 \| 66\.91 \| 74\.79 \| \| ST-SLidR \| 40\.48 \| 40\.75 \| 54\.69 \| 60\.75 \| 67\.70 \| 75\.14 \| \| HVDistill \| 39\.50 \| 42\.70 \| 56\.60 \| 62\.90 \| 69\.30 \| 76\.60 \| \| Seal \| 44\.95 \| 45\.84 \| 55\.64 \| 62\.97 \| 68\.41 \| 75\.60 \| \| Ours \| 50\.09 \| 50\.60 \| 60\.25 \| 65\.07 \| 70\.15 \| 76\.69 \| --- Thanks again for your diligence as a reviewer. It is with great pleasure communicating with you. Please feel free to share any additional comments or feedback on the manuscript. References:\ [R1] Image-to-lidar self-supervised distillation for autonomous driving data.\ [R2] Self-supervised image-to-point distillation via semantically tolerant contrastive loss.\ [R3] Segment Any Point Cloud Sequences by Distilling Vision Foundation Models.\ [R4] HVDistill: Transferring Knowledge from Images to Point Clouds via Unsupervised Hybrid-View Distillation.	Summary: Annotating point clouds with semantic classes can be expensive and time consuming. The authors of this work propose a new pretraining strategy for weakly supervising point cloud segmentation using image-based supervision (i.e., image-to-LIDAR knowledge transfer). The proposed approach improves upon traditional contrastive pretraining strategies by leveraging visual foundation models (VFMs, e.g., SAM) to provide weak supervision for associating LiDAR points with corresponding pixels that have matching semantic classes. The authors also model features using von Mises-Fisher distributions to further encourage feature semantic clustering, and improve upon sampling by incorporating spatial distances of points and class frequency. This approach shows impressive state-of-the-art performance on pretraining across two benchmark datasets (SemanticKITTI and nuScenes). The ablation study also carefully highlights the impact of each of the contributions. This work will likely serve as a healthy addition to the image-to-point knowledge transfer community. Strengths: 1. The authors identify an, evidently, common problem in point and pixel contrastive learning and address this issue with the proposed method. Namely, the authors (or perhaps Mahmoud et al. [36], see limitations section) recognize that prior works do not ensure semantic consistency when performing contrastive learning for image to point cloud knowledge transfer. This issue causes objects of the same class (e.g., car) to be pushed apart in feature space, simply because they are not part of the same super pixel. The authors address this using weakly contrastive distillation to ensure semantic consistency across anchor points. 2. State-of-the-art results by pretraining on nuScenes and SemanticKITTI across a wide range of annotation data limitations. 3. A detailed and thorough evaluation on multiple benchmarks and an extensive ablation study providing insightful results. The ablation study, in particular, shows the impact of weakly-supervised labels, separate projection heads, different distributions for modeling semantic features, and various sampling strategies. 4. The authors are tackling an interesting and challenging problem of improving knowledge transfer across modalities. This research area is of particular importance given the decreased interest and investment in annotating campaigns by the community, and increased interest in self-supervised methods. Weaknesses: 1. The related work section does not adequately differentiate this approach from prior works. - While the related work section does cite relevant works, it does not identify how the shortcomings of any of these works are addressed in this paper. Moreover, the related work section does not isolate how this paper is different, unique, or better than any of the existing approaches at image to point cloud knowledge transfer. - In particular, I found Mahmoud et al.'s [36] approach, for feature similarity from pixels to points and class balancing, sharing many commonalities with the proposed approach, hence, a detailed comparison may be warranted. - Liu et al. [33] also leverage VFMs like SAM for semantic segmentation to improve image to point cloud knowledge transfer, which is strikingly similar to the proposed approach. Detailed comparisons would greatly help clarify these commonalities, and it would strengthen the reader's confidence in the novelty of the proposed approach. Minor: - L148: Which existing methods make the semantic unmatching mistake? While this may have been briefly mentioned [36] in the introduction, there was no clear statement with multiple cited works to support this claim. Consider citing these (uncited) prior works to provide evidence for this claim. - Tables 2 and 3 could be combined, it seems somewhat unnecessary to keep them separated. - Typographical/grammatical errors: L90, L155, etc. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. What is the impact of using Grounded SAM vs other VFMs for this approach? 2. Which existing methods make the semantic unmatching mistake mentioned in L148? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes, the authors clearly describe the limitations of the approach as it pertains to the (1) accuracy of the pseudo-labels derived from the VFM, (2) the diversity of the training data impacting environment adaptation, and (3) the dependency on highly calibrated cameras and LiDAR sensors to ensure knowledge transfer. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your time and effort in reviewing our paper, the valuable comments, and the favorable recommendation. Q: To differentiate this approach from prior works. A: We agree with you that it's necessary to highlight the shortcomings of previous works and the novelty of OLIVINE in the related work section. Here, we would like to clarify the following points: - Previous works [R1-R5] have not solved the self-conflict problem properly. Especially, Seal [R4] generates semantically coherent superpixels for distinct objects and backgrounds in the 3D scene. However, the superpoints and superpixels with the same category may still be mistakenly considered negative pairs during contrastive learning. By contrast, our method explicitly defines the points and pixels with the same semantic labels as positive pairs during weakly-supervised contrastive learning. - Our pipeline performs knowledge distillation on two levels: self-supervised and weakly-supervised contrastive learning. To achieve this, we develop two different heads in both the image and point cloud branches to decouple the learned representation. Previous methods [R1-R5] have only attempted self-supervised contrastive distillation and have not explored using labels to guide contrastive distillation. - The representation of samples in the same class can vary significantly across different batches during the contrastive distillation, so the model will struggle to learn stable semantic features. By making point features of the same class closely aligned, our method aims to create a more consistent and structured feature space. - Existing methods [R2-R5] are highly dependent on the generated superpixels. Superpixels balance asymmetries between areas with denser coverage of points and sparser areas in the contrastive loss. However, we do not need this process at all and ensure a uniform representation of both spatial and categorical dimensions by employing a novel sampling strategy. Q: A detailed comparison with Mahmoud et al [36]. A: Thanks for your suggestion. The main differences between ours and ST-SLidR [R3] are: - ST-SLidR [R3] reduces the contribution of false negative samples based on superpixel-to-superpixel similarity, using 2D self-supervised features to determine semantic similarities between superpixels. By contrast, our method directly estimates the semantic labels of images with VFMs, and defines pixels and points with the same label as positive pairs. - Regarding class balancing, ST-SLidR [R3] assigns higher weights to over-represented anchors that exhibit high similarities to most negative samples. By contrast, our approach directly adjusts the sampling probability of anchor points using easily accessible semantic labels. In summary, our OLIVINE offers a more direct and effective way to mitigate the effect of false negative samples and class imbalance. Q: Liu et al. [33] also leverage VFMs. Detailed comparisons help clarify these commonalities. A: Thanks for your suggestions to compare Seal [R4] and OLIVINE. We would like to clarify the following points: - To avoid over-segmenting semantically coherent areas, Seal [R4] generates superpixels using VFMs instead of the traditional method SLIC. In contrast, our method does not rely on superpixels. Although we also use VFMs, we leverage them to obtain coarse semantic labels for fine-grained contrastive distillation. - In method Seal [R4], the superpoints and superpixels with the same category may still be mistakenly considered negative pairs during contrastive learning. Our method explicitly defines the points and pixels with the same semantic labels as positive pairs during weakly-supervised contrastive learning. - The semantic labels generated by VFMs, rather than superpixels, can be flexibly utilized in multiple aspects of the knowledge transfer process, such as weakly-supervised contrastive distillation, semantic-guided consistency regularization, and category-aware anchor point sampling. These aspects cannot be effectively addressed using only the class-agnostic superpixels. Q: L148: Which methods make semantic unmatching mistake? A: Thank you for your valuable feedback. Existing methods [R1, R2, R4, R5] may mistakenly treat unmatched (super)points and (super)pixels in the same category as negative pairs during contrastive distillation. We will cite these methods to support this claim in the revised manuscript to provide clear evidence. Q: Tables 2 and 3 could be combined. Typo errors... A: We appreciate your attention to detail. We have combined Tables 2 and 3 to streamline the presentation and carefully corrected the typo errors. Q: Impact of Grounded SAM vs other VFMs for this approach? A: Thanks for your question. Our response is as follows: - Grounded SAM supports text prompts by combining Grounding DINO and SAM. Other VFMs that enable text prompts can also be applied in OLIVINE. - The precision of the semantic labels significantly impacts the effectiveness of OLIVINE. Stronger VFMs provide more accurate semantic labels, leading to better learned representations. As shown in the table below, the potential of our method can be further unleashed by using a stronger VFM, namely SEEM [R6]. \|VFMs\|LP\|1%\|5%\|10%\|25%\| \|:-\|:-:\|:-:\|:-:\|:-:\|:-:\| \|Grounded-SAM\|47\.30\|46\.12\|57\.51\|63\.04\|69\.39\| \|Grounded-SAM-HQ\|47\.84\|48\.03\|58\.51\|64\.08\|69\.52\| \|SEEM\|50\.09\|50\.60\|60\.25\|65\.07\|70\.15\| Ref:\ [R1] Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining.\ [R2] Image-to-lidar self-supervised distillation for autonomous driving data.\ [R3] Self-supervised image-to-point distillation via semantically tolerant contrastive loss.\ [R4] Segment Any Point Cloud Sequences by Distilling Vision Foundation Models.\ [R5] HVDistill: Transferring Knowledge from Images to Point Clouds via Unsupervised Hybrid-View Distillation.\ [R6] Segment Everything Everywhere All at Once. --- Rebuttal Comment 1.1: Comment: Hello Authors, I have reread the paper, the other reviews, and the authors’ comments. Thank you for the thorough rebuttal and responses to each of our questions and concerns. The additional tables and experiments are detailed and insightful. My primary concerns related to an incomplete comparison to related work, similarities to R3, and comparisons to other VFMs have been adequately addressed. I am now more confident in maintaining my original rating of weak accept. --- Rebuttal 2: Title: Authors' Response to Reviewer uKMo Comment: Dear Reviewer uKMo, Thanks again for the time and energy you committed and your valuable comments. Your meticulous review and thoughtful critiques truly reflect your deep domain expertise and diligence as a reviewer. It has been a pleasure communicating and exchanging ideas with you. Please feel free to share any additional comments or feedback on the manuscript. Warm regards, Authors	Summary: In this paper, the authors introduced a novel approach for improving 3D representation learning by leveraging VFMs to generate semantic labels for weakly-supervised pixel-to-point contrastive distillation. The proposed method addressed the self-conflict issue in traditional contrastive learning and presented a density and category-aware sampling strategy to ensure balanced learning. This approach showed the better performance over existing methods on nuScenes and SemanticKITTI datasets. Strengths: First of all, the motivation of the paper seems to be meaningful and pragmatic in the perspective of the better semantic understanding by leveraging VFMs and balanced learning by using the density and category frequency. The key idea is very intuitive how to integrate of VFMs with existing multi-modal SSL allowing for generating semantic labels to deal with self-conflict issues. More specifically, the model is trained by three objectives as weakly-supervised contrastive distillation using pseudo labels to identify positive pairs by category, self-supervised contrastive learning applied to randomly sampled point-pixel pairs, and lastly a regularization based on the von Mises-Fisher distribution to ensure semantic consistency. In experimental section, the proposed method achieved SoTA results in two kinds of downstream tasks (segmentation and detection), demonstrating its effectiveness. The ablation study is highly analytical for each level module. Weaknesses: One concern is the validity of the proposed SSL method for better representation learning. First of all, it is not clear whether the reduced effectiveness with larger data is due to the model’s insufficient size (capability) or limitations in the proposed model itself. Also, an explanation is required to determine whether lower detection performance gain is due to the ineffectiveness of the proposed method, despite using object semantics. If necessary, the reasons for varying performance improvements across different downstream tasks should described in terms of the mechanism of the proposed learning pipeline. Experiment analysis and technical description are not specific and descriptive in some extent. The category-aware sampling is not specified in detail. There is not detailed description of performance variation based on sampled data groups or the extent of improvement over existing methods when learning from a sample of the entire dataset. (1%, 5%, 10%,…) Technical Quality: 3 Clarity: 3 Questions for Authors: There are some vague sentences and grammatical errors in the paper. I recommend that the author will revise the paper. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I mentioned all comments including reasons and suggestions in the above sections. I recommend that the author will provide all the concerns, and improve the completeness of the paper. If the rebuttal period resolves the above-mentioned concerns, I will gladly raise my score. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's time and effort in reviewing our paper. Thanks for your valuable comments and recognition of our work. In the following, we will comprehensively address your concerns. --- Comment: The validity of the proposed SSL method. It is not clear whether the reduced effectiveness with larger data is due to the model’s insufficient size or limitations in the proposed model. Response: Thanks for your comments. We would like to clarify the following points to address your concerns: - When the available training data is limited, the benefits of pre-trained model weights on downstream tasks are more pronounced. This phenomenon is widely observed in self-supervised learning. When downstream task data is limited, these representations are crucial because they provide a strong starting point, capturing features that the model wouldn't learn from the small labeled dataset alone. As the amount of labeled data increases, the model can learn these features directly from the labeled data, making the initial representations from pretraining less critical. - We also conducted experiments with a stronger 3D backbone namely WaffleIron [R1] (see the Table B1). The effect of pre-training weights becomes less obvious when training downstream tasks on sufficient data. So the reduced effectiveness with larger data is not due to the capability of backbone. - We can further improve the performance with more accurate semantic labels generated by a stronger SAM like SEEM [R2]. As shown in the Table B2, our method achieves a 2.03\% mIoU improvement on nuScenes with the full training data. - You might mean that the improvement compared to other pretraining methods is not obvious. We believe the main value of self-supervision methods is to improve the performance when the annotation resources are limited. And OLIVINE outperforms existing methods significantly when the annotation is limited. Table B1: Performance for 3D backbone WaffleIron \| Method \| 1% \| 10%\| 100% \| \| :----- \| :----- \| :----- \| :----- \| \| Random \| 33\.26 \| 58\.13 \| 77\.60 \| \| Ours \| 50\.14 \| 66\.43 \| 78\.21 \| Table B2: Comparison of various pre-training techniques. \| Method\| LP \| 1% \| 5% \| 10%\| 25%\| 100% \| \| :-------- \| :----- \| :----- \| :----- \| :----- \| :----- \| :----- \| \| Random\| 8\.10\| 30\.30 \| 47\.84 \| 56\.15 \| 65\.48 \| 74\.66 \| \| Seal\| 44\.95 \| 45\.84 \| 55\.64 \| 62\.97 \| 68\.41 \| 75\.60 \| \| Ours\| 50\.09 \| 50\.60 \| 60\.25 \| 65\.07 \| 70\.15 \| 76\.69 \| --- Comment: An explanation for lower detection performance gain is required. The reasons for varying performance improvements across different downstream tasks should described in terms of the mechanism of the proposed learning pipeline. Response: Thanks for your insightful questions. We provide the following explanations to address your concerns: - We observed a 2.0\% mAP improvement with the SECOND and a 1.5\% mAP improvement with the PV-RCNN, surpassing previous pretraining methods. These improvements were achieved by fine-tuning on full training data, so the enhancements may appear less significant compared to using limited labels. - Compared to the semantic segmentation task, the model architecture for object detection is more complex. Besides the 3D backbone, 3D detectors typically project features to a BEV plane, followed by a 2D convolutional network and RoI operations. These crucial components were not pre-trained, which may limit the overall performance gain from our pre-training approach. - It's important to note that semantic segmentation and object detection use different metrics and scales, making direct performance comparisons improper. The nature of these tasks and their evaluation criteria inherently lead to varying degrees of improvement when applying our proposed method. --- Comment: Experiment analysis and technical description are not specific and descriptive in some extent. The category-aware sampling is not specified in detail. There is not detailed description of performance variation based on sampled data groups from a sample of the entire dataset. Response: Thanks for pointing out this issue. Category-aware and density-aware sampling determine the sampling probability of a point by its category frequency and distance information, respectively. These are part of a hybrid strategy we refer to as density and category-aware sampling (DCAS). Following your suggestion, we have added a comparison of sampling strategies using 1\%, 5\%, 10\%, 25\%, and 100\% of the annotated data from nuScenes. The results are presented in the table below. We found that the density and category-aware sampling strategy consistently achieves the best performance on downstream tasks, effectively leveraging both spatial distribution and category frequency. \|Sampling\|1%\|5%\|10%\|25%\| \|:-------------\|:-----\|:-----\|:-----\|:-----\| \|Random\|44\.91\|56\.01\|62\.58\|68\.74\| \|Density-aware\|45\.33\|56\.6\|62\.74\|68\.96\| \|Category-aware\|45\.74\|56\.98\|62\.89\|69\.18\| \|DCAS (Density and Category-aware)\|46\.12\|57\.51\|63\.04\|69\.39\| --- Comment: There are some vague sentences and grammatical errors in the paper. Response: Thank you for your feedback. We appreciate your attention to detail. We have thoroughly reviewed the manuscript, revised the vague sentences, and corrected the grammatical errors. We genuinely hope that these clarifications address your concerns. Thanks again for your valuable time and feedback. We will include the results and analysis in the revised manuscript. --- References:\ [R1] Puy et al. Using a Waffle Iron for Automotive Point Cloud Semantic Segmentation. ICCV2023.\ [R2] Zou et al. SEEM: Segment Everything Everywhere All at Once. NeurIPS2023. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I think that the author provided feedbacks to address most of my concern, and the additional experiments are informative. After reading other reviewers and author's comments, I keep current rating. --- Rebuttal 2: Title: Authors' Response to Reviewer 54Gn Comment: Dear Reviewer 54Gn, Thank you for your response and for taking the time to carefully review our rebuttal. We greatly appreciate your recognition of our efforts to address your concerns and the value you found in the additional experiments we conducted. Your detailed and thoughtful review demonstrates a profound expertise in this domain. I have thoroughly enjoyed the opportunity to learn from your perspective. Please feel free to share any further comments or suggestions. Warm regards, Authors	Summary: The paper addresses the "self-conflict" issue in contrastive image-to-LiDAR knowledge transfer, where features of semantically similar but unmatched points and pixels are unintentionally dissociated, compromising representation integrity. To solve this, Visual Foundation Models are employed to generate semantic labels for weakly-supervised pixel-to-point contrastive distillation. The method includes structuring the feature space with von Mises-Fisher distributions for consistency and adjusting sampling probabilities to handle spatial and category imbalances. Extensive experiments demonstrate that this approach significantly outperforms traditional methods in various downstream tasks. Strengths: 1. The paper uses Visual Foundation Models to generate semantic labels, resolving the "self-conflict" issue and improving representation integrity. 2. The paper proposes a density and category-aware sampling method, ensuring balanced learning and better representation of minority categories. Weaknesses: 1. The overall architecture is similar to Seal [1], limiting its novelty except for the sampling strategy. Providing more clarification about the differences from Seal would be beneficial. 2. The improvement in fine-tuning results compared to the state-of-the-art is marginal. [1] Segment any point cloud sequences by distilling vision foundation models Technical Quality: 2 Clarity: 3 Questions for Authors: Please refer to the weaknesses section. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The paper discusses the limitations in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's time and effort in reviewing our paper. In the following, we will comprehensively address your concerns. --- Comment: The overall architecture is similar to Seal, limiting its novelty except for the sampling strategy. Providing more clarification about the differences from Seal would be beneficial. Response: Our overall framework significantly differs from the existing method Seal [R1]. We would like to clarify the following points to highlight the novelty of our method: - The purposes of using VFMs in Seal [R1] and our method are completely different. To avoid over-segmenting semantically coherent areas, Seal [R1] generates superpixels using visual foundation models (VFMs) instead of the traditional method SLIC [R2]. In contrast, our method does not rely on superpixels. Although we also use VFMs, we leverage them to obtain coarse semantic labels for fine-grained contrastive distillation. - Although the more precise superpixels generated by VFMs could mitigate the self-conflict problems to some extent, such a method does not solve the problem thoroughly. The superpoints and superpixels with the same category may still be mistakenly considered negative pairs during contrastive learning. Our method explicitly defines the points and pixels with the same semantic labels as positive pairs during weakly-supervised contrastive learning. - Our pipeline performs knowledge distillation on two levels: self-supervised and weakly-supervised contrastive learning. To achieve this, we develop two different heads in both the image and point cloud branches to decouple the learned representation. Previous methods like Seal have only attempted self-supervised contrastive distillation and have not explored using labels to guide contrastive distillation. - We explicitly model the features of each class as a von Mises-Fisher (vMF) distribution, promoting feature consistency within the same category. This approach cultivates a meaningful and structured feature space, an aspect that Seal does not explore. - Existing methods like Seal [R1] are highly dependent on the generated superpixels. Superpixels balance asymmetries between areas with denser coverage of points and sparser areas in the contrastive loss. However, we do not need this process at all and ensure a uniform representation of both spatial and categorical dimensions by employing a novel sampling strategy. We genuinely hope that these clarifications provide a clearer perspective on our research and its merits. Thanks again for your valuable time and feedback. We will further clarify the novelty and the differences with related methods like Seal [R1] detailedly in the revised manuscript. --- Comment: The improvement in fine-tuning results compared to the state-of-the-art is marginal. Response: We agree with you that the improvement in downstream tasks compared to the state-of-the-art is not significant. But, we have to clarify the following points: - As stated in the manuscript, we believe that employing stronger visual foundation models for more precise semantic labels can lead to better 3D representations. Therefore, we obtained coarse labels with a stronger VFM, namely SEEM, and evaluated the learned 3D representation. As shown in the table below, our method outperforms Seal [R1] by a significant margin, achieving an improvement of 5.14\% under the setting of linear probing. - We have completely open-sourced the code for OLIVINE, whereas existing state-of-the-art methods like HVDistill and Seal have NOT yet made their training code available. We believe this contributes positively to the image-to-point knowledge transfer community by promoting transparency and enabling further research. - Our method is compatible with existing techniques. For example, the semantic temporal consistency proposed in Seal [R1] and BEV-based contrastive distillation [R3] can also be integrated into our pipeline. We plan to explore these aspects further once the source code for these works is released. [Table A1] Comparison of various pre-training techniques for semantic segmentation tasks using either finetuning or linear probing. \| Method\| LP \| 1% \| 5% \| 10%\| 25%\| 100% \| \| :-------- \| :----- \| :----- \| :----- \| :----- \| :----- \| :----- \| \| Random \| 8\.10 \| 30\.30 \| 47\.84 \| 56\.15 \| 65\.48 \| 74\.66 \| \| PPKT \| 35\.90 \| 37\.80 \| 53\.74 \| 60\.25 \| 67\.14 \| 74\.52 \| \| SLidR\| 38\.80 \| 38\.30 \| 52\.49 \| 59\.84 \| 66\.91 \| 74\.79 \| \| ST-SLidR\| 40\.48 \| 40\.75 \| 54\.69 \| 60\.75 \| 67\.70 \| 75\.14 \| \| HVDistill\| 39\.50 \| 42\.70 \| 56\.60 \| 62\.90 \| 69\.30 \| 76\.60 \| \| Seal \| 44\.95 \| 45\.84 \| 55\.64 \| 62\.97 \| 68\.41 \| 75\.60 \| \| Ours \| 50\.09 \| 50\.60 \| 60\.25 \| 65\.07 \| 70\.15 \| 76\.69 \| References:\ [R1] Liu et al. Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. NeurIPS2023. \ [R2] Achanta et al. Slic superpixels compared to state-of-the-art superpixel methods. TPAMI2021. \ [R3] Zhang et al. HVDistill: Transferring Knowledge from Images to Point Clouds via Unsupervised Hybrid-View Distillation. IJCV2024. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal. It has addressed most of my concerns. As a result, I will maintain my current rating. --- Rebuttal 2: Title: Authors' Response to Reviewer 2WGA Comment: Dear Reviewer 2WGA, Thank you for taking the time to review our rebuttal and for your constructive feedback throughout the process. We are glad that we could address most of your concerns. We will actively participate in the Author-Reviewer discussion session. Please feel free to share any additional comments or feedback on the manuscript. Warm regards, Authors	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for your time and constructive comments. --- We are glad that the reviewers see the value in our work: 1. "_The paper addresses the self-conflict issue in contrastive image-to-LiDAR knowledge transfer ... significantly outperforms traditional methods in various downstream tasks_" (Reviewer 2WGA); 2. "_the motivation of the paper seems to be meaningful and pragmatic in the perspective of the better semantic understanding_" (Reviewer 54Gn); 3. "_The key idea is very intuitive how to integrate VFMs with existing multi-modal SSL_" (Reviewer 54Gn); 4. "_The ablation study is highly analytical for each level module_" (Reviewer 54Gn); 5. "_This work will likely serve as a healthy addition to the image-to-point knowledge transfer community_" (Reviewer uKMo). --- We would like to emphasize the uniqueness and advantages of our approach over existing ones: 1. Previous works [R1-R5] have not solved the self-conflict problem properly. Especially, Seal [R4] generates semantically coherent superpixels for distinct objects and backgrounds in the 3D scene. However, the superpoints and superpixels with the same category may still be mistakenly considered negative pairs during contrastive learning. By contrast, our method explicitly defines the points and pixels with the same semantic labels as positive pairs during weakly-supervised contrastive learning. 2. Our pipeline performs knowledge distillation on two levels: self-supervised and weakly-supervised contrastive learning. To achieve this, we develop two different heads in both the image and point cloud branches to decouple the learned representation. Previous methods [R1-R5] have only attempted self-supervised contrastive distillation and have not explored using labels to guide contrastive distillation. 3. The representation of samples in the same class can vary significantly across different batches during the contrastive distillation, so the model will struggle to learn stable semantic features. By making point features of the same class closely aligned, our method aims to create a more consistent and structured feature space. 4. Existing methods [R2-R5] are highly dependent on the generated superpixels. Superpixels balance asymmetries between areas with denser coverage of points and sparser areas in the contrastive loss. However, we do not need this process at all and ensure a uniform representation of both spatial and categorical dimensions by employing a novel sampling strategy. 5. ST-SLidR [R3] reduces the contribution of false negative samples based on superpixel-to-superpixel similarity, using 2D self-supervised features to determine semantic similarities between superpixels. By contrast, our method directly estimates the semantic labels of images with VFMs, and defines pixels and points with the same label as positive pairs. 6. The purposes of using VFMs in Seal [R4] and our method are completely different. To avoid over-segmenting semantically coherent areas, Seal [R4] generates superpixels using visual foundation models (VFMs) instead of the traditional method SLIC [R6]. In contrast, our method does not rely on superpixels. Although we also use VFMs, we leverage them to obtain coarse semantic labels for fine-grained contrastive distillation. --- Following the reviewers' valuable comments and suggestions, we have made these efforts: 1. We have achieved further improvements on downstream tasks using semantic labels generated by stronger VFMs, as suggested by Reviewer 2WGA. 2. We have highlighted the novelty of our method and clarified the differences from previous methods, as suggested by Reviewers 2WGA and uKMo. 3. We discussed the reasons for varying performance improvements across different downstream tasks, considering the mechanism of the proposed learning pipeline, as suggested by Reviewer 54Gn. 4. We have supplemented the experiment analysis and provided a technical description of sampling strategies, as suggested by Reviewer 54Gn. 5. We have added relevant citations to support our claim in L148, as suggested by Reviewer uKMo. 6. We have combined Tables 2 and 3 to streamline the presentation and carefully corrected typographical errors, as suggested by Reviewers uKMo and svRz. 7. We have compared the effects of different VFMs for generating semantic labels, as suggested by Reviewer uKMo. 8. We have provided a detailed explanation and theoretical justification for the application of the vMF distribution, as suggested by Reviewer svRz. 9. We have added experiments on six additional LiDAR-based point cloud datasets and one out-of-distribution dataset, as suggested by Reviewer svRz. 10. We have reported the computational cost of the proposed pretraining method, as suggested by Reviewer svRz. --- Finally, we extend our gratitude to the PCs, ACs, and all the reviewers for their dedicated time and effort in this review process. We look forward to engaging in discussions with you over the next few days. --- References:\ [R1] Liu et al. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arxiv2021.\ [R2] Sautier et al. Image-to-lidar self-supervised distillation for autonomous driving data. CVPR2022.\ [R3] Mahmoud et al. Self-supervised image-to-point distillation via semantically tolerant contrastive loss. CVPR2023.\ [R4] Liu et al. Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. NeurIPS2023.\ [R5] Zhang et al. HVDistill: Transferring Knowledge from Images to Point Clouds via Unsupervised Hybrid-View Distillation. IJCV2024. Pdf: /pdf/391c46ea2568c02ca3d95994323776b1956f45e1.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Worst-Case Offline Reinforcement Learning with Arbitrary Data Support	Accept (poster)	Summary: In this submission, the authors propose some new bounds for offline reinforcement learning. Their contribution is twofold. First, they remove the classical data support assumption by solving a relaxed problem, in which any non observed transition is replaced by a transition to an absorbing state. This new MDP yields a lower reward than the original one, hence the worst-case name, but can be learned with the given data. Second, they use a regularized version of the Linear Programming formulation of the optimal policy problem to estimate the good policy. Their improved bounds come from the direct use of the dual variable to obtain the policy, rather than using a policy-improvement step on the primal variable (the value function). Strengths: The approach is interesting with simultaneously a new “worst-case” setting and new way to obtain result with the regularized linear programming approach. The paper is well written, the proofs seem correct (I did not check them carefully). Weaknesses: The paper is very technical and is written in a linear way. I know that the space constraint is strong, but I would have liked a better summary of the main idea at the beginning of the paper so that the position with respect to the relative work would be easier to understand. The notations often differ only by a slight detail. Using larger difference may be beneficial for the readers. Technical Quality: 4 Clarity: 3 Questions for Authors: The bounds obtained are upper bound. Are the author aware of corresponding lower bounds? Typos and misc.: - 167: formulation - 179: Why is $v \geq 0$ so important - 209: $\bar v$ vs $\tilde v$ : hard to distinguish - 259: $\epsilon_\theta$ is often called an excess risk, or something similar. - 306: $n \geq$ rather than $n =$ Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The only limitation is that there is no numerical experiment to support the claim, but this should not be expected from such a theoretical paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your helpful review and feedbacks. In particular, we agree that the manuscript can be improved by adding the position/contribution summary and improving visual aspects of some notation. Please find below the answers for your questions. ---- > Are the author aware of corresponding lower bounds? No, we are not aware of lower bounds for the model-free estimators with funciton approximation. The closest result we are aware of is given by [1], which is a minimax lower bound for the model-based estimators without function approximation (i.e., tabular setting). Since the underlying hypothesis classes are different (i.e., general model-free function class vs tabular environment class), it is not straightforwardly applicable to our setting. ---- > 179: Why is $v\ge 0$ so important We assume you are referring to Line 169. In fact, the nonnegativity of $v$ is the key to make the LP feasibility agnostic to the concentrability condition. This is seen by that, when the concentrability does not hold (i.e., $\mathrm{supp}(d^\pi)\not\subset \mathrm{supp}(\mu)$), the $v$-residual term $D\_V^\pi(v)$ in Eq. (4) can be negative infinity by taking $v(s)\to-\infty$ for $s\in \mathrm{supp}(d^\pi)\setminus\mathrm{supp}(\mu)$. The constraint of $v\ge 0$ prevents the divergence towards negative infinity and necessary to ensure the concentrability-agnostic existence of the saddle points of $L(v,f)$. ---- [1] Li, Shi, Chen, Chi & Wei. (2024). Settling the sample complexity of model-based offline reinforcement learning. The Annals of Statistics. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. It addressed most of my concerns, and I believe the paper meets the acceptance threshold.	Summary: This paper studies offline reinforcement learning with arbitrary data support. More specifically, they truncate the original environment so that the new environment (called truncated environment) is always covered by the offline data. They further show that the optimal policy of the truncated environment can be found by solving a regularized Lagrangian problem and analyze the corresponding sample complexity. As a byproduct, they attain a $O(1/\epsilon^2)$ sample complexity bound for learning an $\epsilon$-optimal policy against any comparator policy with bounded concentrability and realizability. Strengths: 1. The connection between the truncated environment and the regularized Lagrangian problem is very novel and useful in my opinion. 2. The authors improve the sample complexity of learning with single-policy concentrability and realizability to $1/\epsilon^2$, which is a significant contribution. Weaknesses: 1. In the sample complexity of Corollary 6.16.2/6.3, the rates scale with $\tilde{C}_{\infty}$, which is indeed the concentrability bound for all policies within the policy class. This seems stronger than the naive single-policy concentrability. 2. The algorithm requires the knowledge of the behavior policy, which might be hard to estimate in practice. Technical Quality: 3 Clarity: 2 Questions for Authors: Similar to the weaknesses: 1. can we get rid of $\tilde{C}_{\infty}$? 2. can the algorithm work without the knowledge of the behavior policy? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: None. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your helpful review and comments. Please find below the answers to your questions. > can we get rid of $\tilde{C}\_\infty$? Asymptotically, yes. Such extensions are discussed in the appendix, Section E.4. In summary, we can replace the uniform concentrability $\tilde{C}\_\infty$ with either of the two variants of local concentrabilities $\tilde{C}\_0$ and $\tilde{c}\_0$, where the policy class is restricted to near-optimal policies rather than all policies. Please refer to Definition E.1 and E.2 for the precise definitions and Corollary E.2 and E.3 for the corresponding results. That being said, whether similar improvements in a non-asymptotic manner is possible or not remains unclear and open for the future work. > can the algorithm work without the knowledge of the behavior policy? This is also yes. In our algorithm, the behavior policy need not be known since it is estimated as $\beta_\theta$. --- Rebuttal Comment 1.1: Comment: Thank you for the rebuttal! I will keep my score.	Summary: The paper develops a framework for evaluating offline reinforcement learning methods without any data-support conditions. Traditional techniques rely on the concentrability coefficient to ground the difference between the offline data distribution and the induced policy distribution. However, the concentrability condition is often unrealistic in practice, especially when the state-action space is very large, as is frequently the case. This paper proves by construction a sample complexity upper bound of $O(\epsilon^{-2}(1 - \gamma)^{-4}\ln(1/\delta))$ is achievable without relying on the concentrability condition. This improves upon several bounds from previous work, involving similar assumptions. Strengths: - The paper shows some originality in creating a new framework that does not rely on the frequent concentrability condition found in offline RL literature. The associated restriction in this paper is even less-so than the single-policy concentrability condition found in Zhan, Huang, Huang, et al. - The paper simplifies the traditional offline RL evaluation framework by removing the need to directly incorporate pessimism or behavioral cloning via hyperparameters into the analysis. - The paper provides a strong theoretical proof to establish the stated upper bound. Weaknesses: - I feel the proposed truncated environment is overly restrictive and ignores information between similar state-action pairs, even if they're not represented in the support of the offline data distribution. Although the benefit of doing this is that it removes the need to control for pessimism and behavioral cloning, there could be some useful information discarded, which could help improve the bound derived in the paper. UPDATE: this has been addressed by the authors' followup. Technical Quality: 3 Clarity: 4 Questions for Authors: No questions. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: The authors have mentioned that although their results extend to continuous state-action spaces, the concentrability coefficients are not unconditionally finite anymore. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your helpful review and comments. Below, we would like to provide additional discussion on the weakness you mentioned. > I feel the proposed truncated environment is overly restrictive and ignores information between similar state-action pairs, even if they're not represented in the support of the offline data distribution. Although the benefit of doing this is that it removes the need to control for pessimism and behavioral cloning, there could be some useful information discarded, which could help improve the bound derived in the paper. First of all, we agree that there may be more practically reasonable way of introducing pessimism and it is an interesting research topic. However, we would like to emphasize that the effect of ignoring the state/action similarities here is minimal under the conventional RL setting. This can be verified by that the truncation of the transition kernel conserves the (non-pessimistic) policy value $J(\pi)$ under $\pi$-concentrability, as stated by Theorem 4.1. In other words, the seemingly-excessive pessimism of our truncation method only affects the RL outside the data support, which is previously unexplored. Thus, we think it is currently unclear if the conventional pessimism/behavioral cloning methods can be better alternatives here. Also note that our method can take into account the on-support state-action similarity via the smoothness of the function approximators. --- Rebuttal Comment 1.1: Comment: Ahh, that is actually a good point. I've modified my rating to reflect this. Thank you for addressing my concerns.	Summary: This paper studies offline RL with no data support assumptions. To address the data coverage problem, the paper studies a new setting with worst-case policy value. It formulates the RL problem with Lagrangian and shows the instability compared with the regularized Lagrangian. Improved sample complexity is proved in the new framework. Strengths: 1. The proposed new framework is novel and interesting in the study of offline RL. 2. The paper provides analysis of the stability of lagrangian and regularized lagrangian. 3. The paper proved an improved sample complexity. Weaknesses: 1. The paper studies everything in a new framework, where the transition is constrained in the range of the data. It seems unfair to compare the results in the new framework with previous works, as they can also work for out-of-distribution data to some extent. As shown in Theorem 4.1, the loss will become smaller in the new framework. 2. The behavior of the truncated transition kernel (2) is strange. As the indicator function does not contain any information about the next state s’, $\sum_{s’} \tilde T(s’\|s,a) = \chi_{\mu, \beta}(s,a)$, which takes values in 1 or 0 only. It does not match the leaky description in the paper, i.e., it may sum to less than unity. 3. The paper misses comparison with some relevant papers. For example, [1] [2] [1] Rashidinejad et al 2022. Optimal conservative offline rl with general function approximation via augmented lagrangian. In ICLR. [2] Ozdaglar et al. 2023. Revisiting the linear-programming framework for offline rl with general function approximation. In ICML. 4. The superscripts used are too complex and lack explanations, making the writing hard to follow. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Why do you think the realizability assumption in [ZHH+22] is $\pi_n$-realizable? Do you think there is a huge influence on the result? 2. What are the $\bar w$ and $\\|\cdot\\|_{1,\bar w}$ in equation (10)? 3. In equation (17), the empirical loss function is defined from the one-sample loss function. Is it possible to generalize the method using the batch loss function? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have addressed the limitations in section 7. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We really appreciate your helpful review and comments. We hope our response below resolve potential misunderstandings and address your questions. The response consists of two parts: clarifications and answers. ## Clarifications > It seems unfair to compare the results in the new framework with previous works, as they can also work for out-of-distribution data to some extent. As shown in Theorem 4.1, the loss will become smaller in the new framework. Let us clarify one potential confusion here: The comparison of our method to previous work in Table 1 is in fact fair since all the methods in the table bound the sample complexity with respect to the same (conventional) metric, $J(\pi)$. This is, as stated in the last sentence of Section 6, seen by the fact the new metric and conventional metric coincides under the concentrability condition. > The paper misses comparison with some relevant papers. For example, [1] [2] Thank you for pointing out relevant papers. We agree these references should be included in our paper. Below, we discuss their similarities and differences in comparison with our result. First of all, one notable similarity is that [1], [2] and we all studied offline RL based on the LP formulation. On the other hand, the most essential difference is that both [1] and [2] consider the conventional offline RL, unlike us considering the worst-case offline RL making the scopes of our analysis broader than theirs. That being said, their results can be compared to ours within the non-worst-case framework as we did in Table 1. The following is the extension of Table 1 including [1] and [2]. \| method \| concentrability \| realizability \| sample complexity bound \| \|----\|----\|----\|----\| \| [1] \| $\pi^\$ \| $\pi^\$-realizable + "completeness" \| $\epsilon^{-2}(1-\gamma)^{-6}\ln(\mathcal{N}/\delta)$ \| \| [2] \| $\pi^\$ \| $\pi^\$-realizable \| $\epsilon^{-2}C\_{\mathrm{gap}}^{-2}(1-\gamma)^{-6}\ln(\mathcal{N}/\delta)$ \| \| ours \| any comparator \| $\pi^\$-realizable \| $\epsilon^{-2}(1-\gamma)^{-4}\ln(\mathcal{N}/\delta)$ \| In summary, our result strictly improves upon both of theirs. Below are the detailed comments on the table. - [1] achieves the sample complexity bound of $O(1/\epsilon^2)$ leveraging the notion of the occupancy validity, enforced by a new regularization term. However, their bound requires some completeness-type condition ("$u^\star\_w\in\mathcal{U}$ for all $\mathcal{W}$", Theorem 4, [1]) for one of its function approximators $\mathcal{U}$, which is more stringent than the realizability. Besides that, our sample complexity bound improves theirs by factor of $(1-\gamma)^2$. - [2] shows two distinct results: one requiring the completeness-type assumption and the other requiring the realizability and the action gap assumption, in addition to the concentrability. We included the latter to the table. Roughly speaking, the result of [2] is similar to that of [CJ22] except with the difference in the infinite/finite time horizons, requiring the action gap to be bounded away from zero. ## Answers to the questions > Why do you think the realizability assumption in [ZHH+22] is $\pi\_n$-realizable? In Corollary 12, [ZHH+22], their sample complexity bound requires the regularization weight $\alpha$ to be dependent on the error tolerance parameter $\epsilon$. This implies the target policy $\pi\_{\alpha,B\_w}^\$ of the realizability condition (Assumption 6, [ZHH+22]) depends on $\epsilon$ as well, i.e., their method requires $\pi\_\epsilon$-realizability. Since the minimum possible $\epsilon$ depends on $n$, it also implies $\pi\_n$-realizability. > Do you think there is a huge influence on the result? We think whether the influence is huge or not is context-dependent. One potential drawback is that the nonconstancy of $\pi\_n$ would make the nonparametric extension more difficult since the target policy moves around as $n\to\infty$. > What are the $\bar{\omega}$ and $\|\cdot\|\_{1,\bar{\omega}}$ in equation (10)? This is typo: sorry for that and thank you for pointing it out. The correct notation is $\bar{\mu}$ and $\|\cdot\|\_{1,\bar{\mu}}$, which are defined just below Eq. (5). > In equation (17), the empirical loss function is defined from the one-sample loss function. Is it possible to generalize the method using the batch loss function? Our method uses the mean of the one-sample loss functions over the batch data $\mathcal{D}$ (Eq. (16)), which is, in a sense, a batch loss function and our method uses it directly. Please let us know if you have other kind of batch loss function in mind. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed response. I find them quite useful. However, I found my question in weakness 2 was not answered. I will consider increasing my scores if it could be further explained. --- Reply to Comment 1.1.1: Comment: Thank you for the request for explanation. Please find below our comment on the weakness #2. > The behavior of the truncated transition kernel (2) is strange. As the indicator function does not contain any information about the next state s’, $\sum\_{s'} \tilde{T}(s'\|s,a)=\chi_{\mu,\beta}(s,a)$, which takes values in 1 or 0 only. It does not match the leaky description in the paper, i.e., it may sum to less than unity. The said behavior of the truncated transition kernel, $\sum\_{s'} \tilde{T}(s'\|s,a)=\chi_{\mu,\beta}(s,a)$, is intended. That being said, we agree that "it may sum to less than unity" is not the best description and it is better to say like "it may sum to zero". Thank you for pointing it out and we will update it accordingly. Hope this answers your question. If you have any further questions or concerns, please let us know.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
One-shot Federated Learning via Synthetic Distiller-Distillate Communication	Accept (poster)	Summary: The paper introduces FedSD2C, a new one-shot Federated Learning (FL) framework designed to address challenges in existing methods. Previous approaches have used data-free knowledge distillation to improve one-shot FL, but these methods struggle with data heterogeneity and scalability issues. FedSD2C aims to solve these issues by: 1.Introducing a distiller to synthesize informative distillates directly from local data, reducing information loss. 2.Sharing synthetic distillates instead of inconsistent local models to address data heterogeneity. The authors claim that empirical results show FedSD2C outperforms other one-shot FL methods, especially with more complex and real datasets. Strengths: This paper focuses on one-shot Federated Learning (FL), which is an intriguing topic. The paper presents a comprehensive set of experiments, both on model performance and privacy. Weaknesses: Thanks for the authors' efforts in presenting this paper. After carefully reading it, I have the following questions and comments: 1. On line 52, the values 4.21 and 2.06 are not clearly explained. I suggest adding a figure or table to provide more specific details about these numbers. 2. While I can imagine that transmitting synthetic data could improve the global model's performance more easily than transmitting the model itself, this inevitably leads to privacy concerns. Although the authors provide some data reconstruction experiments, the results are not entirely convincing: a) To my knowledge, dataset distillation/coreset selection does not inherently protect privacy [1]. I suggest the authors include additional experiments on Membership Inference Attacks (MIA) (based on results from data-free KD [2], I suspect coresets provide even less privacy protection, since data-free KD doesn't use any 'real' training data ). b) According to the privacy onion concept [3], memorization is relative, then in your method the selected images may be at higher risk of privacy leakage compared to those not selected. 3. Compared to other methods, the proposed method communication overhead seems larger. It requires to transmit both the auto-encoder and a large amount of synthetic data. 4. The method relies heavily on Stable Diffusion's auto-encoder, which seems to utilize very strong prior information. How well would your method perform on datasets where diffusion models are not as proficient, such as medical datasets? 5. To my knowledge, another one-shot FL work[4] that relies on Stable Diffusion achieved results of about 75.0 on ImageNette, which appears significantly better than your 55.90. [1] No Free Lunch in "Privacy for Free: How does Dataset Condensation Help Privacy" [2] Evaluations of Machine Learning Privacy Defenses are Misleading [3] The Privacy Onion Effect: Memorization is Relative [4] Federated Generative Learning with Foundation Models Technical Quality: 3 Clarity: 2 Questions for Authors: see above Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: see above Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: *Q1 On line 52, the values 4.21 and 2.06 are not clearly explained. I suggest adding a figure or table to provide more specific details about these numbers.* We would like to thank the Reviewer Bcom for suggestions. We will include a detailed figure to better illustrate this in the final version. *Q2 While I can imagine that transmitting synthetic data could improve the global model's performance more easily than transmitting the model itself, this inevitably leads to privacy concerns. Although the authors provide some data reconstruction experiments, the results are not entirely convincing: a) To my knowledge, dataset distillation/coreset selection does not inherently protect privacy [1]. I suggest the authors include additional experiments on Membership Inference Attacks (MIA) (based on results from data-free KD [2], I suspect coresets provide even less privacy protection, since data-free KD doesn't use any 'real' training data ). b) According to the privacy onion concept [3], memorization is relative, then in your method the selected images may be at higher risk of privacy leakage compared to those not selected.* Thanks for the insightful comments. As suggestion, we perform Membership Inference Attacks on our methods. We employ an improved version of LiRA[1]and set the raw images of Core-Set as the canary (target data $x$), as this is the most serious cases of our methods. Please note that we consider a semi-honest server, so the victim model for us is a model trained on synthetic data, and for the sharing model method is the client-uploaded local model. The results confirm that our approach does not introduce more privacy risk than the sharing model-based approach, even for the most vulnerable targets. Furthermore, according to Theorem 3.2 of [2], introducing DP-SGD during the distillate synthesis stage can provide theoretical privacy guarantees for our method. \| Method \| TPR@FPR=0.1 \| \| ---------------------------------- \| ----------- \| \| Sharing Model (DENSE, Co-Boosting) \| 22.81 \| \| FedSD2C \| 20.13 \| [1] Aerni, M, et al. Evaluations of Machine Learning Privacy Defenses are Misleading. 2024. [2] Xiong Y, et al. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. 2023. *Q3 Compared to other methods, the proposed method communication overhead seems larger. It requires to transmit both the auto-encoder and a large amount of synthetic data.* Thanks for the comments. In fact, our method does not require server-to-client transmission. The pre-trained Autoencoder is available from public repositories. Instead, in our framework, only the synthetic distillate is transmitted from the client to the server, with no other data or model being transferred. As illustrated in Table 3 of our paper, our method significantly reduces the communication cost to a mere 0.5 MB, in contrast to the 44 MB required by sharing model-based methods. This comparison underscores the superiority of our approach in communication efficiency. *Q4 The method relies heavily on Stable Diffusion's auto-encoder, which seems to utilize very strong prior information. How well would your method perform on datasets where diffusion models are not as proficient, such as medical datasets?* Thanks for the detailed comments. The large-scale datasets used to pre-train the Autoencoder contain diverse data, which is enough to cover the data domains of most clients. For untrained data domains, such as medical dataset COVID-FL [1], we conducted experiments to verify that Autoencoder can be extended to more different data domains. All of these illustrate the practices of employing a pre-trained autoencoder. The experimental results on medical datasets are as follows: \| Method \| 0.1 \| 0.3 \| 0.5 \| \| ------- \| ----- \| ----- \| ----- \| \| DENSE \| 46.15 \| 57.55 \| 62.83 \| \| CoBoost \| 45.07 \| 60.27 \| 65.61 \| \| FedSD2C \| 52.65 \| 62.50 \| 66.68 \| [1] Yan R, et al. Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging. 2023. *Q5 To my knowledge, another one-shot FL work[4] that relies on Stable Diffusion achieved results of about 75.0 on ImageNette, which appears significantly better than your 55.90.* Thanks for bringing a related work to our attention. In fact, it is not absolutely fair to compare our approach to theirs. Firstly, the approach in work [1] demands significantly more computational resources and data. Their method requires 400 times the data volume compared to ours ($ipc=20,000$ vs. $ipc=50$). Our method is more cost-effective in terms of data synthesis. It relies solely on an autoencoder, eliminating the need for Stable Diffusion, which is a computationally intensive process. Secondly, the method from work [1] necessitates prior knowledge of the category name and assumes that the images adhere to the prior distribution of Stable Diffusion. This assumption can be restrictive, as it may not be applicable to diverse datasets or specialized domains such as medical imaging, where Stable Diffusion may not be able to synthesize X-rays or similar images based solely on labels. [1] Zhang J, et al. Federated Generative Learning with Foundation Models. 2023. --- Rebuttal Comment 1.1: Title: response Comment: Thanks for authors' response. As for the experiment on privacy attacks, since your model's performance is too low (far below 90%), it will also lead to biased MIA results. Therefore, I don't recommend claiming in the paper that your method can protect privacy (without any differential privacy guarantee). For another one-shot FL work[1], I think they also tested the method on medical datasets. After reading the rebuttal, I am happy with increasing the score to a 5 or 6 (if there is no overstatement in the final version). [1] Zhang J, et al. Federated Generative Learning with Foundation Models. 2023. --- Reply to Comment 1.1.1: Title: A friendly reminder Comment: Thanks for the feedback of Reviewer Bcom and we are encouraged that most of the concerns and questions are addressed. As mentioned by the reviewer, we will definitely include the results in the rebuttal into our revision. We have noted that the current rating still tends toward the negative. We kindly request clarification on any unresolved issues that might be affecting the reviewer's rating. Please feel free to share any remaining concerns. We are fully committed to addressing these issues during the remainder of the review discussion period. We greatly appreciate your efforts and look forward to your additional feedback. --- Rebuttal 2: Title: Looking Forward to Further Discussions Comment: Dear Reviewer Bcom, Thank you once again for your constructive comments and the effort you put into reviewing our submission. Please let us know if our response has addressed your concerns. We are more than happy to address any further comments you may have. Thanks! --- Rebuttal 3: Title: Thank you for raising the score! Comment: We sincerely appreciate the time and effort you have invested in reviewing our paper, and your insightful feedback has been a critical factor in enhancing the overall quality of our work. Following your comments, we will revise the privacy statement in the abstract and introduction to accurately reflect its limitations and ensure there is no overstatement. Additionally, we will include a discussion on work[1], as well as all other comments. Once more, we appreciate the time and effort you've dedicated to our paper. [1] Zhang J, et al. Federated Generative Learning with Foundation Models. 2023.	Summary: This paper proposes a one-shot FL method (FedSD2C), utilizing V-information to select local core set data and server-pretrained autoencoder and Fourier-domain perturbation to ensure privacy preservation for local "distillate" sharing. In comparison to existing works such as DENSE and Co-Boosting, FedSD2C can reduce information loss in one-shot FL and improve up to 2.7x global model performance. Strengths: The paper tackles an important setting in FL and the one-shot performance outperforms listed related works. Weaknesses: - The assumption that the server holds an autoencoder is pretty strong, as the autoencoder must be trained in the clients' data domain to ensure it works. - The paper only showed that Fourier-based perturbation can 'visually' protect privacy by using PSNR as a metric. Although the paper mentioned MIA, it did not evaluate existing privacy attacks. - The clarity of paper writing can be improved. Essential details are missing to understand the contributions. See more in the Questions section. Technical Quality: 2 Clarity: 3 Questions for Authors: - Can you more rigorously demonstrate the privacy-preservation statement? For example, differential privacy is a widely accepted concept. Could you please show the provable privacy-preservation guarantee like DP? - It is unclear how Figure 1 is generated. The authors should detail the datasets, model, and necessary information. Otherwise, comparing the figures is meaningless. - What is the computational cost for the core-set selection step? - It is unclear which parts of the algorithm contribute to handling the heterogeneous data. It seems that coreset selection could contribute well if intentionally sampling (almost) balanced data. - What is the number of clients for Table 1? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Limitation was discussed Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: *Q1 The assumption that the server holds an autoencoder is pretty strong, as the autoencoder must be trained in the clients' data domain to ensure it works.* Thanks for the detailed comments. The large-scale datasets used to pre-train the autoencoder contain diverse data, which is enough to cover the data domains of most clients. For untrained data domains, such as medical dataset COVID-FL [1], we conducted experiments to verify that Autoencoder can be extended to more different data domains. Moreover, Thanks to the growing open source community, autoencoders pre-trained for different data domains are also readily available. All of these illustrate the practices of employing a pre-trained autoencoder. The experimental results on medical datasets are as follows: \| Method \| 0.1 \| 0.3 \| 0.5 \| \| ------- \| ----- \| ----- \| ----- \| \| DENSE \| 46.15 \| 57.55 \| 62.83 \| \| CoBoost \| 45.07 \| 60.27 \| 65.61 \| \| FedSD2C \| 52.65 \| 62.50 \| 66.68 \| [1] Yan R, et al. Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging. 2023. *Q2 The paper only showed that Fourier-based perturbation can 'visually' protect privacy by using PSNR as a metric. Although the paper mentioned MIA, it did not evaluate existing privacy attacks. Can you more rigorously demonstrate the privacy-preservation statement? For example, differential privacy is a widely accepted concept. Could you please show the provable privacy-preservation guarantee like DP?* Thanks for the comments. Our paper emphasizes the empirical contributions of using the Fourier transform to enhance the privacy of synthetic data. In this regard, our paper performed Model Inversion Attacks to validate that our approach provides the best trade-off between privacy and performance. To further validate the effectiveness of our method, we employ an improved version of LiRA[1] to conduct Membership Inference Attacks on our methods. We set the raw images of Core-Set as the canary (target data $x$), as this is the most serious case of our methods. The results confirm that our approach does not introduce more privacy risk than the sharing model-based approach, even for the most vulnerable targets. Furthermore, according to Theorem 3.2 of [2], introducing DP-SGD during the distillate synthesis stage can provide theoretical privacy guarantees for our method. \| Method \| TPR@FPR=0.1 \| \| ---------------------------------- \| ----------- \| \| Sharing Model (DENSE, Co-Boosting) \| 22.81 \| \| FedSD2C \| 20.13 \| [1] Aerni, M, et al. Evaluations of Machine Learning Privacy Defenses are Misleading. 2024. [2] Xiong Y, et al. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. 2023. *Q3 Unclear statement of Figure 1 and What is the number of clients for Table 1?* We apologize for the unclear statement. We employ one of the local models (ResNet18) to extract features from its corresponding local dataset, and the synthetic data of DENSE, Co-Boosting and our proposed method. The feature is extracted from the final layer (before the classifier), and we use t-SNE plots to illustrate the feature distribution. This visual comparison aims to demonstrate the effectiveness of our method in capturing the data distribution. For all experiments presented in our paper, the default number of clients involved is 10, unless otherwise specified. We will include these details in our final version. *Q4 What is the computational cost for the core-set selection step?* Thanks for the valuable question. In the Core-Set selection stage, the client runs $K$ (number of patch) inferences on each image $x_i\in X_t$ ($X_t$ indicates local dataset), i.e. the computational cost = $\mathrm{FLOPs}\times \\|X_t\\| \times K$ *Q5 It is unclear which parts of the algorithm contribute to handling the heterogeneous data. It seems that coreset selection could contribute well if intentionally sampling (almost) balanced data.* Thanks for the detailed comments. Data heterogeneity leads to inconsistencies in local models (known as client drift), which is a significant challenge for sharing model-based methods. Our method addresses this issue by transmitting synthetic distillate, which avoids the necessity for ensembling inconsistent local models and, consequently, mitigates the effect of client drift. Moreover, a pre-trained autoencoder is introduced to optimize distillate in its latent space. This manner prevents overfitting to patterns that are only recognized by the local model, which further mitigates the impact of data heterogeneity. --- Rebuttal 2: Title: Looking Forward to Further Discussions Comment: Dear Reviewer 3gL8, We would like to thank you once again for the insightful feedback and great efforts in reviewing our paper. Please let us know if you have follow-up concerns, and we are eager to engage in any further comments. Thanks! --- Rebuttal 3: Title: Thank you for the rebuttal Comment: Thank you for the rebuttal, which partially addresses my original questions. It would be highly beneficial to delve deeper into the required capacity of an autoencoder, especially when it is trained on a domain different from the one being applied. The exploration of circumstances under which adaptation might fail, along with potential techniques to mitigate such failures or indicators that could signal these failures, is crucial. A deeper understanding of these underlying assumptions would significantly enhance the practical value of the proposed method. Additionally, I share the same concerns as Reviewer Bcom regarding the original submission's emphasis on privacy. Given that communication efficiency and privacy are highlighted as key contributions (contribution 2), revising the privacy-related statements could necessitate substantial changes to the original submission and might reduce overall contributions. I appreciate the inclusion of the new experiment with LiRA in the rebuttal. To further substantiate the privacy claims, it would be advisable for the authors to incorporate a more comprehensive empirical evaluation. Additionally, the rebuttal mentions that integrating differential DP into data generation could bolster privacy. However, it would be valuable to understand how this integration might impact model performance in practical scenarios. To clarify, my follow-up questions are not intended to request additional experiments but rather to better understand the merits and limitations of the work and to have a clearer picture of the revision plan. --- Rebuttal 4: Title: Response Comment: Dear Reviewer 3gL8, Thank you for your constructive feedback. 1. Pre-trained Autoencoder. We acknowledge that the effectiveness of a pre-trained autoencoder can be significantly reduced when applied to different data domains. When validating on the medical dataset COVID-FL, we set the number of iterations for distillate synthesis to $T_{syn}=1000$, which is much more than the $T_{syn}=50$ used for natural dataset, and observed a severe performance degradation when set to $T_{syn}=50$ for COVID-FL. We believe the primary impact may be on convergence speed. To illustrate, we conducted an experiment where we replaced the encoder and decoder of the pre-trained autoencoder with randomly initialized downsample and upsample layers, respectively, during distillate synthesis. The experiment is conducted on ImageNette with ResNet18 ($\alpha=0.1,ipc=80$). As illustrated in the table below, random initialization achieved comparable performance to the pre-trained Autoencoder after twenty times more iterations. The results illustrate two conclusion: 1) our method is not sensitive to the data domain and can achieve good results with more optimization iterations, and 2) the image prior in the pre-trained Autoencoder can speed up the optimization of distillate synthesis, and a slow convergence speed may be a potential indicator of the failure of a pre-trained autoencoder. \| Method and Dataset \| Accuracy \| Iteration \| \| ----------------------------------- \| -------- \| --------- \| \| Pre-trained AE, COVID-FL \| 46.45 \| 50 \| \| Pre-trained AE, COVID-FL \| 52.65 \| 1000 \| \| Pre-trained AE, ImageNette \| 56.13 \| 50 \| \| Random Initialization, ImageNette \| 54.54 \| 1000 \| 2. Privacy. We acknowledge that our method does not provide provable privacy protection. Our primary emphasis is on empirical contributions, demonstrating that our method is superior in empirical evaluation compared to other baselines. As suggested by Reviewer nVDn, we have included the SSIM metric to further validate our method. Additionally, we perform experiments of integrating DP-SGD in our method on Tiny-ImageNet with ResNet18 ($\alpha=0.1,ipc=50$) to provide a clear view of the trade-offs involved. The results are shown below: \| \| $\epsilon=1$ \| $\epsilon=4$ \| $\epsilon=8$ \| $\epsilon=\infty$ \| \| ------- \| ------------ \| ------------ \| ------------ \| ----------------- \| \| FedSD2C \| 22.92 \| 25.13 \| 26.01 \| 26.83 \| Thank you once again for your valuable comments. Your insights have been invaluable in refining our approach and understanding the practical implications of our work. We will include all these discussions in our final version.	Summary: The paper presents FedSD2C, a novel one-shot federated learning (FL) framework that aims to improve communication efficiency, privacy preservation, and model performance. The approach addresses issues with data heterogeneity and information loss by synthesizing informative distillates from local data and sharing these instead of local models. Empirical results show that FedSD2C significantly outperforms existing one-shot FL methods. Strengths: 1. One-shot FL is a potential direction that can significantly minimize communication costs in FL. 2. This approach does not rely on sharing private data, which protects the data privacy. 3. Using V-information and fourier transform perturbation is interesting. 4. Experimental results show significant improvements. Weaknesses: 1. A critical question on the optimal observer of approximating the V-information is that, the local model is trained locally, which may not indicate good V-information of the global datasets, i.e., all local dataets. 2. The privacy is not strictly guaranteed. The perturbation with 3, 4, 5 cannot ensure data privacy. 3. Figure 2 shows the reconstruction is very similar with the original images. 4. There is no theoretical analysis on privacy and generalization of the proposed method. 5. The pre-trained autoencoder plays a key role in the framework. But the experimental study doesn't investigate the impact of the pre-trained autoencoder on the performance. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. What are the key hyper-parameters used in the core-set selection algorithm? 2. According to Algorithm 2, the core-set includes patches with different scales. How does the encoder handle patches with different scales? 3. There is still a huge performance gap between FedSD2C and Central in Table 1. What are the main factors that result in the performance drop? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors have pointed out the computational overhead on local devices as the limitation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: *Q1 A critical question on the optimal observer of approximating the V-information is that, the local model is trained locally, which may not indicate good V-information of the global datasets, i.e., all local datasets.* Thanks for the detailed comments. In the context of one-shot federated learning, where global datasets are inaccessible, our method aims to distill optimal local datasets. It's essential to recognize that while local models are indeed tailored to their respective local data domains, this specialization is not a limitation but rather an advantage. This localized understanding positions the local model as the most effective observer for the Core-Set selection. *Q2 The privacy is not strictly guaranteed. The perturbation with 3, 4, 5 cannot ensure data privacy. Figure 2 shows the reconstruction is very similar with the original images. There is no theoretical analysis on privacy and generalization of the proposed method.* Thanks for the detailed comments. Our paper emphasizes the empirical contributions of using the Fourier transform to enhance the privacy of synthetic data. In this regard, our paper performed Model Inversion Attacks to validate that our approach provides the best trade-off between privacy and performance. The image perturbations applied are designed to protect individual private information while still allowing for the synthesis of key image patterns. Therefore, similar image styles do not mean that private information has been compromised. To further validate the effectiveness of our method, we employ an improved version of LiRA[1] to conduct Membership Inference Attacks on our methods. We set the raw images of Core-Set as the canary (target data $x$), as this is the most serious case of our methods. The results confirm that our approach does not introduce more privacy risk than the sharing model-based approach, even for the most vulnerable targets. Furthermore, according to Theorem 3.2 of [2], introducing DP-SGD during the distillate synthesis stage can provide theoretical privacy guarantees for our method. \| Method \| TPR@FPR=0.1 \| \| ---------------------------------- \| ----------- \| \| Sharing Model (DENSE, Co-Boosting) \| 22.81 \| \| FedSD2C \| 20.13 \| [1] Aerni, M, et al. Evaluations of Machine Learning Privacy Defenses are Misleading. 2024. [2] Xiong Y, et al. FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning. 2023. *Q3 The pre-trained autoencoder plays a key role in the framework. But the experimental study doesn't investigate the impact of the pre-trained autoencoder on the performance.* Thanks for the detailed comments. In fact, we have explored the impact of pre-trained Autoencoders on communication efficiency in Table 3 of our paper. The results show that the introduction of pre-trained Autoencoders reduces the communication cost and achieves better results. To further validate its effectiveness, we perform experiments w/ and w/o Autoencoder during distillate synthesis ($ipc=50,\alpha=0.1$). The table clearly illustrates the performance gain, highlighting the advantage of pre-trained Autoencoders in our method. \| Method \| TinyImage \| ImageNette \| \| -------------- \| --------- \| ---------- \| \| FedSD2C w/o AE \| 24.35 \| 46.43 \| \| FedSD2C \| 26.83 \| 47.52 \| In addition, we conducted experiments on the medical dataset COVID-FL [1] to verify that the employment of Autoencoder can be extended to more different data domains. The results are as follows: \| Method \| 0.1 \| 0.3 \| 0.5 \| \| ------- \| ----- \| ----- \| ----- \| \| DENSE \| 46.15 \| 57.55 \| 62.83 \| \| CoBoost \| 45.07 \| 60.27 \| 65.61 \| \| FedSD2C \| 52.65 \| 62.50 \| 66.68 \| [1] Yan R, et al. Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging. 2023. *Q4 What are the key hyper-parameters used in the core-set selection algorithm? According to Algorithm 2, the core-set includes patches with different scales. How does the encoder handle patches with different scales?* We apologize for the unclear statement. For each image $x_i$, we employ the `torchvision.transform.RandomResizeCrop` $K$ times to generate a collection of patches. Each patch with different scales will be resized to the resolution of its original image. We will revise this in the final version. *Q5 There is still a huge performance gap between FedSD2C and Central in Table 1. What are the main factors that result in the performance drop?* Thanks for the question. This gap can be attributed to the inherent challenges of one-shot federated learning. For example, the data heterogeneity across different clients can lead to the generation of noisy soft labels, which can impede the server model from extracting accurate knowledge from the synthetic data. There is a natural trade-off between communication efficiency, which is crucial in federated learning to minimize communication overhead, and the performance of the server model, which requires sufficient information to learn effectively. Despite these challenges, it is important to recognize our method's strength. Our approach demonstrates the lowest performance gap and best communication efficiency among comparable methods, thereby offering optimal practical utility for one-shot federated learning environments. We believe that the trade-offs our approach entails are justified and favorable, given the current state of other methods and the inherent constraints of one-shot federated learning. --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: Thanks for the clarification and responses. I will rasie my rating to 5. --- Rebuttal 2: Title: Looking Forward to Further Discussions Comment: Dear Reviwer EquD, We would like to thank you again for your constructive comments and kind effort in reviewing our submission. We kindly ask you to inform us if our replies have successfully resolved your concerns, and we are more than happy to address any further comments. Thanks! --- Rebuttal 3: Title: Thank you for raising the score! Comment: Thank you very much for your acknowledgment of our rebuttal. We will include the results into our revision.	Summary: This paper proposes a one-shot federated learning approach designed to enhance privacy protection, communication efficiency, and model performance. Firstly, the authors introduce a Core-Set selection method based on V-information to extract the most informative data from the original dataset. The amplitude spectrum of the images in the Core-Set is then perturbed using a Fourier transform, and these perturbed images are input into an Autoencoder to obtain their representations. Finally, these representations are transmitted to the server, which decodes the image information from the representations for training. Strengths: The one-shot learning method proposed in this paper enhances privacy protection, transmission efficiency, and model performance compared to previous methods. The approach appears innovative, employing theoretically grounded techniques such as the Core-Set selection method, and using an Autoencoder and decoder for data transmission, which improves both transmission efficiency and model performance. Weaknesses: 1. In the Core-Set Selection stage, the authors do not explicitly define the patches used in Level 1: identifying the most informative image segments. I have the following questions: Are the patches the same size across different datasets? How many patches are extracted from each image? Does the size or number of patches affect the results? If so, how can one determine the optimal patch size? 2. In the perturbation stage, the authors do not provide sufficient reasoning for using the Fourier transform. Why must the Fourier transform be used instead of other transformations? For example, does using a wavelet transform and merging its low-frequency components achieve similar effects? Additionally, the authors mention that the merged image can be random noise. Is this too idealistic? Does using particularly weak random noise also achieve similar effects? 3. During the transmission phase using an Autoencoder, does the performance of the model depend on the pre-training of the Autoencoder? If a pre-trained model is not used, does this method still work effectively? 4. It would be beneficial for the authors to create a framework diagram of the entire method to enhance readability. 5. The paper contains many capitalization and punctuation errors, such as in line 50 and line 229. Additionally, please carefully review the citations in the article as they are quite disorganized, including issues with formatting, capitalization, and more. Please ensure consistency throughout. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. The paper utilizes several large datasets, but is the method effective for low-resolution datasets such as CIFAR-10? 2. The authors should provide the model performance using only the Core-Set, where clients send all their original Core-Set data to the server for centralized training. 3. Please provide additional visual metrics beyond PSNR to make the results more convincing, such as SSIM. 4. The authors should conduct ablation experiments to demonstrate the performance of each part of the proposed method. For example, evaluating the model performance without the Core-Set data selection step. 5. Since only representations are ultimately transmitted, and attackers cannot recover images from representations, I wonder if the image perturbation stage is still necessary? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Yes, have discussed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: *Q1 Patch size, number of patch and influence of number of patch* A1 We apologize for the unclear statement. For each image $x_i$, we employ the `torchvision.transform.RandomResizeCrop` $K$ times to generate a collection of patches. For patch size, we set the `scale=(0.08, 1.0)`, which is to collect diverse image patches, so we don't set a fixed size. For the number of patches, we perform experiments to determine the best $K$ empirically. The table below indicates that performance improves with an increased $K$, stabilizing when $K$ reaches 10. Consequently, we have empirically set $K=10$ based on these observations. \| $K$ \| 1 \| 3 \| 5 \| 10 \| 20 \| 30 \| \| ------------ \| ----- \| ---- \| ----- \| ----- \| ----- \| ----- \| \| TinyImagenet \| 16.58 \| 19.8 \| 21.66 \| 23.34 \| 23.38 \| 23.29 \| \| ImageNette \| 52.89 \| 54.8 \| 55.06 \| 55.13 \| 54.06 \| 53.45 \| *Q2 Motivation for Fourier transfomation. Merging random noise* We would like to thank Reviewer nVDn for the suggestion. The employment of the Fourier transform is for its ability to balance privacy preservation with information retention. Indead, wavelet transform is also suitable for our proposed methods. But our decision was driven by its simplicity. We appreciate the reviewer's insight and will consider exploring the wavelet transform and its potential benefits in future research. For image perturbations, we focus on preserving the frequency components of the Core-Set sample. The information in the amplitude component can be reconstructed in the latent space of the Autoencoder based on the image priors during synthesis. Since the noise does not follow the image a priori, merging with it will have a performance impact, but it can also lead to greater privacy protection. In response, we conducted an experiment to replace the merged images with Gaussian noise (ResNet18, $\alpha=0.1,ipc=50$, Tiny-ImageNet). The experimental results show that noise merging can be used as a supplement to the need for stronger privacy protection. \| Method \| Acc. \| PSNR \| \| ---------------------- \| ----- \| ----- \| \| FedSD2C (Merging with Gaussian noise) \| 22.21 \| 12.91 \| \| FedSD2C \| 26.83 \| 16.95 \| Q3 Effectiveness of Autoencoder, only Core-Set and performance of Core-Set selection step 1) The introduction of pre-trained Autoencoders is crucial as it provides generalized image priors. This image prior helps synthetic distillate to prevent overfitting to localized data patterns, thus reducing the negative effects of data heterogeneity. The lower resolution of the latent space representations can significantly boost communication efficiency. Our experimental results presented in Table 3 of our paper, confirm the benefits of employing a pre-trained Autoencoder. Moreover, we conduct experiments w/o Autoencoder under the same setting (ResNet18, $\alpha=0.1,ipc=50$). The table clearly illustrates the effectiveness of our method even when not using a pre-trained model. However, the performance gain is substantial, highlighting the advantage of pre-training in our method. 2) Thanks for your feedback. We now add the results of only using Core-Set (ResNet18, $\alpha=0.1,ipc=50$) and perform an ablation study by replacing Core-Set selection with random selection. As depicted in the table, better performance can be achieved by simply transferring Core-Set, but this comes at the cost of privacy compromise. Random selection struggles to capture the necessary data diversity and representativeness, resulting in the poorest performance. \| Method \| Tiny-ImageNet \| ImageNette \| \| --------------------------------------------- \| --------- \| ---------- \| \| Core-Set \| 31.01 \| 60.54 \| \| FedSD2C w/o AE \| 24.35 \| 46.43 \| \| FedSD2C w/o Core-Set (Random selection) \| 23.32 \| 42.06 \| \| FedSD2C \| 26.83 \| 47.52 \| *Q4 Experiments on CIFAR-10* Thanks for the pertinent comments. Our approach focuses on efficiency on large data rather than low-resolution datasets. Core-Set selection stage only requires inference, and distillate synthesis effectively reduces the number of parameters by optimizing in the latent space. This efficiency-tailored approach results in the synthesis of a more compact data compared to the DFKD-based methods. We have conducted a thorough evaluation of our approach on the CIFAR-10 dataset. As depicted in Table 2 (PDF), there is an initial performance discrepancy at the standard setting of $ipc=50$. However, upon increasing the amount of synthetic data ($ipc=500$), our method achieves results comparable results. *Q5 Lack of SSIM.* As suggestions, we include the SSIM in our privacy evaluation. The results indicate that our method achieves the best tradeoff between performance and privacy protection. \| Privacy-preserving \| Acc. \| PSNR \| SSIM \| \| ------------------------- \| ----- \| ----- \| ----- \| \| ours ($\lambda=0.8$) \| 20.85 \| 16.95 \| 35.89 \| \| $Gaussian (s=0.2, p=0.2)$ \| 19.32 \| 23.52 \| 68.56 \| \| $Gaussian (s=0.2, p=0.1)$ \| 21.48 \| 27.51 \| 78.90 \| \| FedMix \| 13.86 \| 16.26 \| 56.91 \| *Q6 The necessity of image perturbations* Although it is not possible for attackers to directly reconstruct images from their representations, there remains a risk of a semi-honest server attempting to infer private information from the data. By introducing image perturbation, we make it significantly more challenging for any adversary to deduce private information from synthetic distillate. Empirical results on the Model Inversion attack and Membership Inference attack (Table 2 of PDF file) demonstrate that our method achieves a superior balance between privacy protection and utility. *Q7 Punctuation errors and framework diagram* We would like thank the Reviewer nVDn for suggestions. We will revise them in final version. --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: Dear authors, Thanks for your response. You have addressed most of my concerns, so I will raise the score to 5. However, one remaining concern is that you need to compare the Fourier transform and wavelet for different scenarios. For certain scenarios, wavelet may outperform the Fourier transform. --- Reply to Comment 1.1.1: Title: Response Comment: Dear Reviewer nVDn, We would like to thank you for your constructive feedback. As suggestion, we perform experiments using the Wavelet Transform on Tiny-ImageNet with ResNet18($\alpha=0.1,ipc=50$). The results indicate that Wavelet Transform offers greater scalability in privacy protection. By increasing $\lambda$, the PSNR/SSIM can be reduced to as low as 12.90/15.30. When the accuracy is comparable to that of Fourier transform (Wavelet $\lambda=0.5$ vs. Fourier $\lambda=0.8$), the PSNR/SSIM of Wavelet transform is lower. We sincerely appreciate your innovative comments and will include this discussion in our final version. \| \| Acc. \| PSNR \| SSIM \| \| ---------------------- \| ----- \| ----- \| ----- \| \| Wavelet($\lambda=0.1$) \| 28.05 \| 18.86 \| 44.76 \| \| Fourier($\lambda=0.1$) \| 28.22 \| 20.54 \| 51.50 \| \| Wavelet($\lambda=0.5$) \| 26.91 \| 15.22 \| 27.34 \| \| Fourier($\lambda=0.5$) \| 28.09 \| 18.06 \| 43.26 \| \| Wavelet($\lambda=0.8$) \| 26.06 \| 12.90 \| 15.30 \| \| Fourier($\lambda=0.8$) \| 26.83 \| 16.95 \| 35.89 \| --- Rebuttal 2: Title: Looking Forward to Further Discussions Comment: Dear Reviewer nVDn, We sincerely appreicate your great efforts in reviewing our submission. Your constructive comments really help improve our paper. Please do let us know if our response has addressed your concerns, and we are more than happy to address any further comments. Thanks!	Rebuttal 1: Rebuttal: Dear Reviewers and ACs, We would like to thank the reviewers' insightful reviews and constructive comments on our manuscript. We have carefully considered all the suggestions and made the following changes: 1. We have included an additional datasets. By doing so, we aim to demonstrate our methods can adapt to low-resolution datasets (CIFAR-10) and vairous data domains (medical dataset COVID-FL[1]) 2. To further substantiate the privacy-preserving capabilities of our proposed method, we have performed Membership Inference Attack and provided comprehensive explanations to mitigate privacy concerns. Thank you once again for your valuable feedback. [1] Yan R, et al. Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging. 2023. Pdf: /pdf/f8ad943a903716039c1a7a65adaa0337ebaca778.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks	Accept (spotlight)	Summary: This paper presents two novel algorithms for computing the Lipschitz constant of feedforward neural networks (NN). The starting point is a previously-known semi-definite programming (SDP) problem which enables to compute the Lipschitz constant. The paper proposes a decomposition of this SDP in sequential subproblems over layers, then relax the subproblems to enable iterative computations across layers, instead of solving a joint problem on all layers. This approach scales much better with both width and depth, as demonstrated by experiments on neural networks at initialization and after training on MNIST. Strengths: - The paper is overall well-written (see caveat in the weakness section). - The question of computing the Lipschitz constant of neural networks is important for a number of downstream tasks. The proposed method provides estimates that are experimentally on-par with approaches based on SDP methods, while being order of magnitude faster. Weaknesses: [EDIT (Aug.7): the rebuttal answered my questions adequately. In particular, the method does provide provably-correct upper-bounds.] - Theoretical results in Section 3.3 are a bit hard to follow, because the section gives the story behind the proposed relaxation, as well as geometric interpretation, but does not provide a main result summarizing the theoretical guarantees of the proposed approach. This is a key point, because provably correct upper-bounds on the Lipschitz constant are of course much preferable. Although it is suggested that the proposed algorithms are indeed provably correct, it is not clearly stated in the paper. So the paper would highly benefit from a clear statement on this fact, as well as a summary of the theoretical results into a theorem (see also Questions). - The comparison with methods in the literature is limited to SDP methods, which is OK given that the main contribution is to provide a clever relaxation of these methods, but still a broader comparison would have been interesting. - The approach only applies to feedforward NN. Technical Quality: 3 Clarity: 2 Questions for Authors: - I believe that Proposition 4 shows that ECLipsE-Fast provably gives an upper-bound on the true (unknown) Lipschitz constant of the neural network. Is this correct? - Does ECLipsE also always give an upper-bound on the true Lipschitz constant? If so, are there assumptions for this to hold? If not, at which steps are approximations made? I guess this should more or less follow from Lemmas 1 and 2, and Propositions 2 and 3, but it is not clearly stated in the paper. Minor remarks: - Authors could consider typesetting their algorithms as “Eclipse” and “Eclipse-Fast” to improve readability. - line 199: “such” missing. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations are adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your encouraging evaluation of our work and thoughtful questions. We address all of them in detail as follows. >Theoretical results in Section 3.3 are a bit hard to follow, because the section gives the story behind the proposed relaxation, as well as geometric interpretation, but does not provide a main result summarizing the theoretical guarantees of the proposed approach. This is a key point, because provably correct upper-bounds on the Lipschitz constant are of course much preferable. Although it is suggested that the proposed algorithms are indeed provably correct, it is not clearly stated in the paper. So the paper would highly benefit from a clear statement on this fact, as well as a summary of the theoretical results into a theorem (see also Questions). Theorem 2 provides a provable guarantee for the upper-bounds on the Lipschitz constant estimates. Specifically, we show that as long as there exist $\Lambda_i>0$, $i\in \mathbb{Z}_{l-1}$, such that the inequalities in (4) hold, the Lipschitz estimate we obtain is a strict upper bound for the Lipschitz constant. Thus, the existence of positive $\Lambda_i$s and the conditions in (4) already provide the theoretical guarantee. Then, we proceed to develop two algorithms to find the positive $\Lambda_i$s that will satisfy the provable guarantees in Theorem 2. Finally, the theory and intuition for finding good $\Lambda_i$s are detailed in Section 3.3. We further note that the relaxation between ECLipsE and ECLipsE-Fast solely pertains to finding positive $\Lambda_i$s while trading off computational speed for accuracy, while the upper bound on the Lipschitz constant is still strict for both algorithms. Finally, as we discuss in the General Response - point III, we will move the pseudo-code of the algorithm, and additional theoretical results from the Appendix to the main text for clarity of exposition as suggested by the reviewers. >The comparison with methods in the literature is limited to SDP methods, which is OK given that the main contribution is to provide a clever relaxation of these methods, but still a broader comparison would have been interesting For non-SDP based methods, we have included a benchmark method CPLip (green dash line in Fig. 3, 4) in our experiments, which turns out to be exponentially expensive in computational cost and is therefore not scalable to the deep neural networks we consider (see Fig. 4). Of course, as the reviewer rightly points out, we do indeed focus most of our comparisons on SDP methods. >The approach only applies to feedforward NN. We thank the reviewer for this question. Since this was also raised by other reviewers, we have answered point I of the General Response. >I believe that Proposition 4 shows that ECLipsE-Fast provably gives an upper-bound on the true (unknown) Lipschitz constant of the neural network. Is this correct?\ Does ECLipsE also always give an upper-bound on the true Lipschitz constant? If so, are there assumptions for this to hold? If not, at which steps are approximations made? I guess this should more or less follow from Lemmas 1 and 2, and Propositions 2 and 3, but it is not clearly stated in the paper. Propostion 4 states the closed form solution for $\Lambda_i$s and guarantees the positive definiteness of $M_i$ at each stage. As discussed in the first point of this review response above, the Lipschitz estimate given by ECLipsE and ECLipsE-fast are thus both provably strict upper bounds by Theorem 2. We will further clarify these points, as well as address typos and minor suggestions from the reviewer in the final version of our paper. --- Rebuttal Comment 1.1: Title: Thank you for the rebuttal Comment: I thank the authors for their precise rebuttal, which answers adequately my questions. I raised my score accordingly. --- Reply to Comment 1.1.1: Comment: Thank you for considering our response and for increasing your evaluation. We appreciate your feedback on the paper!	Summary: This paper tackles the problem of computing the Lipschitz constant of a neural network. Since computing the exact Lipschitz constant is NP-hard, efforts have been made to obtain tight upper bounds on the Lipschitz constant. This paper builds on the work of LipSDP [1], which involves solving a large matrix verification problem. Since that the large matrix verification problem grows significantly for both deeper and wider networks, this paper proposes a compositional approach to estimate the Lipschitz constants of deep feed-forward neural networks more efficiently. First, the authors obtain an exact decomposition of the large matrix verification problem into smaller sub-problems and, then, exploiting the underlying cascade structure of the network the authors develop two algorithms to compute a bound on the Lipschitz constant: - The first algorithm explores the geometric features of the problem and provides a tight estimate of the Lipschitz constant by solving small semidefinite programs (SDPs) that are only as large as the size of each layer. - The second algorithm relaxes these subproblems and provides a closed-form solution to each subproblem for extremely fast estimation, altogether eliminating the need to solve SDPs altogether. Finally, the authors provide extensive experiments to show the different levels of tradeoffs between efficiency and accuracy of the two algorithms. They show that their approach provides a steep reduction in computation time while yielding Lipschitz bounds that are very close to, or even better than, those achieved by state-of-the-art approaches. Strengths: - The paper is clear and well written. The problem of providing a scalable algorithm for computing the Lipschitz of neural networks is interesting and important. - The exact decomposition of the large matrix verification problem into smaller subproblems is very interesting. - The two algorithms for computing the sequence of inequalities provide interesting trade-offs. The first algorithm (ECLipsE) looks, if I understand correctly, like a direct improvement of LipSDP, since ECLipsE provides the same value as LipSDP in a more efficient way. - The second algorithm also looks interesting as it provides a way to compute the Lipschitz constant without SDPs. Weaknesses: - it looks like the approach is restricted to a very limited set of neural networks (feedforward neural networks), can the approach be used for convolutional neural networks? - In the experiments, the authors use randomly generated neural networks for their first set of experiments, in my experience it is usually easier to compute SDP on random weight matrices than on trained weight matrices due to conditioning, could the authors provide results of these experiments with trained networks? - Can the authors clarify if ECLipsE has to compute all subproblems at the same time or if a sequential approach is possible? - I assume that the ECLipsE algorithm uses Matlab for SPD optimization, have the authors tried using a deep learning framework (e.g. PyTorch) for ECLipsE-Fast? Could ECLipsE-Fast be used during training, e.g. for regularization? Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your encouraging evaluation of our work and thoughtful questions. We address all of them in detail as follows. >It looks like the approach is restricted to a very limited set of neural networks (feedforward neural networks), can the approach be used for convolutional neural networks? Since this was also raised by other reviewers, we have answered it in point I of the General Response. >In the experiments, the authors use randomly generated neural networks for their first set of experiments, in my experience it is usually easier to compute SDP on random weight matrices than on trained weight matrices due to conditioning, could the authors provide results of these experiments with trained networks？ We do in fact have experiments for both cases. Section 4.1 considers randomly generated neural networks and Section 4.2 considers neural networks trained for MNIST tasks. Note that our algorithms estimate the Lipschitz constant for a given model, that is, all the parameters are fixed and the algorithms are implemented out after the model is trained. Therefore, we did not observe any additional difficulties in estimating the Lipschitz constant for trained weight matrices. Please see Section 4.2 for detailed results on trained weight matrices. >Can the authors clarify if ECLipsE has to compute all subproblems at the same time or if a sequential approach is possible? ECLipsE computes the subproblems by sequence. Starting with $i=1$, the SDP problem as expressed in (6) requires the information matrix $M_{i-1}$ passed on from the computation at the $(i-1)$-th stage. Therefore, it is a sequential approach, aligning with the cascaded neural network structure. >I assume that the ECLipsE algorithm uses Matlab for SPD optimization, have the authors tried using a deep learning framework (e.g. PyTorch) for ECLipsE-Fast? Could ECLipsE-Fast be used during training, e.g. for regularization? Yes, ECLipsE uses Matlab with solver Mosek to solve the SDPs. For ECLipsE-Fast, the solutions for $\Lambda_i$s at each step are actually obtained in closed-form as stated in Proposition 4, and at the last step the Lipschitz estimate is also in closed-form as given in Proposition 1. Thus, there is no need to solve any SDPs at all for ECLipsE-Fast; we obtain fast estimates by only evaluating closed-form analytical expressions without the need to use any deep learning framework. Lastly, we believe our efficient, scalable and accurate methods like ECLipsE-Fast will facilitate robust training for neural networks in the future, although this is beyond the scope of the present paper. --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications. I agree that the unrolling approach could be used in the context of this paper - although it might not be the most scalable approach (unrolling a large convolution leads to extremely large matrices). I agree that the exploration of other types of architectures could be left to future work, and I am in favor of accepting this paper and will raise my score accordingly. --- Reply to Comment 1.1.1: Comment: Thank you for considering our response, and raising your evaluation! We also appreciate your feedback on improving our manuscript! We agree that unrolling the convolutional layer and applying FNN based methods may not be the most practical solution. Further study is necessary to develop scalable algorithms for other network architectures, which will be the subject of future work.	Summary: The paper proposes two algorithms, ECLipsE and ECLipsE-Fast, to estimate the Lipschitz constant of a feed-forward neural network. The estimation of the Lipschitzness plays a crucial role in certifying the robustness of neural networks and is known to be an NP-hard problem. The proposed algorithms are based on the LipSDP of Fazlyab et al. (2019), which describes the semidefinite program (SDP) that an upper bound of such a Lipschitz constant should generally satisfy. The authors decompose the original large SDP into smaller layer-wise SDPs to improve the scalability of the original approach. The validity of the resulting methods, ECLipsE and ECLipsE-Fast, is supported by theoretical analyses. Experiments show a steep reduction in computation time while maintaining competitive accuracy compared to the LipSDP. Strengths: Originality: The proposed algorithm is novel and clearly distinguishes itself from prior works. Quality: The motivation of the work is clearly stated and explained. Prior works are also well-discussed. The authors provide thorough and sound mathematical justification and geometrical intuitions for the two algorithms. Clarity: The paper is well-written including the methodology, and the motivation is clear. Significance: Compared to LipSDP, the proposed algorithm provides significant improvement in terms of efficiency and addresses concerns stated in the motivations of the paper in the beginning. Overall, I feel this is a good paper with a promising approach equipped with a thorough and interesting mathematical justification. Weaknesses: Overall, in my view, the main weakness of this paper is that there is no explicit comparison with other works trying to improve the scalability of LipSDP (such as [20]). As a result, while the paper indeed improves the original LipSDP in a new way, it is unclear how significant this work is taking into account existing literature. In addition, the limitations should be more carefully discussed. See below for further detailed comments and advice. ### Quality (W-Q1) There are some typos: l.60 (constans), p.3 eqation (3) (index of the bottom right element is i+1 but should be l), l.172 (in the matrix WMW there are 2 unnecessary “L”), l.264 (computatinoal), between l.420 and l.421 (i\in \mathbb{R}^n), p.13 equation (18) (if i=0 should be if i=1?), l.477 (functionsare), l.480 (the is norm), (W-Q2) I feel that some use of words is misleading. 1. l.60 “We develop a sequential Cholesky decomposition technique to obtain […]“: If this is about Theorem 2, it directly uses the result of Agarwal et al., so maybe “use, employ” would be better than “develop”. 2. l.10 “The first algorithm [...] enables us to provide a tight Lipschitz constant”, l.56 “ algorithm [...] enables us to provide an accurate Lipschitz”... What do you mean by “tight” and “accurate”? Since the proposed algorithms are using some simplifications, I think that those adjectives should be relative, i.e., only used in comparison with something else. Notably, the experiments (e.g., Figure 3) show that CPLip is far more precise than the proposed algorithms so the authors should clarify the meaning of “tight” and “accurate” (or delete them if there is no justification) when describing their own algorithms. 3. Between l.462 and l.463, you write “$N/ci(M_i)^{-1}\ge0$”. Semi-positive definiteness was only defined for symmetric matrices but this one is not necessarily symmetric. There should be an easy fix, but it is ambiguous what you mean by this sign. (W-Q3) In Proposition 3, the authors simplify Theorem 2 by setting $M_i$ to $c_iW_{i+1}^\top W_{i+1}+N$ without any proper discussion about this choice (Q2). (W-Q4) The proof of positive definiteness of $M_i$ is missing in Proposition 3 (Q3). (W-Q5) (minor) Some references should be adjusted: [2] was accepted at ICLR2018, some capital letters are missing (L of lipschitz)... ### Clarity (W-C1) Some parts may require clarifications (See also Questions): 1. (minor) l.111 “[30] provides a counterexample to the most accurate approach”: The concrete property of the most accurate approach disproved by [30] could be explained in a few words. 2. (minor) A mathematical comparison of the computational complexity between your approach and LipSDP will largely help the reader to quickly understand the difference in scalability. (W-C2) While limitations of proposed algorithms are all stated, they are dispersed throughout the paper. A short subsection summarizing them would be useful. (See also Limitations) (W-C3) It was a bit difficult to understand the idea of the proposed algorithms from Subsection 3.2. Perhaps showing Algorithm 1 in the main text would be better. ### Significance (W-S1) (major) One of the main contributions of this work is the scalability of the proposed algorithms. However, the authors do not compare their method with other accelerations of LipSDP. Therefore, it is difficult to situate their work within the broader context of efforts to improve the scalability of LipSDP. At least, comparing the algorithms with that of Wang et al. (2024) [20] is important to clarify these points. (W-S2) The algorithms were run on medium-scale neural networks, and it is difficult to imagine the scalability of ECLipsE and ECLipsE-Fast. Experiments with even larger architectures (for example, those for training on CIFAR or ImageNet) would be more convincing. (W-S3) I feel experiments of Subsection 4.2 are a little bit redundant (Q4). (W-S4) Limitations of the algorithms should be discussed in more detail and more explicitly. Technical Quality: 3 Clarity: 3 Questions for Authors: Q1: Is Theorem 1 necessary and sufficient? If not, this should be added to the limitations of the work. Q2: Why did the authors set $M_i$ to $c_iW_{i+1}^\top W_{i+1}+N$ in Proposition 3? Q3: How is the positive definiteness of $M_i$ guaranteed in Equation (6) and Proposition 3? Q4: What was the motivation to run experiments of Subsection 4.2? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: There are several limitations of the algorithms that are worth mentioning: 1. The algorithms (Lemma 1) need that the last weight matrix is full row rank. So, we cannot blindly apply them to any feed-forward neural network. 2. There is a simplification when transforming Theorem 2 into Proposition 3 by limiting the expression of $M_i $ to $c_iW_{i+1}^\top W_{i+1}+N$. This may lead to looser bounds than the original LipSDP. 3. The algorithms cannot be applied to CNNs and residual networks. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for a thorough reading of our manuscript and providing several suggestions to improve the clarity of our presentation (see General Response - point III). We individually address the technical questions raised by the reviewer below, and provide additional experiments benchmarking our algorithms (results in attached PDF of the General Response and discussion below). >One of the main contributions of this work is the scalability of the proposed algorithms. [..] At least, comparing the algorithms with [..] [20] is important. While [20] was too recent to reproduce at the time of our submission (<1 month), we can now provide additional experiments using their open-source code and considering the same NNs in Section 4.1 Case 1 and Case 2. The results in the attached PDF (in the General Response) demonstrate that (i) ECLipsE and [20] have similar computation times for smaller networks; however, the computation time for [20] grows more rapidly for both deeper and wider networks, and (ii) ECLipsE-Fast remains orders of magnitude faster than all algorithms while providing Lipschitz estimates that are very close to those achieved by LipSDP-Layer, and (iii) importantly, [20] provides inadmissible estimates for moderate networks, returning as much as $10^4-10^6$ times and 10-100 times the trivial bound in Tables 2a and 1a respectively. Note that all the estimates are normalized with respect to trivial upper bounds. Another work on improving the scalability of LipSDP is [33]. However, as acknowledged by the authors in their footnote, their acceleration depends on LipSDP-Network, which is proved to be invalid by [30]. Therefore, we do not include it as a benchmark. >The proof of positive definiteness of $M_i$ is missing in Proposition 3. Thank you for raising this question. The first part of the proof of Proposition 3 (l.463-465) showing the non-emptiness of the feasible region implies the positive definiteness of $M_i$ for the next stage. This is because $M_i=\Lambda_i-\frac{1}{4}\Lambda_iW_i(M_{i-1})^{-1}W_i^T\Lambda_i$. From (10), $M_i>c_iW_{i+1}^TW_{i+1}>0$ as $c_i>0$ and $W_{i+1}^TW_{i+1}\geq 0$. Then, with $M_0>0$, $M_i$ is guaranteed to be positive definite at each step. We will supplement the proof for clarity. >In Proposition 3, the authors simplify Theorem 2 by setting $M_i$ to $c_iW_{i+1}^TW_{i+1}+N$ without any proper discussion about this choice. There is no simplification here, and we can write $M_i$ exactly as $c_iW_{i+1}^TW_{i+1}+N$. We briefly summarize the proof here. Since $M_i$ is positive definite (see response above), and $W_{i+1}^TW_{i+1}$ is positive semidefinite, there exists constant $C$ such that for any $c\in[0,C]$, $M_i-cW_{i+1}^TW_{i+1}\geq 0$. Now, for the $i$-th layer, let $c_i$ be the largest possible $C$ such that $M_i-cW_{i+1}^TW_{i+1}\geq 0$ holds. Then, $N=M_i-c_iW_{i+1}^TW_{i+1}$ is a positive semidefinite matrix and also a singular matrix. We show this by contradiction in the proof of Proposition 3; see lines 465-468, starting at `we now prove by contradiction..'. (we do not include this here for brevity). Therefore, we can equivalently write $M_i=c_iW_{i+1}^TW_{i+1}+N$, where $N$ is a positive semidefinite matrix and also a singular matrix. >The algorithms were run on medium-scale neural networks, and it is difficult to imagine the scalability of ECLipsE and ECLipsE-Fast [..] ECLipsE-Fast has closed-form solution and is therefore scalable, no matter the network size. We did not present larger networks because the existing benchmarks already exceed the cutoff time of 30 min for medium-sized networks. Further, from existing experiments, our acceleration is already pronounced with promising accuracy, and the advantage will be even more significant for larger networks. >What was the motivation to run experiments of Subsection 4.2? The motivation for Section 4.2 is two-fold. First, we show that our algorithms apply not only to randomly generated weights but also to those trained for some specific tasks. Second, although ECLipsE provides a tighter estimate in general, we show an interesting case where ECLipsE-Fast is more favorable compared to ECLipsE (with similar accuracy but much faster speed). >[..] need that the last weight matrix is full row rank [..] cannot blindly apply them to any feed-forward neural network. Since this was also raised by another reviewer, we address it in point II of the General Response. In short, this assumption is not necessary, and both algorithms apply even when it is not satisfied. >[..] cannot be applied to CNNs and residual networks. Since this question was also raised by other reviewers, we have answered it under General Response - point I. >(a) l.10 [..] What do you mean by “tight” and “accurate”? \ (b) Between l.462 and l.463, you write “$N/c_i(M_i)^{-1}\geq 0$ ”. Semi-positive definiteness was only defined for symmetric matrices but this one is not necessarily symmetric [..] (a) In the Introduction, "tight" and "accurate" are used in a general sense, to express that our method provides Lipschitz estimates that are comparable to existing methods while achieving significantly enhanced scalability. We will rephrase in response to the reviewer's suggestions. (b) By the definitions of $N$ and $M_i$, they are indeed guaranteed to be symmetric. >(minor) (a) l.111 “[..] the most accurate approach disproved by [30] could be explained [..]\ (b) A mathematical comparison of the computational complexity between [..] LipSDP will largely help [..] (a) [30] gives a counterexample showing that the Lipschitz estimate from LipSDP-Network is not a strict upper bound. (b) Please see point (1) of our response to Reviewer kbMy for a computational complexity analysis (not repeated here due to space constraints). --- Rebuttal Comment 1.1: Comment: Thank you for these clarifications and additional experiments. The mathematical justification of the algorithms now seems valid to me. As you suggested, a clearer explanation of the points in the proof I asked for clarification on may be beneficial for the updated version of your paper. I am in favor of accepting this paper and will raise my score accordingly. Still just one detail: >(b) By the definitions of $N$ and $M_i$, they are indeed guaranteed to be symmetric. I agree that $N$ and $M_i$ are both symmetric, but the product of symmetric matrices (e.g., $N/c_i (M_i)^{-1}$) is not necessarily symmetric. That is why I pointed out that $N/c_i (M_i)^{-1}\geq 0$ was not well-defined as $\geq$ was only defined for symmetric matrices in the Notation. Shifting the focus of the discussion in l.462-463 on the singular values without using $\geq 0$ should solve the problem anyway. --- Reply to Comment 1.1.1: Comment: We thank you for considering our response, and for raising your evaluation. We are truly grateful for the thorough and detailed suggestions on improving our manuscript! Yes, the product of two symmetric matrices is not necessarily symmetric. We can instead focus on the discussing eigenvalues of $N(M_i)^{-1}$ and prove that it can only have non-negative eigenvalues. We will edit the proof accordingly in our final version.	Summary: The paper proposes two novel Lipschitz constant estimation algorithms ECLipsE and ECLipsE-Fast. They are supported by a new decomposition theory developed for the Lip-SDP framework, derived by applying an existing theory (Lemma 2 of [31]). Experiments demonstrate the estimation accuracy and acceleration using toy radom networks and networks trained on MNIST data, by comparing with classical Lipschitz constant estimation method. Strengths: The targeted research problem is important and useful. The decomposition theory is new and the two estimation algorithms are novel. I highly appreciate the beauty of the application of Lemma 2 of [31] in the proposed theory development. The achieved result improvement is satisfactory for deep networks. Weaknesses: (1) The proposed algorithm is efficient at addressing network depth, but does not look at the width. The theory behind self-explains the success of its capability of handling depth. However, layers with very high numbers of neurons still pose challenges for the sub-problems, e.g., solving for Eq. (6) and Eq. (7). There is a lack of mentioning of this. Also the experiments only studied some modest widths up to 100 neurons. It would be good to see more empirical results with higher neuron numbers in each layer, to understand the limit of the proposed algorithms on network width. (2) In experiments, there seems a missing comparison with the “parallel implementation by splitting” version of Lip-SDP as reported in their paper [14], which was proposed to address the depth issue. (3) The method description can be improved, e.g., being more organised. The paper can present information that is more important and helpful to practitioners and general readers’ understanding in main paper, while leave some analysis and supporting lemmas to appendix. Personally, I find it helpful to see in the main paper a description of the existing Lemma 2 in [31] and the pseudo code of the proposed algorithm. Technical Quality: 3 Clarity: 2 Questions for Authors: The authors are invited to address my comments (1) and (2) in weakness section. If possible, it would be good to see some added results to help understand more the width capacity of the proposed algorithms. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Discussion on discussion is pretty limited. For instance, it can be improved around the network width issue. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your encouraging evaluation of our work and for the insightful questions on our experiments. We address all your questions as follows, and provide additional experiments to show the strength of our method for wide networks. >(I). The proposed algorithm is efficient at addressing network depth, but does not look at the width. The theory behind self-explains the success of its capability of handling depth. However, layers with very high numbers of neurons still pose challenges for the sub-problems, e.g., solving for Eq. (6) and Eq. (7). There is a lack of mentioning of this. (II). Also the experiments only studied some modest widths up to 100 neurons. It would be good to see more empirical results with higher neuron numbers in each layer, to understand the limit of the proposed algorithms on network width. (III). Discussion on discussion is pretty limited. For instance, it can be improved around the network width issue. We thank the reviewer for these important questions, and answer them here. While we acknowledge that our computational advantage is more pronounced with respect to network depth, the speedup for wide networks is also significant, especially when the network is also deep. To see this, we can assess the computational complexity for ECLipsE and LipSDP-Neuron. Suppose a neural network has $n$ hidden layers with $m$ neurons. Then, the large matrix in Theorem 1 has dimension $nm+O(1)$ and the decision variable is of size $nm+O(1)$. The computational complexity for solving an LMI with decision variables of matrix size $A$ and $B$ is $O(A^3+A^2B^2)$. Therefore, the computational cost for LipSDP (solving SDP involving the large matrix) is $O((nm+O(1))^3+(nm+O(1))^2(nm+O(1))^2)=O(n^4m^4)$. Contrarily, ECLipsE solves $n$ sub-problems as in Eq. (6), with each involving matrix of size $O(m)$ and $m$ decision variables. The corresponding total computational cost is $n\times (O(m^3+m^2m^2))=O(nm^4)$. We can see that the complexity is significantly decreased in terms of the depth, but is the same in terms of the width, immediately indicating the advantage for deep networks. Nevertheless, as $m$ grows, the difference between $O(n^4m^4)$ and $O(nm^4)$ are still enlarged drastically, especially with large $n$. More importantly, for ECLipsE-Fast, we note that we do not need to solve any SDPs as the solutions are all provided in closed-form from Propositions 1 and 4. Thus, the computational cost drops to $n\times O(m^3)=O(nm^3)$. This is the fastest one can expect if the weights on each layer are treated as a whole. Admittedly, if a neural network is considerably wide, it can still pose challenges to the sub-problems regardless of all the accelerations we have achieved, in which case we will have to apply methods that split the weights themselves, introducing some unavoidable conservativeness. From the experiment side, we initially considered networks with up to 100 neurons, since benchmark methods like LipSDP-Neuron with 50 neurons and LipSDP-Layer with 70 neurons already fail to return estimates within the cutoff time of 15 min (see Figure 4). Meanwhile, both ECLipsE and ECLipsE-Fast still work well in these settings, demonstrating our advantages regarding width. To further illustrate the strengths and limitations, we consider randomly generated NNs with 50 layers as shown below, and find that (i) ECLipsE-Fast is extremely fast even for very wide networks, with a running time of only 15.63 seconds for a width of 1000, while the computation time for LipSDP-Layer grows significantly, and (ii) ECLipsE is comparable to LipSDP-Neuron split into 5 sub-networks in terms of time performance (note that LipSDP-Neuron cannot return estimates for any of the cases without splitting, which slightly decreases its accuracy with respect to ECLipsE). We will include these additional results and discussions in the final version. Note that all the estimates are normalized with respect to trivial upper bounds. Normalized Lipschitz Estimates for Randomly Generated Neural Network with 50 Layers \| Neuron \| ECLipsE \| ECLipsE-Fast \| LipSDP-Neuron Split by 5 \| LipSDP-Layer Split by 5 \| \|--------\|-------\|--------------\|--------------\|----------------------\| \|150\|0.743745\|0.867548\|0.758217\| 0.87342\| \|200\| 0.773494 \| 0.883758\|0.785171\| 0.888306\| \|300\| >30min\|0.897008\| >30min\|0.899164\| \|400\| \|0.899916\| \| >30min\| \|500\| \|0.903529\| \| \| \|1000\| \|0.912093\| \| \| Time Used for Randomly Generated Neural Network with 50 Layers (Seconds) \| Neuron \| ECLipsE \| ECLipsE-Fast \| LipSDP-Neuron Split by 5 \| LipSDP-Layer Split by 5 \| \|--------\|---------\|--------------\|--------------------------\|-------------------------\| \| 150 \| 387.7\| 0.387262 \| 451.07 \| 93.129 \| \| 200 \| 1386.6\| 0.584115 \| 1377.9 \| 210.16 \| \| 300 \| >30min\| 1.321177 \| >30min \| 612.47 \| \| 400 \| \| 2.657505 \| \| 2110.9 \| \| 500 \| \| 3.7435 \| \| >30min \| \| 1000 \| \| 15.63342 \| \| \| >In experiments, there seems a missing comparison with the “parallel implementation by splitting” version of Lip-SDP [...] to address the depth issue. The ``parallel implementation by splitting'' version of LipSDP is implemented in Section 4.1 Case 4.3 directly using the code provided by [32], where we compare 3 ways of splitting, namely into 3, 5, and 10 layers respectively. The results are promising as discussed in Case 4.3: ECLipsE-Fast is the fastest algorithm and outperforms LipSDP-Layer regardless of how we split the neural networks. ECLipsE is also shown to be relatively more accurate and efficient than all LipSDP methods, no matter the split. >The method description can be improved, e.g., being more organised [...] I find it helpful to see in the main paper a description of the existing Lemma 2 in [31] and the pseudo code of the proposed algorithm. We thank the reviewer for their suggestions on improving the clarity of our paper, and will revise accordingly - please see General Response point III. --- Rebuttal Comment 1.1: Comment: I thank the authors for their very clear explanation on their algorithm complexity with respect to neural network depth and width, and providing experiments with higher width to demonstrate algorithm capacity, while acknowledging the limit/boundary. I am happy to see the paper to be accepted, therefore will increase my score to 7. Meanwhile, I recommend the authors to discuss around "width" in their discussion/limitation section. --- Reply to Comment 1.1.1: Comment: We thank you for your feedback on our manuscript, particularly on wider vs deeper networks, and for raising your score! We will add the experiments on width and a discussion section on limitations.	Rebuttal 1: Rebuttal: We are extremely grateful to the reviewers for their detailed, thorough, and constructive feedback. We are glad to read that the reviewers found paper to be interesting, novel, practical and well-written. We appreciate the suggestions from Reviewers kbMy, 3iDD on enhancing the clarity of writing, Reviewer kXCT on providing additional experiments benchmarking to the state-of-the-art, and Reviewers qsZt, kXCT, LaVm, 3iDD on the generalization of our algorithms to other neural network architectures. We address all the reviewers' concerns and questions in individual responses, and provide additional experiments to illustrate the strength of our methods. First, we address common questions raised by multiple reviewers and ones that provide additional benchmarking experiments here. I. On the applicability of our algorithms to other neural network architectures such as CNNs (Reviewer qsZt, Reviewer kXCT, Reviewer LaVm, Reviewer 3iDD): Several reviewers raise the question of whether our algorithms are applicable beyond feedforward neural networks (FNNs) to other classes such as convolutional neural networks (CNNs) and residual networks. While our work exploits the mathematical structure of the underlying matrices arising from cascaded architectures to develop fast algorithms, the applicability of our algorithms is not restricted to only FNNs. In the case of CNNs, we can adopt a strategy similar to LipSDP where the CNN can be unrolled into a large fully connected neural network, following which we can apply both ECLipsE and ECLipsE-Fast. While our current study focuses on significantly accelerating the computation of Lipschitz constants for FNNs, future work will involve exploring the mathematical structures of other architectures such as residual networks to develop similarly fast algorithms for Lipschitz constant estimation. II. On the full row rank assumption on the last weight matrix in Lemma 1 (Reviewer qsZt, Reviewer kXCT): We thank the reviewers for this insightful question. First, we would like to clarify that both ECLipsE and ECLipsE-Fast are still valid even if the full row rank assumption is not satisfied. This is due to the fact that at the last stage, the Lipschitz estimate is in fact given by a closed-form expression with no requirement on the row rank of $W_l$ as Proposition 1. We note that the reason that this assumption was made is solely for ease of exposition of the intuition arising from the geometric features of the problem. Specifically, as discussed after Lemma 2, if the weight matrix is full ranked, then minimizing $\sigma_{max}(F_i)$ aligns with minimizing $\sigma_{max}\left(W_l^TW_l(M_{l-1})^{-1}\right)=\sigma_{max}\left(W_l(M_{l-1})^{-1}W_l^T\right)=\sigma_{max}(F_{l})$ at the last stage, from a geometric perspective. However, we note that the algorithms themselves do not rely on this fact due to the closed-form expression in Proposition 1. Also, practically speaking, it is common to set the dimension of the last hidden layer to be much larger than the output dimension. Thus, $W_l$, as a fat matrix, is almost always full row ranked (see the discussions after Lemma 1). We acknowledge that this choice of presentation may have caused some ambiguity regarding the necessity of this assumption. We will clarify this in the final version of the paper. III. On the organization of the algorithm description in the paper (Reviewer kbMy, Reviewer kXCT, Reviewer 3iDD): Several reviewers have suggested moving the existing Lemma 2 in [31] and the pseudo code of the proposed algorithm to the main paper for better organization and presentation of the theoretical results in the paper. We thank the reviewers for this suggestion. Given the additional page available for the camera-ready version, we will incorporate them into the main body of the final paper. We will also fix all typos and take into account editorial suggestions from various reviewers in the final version. IV. Additional Benchmarking Experiments (Reviewer kXCT): Reviewer kXCT suggested that our algorithms should be benchmarked against those in [20]. While [20] was too recent to reproduce at the time of our submission (<1 month), we now provide additional experiments (see attached PDF) benchmarking our algorithms with respect to [20], considering the same NNs in Section 4.1 Case 1 and Case 2. In short, (i) ECLipsE and [20] have similar computation times for smaller networks; however, the computation time for [20] grows more rapidly for both deeper and wider networks, and (ii) ECLipsE-Fast remains several orders of magnitude faster than all algorithms while providing Lipschitz estimates that are very close to those achieved by LipSDP-Layer, and (iii) importantly, [20] provides inadmissible estimates for moderate networks, returning as much as $10^4-10^6$ times and 10-100 times the trivial bound in Tables 2a and 1a respectively. Pdf: /pdf/f2107ba17811ae6dac301b60111506577301d059.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The authors are able to decompose a particular case of LipSDP-neuron exactly into a series of sub-problems leading to the proposed algorithm ECLipSE. In the case of the relaxed LipSDP-layer, it can be shown that each sub-problem can be solved analytically and eliminate the need for solving an SDP. The proposed algorithm ECLipSE-fast can provide Lipschitz estimates of deep-NN quite fast at the expense of increased conservativeness. Strengths: -The paper is well written and the theoretical arguments are well motivated and connected to the the numerical experiments. The paper mainly builds upon LipSDP, but I believe the theoretical insights to be novel. -The insight that LipSDP may be simplified into sub-problems is a significant contribution and advances the practical value of LipSDP. In particular, the proposal of ECLipsE-Fast which is an analytical solution to a relaxed sub-problem shows a significant improvement over the previous naive product bound and is very scalable. Weaknesses: -It seems that the approach doesn’t yet apply to residual networks or CNN commonly found in state-of-the-art vision models. For this reason, it seems difficult to show improved certified robustness on common image classification benchmarks which would benefit the most from increased scalability (CIFAR10-100, Tiny-imagenet, etc). -Certified robustness is mentioned as an application at several points in the paper. While the scalable and tighter Lipschitz estimates on random networks and MNIST certainly suggest some improvements, no practical measures of certified robustness are presented (e.g. robust accuracy). For certified robust accuracy on MNIST, it is common to use 1-Lipschitz parameterization (SLL, AOL, other direct parameterizations) which eliminates the need for Lipschitz estimation. Would applying ECLipSE to composed 1-Lipschitz networks provide significantly tighter estimates in this case? Technical Quality: 4 Clarity: 3 Questions for Authors: -How does the full row rank assumption limit the applications of ECLipSE’s? Is it possible to relax this assumption? It seems like networks considered in this paper are only of constant width or decreasing in output dimension in the case of MNIST which will usually satisfy the full row-rank assumption. -Can you comment on the slight gap between the Lipschitz constant estimates shown in Appendix E between ECLipsE and LipSDP-Neuron? Is it that LipSDP is not as accurate in larger settings or is ECLipsE somehow slightly conservative? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Limitations are adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments and positive evaluation of our work. We address your concerns as follows. >It seems that the approach doesn't yet apply to residual networks or CNN commonly found in state-of-the-art vision models. For this reason, it seems difficult to show improved certified robustness on common image classification benchmarks which would benefit the most from increased scalability (CIFAR10-100, Tiny-imagenet, etc). We thank the reviewer for this question; since this was also raised by other reviewers, we have answered point I of the General Response. >Certified robustness is mentioned as an application at several points in the paper. While the scalable and tighter Lipschitz estimates on random networks and MNIST certainly suggest some improvements, no practical measures of certified robustness are presented (e.g. robust accuracy). We show in experiments that for both randomly generalized networks, as well as ones trained for specific tasks, our algorithms consistently give promising results. Thus, we believe our method will benefit studies that certify robustness by building a relationship between the Lipschitz constant and robustness metrics (e.g. robust accuracy). For instance, [19] builds Lipschitz-based surrogates for certified radius, which is a classical robustness measure for classification tasks. We believe that such applications will benefit from our efficient Lipschitz estimation algorithms. >For certified robust accuracy on MNIST, it is common to use 1-Lipschitz parameterization (SLL, AOL, other direct parameterizations) which eliminates the need for Lipschitz estimation. Would applying ECLipSE to composed 1-Lipschitz networks provide significantly tighter estimates in this case? We thank the reviewer for this question. While we consider a very general FNN structure, with only a slope restrictedness assumption on the activation functions, we anticipate that incorporating side information corresponding to specific parametrizations will yield better estimates in any algorithm. However, this is the subject of future work. Moreover, to the best of our knowledge, there is no theoretical guarantee of obtaining a tighter estimate for 1-Lipschitz networks. We now turn to the question of whether Lipschitz constant estimation is necessary, given developments like 1-Lipschitz networks. While 1-Lipschitz parameterization is commonly used in robust training and the robustness is guaranteed by way of parameterization, the 1-Lipschitz parameterization has limited expressive power. For example, AOLs restrict the output to be a sum of individual contributions from inputs, which may not be sufficient for some complex tasks. In contrast, our theory applies to FNNs, which are a very general network structures with universal expressive power adopted in various network designs. Moreover, our theoretical development only requires the rather mild assumption of slope-restrictedness of the activation function, which is satisfied for most cases. Thus, our work facilitates fast Lipschitz constant estimation for more general structures, while still being applicable to direct parametrizations. > How does the full row rank assumption limit the applications of ECLipSE’s? Is it possible to relax this assumption? It seems like networks considered in this paper are only of constant width or decreasing in output dimension in the case of MNIST which will usually satisfy the full row-rank assumption. We thank the reviewer for this question. Since this was also raised by another reviewer, we have answered it under the General Response section - see part II therein for a detailed response. In short, this assumption is not necessary, and both algorithms apply even when it is not satisfied. >Can you comment on the slight gap between the Lipschitz constant estimates shown in Appendix E between ECLipsE and LipSDP-Neuron? Is it that LipSDP is not as accurate in larger settings or is ECLipsE somehow slightly conservative? The slight gap in performance, in fact, exhibits the high accuracy of ECLipsE. While LipSDP-Neuron provides en efficient approach to estimate Lipschitz constants, it is practically not as scalable as our algorithms for large networks, and yields unacceptably long running times (>15min for 50 layers with only 60 neurons). Therefore, to implement LipSDP-Neuron, as suggested in [14], the NN must be split into several small sub-networks. However, the different sub-networks are treated independently, and their Lipschitz constants are multiplied at the end, thus completely cutting off the relationship among sub-networks. In contrast, our method always keeps the information from previous layers due to the exact decomposition in Theorem 2. This explains why LipSDP-Neuron yields less accurate results compared to EClipsE for the large networks discussed in Section 4.1 Case 3, and is precisely where our advantages lie. Theoretically, LipSDP is a centralized method where solving a large matrix SDP is unavoidable. In this context, we note that ECLipsE, being a distributed algorithm, will naturally result in a trade-off between speed and accuracy, yielding more conservative estimates than centralized algorithms like LipSDP. However, as we demonstrate in our experiments, the Lipschitz constant estimates from both ECLipsE and ECLipsE-Fast are fairly close to those obtained using LipSDP, while providing a significant computational advantage. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my concerns about the row-rank assumption and the gap of the estimates of Appendix E. I think ECLipsE is an interesting result with high impact in the neural-network robustness community. I have raised my score. --- Reply to Comment 1.1.1: Comment: Thank you for your positive evaluation of our work, and for raising your score! We appreciate your feedback on the paper!	null	null	null	null	null	null
VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time	Accept (oral)	Summary: The paper present VASA, a new method for talking head video generation. The method is build as a diffusion model using Transformer. To increase the performance of the model and allow more control on the generated video the authors decided to use the representation from [1] and learn to disentangle its components. With this the model can control the gaze, expression and camera position. The method achieve impressive qualitative and outperforms the methods they compare with. [1] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022 Strengths: The qualitative results are very impressive and probably better than the state-of-the-art. The idea of disentangling the existing representation of [1] is nice. The model is very fast at inference on costumer grade GPU. The paper is well written and easy to understand. [1] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022 Weaknesses: My main issue with the paper is that despite the work being presented as reproducible in the checklist I feel that many too many details are missing to actually reproduce the method or the experiments : - The new proposed CAPP score lack details for reproducibility e.g training procedure. Sharing this model would also be of interest to the community (and do not raise ethical concern which could justify keeping the code private). As of now it is difficult to know if the metric is actually sound. - It is unclear if the entire voxceleb dataset is used and how many clips remain after preprocessing ? - Lack of details of the new OneMin-32 dataset : size, type of videos, resolution, origins of the videos... - The paper says that the model is trained on "4 NVIDIA RTX A6000 GPUs" and that the model "train on massive talking face videos from a large number of identities". How long does the training take and how much data is actually used. Because if all of VoxCeleb + the new unreleased dataset are used the training could be very long. More details are required here. - It is not clear how the condition are used ion the network. Are they simply concatenated to the motion latent or used in cross Attention inside the Transformer - It is not entirely clear if the architecture of [2] used out of the box to obtain the facial latent or if it was modified for the disentanglement. Assuming that the dataset used is very large (>10e6 samples) is the comparison against the other methods that use 50k-100k samples for training fair. An ablation with a training on a dataset of that scale would have been interesting. Without it the impressive qualitative results of the method could simply be due to the huge amount of data. The novelty is limited the paper mostly reuse existing modules and innovation is mainly in the disentanglement. The comparison against sota is limited, the most recent method, SadTalker, is from 2022. The FVD score on VoxCeleb should have been shown anyway, other method from the literature present it. The method apparently used [1] for gaze direction estimation. However [72] appears to be a method to classify gaze between different modes (fixed, quick motion...). Was the method modified to obtain gaze direction g that is used in the paper ? The ablation only present results on the gaze and audio condition. It would have been interesting to also see the effect of the expression condition. [1] Raimondas Zemblys, Diederick C Niehorster, and Kenneth Holmqvist. gazenet: End-to-end eye-movement event detection with deep neural networks. Behavior research methods, 51:840–864, 2019. [2] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022. Technical Quality: 3 Clarity: 4 Questions for Authors: How do the authors deal with the resolution difference between methods when computing FVD ? With CAPP score are the head motions related to speech semantics measured ? (e.g head shake when saying no.) to finish Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The authors address the limitation in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to W1 (CAPP model sharing): We will soon release the CAPP model, which we believe fills the missing piece of audio-pose alignment metric in talking face generation research and will be valuable to the community. Response to W2 (Voxceleb training data): We used the entire training set of Voxceleb2 for our model training. After filtering out invalid or low-quality videos, we ended up with approximately 170K training clips. Response to W3 (OneMin-32 data details): It contains 32 one-minute video clips of 17 subjects (Line 229-232). They are mostly educational lectures and coaching sessions sourced from online video sharing platforms. The resolution is 512x512. Response to W4 (training time): Our face latent model (encoder and decoder) takes around 7 days' training on a 4xA6000 workstation, and the diffusion transformer takes ~3 days. The total data used for training is ~500k clips (about 2-10 seconds each). Response to W5 (condition usage): Yes, the condition signals are simply concatenated with noise along the temporal dimension as the input to the transformer. We'll clarify this in the paper. Response to W6 (change to [2]): Our latent model architecture is same to [2]. We did not change the architecture but modified the training loss functions (Line 149-160) which are critical to achieve disentanglement (Figure A.8 and Line 344-350). Response to W7 (data scale ablation): As mentioned above, our data size is ~500K (not ">10e6"). In the attached one-page PDF, we add a data size ablation study and comparison with other methods, as per the reviewer's suggestion. We trained a model using only 10% of the data (i.e., 50k clips). As shown in Table I, this model achieve comparable audio-lip and audio-pose synchronization to the full-data model, though the FVD and $\Delta$p metrics are not as good. This shows that our method performs well even with much less data, and more data enhances the motion diversity. Moreover, it still significantly outperforms other methods in all metrics assessing synchronization, motion intensity, and video quality. Response to W8 (novelty): First, our motivation in the first place is, to model the human conversational behavior (facial dynamics and head movement ) holistically using diffusion model in a latent space that is agnostic to ID and appearance. This is our core innovation and, to the best of our knowledge, no previous methods have done this (it differs from the trend of further factor disentanglement and direct image generation; see our discussion in Line 47-55, 103-111). Second, in pursuit of the aforementioned goal, we did find the 3D-aided representation to be promising especially in terms of expressiveness and hence chose to leverage them. However, they can NOT meet our requirement of effective disentanglement. We made some insightful and provably critical modifications (Line 149-160, 344-350, and Figure A.8), without which we can never reach the current generation quality, esp. the liveliness with nuanced emotions. We perhaps have underemphasized the importance and contribution of such modifications, and will revise our presentation in the revision. Apart from the two main contributions, our paper also offers other ones such as the design of face-factor-conditioned diffusion training, the CAPP model for filling the missing piece of pose-audio alignment metric, etc., which are also novel and valuable to the community. Response to W9 (limited comparison to sota?): To the best of our knowledge, there are no other published methods which can generate both audio-driven head poses and facial movements with single images. We mentioned some concurrent unpublished works in our paper, and have added the visual comparison with a concurrent work EMO in the one-page PDF. We'd appreciate it if the reviewer can point out some specific papers that we should compare with. Regarding the added comparison with EMO, we provide our results on some samples from EMO's official website (we are unable to provide video links per the rebuttal policy). As shown in Figure I, our method works consistently well and delivers vivid talking head videos. It is obvious that EMO has smaller head motion compared to us. Also, EMO seems less robust than ours in some cases, with artifacts – such as abrupt expression change, inaccurate lip sync, and subtle texture detail flickering – occasionally appear upon close inspection (note that their reported average lip-sync score is significantly lower than ours). On the other hand, however, EMO's video quality is slightly higher than ours in terms of sharpness, owing to their use of the large and powerful image generation foundation model. Response to W10 (FVD on Voxceleb2) : As shown in Table I of the attached PDF, we provide the FVD scores of different methods on Voxceleb2. However, it should be noted that the video quality of Voxceleb2 varies widely and is often low (see Figure II of the PDF). Hence the FVD score may not accurately reflect the true generation quality, as mentioned in our paper. Response to W11 (gaze estimation) : Thanks for your careful reading. We found that we inadvertently cited the wrong paper: we actually used L2CS-Net [a] to extract gaze direction. Will fix this error in our revised paper. [a] Ahmed A.Abdelrahman, Thorsten Hempel, Aly Khalifa and Ayoub Al-Hamadi, L2CS-Net: Fine-Grained Gaze Estimation in Unconstrained Environments, 2022 Response to Q1: The results of different methods are resized to the same resolution (224x224) for FVD evaluation. Response to Q2: This is an interesting question. For now we did not conduct in-depth analysis on whether the CAPP score captures the semantic relationship between speech and pose. We'll further explore this in our future work and thank you for the suggestion. --- We hope we have addressed your questions. If not, it would be great to let us further know your concerns during discussion. --- Rebuttal 2: Title: Rating after rebuttal Comment: After reading the extensive rebuttal I see that the authors responded to most of my concerns. I see no reason to reject this paper and change my rating to accept. Some of the explanation of the rebuttal should be included in the final version. I would have been interesting to see comparison with more recent method even if they don't generate head pose. If the head pose is controllable shouldn't it be possible to freeze it to match that of the other methods ? --- Rebuttal 3: Comment: Thank you for your acknowledgment of our response and the additional comments. Yes, our method can be easily adapted to generate facial dynamics only. Another easier way to achieve this is to directly replace the generated head poses with predefined ones before face image decoding. However, we shall point out that if the given head poses do not match the emotion or rhythm of the audio, the realism of the generated talking face video could degrade significantly (e.g., a calm head movement with intense speech or a rhythmic nodding with smooth speech would look weird). Generating realistic poses is one of the key contributing factors to achieve our high-quality results. That being said, we will try to add comparisons and more discussions about this type of methods in our revised paper and thank you again for the suggestion.	Summary: The paper presents a method for generating highly realistic talking head avatars that combines the diversity of facial expressions with the real-time generation speed. It provides a practical and commercially valuable approach to the field of talking head generation. Strengths: 1. The overall structure of the paper is very clear and coherent, with a well-defined problem statement. 2. By decoupling the information in dynamic faces, better control over the expressiveness of the generated faces can be achieved, meeting the needs of the users. 3. The visual presentation of the video is excellent and leaves a lasting impression. Weaknesses: 1. The paper does not explain why the proposed method can achieve real-time generation; the use of a diffusion transformer structure might actually lead to a decrease in speed. 2. There are some unclear configurations in the implementation section of the method, such as the scale of the video used for training. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Will this project be open-sourced? As it could actively promote progress in the field of talking head generation. If it is not open-sourced, it is suggested to provide more implementation details. 2. Why is there no comparison with the recent EMO[1] method, for which there are already corresponding implementations in the open-source community? [1] Tian L, Wang Q, Zhang B, et al. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions[J]. arXiv preprint arXiv:2402.17485, 2024. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: 1. Although VASA-1 has made significant progress in generating realistic facial dynamics and head movements, the paper mentions that it currently only handles the portrait area up to the torso and does not extend to the full body. The coordination of full-body movements is necessary to achieve more natural and realistic virtual characters. 2. The paper mentions that, despite using a 3D latent representation, a more explicit 3D facial model is not employed, which may lead to some artifacts caused by neural rendering, such as texture adhesion issues. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to W1: We achieve real-time efficiency because of our framework design, i.e. (diffusion-based) motion generation in latent space + (CNN-based) decoding in image space. Our diffusion transformer works in the latent space and is small (only 29M parameters), so it runs very fast. The CNN image decoder is also small (55M parameters) and runs efficiently. (Note that we only need to run the face encoder once for generating a video and thus its time can be neglected or counted into the starting latency.) We simply evaluated the running efficiency with our whole method naively deployed in PyTorch without any special speed-up strategy. We believe there's still room for improvement with sophisticated implementation optimization. Response to W2: Regarding the scale of video data, we trained our model on approximately 500k clips (2-10 seconds each). In the attached one-page PDF, we also provide an additional ablation study for training data scale. As shown in Table I, the model trained with 10% of the data achieves comparable audio-lip and audio-pose synchronization to the full-data model, though the FVD and $\Delta$p metrics are not as good. This shows that our method performs well even with much less data, and more data enhances the motion diversity. Response to Q1: We'll try our best to release the source code of our project in the future. However, we hope you understand that due to significant concerns regarding the potential risks, particularly those related to deepfake and fraud, we (as well as the community) do need to be very cautious with releasing a powerful model. In fact, due to RAI considerations, our team faced great difficulties in getting approved by our organization for open source, unlike any other projects we did before. While we are seeking for the possibility of open source, we'll also add more implementation details such as those suggested by the reviewers into the revised paper. Also note that we will soon release the CAPP model, which we believe fills the missing piece of audio-pose alignment metric in talking face generation research and will be valuable to the community. Response to Q2: EMO is a concurrent work (published Feb 27th on arXiv) at the time of our submission (May 22th) and there's no public implementation, so we did not compare with it. However, we did mention it with some discussions in our paper (Line 41-44, 115-118). EMO uses an image diffusion model based on StableDiffusion to generate talking face videos, which is a significantly different technique. It can generate high-quality videos but suffers from heavy computation and slow generation speed compared to ours. In the attached one-page PDF, we provide our results including animations on some samples from EMO's official website (we are unable to provide video links per the rebuttal policy). As shown in Figure I, our method works consistently well on EMO's demonstrated cases and delivers vivid talking head videos. It is obvious that EMO has smaller head motion compared to ours, perhaps due to the constraint of the face region mask used by it. Also, EMO seems less robust than ours in some cases, with artifacts – such as abrupt expression change, inaccurate lip sync, and subtle texture detail flickering – occasionally appears upon close inspection (note that their reported average lip-sync score is significantly lower than ours). On the other hand, however, EMO's video quality is slightly higher than ours in terms of sharpness, owing to their use of the large and powerful image generation foundation model. Response to Limitations: Thank you for the comments. We plan to handle upper-body/full-body and explore more explicit 3D representations in our future work (both projects are on-going). --- We hope we have addressed your questions. If not, it would be great to let us further know your concerns during discussion. --- Rebuttal Comment 1.1: Comment: After reading the author's rebuttal, most of my doubts are eliminated. I would like to ask how scalable is the VASA1 method and whether it can be applied to full body generation? Compared to face generation, generating a natural full body is more complex and difficult. --- Reply to Comment 1.1.1: Comment: Thank you for the further comments; we are glad to see our response eliminated your doubts. Regarding scaling VASA-1 to body generation, the problem is indeed more complex and difficult. But we believe the idea of VASA, i.,e., generating conversational human behavior holistically in a compact, ID-agnostic latent space and then generating the images, applies to body as well. We will work on this in our future work and keep the community updated on progresses and milestones.	Summary: This paper introduces a two-stage talking head method that can generate impressive talking faces. It includes 1) A diffusion-based model to generate implicit facial dynamics and head movements from audio and additional conditions. 2) A modified 3D-aided face reenactment decoder for generating faces from latent space. This method delivers high video quality with fast inference speed Strengths: 1. Although many works utilize diffusion models to map audio to intermediate facial features, VASA-V1 demonstrates excellent engineering and generation capabilities, achieving appealing results. 2. This method surpasses real-time speed at a resolution of 512x512, with fast startup cost and ID switching speed, leaving an impressive effect. 3. The method outperforms existing comparative methods in terms of visual effects and numerical results for video realism and audio-visual synchronization. Weaknesses: 1. The 3D-aided face reenactment framework stage should be crucial for the overall method. However, some descriptions are too brief and vague, making them hard to follow. 2. The paper's explanation of the fusion method for condition signals in the Diffusion Transformer is confusing and needs more specific details. 3. The comparison methods in the paper lack implementation details. Considering the different scales of training data for various methods, are the comparison results in the table fair? Technical Quality: 4 Clarity: 1 Questions for Authors: 1. How does the number of layers in the 8-layer transformer encoder affect the results in the paper? 2. Will the proposed CAPP in the paper be open-sourced? 3. Does the 3D-aided face reenactment part of the method use distillation to speed up? How can the inference speed of MegaPortraits be accelerated? 4. What are the parameter counts for each stage of the model? Confidence: 4 Soundness: 4 Presentation: 1 Contribution: 3 Limitations: The supplementary materials of the paper include relevant discussions. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to W1: Thank you for the comment. We will add more details of the 3D-aided face latent model into our main paper and appendix. Our network architecture follows MegaPortraits [18] where details can be found. To achieve disentanglement, we modified the loss functions and incorporated the cross-transfer consistency loss $L_{consist}$ and ID-disentanglement loss $l_{cross-id}$. (Line 149-160). These loss modifications are critical (Line 344-350 and Figure A.8), without which we can never reach the current generation quality, esp. the liveliness with nuanced emotions. Again, we'll add more details and further improve the clarity as per your suggestion. Response to W2: Different conditional signals are directly concatenated with noise along the temporal dimension as the input to the transformer. Response to W3: First, please note that different methods in the literature may have used different data for training. We are unable to compare with these methods using exactly the same data and have simply followed the practice of running the trained models for comparison. In terms of training data scale, our model is trained on ~500k clips (2-10 seconds each). To validate the data scale influence and compare it with previous methods at similar scales, we additionally trained a model using only 10% of the data (i.e., 50k clips). As shown in Table I in the attached PDF, the model trained with 10% of the data achieves comparable audio-lip and audio-pose synchronization to the full-dataset model, though the FVD and $\Delta$p metrics are not as good. This shows that our method performs well even with much less data, and more data enhances the motion diversity. Regarding the compared methods, Audio2Head used >70k data clips for training but clearly underperformed compared to our model with 50k clips. MakeItTalk and SadTalker used very small subsets of Voxceleb for training, but there's no clear evidence that increasing their data would improve their performance significantly or even bring any positive consequence - we explain the reasons as follows. MakeItTalk uses an LSTM to map audio features to landmark offsets deterministically, which may struggle with modeling complex data distributions and one-to-many mappings as training data increases. SadTalker assigns a style code for each identity to generate head poses, but more data will introduce more diverse head motion patterns for the same identity, which a shallow VAE with a condition code might not be able to effectively model. Our model with 10% data still significantly outperforms these methods in all metrics assessing synchronization, motion intensity, and video quality. Response to Q1: We set the transformer layer number to 8 as we found it produces good results while enabling the whole algorithm to run in real time on a consumer-grade GPU. We didn't explore more layers or larger model size because real-time efficiency is a key factor we want to achieve. We presume that a larger model size will further improve the performance because our current model is still small, and we'll further explore this in our future work. Response to Q2: Yes, we will soon release the CAPP model, which we believe fills the missing piece of audio-pose alignment metric in talking face generation research and will be valuable to the community. Response to Q3: No, we did not use distillation or any other strategies to speed-up the 3D-aided face encoder and decoder. These models are naturally small and run very fast. Note we only need to run the encoder once so essentially only the decoder needs to run for generating each video frame. Response to Q4: The parameter counts of our 3D-aided face latent model and diffusion transformer model are about 200M and 29M respectively. --- We hope we have addressed your questions. If not, it would be great to let us further know your concerns during discussion.	Summary: This paper aims to effectively and efficiently generate high-fidelity audio-driven talking head videos. To improve performance and efficiency, the authors have designed a Diffusion Transformer model within the latent space of motion signals, encompassing facial dynamics and head movements. Additionally, they propose a data-driven metric named Contrastive Audio and Pose Pretraining. Strengths: - The paper applies the diffusion model to the task of generating audio-driven talking head videos, innovatively defining the diffusion model within the latent features of motion rather than those of the image, which is quite interesting. - The paper is well-written and easy to follow, with detailed experiments that convincingly demonstrate the effectiveness of the proposed method. Weaknesses: The primary concern is the paper's contribution, as the realism and liveliness of the generated videos could be attributed to the performance of MegaPortraits. MegaPortraits' encoders effectively learn latent motion and appearance representations, supported by robust 3D warping generators and an image generator that ensures high-quality outputs. VASA-1, in a way, learns to generate latent motion representations akin to those in MegaPortraits through audio inputs. Despite this dependency, the method performs well overall. Therefore, my overall assessment leans towards accepting it, albeit with some reservations. Technical Quality: 4 Clarity: 4 Questions for Authors: - At inference time, how does the model generate condition signals like the main eye gaze direction and head-to-camera distance, given that the driving signal is only audio? - This does not decrease the novelty of this work. However, a quantitative comparison between VASA-1 and EMO would be quite interesting. Given that EMO's code is unavailable, leveraging the image and audio from their officially provided videos for comparison is encouraged. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to W1 (contributions): Thank you for the comment. In response to your question on our technical contributions and relationship with MegaPortraits, we'd like to emphasize two aspects. First, our motivation in the first place is to model the human face conversational behavior (facial dynamics and head movement) holistically using a diffusion model, in a latent space that is agnostic to ID and appearance. This is our core innovation and, to the best of our knowledge, no previous methods have done this (it differs from the trends of further factor disentanglement and direct image generation; see our discussion in Lines 47-55, 103-111). Second, in pursuit of the aforementioned goal and building a disentangled latent space, we did find the 3D-aided representation to be promising especially in terms of expressiveness and hence chose to leverage them. However, they can NOT meet our requirement of effective disentanglement. We made some insightful and provably critical modifications (Line 149-160, 344-350, and Figure A.8), without which we can never reach the current generation quality, esp. the liveliness with nuanced emotions. We perhaps have underemphasized the importance and contribution of such modifications, and will revise our presentation in the revision. Apart from the two main technical contributions, our paper also offers other ones such as the design of face-factor-conditioned diffusion training, the CAPP model for filling the missing piece of pose-audio alignment evaluation, etc., which are also novel and valuable to the community. Response to Q1: The extra condition signals such as the main eye gaze direction and head-to-camera distance are optional and they are provided by users. If not given, we can either set them to some default parameters (e.g., a forward-looking eye gaze and the average head-to-camera distance of the training data; see Line 220-222), or just leave them blank for unconditional generation. Response to Q2: Thank you for the suggestion. We have run our method on the images and audios from EMO's official website as per your suggestion. Some visual results including animations can be found in the one-page PDF provided on this page (we are unable to provide video links per the rebuttal policy). EMO is a concurrent work which we mentioned and discussed in the related work section. It uses an image diffusion model based on StableDiffusion to generate talking face videos, which is a significantly different technique. EMO can generate high-quality videos but suffers from heavy computation and slow generation speed compared to ours. As shown in Figure I of the attached PDF, our method works consistently well on EMO's demonstrated cases and delivers vivid talking head videos. It is obvious that EMO has smaller head motion compared to ours, perhaps due to the constraint of the face region mask used by it. Also, EMO seems less robust than ours in some cases, with artifacts – such as abrupt expression change, inaccurate lip sync, and subtle texture detail flickering – occasionally appear upon close inspection (note that their reported average lip-sync score is significantly lower than ours). On the other hand, however, EMO's video quality is slightly higher than ours in terms of sharpness, owing to their use of the large and powerful image generation foundation model. --- We hope we have addressed your questions. If not, it would be great to let us further know your concerns during discussion. --- Rebuttal Comment 1.1: Title: Final Rating Comment: Thank you to the authors for their feedback and efforts. After reviewing the rebuttal, I note that the authors have addressed some of my concerns, which leads me to maintain my initial rating. However, I recommend that the final version of the paper include more detailed explanations, particularly regarding the contributions and the reasons behind the statement, "We did find the 3D-aided representation to be promising." These details will enhance the clarity and impact of the work. --- Reply to Comment 1.1.1: Comment: Thank you for your further comment and suggestion. We will incorporate more details including those suggested by the reviewers.	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for the valuable comments and suggestions. We are encouraged by the reviewer's acknowledgment that our paper: "innovatively defining the diffusion model within... which is quite interesting"; "convincingly demonstrate the effectiveness.." (Reviewer WJD9); "demonstrates excellent engineering and generation capabilities", "leaving an impressive effect" (Reviewer wMXj); "visual presentation is excellent and leaves a lasting impression" "has made significant progress ..." (Reviewer 1diB); "qualitative results are very impressive" (Reviewer xKzZ). We'd like to reiterate our novelty and contributions here: - We propose diffusion-based holistic human face conversational behavior modelling (facial dynamics and head movement), in a latent space that is agnostic to ID and appearance. This is our core innovation and, to the best of our knowledge, no previous methods have done this. It is a new approach which differs from the recent trends of further facial factor disentanglement and direct image generation (see our discussion in Lines 47-55, 103-111, 115-118). - We build a highly disentangled latent space to achieve the aforementioned goal. Although we leveraged existing 3D-aided representation and models due to their high expressiveness, they can NOT meet our requirement of effective disentanglement. We made some insightful and provably critical modifications (Line 149-160, 344-350, and Figure A.8), without which we can never reach the current generation quality, esp. the liveliness with nuanced emotions. - We offer a few other supporting contributions including a controllable diffusion framework that enables flexible control of different face properties, and a new data-driven metric CAPP score for evaluating the alignment between audio and head pose. - We advance audio-driven talking face generation to a new level of realism and liveliness not achieved before. Our work marks the dawn of real-time lifelike avatars which have the potential to reshape human-human and human-AI interactions across broad application domains. We address each reviewer's questions and concerns under their respective reviews. The attached one-page PDF contains the following figure and table contents: - Visual comparison with EMO on EMO's official videos (Figure I) - Sampled images from VoxCeleb2 to demonstrate the varied video quality and explain why we did not evaluate the FVD on it (Figure II) - Training data scale ablation of our method and the requested FVD score on VoxCeleb2 (Table I) Pdf: /pdf/d889d00df489eaa4601dde39c7694a1f32e23c4d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Federated Model Heterogeneous Matryoshka Representation Learning	Accept (poster)	Summary: This paper introduces FedMRL, a method based on distillation to mitigate the model heterogeneity issue in Federated Learning (FL). FedMRL operates by learning a small proxy homogeneous global model in a federated manner and distilling knowledge from it to heterogeneous client models. To enhance representation knowledge interaction between the homogeneous global model and the heterogeneous client local model, the authors employ a Matryoshka Representation Learning (MRL) approach, generating multi-dimensional and multi-granular representations. Theoretical analysis and experiments demonstrate the effectiveness of FedMRL. Strengths: 1. The fusion of representations from the global and local models into a single representation vector, followed by their detachment in a Matryoshka manner, is intriguing and inspiring. 2. The writing is good and easy to follow. 3. Transmitting a global model with a relatively lower feature dimension is promising and can reduce communication overhead compared to using a similar dimension as local models. Weaknesses: 1. The authors highlight the limitations of leveraging the training loss between server and client models (incurring high communication and computation costs) and reference the papers FedKD and FML in lines 39-41. However, the proposed FedMRL also falls within this category by sharing a proxy global model. Refer to Figure 4 for an illustration of FedMRL's high communication costs in the MHeteroFL domain, which also conflicts with the statement "low communication costs" in line 68. Additionally, there is a lack of numerical results regarding communication and computation costs between FedMRL and similar methods (FedKD and FML). 2. Two datasets for only image tasks are insufficient in FL. 3. The client models utilized in the experiments are not sufficiently heterogeneous, as they consist of CNN networks with identical numbers of Conv and FC layers. The variations are limited to the channels in Conv2 and the neural count in FC1. This setup lacks the persuasiveness needed to demonstrate FedMRL's effectiveness in MHeteroFL, especially considering that model architectures can significantly differ in size and structure, as noted in [1]. Can FedMRL accommodate settings involving CNNs and Vision Transformers (ViTs) on clients? Additionally, the considered CNNs are overly simple and small for a comprehensive evaluation. 4. There is only one baseline for the model split category, and it's worth noting that FedGH[2] also falls within this category. 5. The details of computing FLOPs are missing. 6. In the "Proof of Theorem 2" section, obtaining Eq. (31) directly from Eq. (30) is not feasible, as the right side is $\frac{\Delta}{T}$, not 0 as suggested by Eq. (30). Additionally, the existence of solutions for $\eta$ in Eq. (32) may be compromised if $\epsilon < \delta^2$, contradicting the conclusion in line 458 and potentially undermining the convergence guarantee. 7. The privacy analysis presented in lines 486-490 lacks sufficient substantiation. It would benefit from either theoretical analysis or experimental results. Without further analysis, it's challenging to accept the claim that "representation splicing enables the structures of the homogeneous global model and the heterogeneous local model to be not related," especially considering that the global and local models are trained together with a shared representation projector. [1] Zhang, Jianqing, et al. "Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 15. 2024. [2] Yi, Liping, et al. "FedGH: Heterogeneous federated learning with generalized global header." Proceedings of the 31st ACM International Conference on Multimedia. 2023. Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: - W1.a: Apologize for the confusion. Our comment on the limitations of existing strategies, e.g., FedKD and FML, where they communicate the homogeneous proxy model. Although FedMRL falls in this category, it improved communication and computation efficiency by reducing the representation dimension of the homogeneous proxy model without sacrificing the performance through multi-dimension Matryoshka representation learning. - W1.b: For conciseness, we only compared with the best-performed baseline FedProto wrt averaged accuracy in Sec 5.2. First, we want to clarify that, although FedProto achieved lower communication costs, it requires more communication rounds and computational FLOPS in Figure 4. Also, it underperformed FedMRL by 3.36% and 8.48% on CIFAR-10 and CIFAR-100 (Table 1). Second, we wish to highlight that compared with the other baselines, FedMRL achieves relatively lower communication costs. As shown in Table X2, FML, FedKD, and FedAPEN belonging to the same category as FedMRL either do not converge or consume both higher communication and computation costs and also perform lower accuracy than FedMRL. - W1.c: We did not include the comparison with FML and FedKD as we only considered the methods that reach 90% and 50% target accuracy on CIFAR-10 and CIFAR-100 in lines 253-254, but they did not reach the targeted accuracies. We provide extra numerical results when reaching lower 70% and 30% target accuracy on CIFAR-10 and CIFAR-100 in Table X3. FedMRL still performs significantly lower communication and computational overheads than these same-category methods. W2: We selected two standard benchmark datasets to demonstrate the effectiveness of our methods following the existing literature [1,2] for fair comparison. We are grateful that our efforts on extensive and abundant experiments have been recognized by Reviewer 6thni and DqWb. To address your concerns and demonstrate the generalizability of our FedMRL to different datasets, we also add experiments for a next-word prediction (NLP) task with a large real-world non-IID dataset - Stack Overflow and heterogeneous LSTM models across 100 clients. As shown in Table X4, our method improves over the baseline methods by a significant margin. W3: We followed [1,2] to design the model heterogenities. And Yes, FedMRL allows an arbitrary heterogeneous client model structure. To demonstrate the versatility of FedMRL, we add experiments on CIFAR-100 with 100 clients including ResNet-{4,6,8,10,18,34,50,101,152} following FedTGP, CNN, and ViT. Table X1 again shows the best model performance of FedMRL. W4: FedGH needs clients to upload labels to the server for training the shared global prediction header, which may be prohibited in some label-privacy-sensitive FL scenarios [3], so we did not compare it in the submission version. Nevertheless, we have discussed it in the related work section. We add experiments for FedGH in Table X5, which shows that FedMRL still maintains the best model performance. W5: We calculate the average FLOPs of heterogeneous CNN models consumed by forward inference and backwards updating in one iteration of local training, following the traditional FLOPs calculation rules of convolutional and linear layers. We then record the communication rounds required to reach the specified target accuracy and use the product between them as the total computational FLOPs. W6: There are some misunderstandings. The two steps of theoretical derivations definitely hold. Here, we give a detailed analysis. - As we stated $T>0$ and $\Delta>0$, then $\frac{\Delta}{T}>0$, so the derivation from Eq. (30) to Eq. (31) holds. - For $\epsilon>\delta^2$, we supplement additional analysis. Assumption 3 assumes that the parameter variations of the homogeneous small model $\theta_k^t$ and $\theta^t$ before and after aggregation are bounded by $\|\theta^t-\theta_k^t\| _ 2^2 \leq \delta^2$. $\theta_k^t=\theta^{t-1}-\eta\sum_{e=0}^{E-1}g_{\theta^{t-1}}$, so $\|\theta^t-\theta_k^t\| _ 2^2=\|\theta^t-\theta^{t-1}+\eta \sum_{e=0}^{E-1} g_{\theta^{t-1}}\| _ 2^2 \approx \eta^2 \sum_{e=0}^{E-1}\|g_{\theta^{t-1}}\| _ 2^2$, considering that the global homogeneous small models during two consecutive rounds have relatively small variations compared with parameter variations between the local and global homogeneous model. Eq. (28) and (29) define $\epsilon$ as the upper bound of the average gradient of the local training whole model (including homogeneous small model, heterogeneous client model and the local representation projector) during $T$ rounds and $E$ epochs per round, i.e.,$\frac{1}{T} \sum_{t=0}^{T-1} \sum_{e=0}^{E-1}\|\mathcal{L} _ {t E+e}\| _ 2^2<\epsilon$, we can simplify it to $\sum_{e=0}^{E-1}\|\mathcal{L} _ {t E+e}\| _ 2^2<\epsilon$. Since the homogeneous model $\theta$ is only one part of the local training whole model, so $\epsilon>\sum_{e=0}^{E-1}\|\mathcal{L} _ {t E+e}\| _ 2^2>\sum_{e=0}^{E-1}\|g_{\theta^{t-1}}\| _ 2^2$. Since we use the leaning rate $\eta\in(0,1)$, $\eta^2\in(0,1)$, so $\epsilon>\sum_{e=0}^{E-1}\|\mathcal{L} _ {t E+e}\| _ 2^2>\sum_{e=0}^{E-1}\|g_{\theta^{t-1}}\| _ 2^2>\eta^2 \sum_{e=0}^{E-1}\|g_{\theta^{t-1}}\| _ 2^2$. Since $\delta^2$ is the upper bound of $\eta^2 \sum_{e=0}^{E-1}\|g _ \theta^{t-1}\| _ 2^2$, so $\epsilon>\delta^2$. W7: Sorry for this vague description. To make a clear understanding, we re-write this sentence to “we do not limit the representation dimensions $d_1, d_2$ of the proxy homogeneous global model and the heterogeneous client model are the same, so sharing the proxy homogeneous model does not disclose the representation dimension and structure of the heterogeneous client model. ” [1] FedGH: Heterogeneous Federated Learning with Generalized Global Header. [2] FedAPEN: Personalized Cross-silo Federated Learning with Adaptability to Statistical Heterogeneity. [3] One-Shot Federated Learning with Label Differential Privacy. --- Rebuttal 2: Title: Reply to authors Comment: Thank you for your detailed responses, especially for the additional analysis to prove $\epsilon > \delta^2$. I have raised the score to the positive side. --- Rebuttal Comment 2.1: Comment: Thank you very much for your supportive feedback on our response. We indeed highly appreciate your in-depth thought and your valuable time. Always happy to duscuss if additional clarification is required.	Summary: The authors study the model heterogeneity challenge in federated learning using Matryoshka representation learning. It requires that the global model and the local models share one common part inspired by the two key modules: adaptive representation fusion and multi-granularity representation learning. They provide the experiment results and derive the non-convex convergence rate of the algorithm. Strengths: 1, The model heterogeneity is one of the emerging challenges in the federated learning domain. The proposed approach avoids releasing the local model directly and solve the heterogeneous model cooperation issue. 2, This work covers comprehensive related works and the presentation is easy to follow. 3, They provide theoretical analysis in the paper and appendix. Weaknesses: 1. The idea of sharing common parts of models in FL was introduced in FedGH[1], which all the clients and the server share an identical header. The idea of exchanging the small shared model was introduced in ProxyFL[2]. The authors are suggested to consider discussing these two works and emphasize the main novelty of contributions of the proposed approach. 2, This approach increase extra computational cost at the client side. The authors are suggested to provide some qualitative results and some quantitive analysis. 3, The CNN model structures are hand-crafted. The reviewer is wondering if the approach can work with other common-used CNN models and how they would perform. [1] FedGH: Heterogeneous federated learning with generalized global header [2] Decentralized federated learning through proxy model sharing Technical Quality: 3 Clarity: 3 Questions for Authors: 1, Could you please address the concerns at the weakness part? 2, Could you please emphasize the motivation that why you combine Matryoshka representation learning with federated learning? 3, At the local client, is the optimization conducted in a end-to-end way or step-by-step way? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Multiple-run experiment results with basic stats would be more helpful to demonstrate the effectiveness of the proposed algorithm. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: FedGH achieves FL collaboration across clients with heterogeneous local models by sharing a co-training homogeneous prediction header at the server. It can be categorized into the model split branch. However, sharing one part of the heterogeneous client model may result in insufficient generalization for the local complete model and overfitting for the local remaining part, leading to model performance degradation and also disclosing partial model structure privacy. ProxyFL enables clients with heterogeneous models to exchange knowledge by sharing a proxy homogeneous model, it belongs to mutual learning-based model-heterogeneous FL methods like FML, FedKD, and FedAPEN. These methods utilize mutual loss or distillation loss calculated by the distance between the predictions of the shared proxy homogeneous model and the heterogeneous client model to train two models alternatively. The two models only exchange limited knowledge through the loss, resulting in model performance bottlenecks. FedMRL adds a proxy homogeneous model shared by clients with heterogeneous models for federated learning. Its innovative contributions mainly include the following two points. Owing to these designs, FedMRL can perform good model performance while effectively protecting heterogeneous model structure privacy, compared with FedGH and ProxyFL. - Adaptive Representation Fusion: We designed a lightweight personalized representation projector to fuse the global generalized representation extracted by the shared global proxy homogeneous model and the local personalized representation extracted by the heterogeneous client model. The local projector and the two models are trained in an end-to-end manner, so the projector is adaptive to local non-IID data distribution, implementing personalized adaptive representation fusion. - Multi-Perspective Matryoshka Representation Learning: Based on the fused representation which includes both global generalized and local personalized feature information, we construct multi-dimension and multi-granularity Matryoshka representations and improve model learning capability through Matryoshka representation learning. After local model training, only the proxy homogeneous models are transmitted between the server and clients and complete heterogeneous client models are always stored within clients, protecting client model structure privacy. W2: The extra computational cost at clients is increased by additionally training a small proxy homogeneous model and a lightweight one-linear-layer representation projector. Since the additional two models constitute only about 9% of the entire model, the additional computational cost in our FedMRL is minimal. In comparison to baseline methods in the same category that utilize proxy models, such as FML, FedKD, and FedAPEN, FedMRL's additional computational cost is lower, specifically, 9.21MB and 6.49MB FLOPs in Table X2. This is primarily due to our method's reduction in representation dimension. Our experiments, as illustrated in Figure 4, demonstrate that FedMRL maintains efficient computation while achieving significant improvements in model accuracy compared to other baselines, as shown in Table 1. Considering the significant performance improvement, we believe the minimal additional computational cost is a worthwhile compromise. The qualitative analysis was illustrated in Appendix D, lines 494-500, and the quantitive analysis was reported in Section 5.2.4. Both the qualitative and quantitive analysis demonstrate that although FedMRL increases slight extra computational overheads in one communication round, the adaptive representation fusion and multi-perspective Matryoshka representation learning enhance model generalization and personalization, speeding up model convergence. Therefore, as shown in Figure 4, FedMRL needs fewer communication rounds to reach the specified target model accuracy than the best baseline, and it also consumes lower total computational costs for reaching the target accuracy due to faster model convergence. Hence, FedMRL is also computationally efficient. W3: FedMRL can be applied to FL clients with arbitrary model structures since it fuses representations with an adaptive representation projector with a dimension that is adjusted freely for clients. We supplement extra experiments on the CIFAR-100 dataset with 100 clients under more complicated model heterogeneity including ResNet-{4, 6, 8, 10, 18, 34, 50, 101, 152}, CNN, and a ViT model. The results in Table X1 again validate the state-of-the-art performance of FedMRL. ‘-’ denotes failure to converge. Q1: Thanks for your valuable suggestions. We have tried our best to address the above concerns. Q2: Matryoshka representation learning has been substantiated to be an effective and efficient method to improve model learning capability. Inspired by it, we construct multi-dimension and multi-granularity Matryoshka representations processed by the global proxy homogeneous model’s prediction header and the local heterogeneous client model’s prediction header, respectively. The Matryoshka representation learning enables clients to learn the global generalized knowledge and local personalized knowledge from multiple perspectives, enhancing both model generalization and personalization and hence improving model performance. Q3: The optimization at the client is conducted in an end-to-end way. Limitation: Thanks for your advice. We conducted 3 trials of each experiment and only reported the average result. We will supplement the corresponding variances in the revised version. --- Rebuttal Comment 1.1: Comment: Thanks for your reply. I will keep my score positive as it is. Thank you. --- Reply to Comment 1.1.1: Comment: Thank you so much for your positive feedback and continued support. We greatly appreciate your valuable comments!	Summary: The paper proposed a FedMRL method for model-heterogeneous FL, which adapted Matryoshka Representation Learning to learn representations from multiple granularities. Strengths: 1. The proposed method is a new way to tackle the heterogeneous challenge of federated learning. 2. The paper is well-organized and easy to follow. 3. Abundant experiments and theoretical analysis demonstrate the effectiveness of the proposed methods. Weaknesses: 1. The novelty of the proposed solution needs a stronger justification. There are some representation fusion-based solutions to tackle heterogeneous federated learning challenges, for example, Federated Self-supervised Learning [12] and Federated Contrastive Learning [2]. [1] Weiming Zhuang, et al., DIVERGENCE-AWARE FEDERATED SELF-SUPERVISED LEARNING, ICLR 2022 [2] Qinbin Li, et al., Model-Contrastive Federated Learning, CVPR 2022 [3] Yue Tan, et al., Federated Learning from Pre-Trained Models: A Contrastive Learning Approach, NeurIPS 2022 2. The theoretical analysis shows no significant relevance to the Matryoshka Representation Learning. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. In Figure 1, the Matryoshka Representation is a key component for performance improvement. Does the Matryoshka Representation still capable of improving performance in a deeper CNN, or other neural architectures, e.g. ResNet, UNet and Transformer? 2. Are there any other Multi-Granularity Representation Learning methods other than Matryoshka Representation Learning? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: We acknowledge the pointed related work heterogeneous federated learning. However, our approach, FedMRL, introduces significant innovations that differentiate it from existing methods. We appreciate the oppporutnity to highlight our novelty. Frist, The referenced works do not fuse representation. The referenced 3 FL contrastive learning methods use the representation distance of the shared global model and the local model as the training loss to enhance the generalization of the local model, which is only suitable for clients with homogeneous models since the server is required to aggregate them. Second, Innovative Contributions in FedMRL. For representation learning in FedMRL, we propose two key innovative contributions: - Innovative Representation Fusion: We fuse the global generalized representations extracted by the proxy homogeneous model and the local personalized representations extracted by the heterogeneous client model through a training local personalized representation projector. The representation projector and the two models are simultaneously trained in an end-to-end manner, so the representation projector achieves a personalized representation fusion which is adapting to non-IID local data distributions. - Multi-Perspective Representation Learning: Based on the fused representation, we construct multi-dimension and multi-granularity Matryoshka representations. Each embedded dimension of Matryoshka representations is respectively processed by the proxy homogeneous model’s prediction head and the heterogeneous client model’s prediction head to output predictions and then compute loss with the ground-truth label, the loss sum is as the final loss to update all models in an end-to-end manner. In short, FedMRL innovatively utilizes multi-perspective Matryoshka representation learning to learn the global generalized feature and the local personalized feature from multiple perspectives, which is beneficial to improve model generalization and personalization simultaneously. Owing to these designs, FedMRL can be freely applied to more practical FL scenarios where clients may hold structure-heterogeneous models. W2: We would like to clarify that our theoretical analysis serves as a pivotal motivation for proposing suitable techniques. In our theoretical analysis, we derive the non-convex convergence rate of FedMRL based on a complete communication round. One key component impacting the overall convergence is the local training of the whole model (the proxy homogeneous model, the heterogeneous client model, and the local representation projector) at clients. The proposed Matryoshka Representation Learning method is known for better convergence and generalization in model training, as evidenced by [1]. This implies its ability to achieve better local model training. As shown in our Theorem 2, our convergence analysis indicates that the overall convergence benefits from improved local model training convergence. Therefore, we introduced Matryoshka Representation Learning in our framework to enhance this aspect. Additionally, our empirical results in Figure 3 confirm that FedMRL converges to higher model accuracy with a faster convergence speed compared to state-of-the-art baselines. This empirical evidence supports the theoretical claims and highlights the significant relevance of Matryoshka Representation Learning to the observed convergence behavior. Q1: Yes, the Matryoshka Representation can still improve model performance for deeper models. The core idea of the Matryoshka Representation is to add multiple prediction heads to process the embedded Matryoshka representations ranging from low-to-high dimensions and coarse-to-fine granularities to improve the learning capability of the encoder (i.e., feature extractor, model layers before the prediction head). This idea is inspired by the insight that people often first perceive a coarse outline of an observation objective and then carefully see the fine details, so multiple-perspective observations can enhance the understanding of one thing. For arbitrary shallow or deep models, all of them can construct Matryoshka Representations and append corresponding multiple prediction heads to improve model performance. Q2: Thank you for your insightful question. While there might be alternative solutions for Multi-Granularity Representation Learning, as far as we know, the Matryoshka Representation Learning method stands out as both effective and efficient, substantiated by extensive experiments. We chose the Matryoshka Representation Learning method for several reasons: - Effectiveness: The method has demonstrated significant improvements in model accuracy and performance through rigorous testing. - Efficiency: It allows for simultaneous training of both global and local models in an end-to-end manner, which is crucial for handling the non-IID data distributions in federated learning. FedMRL is the first to explore this method specifically in the FL domain, especially in the context of model-heterogeneous federated learning, addressing both generalization and personalization challenges. While we acknowledge that other methods may exist or emerge, we believe that Matryoshka Representation Learning currently provides a robust solution for the challenges at hand. [1] Matryoshka Representation Learning, NeurIPS 2022.	Summary: This paper focus on model heterogeneous federated learning. Existing distillation-based learning results in limited knowledge transfer. In order to mitigate this challenge, the authors propose FedMRL. In FedMRL, each client trains an extra shared global auxiliary homogeneous small model such that the server can directly learn local data distribution from the auxiliary model. The authors provide theoretical convergence analysis for FedMRL. Experiments on benchmark datasets demonstrate the effectiveness of the proposed FedMRL. Strengths: * The idea of utilizing a small homogeneous model is novel and interesting. * The writing is clear and easy to follow. * Convergence analysis is provided. * The experiments are extensive and can validate the effectiveness of the proposed method. Weaknesses: * FedMRL can add extra burden to computation power of client devices. * The authors did not compare the theoretical convergence rate with traditional distillation based model heterogeneous federated learning. Technical Quality: 4 Clarity: 4 Questions for Authors: Please refer to weaknesses. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: There is no potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: We would like to clarify that the additional computational cost in our proposed FedMRL is minimal. Specifically, it involves a small proxy homogeneous model and a one-linear-layer representation projector for the clients. Notably, the parameters of this additional component constitute only about 9% of the entire model. In comparison to baseline methods in the same category that utilize proxy models, such as FML, FedKD, and FedAPEN, FedMRL's additional computational cost is lower, specifically, 9.21MB and 6.49MB FLOPs in Table X2. This is primarily due to our method's reduction in representation dimension. Our experiments, as illustrated in Figure 4, demonstrate that FedMRL maintains efficient computation while achieving significant improvements in model accuracy compared to other baselines, as shown in Table 1. Considering the significant performance improvement, we believe the minimal additional computational cost is a worthwhile compromise. W2: We first restate our theoretical analysis. We prove the non-convex convergence rate through a complete communication round. To this end, we rely on the assumptions detailed in Section B in the Appendix introduce the error bounds associated with local training with hard loss (Lemma 1), and model aggregation (Lemma 2) and then derive the error bound of a once complete round of FL (Theorem 1). The convergence rate (Theorem 2) is based on training one complete round of FL by communicating and training small global homogenous proxy models and client heterogeneous models using hard loss. Note that, the baseline traditional distillation-based models (FD and FedProto) use additional knowledge distillation techniques that aggregate the output logits or representations of client models to generate the global logits or representations which are used in the next round’s local model training, such that additional loss and training schemes involved but do not fit in our theoretical framework due to their different model training and aggregation manners. Nevertheless, we have empirically demonstrated the better performance of FedMRL compared with the public data-free distillation-based methods FD and FedProto. --- Rebuttal Comment 1.1: Comment: Thanks for your response. I have raised my score. --- Reply to Comment 1.1.1: Comment: Thank you for your positive feedback and raising our score. Your valuable comments are greatly appreciated!	Rebuttal 1: Rebuttal: Please see Tables X1-X5 for rebuttal from the attached pdf file. Pdf: /pdf/16a0e4ccd0335ae1e1440eae19c2db1f79ed68cd.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication	Accept (poster)	Summary: This paper proposes to combine physical standability with existing diffusion-based text-to-3D generators in order to synthesize objects that not only follow the text description but also can stand on their own without falling. To this end, the paper proposes a physical loss function and uses it with score distillation sampling methods that generate 3D objects from texts. The paper shows that their generated model show better standability both in simulation and after physical realization. Strengths: - Good exposition and writing, paper is rather easy to follow. - The validation of results especially the 3D printing of generated models. - The standability and stable equilibrium losses are elegant. Weaknesses: First, there are good methods, such as 'make it stand' that could be used as post-processing tools for making text-based generated models stand. Why should we go through the paper's approach in the light of these methods? Some important ablations are missing. For example, - Separating of how much each loss is contributing to standability. For example, what if we use Magic 3D outputs and smoothen the geometry? - What are the pros and cons of using the physical loss as a post-processing or jointly with SDS? The claims are more broad than the actual contributions. - Take "fabrication-ready" in line 62. The paper is concerned with stabdability and it should adapt the text (and even the title). In 3D printing, "self supporting" means that there is no need to support materials. The paper doesn't cover this. - The relationship to robotics is not well-founded except an example that uses robot grippers to do the standability tests, which is not really related to benefits to robotics. The paper is heavily relied on score distillation sampling methods but does not explain them very well. I understand that the paper refers to the original papers in this area but it is worth including a paragraph about SDS. I look forward to see a justification in the rebuttal. Technical Quality: 2 Clarity: 3 Questions for Authors: - The paper mentions too many steps of integration and back propagation in Section 4.1.1. Does it still do this? If not, where exactly the differentiable physics simulation is used? - What is the default intimal physical state? - How does the paper figures out the upright position? - The paper mentions mesh topology (Line 246) but doesn't show it. A zoom in mesh is necessary for this. - What are the mesh sizes and dimensionalities of printed objects? - Are Figure 7 results printed? - x_{com} is not defined in Eq. 7. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: A main limitation is the mesh size. The paper has mentioned it. This is important in the light of 3D printing that require sometimes huge resolutions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for recognizing the effectiveness and elegance of our pipeline. Below we clarify your questions. Q1: First, there are good methods, such as 'make it stand' that could be used as post-processing tools[...]. Why should we go through the paper's approach in the light of these methods? We provide additional comparisons with cutting the mesh and make-it-stand [59] in the attached PDF. In our global response, we discuss in detail the drawbacks of such post-processing methods compared to our joint optimization approach. Please refer to the global response for more details. Overall, post-processing methods overlook semantics during optimization, potentially leading to degraded text alignment. Also, make-it-stand [59] assumes the contact region is unchanged. Such inflexibility may result in undesired outcomes, as demonstrated in Fig. 2. In addition, To deform the underlying mesh for improving standability, make-it-stand [59] requires manually placing a few handles, whereas our method requires no human intervention and can thus be used for batch generation, as we did in Fig. 8 in the paper. Hence, we believe that our method is a streamlined plug-in tailored for 3D generative models. Q2: Separating of how much each loss is contributing to standability. Thank you for your suggestion! We have added more ablation studies to demonstrate the necessity of our proposed losses. Please refer to Fig. 3 and Table 1 in the attached PDF and the global rebuttal for more details. Additionally, we would like to point out that for all our comparisons with Magic3D outputs, we applied the same scale of normal smoothing to both our method and the baseline method. Q3: What are the pros and cons of using the physical loss as a post-processing or jointly with SDS? As stated in our response to Q1, post-processing methods overlook semantics during optimization, which can lead to degraded text alignment. We also applied our proposed physical loss in a post-processing manner, as shown in Fig. 3(e) in the attached PDF. While the generated results are still able to stand, the text alignment is significantly compromised. Q4: The paper is concerned with standability and it should adapt the text (and even the title). In 3D printing, "self supporting" means that there is no need to support materials. Thanks for pointing out this! Our definition of self supporting is different from the one in 3D printing. Specifically, we consider self supporting as the condition where, once the upright pose is determined by the text-to-3D model, the generated 3D model should satisfy the standability criteria defined in Eq. (3). We will clarify this in the revised version. In this work, we mainly use 3D printing as a verification in the real world settings. Q5: The relationship to robotics is not well-founded except an example that uses robot grippers to do the standability tests [...]. The relevance of our work to robotics lies in our capability to automatically generate standable 3D assets. A simple scenario would involve training a robot to lift, move, and place objects without tipping them over. Our method ensures that the generated contents can stand stably, as is necessary for many real-world objects. We aim to promote the generation of 3D assets with automatic incorporation of physical standability, and we believe this property can effectively help robots practice accurate interaction with objects, such as grasping, picking, and placing, both in simulation and in the real world. Q6: The paper is heavily relied on score distillation sampling methods but does not explain them very well. [...] it is worth including a paragraph about SDS. We will revise Section 3.1 and include a more detailed description of SDS in the appendix. Q7: The paper mentions too many steps of integration and back propagation in Section 4.1.1. Does it still do this? If not, where exactly the differentiable physics simulation is used? In line 166, we mentioned that applying physical simulation at every optimization step can be time-consuming due to the many steps of forward time-stepping and backward propagation. In all our experiments, we apply the standability loss (physical simulation) once every 10 iterations, and we find that this approach is sufficient to ensure significant loss reduction and does not notably increase computational overhead (see line 226). Q8: What is the default intimal physical state? We treat the meshes as solid, uniform objects. We outlined the default physical parameters and simulation settings in Appendix A.1. Q9: How does the paper figures out the upright position? As elaborated in our global response, the upright direction is semantics-driven and generative models inherently learn these semantics from their training data. We thus define the upright pose by defining the upward axis to coincide with the upward axis from the default 3D-generated result. We will make this point clearer in the paper. Q10: The paper mentions mesh topology (Line 246) but doesn't show it. A zoom in mesh is necessary for this. We included zoom-in views of textured geometry in Fig. 4 to illustrate local topology changes. We will provide additional zoom-in views of mesh geometry in the appendix. Q11: What are the mesh sizes and dimensionalities of printed objects? The average number of vertices/elements per generated mesh is 27k/54k. We remark that our mesh resolution, determined by the underlying implicit representation, is adjustable based on user needs. Additionally, we can optimize a low-resolution mesh and output a final mesh with higher resolution. Q12: Are Figure 7 results printed? We did not print them. Q13: x_{com} is not defined in Eq. 7. It denotes the position of the center of mass of the underlying (solid with uniform density) geometry; we’ll clarify this in the revised version. --- Rebuttal Comment 1.1: Comment: Thank you for your explanations. I changed my score and agree with accepting the paper.	Summary: This paper introduces a method to make 3D assets produced by SDS-based generative models stand on their own. On top of the SDS loss, it proposes a “standability" loss that encourages the rigid-body simulation result to be rotation-free, and a “stable” loss to encourage the generated shape to be a “local minimum” of height for the center of mass. The evaluation is conducted both in simulation and 3D-printed shapes in the real world. Strengths: - This paper addresses an interesting problem of making 3D-generated shapes stand. Although this problem has been studied from the perspective of 3D printing, this paper introduces new methods tailored to the new setting of SDS-based 3D generation. - The proposed method seems to be effective from the various presented results, with a significant improvement in the success rate of standing. Weaknesses: - The proposed method involves solving an optimization that involves multiple loss terms. No ablation study is provided to prove the necessity of each term. In particular, are both standability loss and stable loss required? How are their influence on the optimization results respectively? Are both normal smoothing Eq (8) and Laplacian smoothing Eq (9) required and why is Laplacian smoothing only applied to the bottom of the shape? - The traditional balancing method in computational fabrication [59] is only discussed but not actually compared in experiments. - What are the loss weights for each term in Eq (10)? These numbers may be tricky to set, and thus make the paper hard to reproduce. - How is $T$ set in Eq (5)? How do different choices of $T$ influence the results? - The success rate of the baseline method is not reported in Fig. 6 for the audience to understand the improvement. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses. These are important questions to answer for a well-rounded paper. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Seems adequate. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for recognizing the novelty and effectiveness of our work. Indeed, our work is not directly targeting 3D printing, but an automated 3D generation pipeline without manual tuning. Q1: No ablation study is provided to prove the necessity of each term. In particular, are both standability loss and stable loss required? How are their influence on the optimization results respectively? Are both normal smoothing Eq (8) and Laplacian smoothing Eq (9) required and why is Laplacian smoothing only applied to the bottom of the shape? We perform additional ablation studies, shown in Fig. 3 in the attached PDF, to justify the necessity of each term. In particular, our standability loss is necessary for ensuring successful self-support, as it uses physical simulation for verification (Fig. 3 (b)); the stability loss is included to ensure that the object remains stable under small perturbations (Table 1), since, in real-world applications, objects may not be placed in the exact same initial pose; our Laplacian smoothing is applied to avoid artifacts such as irregular shapes on the contact surface (Fig. 3 (d)), and is therefore only applied at the bottom; while the normal smoothing loss is used to improve the overall appearance of the shape, as the mesh generated in the coarse stage is usually very rough. Q2: The traditional balancing method in computational fabrication [59] is only discussed but not actually compared in experiments. We provide additional comparison experiments with make-it-stand [59], as shown in Fig. 2 in the attached PDF. In our global response, we discuss in detail the drawbacks of such post-processing methods compared to our joint optimization approach. Please refer to the global response for more details. Overall, post-processing methods may overlook semantics during optimization, leading to degraded text alignment. In particular, make-it-stand [59] assumes the contact region is unchanged. Such inflexibility may result in undesired outcomes, as demonstrated in Fig. 2. In addition, make-it-stand [59] requires more human intervention, such as selecting multiple thresholds and placing handles, whereas our approach is more streamlined. Q3: What are the loss weights for each term in Eq (10)? These numbers may be tricky to set, and thus make the paper hard to reproduce. In our experiments, we use the following weights for our loss terms by default: {lam_sds = 1, lam_normal = 1e4, lam_stand = 1e5, lam_stable = 1e5, lam_blap = 1e7}. For a few examples, we tune these weights within the following ranges: {lam_stand = 1e5-5e5, lam_stable = 1e5-5e5, lam_blap = 1e6-1e7}. Our heuristic intuition is to keep the SDS and physical loss terms roughly on the same scale. For the regularization terms, we scale them to around 1/1000 to 1/100 of the SDS and physical loss terms. We will release our code upon acceptance so the community can reproduce our results. Q4: How is (T) set in Eq (5)? How do different choices of (T) influence the results? We empirically set $T = 2$ seconds, which in practice is sufficient to generate standable results. Setting $T$ to a larger value will not alter the simulated pose at the end time if the model has reached a steady state, and thus will not affect the resulting loss. However, it will significantly increase the simulation and the subsequent loss backward time, reducing computational efficiency. Conversely, setting $T$ to a smaller value may not be long enough to capture the final steady state, resulting in insufficient penalization of instability and ultimately leading to unstandable 3D models. Q5: The success rate of the baseline method is not reported in Fig. 6 for the audience to understand the improvement When verified in simulation, none of the generated results from the baseline method can stand stably when placed straight up even without perturbation (as mentioned in lines 276-279). In other words, the success rate of baseline methods under perturbation is zero.	Summary: The paper introduces a differentiable simulation-based loss to refine the existing SDS(Score Distillation Sampling)-based text-to-3D frameworks. Concretely, it relies on a differentiable simulator Warp to provide gradients for keeping the rotation of the generated mesh unchanged after a period of time. Besides, the authors also introduce a stable equilibrium loss, which favors a flat contact surface. Experiments show that the proposed method can generate self-supporting 3D models given text prompts, verified by simulation and real-world results with 3D-printed objects. Strengths: The paper introduces an effective and flexible method to refine the existing SDS(Score Distillation Sampling)-based text-to-3D frameworks, like Magic3D and MVDream. The authors also provide experiments on real-world validation. Weaknesses: 1. The paper lacks enough baselines to show that the task the paper tackles is non-trivial. For example, the claim that "Directly integrating these methods with 3D generative AI as a postprocessing module is suboptimal" (L55-L56) is not verified by the experiments. If a self-supporting mesh is needed, the simplest baseline can be cutting the mesh by a flat plane moderately higher than the lowest vertex, as a postprocessing step. 2. The physical constraints used in this work are relatively limited. The standability loss only keeps the rotation unchanged. It can not recover some common cases like "a standing horse" (e.g., Napoleon Crossing the Alps), although those cases may be beyond "self-supporting" defined in this paper. It seems that the simulator has not been fully leveraged, since a simple post-processing baseline (e.g., flatting surfaces that contact with the ground) may also achieve good standability and stability. 3. The current method is only applied to SDS-based methods only (already pointed out by the authors in the limitations) Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can the authors provide more baselines, especially any existing post-processing method with the help of differentiable simulators, to enhance their technical contributions? One example is given in the Weaknesses section. 2. Can the authors give a clear definition of "self-supporting"? It is a little ambiguous, since one can replace the initial pose with the pose after simulation so that Eq. (3) about standability can be satisfied, as long as the limit exists. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The paper has addressed its limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for recognizing the effectiveness and flexibility of our framework. Indeed, our pipeline can work with different 3D generative models and different differentiable physical simulators, allowing potential variants and extensions. Q1: The paper lacks enough baselines to show that the task the paper tackles is non-trivial [...], the simplest baseline can be cutting the mesh by a flat plane moderately higher than the lowest vertex [...] Can the authors provide more baselines? We show more baseline results in the attached PDF. Cutting the mesh by a flat plane moderately above the lowest vertex cannot robustly generate a standable output. As shown in Fig. 1 in the attached PDF, for the case of “a standing goose,” horizontally cutting the mesh from the bottom at various heights all result in failure especially when the center of mass is outside the supporting region. We additionally compare with make-it-stand [59]. As shown in Fig. 2 in the attached PDF, make-it-stand [59] holds the supporting surface unchanged, potentially leading to distorted results. Additionally, its results suffer from worsened text alignment due to the overlook of semantics. Q2: The standability loss only keeps the rotation unchanged. [...] It seems that the simulator has not been fully leveraged, since a simple post-processing baseline (e.g., flatting surfaces that contact with the ground) may also achieve good standability and stability. In this work, we consider the 3D models as rigid bodies whose dynamical states can simply be represented by rotation $R$ and translation $T$. We only consider the rotation $R$ in the standability loss as real-world instability mostly leads to rotational deviation from the initial state. Motion related to translation, such as falling due to gravity, is irrelevant to standability and is therefore disregarded. This point is explained in Line 159-161 of our paper. For future exploration involving other types of materials, such as soft bodies, additional physical state parameters like deformation may need to be considered. As shown in Fig. 1 and 2 in the attached PDF, compared with the post-processing baselines, like cutting the bottom and make-it-stand [59], our joint optimization with simulation loss can dynamically adjust the center of mass and the contact region without compromising the text alignment. Please refer to our global response for more details on joint optimization versus post-processing. Q3: Can the authors give a clear definition of "self-supporting"? It is a little ambiguous, since one can replace the initial pose with the pose after simulation so that Eq. (3) about standability can be satisfied, as long as the limit exists. As elaborated in our global response, we set the initial pose (upright direction) based on the output of text-to-3D generative models, as the upright direction is semantics-driven and generative models inherently learn these semantics from their training data. For the definition of self-supporting, we consider it as the condition where, once the upright pose is determined by the text-to-3D model, the generated 3D model should satisfy the standability criteria defined in Eq. (3). --- Rebuttal Comment 1.1: Comment: First of all, thank you for the answers in the rebuttal. For Q1, I appreciate that the authors provide comparison on certain, perhaps representative, cases qualitatively. However, it is better to compare with other methods quantitatively, which will make the paper more convincing. I would like to keep my rating.	Summary: This paper addresses the problem of generating 3D models from text that are visually appealing but often physically unstable in simulations or when 3D printed. The authors incorporates a differentiable simulation-based loss function and physically inspired regularization to ensure generated models are stable under perturbation. The proposed method involves a two-stage training process where a coarse model is first generated from text prompts and then refined with physical constraints to ensure stability. Experiments show that models created with Atlas3D maintain stability better than those from existing methods. The optimized models are also 3D fabricated to verify the stability in real world. Strengths: The paper proposes an original and important solution to the issue of physical stability in text-to-3D generation, combining differentiable simulation-based loss functions with physically inspired regularization. This approach significantly enhances existing methods, reducing the need for manual post-processing and making the models immediately practical for various applications. The method is sound, with clear and concise writing that guides the reader effectively and should be straightforward to implement for someone with a background in physical simulation. The experiment quality is high, validated through extensive experiments in both simulations and real-world scenarios, demonstrating the method's robustness and versatility. The results confirm the method's effectiveness in ensuring the stability of 3D models. Weaknesses: The paper is strong overall, and I did not identify any significant weaknesses. However, to further support future research in this area, I would suggest releasing the code. Providing the implementation would greatly benefit the community, enabling others to replicate the results and build upon this work. Technical Quality: 4 Clarity: 4 Questions for Authors: I don't have any specific questions at this time. The paper is clear and well-executed, addressing an important issue effectively. Implementation is also described in detail. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: I am curious about the failure modes of Atlas3D. Could you provide more details on the scenarios or conditions under which the method might not perform as expected? Understanding these limitations would be helpful for future research and applications. Additionally, addressing these potential failure modes could further strengthen the robustness of your approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for recognizing the novelty and effectiveness of our method. To the best of our knowledge, our method is the first to bring standability to large generative models via joint optimization of standability loss and score distillation sampling loss. Q1: I am curious about the failure modes of Atlas3D. Could you provide more details on the scenarios or conditions under which the method might not perform as expected? Our approach may fail under certain cases. Firstly, some text prompts inherently contradict the concept of standability, such as “a swan and its cygnets swimming in a pond” or “a beautiful rainbow fish.” In these cases, the SDS loss may conflict with the physical loss, leading to unsatisfactory results. Additionally, since we assume models are solid with uniform density, under certain geometries where the projected position of the center of mass onto the horizontal plane is way off the contact region, our approach may not converge to the global minimum of the standability loss (which is zero). Q2: I would suggest releasing the code. Thank you for the suggestion. Our code has been cleaned up and is ready for release upon acceptance. --- Rebuttal Comment 1.1: Comment: Thank you for the response. I do not have further questions.	Rebuttal 1: Rebuttal: Dear Reviewers, We would like to thank all reviewers for your insightful and constructive feedback. We are encouraged by the recognition that our paper: - Addresses the interesting problem of physical stability in text-to-3D generation and provides an effective and important solution [Reviewer Nrjn, MwTq, XTuH] - Introduces elegant standability and stable equilibrium losses [Reviewer iQck] - Includes well-validated experiments in both simulation and real-world settings [Reviewer Nrjn, MwTq, XTuH, iQck] Below, we first address some common questions raised by the reviewers. ## Upright Direction We believe that the choice of the upright direction is semantically driven and requires human judgment. Different people might have different interpretations of what an upright pose should be. Our objective is to enable the automatic generation of self-supporting 3D models without human intervention, including manually selecting the upright direction. Existing large pre-trained 3D generative models are based on human-made assets, typically crafted to be upright by artists. By learning from such data, these models inherently have an understanding of the upright direction. Thus we designate the vertical direction from the default 3D generation output as the upright direction. This also allows batch evaluation of our method, as shown in Fig. 8 in our paper.. ## Joint Optimization v.s. Post-processing While directly post-processing 3D generated models is a straightforward and effective approach to achieving physical stability, it may result in outcomes that do not align with the text prompt, as it overlooks semantics. One simple post-processing method, as pointed out by Reviewer MwTq, is to cut the mesh by a flat plane slightly higher than the lowest vertex. However, this method will fail when the projection of the center of mass lies outside the contact region, as shown in Fig. 1 in the attached PDF. Additionally, determining the cutting height is another parameter that must be manually set or optimized, and this may also degrade the overall appearance. Another post-processing method, make-it-stand [59], offers an effective way to relocate the center of mass to achieve standability. However, it assumes that the supporting surface is fixed during optimization, which can lead to distorted results due to the imperfect quality of text-to-3D generated models. We provide two examples in Fig. 2 in the attached PDF to illustrate this point. For the goose example generated by Magic3D, one leg is shorter than the other. Make-it-stand only treats one foot as the supporting surface and ignores the other due to its post-processing nature, whereas our joint-optimization pipeline enables stable standing with two legs on the ground. A similar issue also happens to the kangaroo example. More importantly, the text alignment degrades as semantics are overlooked during optimization. In contrast, our proposed joint optimization method preserves the text alignment and dynamically adjusts the center of mass as well as the supporting surface configuration. In Fig. 3(e) in the attached PDF, we also apply our proposed losses in a post-processing manner. Similarly, while it is able to optimize the 3D models to make it standable, the text alignment is compromised. Overall, compared to the post-processing method, joint optimization is a more robust way which better balances text alignment and physical constraints. ## Ablation Study We perform additional ablation studies in the attached PDF to demonstrate the necessity of our proposed losses. It can be observed in Fig. 3(b), that without standability loss $L_\text{stand}$, the figure fails to stand. While the figure can still stand without stable equilibrium loss $L_\text{stable}$, as demonstrated in Fig. 3(c), it is less stable under perturbation. As shown in Tab. 1, introducing the stability loss consistently increases the success rate of standing under different scales of perturbations. Additionally, we show the effectiveness of geometry regularization loss term $L_\text{b-lap}$ in Fig. 3(d), which helps smooth the geometry and avoid spiky artifacts on the surface. We will incorporate these results into our revised version. We believe that our pipeline is lightweight and is applicable to a wide range of 3D generative models. We will release our code upon acceptance, as suggested by Reviewer Nrjn. We also provided detailed responses to each reviewer separately. Pdf: /pdf/c0510d99985688d60ef4c2c31e9f2360dbd98049.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Efficient Large Multi-modal Models via Visual Context Compression	Accept (poster)	Summary: This paper shows that visual tokens are redundant in MLLM and can be compreseed by a large ratio without significantly hurting the model performance. Based on this observation, this paper studied several different approaches to compress visual tokens and identified that the simple average pooling method is the most effective one. Based on this, the paper further studied several different stage-wise MLLM training strategies, where the training starts with heavy compression and ends with no compression. The proposed training strategy could save the training cost by 16% while achieving even better performance than the baseline. Strengths: 1. The observation that visual tokens are redundant and can be largely compressed without significantly hurting the model performance is good. It could serve as a direction for future works. 2. The paper conducted thorough empirical studies about different compression methods, and proposed several training methods based on visual token compression. 3. The proposed method shows better performance than the baseline while reducing the training cost by 16%. Weaknesses: 1. It would be better to provide more discussions and investigations on why other visual compressors are significantly worse than average pooling. This may provide more insights on how to design good compressor. 2. The paper only studied LLaVA-1.5-7B. It would be better to show that if the method could scale up to larger models like 13B, 34B, or other structures such as Mini-Gemini [1]. Showing the efficiency on those large variants can better demonstrates the effectiveness of the method. 3. The 16% training efficiency improvement is marginal in practice, especially given that the four-stage training would be more cumbersome compared to the baseline. [1] Li, Yanwei, et al. "Mini-gemini: Mining the potential of multi-modality vision language models." arXiv preprint arXiv:2403.18814 (2024). Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the weakness section Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive suggestions. We address the concerns raised on a point-by-point basis, including additional benchmarks on the global response PDF. We will include the new results in our revised manuscript. For weakness 1: "more discussions and investigations on why other visual compressors are significantly worse than average pooling", we thanks for the reviewer's suggestion. In line 262-263, we analyzed that "they are ineffective when applied to training because the in-training attention scores are unstable.". Our insight is that while advanced compressors (e.g., attention-based token pruning) excel in inference-only scenarios (See bottom rows in Table 2), the simple pooling method performs better during training (See top rows in Table 2). We hypothesize that this is because training advanced compressors, such as attention-based pruning, necessitates (1) differentiable token selection and (2) stable attention mechanisms of LM Transformers. For weakness 2: "scale up to larger models like 13B, 34B, or other structures such as Mini-Gemini [1]", we thanks for the reviewer's suggestion. To show the scalability, we scale the method up to 13B model, and observe consistent performance and efficiency improvements. \| Scheme \| #Token \| CR \| train-time \| GQA \| MM-Vet \| SQA \| MME \| VQA^T \| POPE \| MMBench \| MMB-CN \| VQAv2 \| LLaVA^W \| VisWiz \| SEED^I \| MMMU \| Avg. \| \|---------------\|---------\|------\|---\|---\|--------\|------\|------\|-----------\|------\|------\|-----------\|------------\|-------------\|--------\|---------\|------\|------\| \| LLaVA-13B \| 18432 \| 100% \| 21.1h \| 63.0 \| 35.0 \| 74.1 \| 1503 \| 57.0 \| 86.6 \| 68.2 \| 63.5 \| 79.6 \| 71.0 \| 53.6 \| 66.4 \| 37.9 \| 63.9 \| \| Ours-13B \| 10863 \| 170% \| 17.6h \| 63.0 \| 35.4 \| 74.2 \| 1502 \| 56.7 \| 86.8 \| 68.0 \| 63.3 \| 79.7 \| 71.3 \| 53.8 \| 66.4 \| 37.8 \| 64.0 \| To show the generalizability to different structures, we supplement the experiment on Mini-Gemini (MGM-2B). Our four-stage compression training delivers comparable results with a 17% increase in training efficiency and significantly fewer vision tokens as input. We plan to incorporate additional structures in future versions. \| Scheme \| #Token \| CR \| GQA \| MM-Vet \| SQA \| MME \| VQA^T \| POPE \| MMBench \| MMB-CN \| VQAv2 \| LLaVA^W \| VisWiz \| SEED^I \| MMMU \| Avg. \| \|---------------\|---------\|------\|------\|--------\|------\|------\|-----------\|------\|------\|-----------\|------------\|-------------\|--------\|---------\|------\|------\| \| MGM-2B \| 18432 \| 100% \| 60.7 \| 30.1 \| 62.7 \| 1327 \| 57.1 \| 86.0 \| 61.9 \| 50.6 \| 76.3 \| 65.9 \| 48.3 \| 63.8 \| 28.1 \| 58.3 \| \| Ours \| 10863 \| 170% \| 58.8 \| 30.2 \| 62.2 \| 1325 \| 54.3 \| 87.0 \| 62.5 \| 52.5 \| 76.3 \| 65.7 \| 48.9 \| 63.1 \| 27.3 \| 58.1 \| We thanks for the reviewer's comments on Weakness 3. It's worth noting that our method not only improves training efficiency by 16% but also enhances average performance by 0.5\% across 13 benchmarks. To address the cumbersomeness, we are exploring the development of a simple training schedule (linearly decreasing the compression ratio over time; like learning rate schedule), aimed at balancing training efficiency and model simplicity. We sincerely hope our rebuttal addresses your concerns, and we look forward to your feedback. Your response will motivate us to refine our paper into a more solid version. Kind regards, Authors of Paper 2615	Summary: The paper presents a novel approach to reducing redundancy in visual tokens within MLLMs by introducing the Visual Context Compressor (VCC) and Stage-wise MLLM Training. The VCC uses simple average pooling to compress visual tokens during training, enhancing efficiency without sacrificing performance. The Stage-wise MLLM Training progressively reduces compression as training proceeds, ensuring no loss of information during testing. The proposed methods demonstrate significant improvements in training efficiency and performance across multiple benchmarks. Strengths: 1. The paper addresses an underexplored area in MLLMs by focusing on the redundancy of visual tokens and introducing effective compression techniques. 2. The proposed methods achieve significant reductions in training costs while maintaining or improving performance on various benchmarks. 3. The paper provides thorough experimental validation across multiple benchmarks, demonstrating the effectiveness of the proposed methods. Weaknesses: 1. While the paper shows improvements across several benchmarks, the evaluation is primarily focused on visual question answering tasks. Additional evaluation on other multi-modal tasks could strengthen the claims. 2. The paper mentions potential information loss due to compression but does not provide a detailed analysis/visualization of how this might affect tasks requiring dense visual information. 3. This paper lacks some essential experiments. (1) The paper does not explore the effect of using larger input resolutions. For example, evaluating the method with 448×448 input images, which contain 2.25 times more visual tokens than the 336×336 input, could provide insights into the method's scalability. Designing a compression setting with a ratio of 225% for this input size and comparing it with the original LLaVA setting (compression ratio of 100%) would be valuable. (2) Testing the method on larger models such as LLaVA-13B or models with more input visual tokens (e.g., LLaVA-Next) could solidify the experimental section and demonstrate the robustness of the proposed approach. 4. The best stage-wise MLLM training scheme looks difficult to transfer to other model & data settings. The training scheme has too many options and variables. If the model size and dataset size increase significantly, the time cost and computation cost for finding the best scheme can become prohibitively large. Technical Quality: 4 Clarity: 4 Questions for Authors: Please see the weakness. Minor issues: 1. What does the #Tokens mean in the tables? Is it inversely proportional to CR? 2. How do you compute the average performance in Table 5? 3. Some suggestions for future improvements: (1) multi-stage compression: use different compressors’ strides in the different positions of the LLM. For example, layer 1-3: stride=1; layer 4-12: stride=2; layer 13-24: stride: 4. (2) extend this technique to other modalities, such as video understanding. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have adequately discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive suggestions. We address the concerns raised on a point-by-point basis, including additional benchmarks on the global response PDF. We will include the new results in our revised manuscript. To address weakness 1, we follow the reviewer's suggestion, and add the evaluation of multi-modal tasks in Table 1 of Global Response, especially MMBench and MMMU (A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI)[49]. For weakness 2, we thanks for the reviewer's comment on information loss. We have discussed this in the limitation section of original paper. The naive compression may not be ideal for tasks requiring dense comprehension, grounded reasoning, and OCR/text capabilities. Nonetheless, our staged training approach, which does not involve compression in the final stage, is capable of managing tasks that require dense visual information. We provide analysis and visualization in Global Response. To address weakness 3.1, we try to implement the experiment of larger 448×448 input resolution. However, we note that the default openai-CLIP-336 encoder can only input 336x336 and interpolating positional embedding to 448x448 degrades the performance, resulting in -9% drop in GQA. We are experimenting the dynamic high resolution of LLaVA-NeXT (its training code has not yet been made public as of Aug/8) by employing openai-CLIP-224 to encode four 224x224 images derived from a 448x448 image. We plan to include the large input experiment in the revision once our new results are available. For weakness 3.2, we thanks for the reviewer's suggestion. As LLaVA-NeXT does not make the training code public upon the submission of rebuttal, we are in the progress of reproducing the code. We supplement the experiment of testing the method on larger model of LLaVA-13B in Table 3 of Global Response, and we can achieve 177% compression rate and accelerate training efficiency by 15% while keeping the performance. We will include the new results in our revision. For weakness 4, we thanks for the reviewer's suggestion. We aimed at providing comprehensive analysis of design space. We will provide a simple version for the practical use in the near future. For instance, we might implement a universal training schedule that gradually reduces the compression ratio over time. We thanks for the reviewer's comments for the minor issues. We address them as follows. Q1. We thanks for the reviewer's comment. #Tokens means the summation of number of visual tokens in each LLM transformer layer. It is inversely proportional to CR. Q2. We mentioned this in line 242: "When reporting average performance in Tab. 5, the score of MME is normalized by 2000, as its range is from 800 to 2000". The average performance is simply the mean of the (normalized when it is MME) individual performances. Q3 (1) multi-stage compression. We thanks for the reviewer's suggestion. The multi-stage compression is a great idea. We follow the suggestion and experiment with the setting of multi-stage (layer 0-3: stride=1; layer 4-11: stride=2; layer 12-23: stride: 4; layer 24-31: stride: 8), resulting in a CR of 267%. We compare this to single-stage compression (on layer 8 stride 8) with a CR of 266%. \| Compressor \| Phase \| CR \| GQA \| MM-Vet \| SQA \| MME \| VQA^T \| POPE \| MMBench \| MMB-CN \| VQAv2 \| LLaVA^W \| VisWiz \| SEED^I \| MMMU \| Avg. \| \|---------------\|---------\|------\|------\|--------\|------\|------\|-----------\|------\|------\|-----------\|------------\|-------------\|--------\|---------\|------\|------\| \| layer=8 stride=8 \| inference \| 266%\| 57.8\|25.3\|70.2\|1337\|52.1\|86.0\|60.4\|52.2\|74.6\|56.0\|48.1\|58.3\|33.3\|57.0\| \| layer=8 stride=8\| training\| 266%\| 60.7 \| 30.7 \| 71.3 \| 1456 \| 56.9 \| 86.4 \| 64.6 \| 58.0 \| 77.9 \| 67.0 \| 48.8 \| 66.0 \| 35.3 \| 61.3 \| \| multi-stage \| inference\| 267% \| 60.7 \| 28.9 \|70.3 \| 1403 \| 55.4 \| 85.1 \| 65.2 \| 57.1 \| 77.7 \| 60.6 \| 49.1 \| 64.8 \| 35.2 \| 60.0 \| \| multi-stage\| training\| 267% \| 60.9 \| 29.5 \|70.5 \| 1408 \| 55.9 \| 84.8 \| 65.4 \| 57.4 \| 76.6 \| 61.1 \| 48.9 \| 64.7 \| 34.9 \| 60.2 \| The multi-stage compressor exhibits strong performance when applied directly during inference, outperforming layer 8 stride 8 by 3% under same CR. However, it's surprising that training the multi-stage compressor yields only a marginal average performance improvement of 0.2%. We analyze that the complexity of multi-stage operations makes the LLM more challenging to train. We will have more future analysis. Q3 (2) video extension. We thanks for the reviewer's suggestion. For extension of modalities, we have experimented on video-language understanding (based on video-LLaVA). We observe consistent enhancements (an average improvement of +0.4\% across three video datasets and a 9% reduction in training time) over the baseline using our new training setting. We will include the new results in our revision. \| Scheme \| #Tokens \| CR \| TFLOPs \| Train-time \| MSVD-QA Score \| MSVD-QA Acc \| MSRVTT-QA Score \| MSRVTT-QA Acc \| ActivityNet-QA Score \| ActivityNet-QA Acc \| Average Score \| Average Acc \| \|------------------\|-------------------\|--------------\|------------------\|------------\|---------------\|-------------\|-----------------\|---------------\|----------------------\|--------------------\|---------------\|-------------\| \| Video-LLaVA-7B \| 147456 \| 100% \| 29.68 \| 40.7h \| 3.69 \| 69.1 \| 3.48 \| 56.8 \| 3.28 \| 47.5 \| 3.48 \| 57.8 \| \| Ours \| 86904 \| 170% \| 18.64 \| 37.1h \| 3.74 \| 69.8 \| 3.49 \| 56.9 \| 3.27 \| 47.8 \| 3.50 \| 58.2 We sincerely hope our rebuttal addresses your concerns, and we look forward to your feedback. Your response will motivate us to refine our paper into a more solid version. Kind regards, Authors of Paper 2615	Summary: The paper presents a compelling study on redundancy of visual tokens in MLLMs and practical approaches to reduce them. The paper first verifies that one could eliminate up to 70% of visual tokens at testing time via simple average pooling with minial performance degradation. Then they experiments with several approaches for compressing the visual tokens, finding that a simple average pooling works the best. They also propose a staged training recipe, where we could save computation during early training stages and gradually remove compression. Strengths: 1. The idea is clean and simple. It is nice to empirically verify the redundancy in visual tokens. 2. The proposed staged training and the discussion around wider then deeper v.s. deeper then wider is interesting. Weaknesses: - There seems to be an easy baseline missing from the discussion, such as 2D conv (as in Honeybee). Honeybee: Locality-enhanced Projector for Multimodal LLM - There are some ambiguous experimental details. Please see questions. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. When doing average pooling, why choose the 1D average pooling instead of 2D average pooling over the grid features? 2. What is K in table 2? I cannot seem to be able to find discussion on tuning k v.s. stride in Section 4.2. 3. For ablation, why report only the performance on the 4 benchmarks? I suspect that the compression performance will also be quite dataset-dependent. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive suggestions. We address the concerns raised on a point-by-point basis, including additional benchmarks on the global response PDF. We will include the new results in our revised manuscript. For weakness 1, we have included 2D-Conv and C-Abstractor (from Honeybee) as a baseline in Table below. The stride of 2D-Conv and C-Abstractor are the same with other compressor to ensure fair comparison under same compression ratio (CR). However, we found that (1) The performance of convolution-based compressors is significantly lower, with 2D-Conv achieving 55.1\%, C-Abstractor achieving 50.6\%, and 1D Pool reaching 60.4\% in average accuracy across 13 benchmarks. (2) there is overhead parameters of conv. We analyze that it is hard to optimize the extra convolution kernel within LLM at the same time (this is an underexplore area). \| Compressor \| #Tokens \| CR \| GQA \| MM-Vet \| SQA \| MME \| VQA^T \| POPE \| MMBench \| MMB-CN \| VQAv2 \| LLaVA^W \| VisWiz \| SEED^I \| MMMU \| Avg. \| \|---------------\|---------\|------\|------\|--------\|------\|------\|-----------\|------\|------\|-----------\|------------\|-------------\|--------\|---------\|------\|------\| \| 2D-Conv \| 2232 \| 826% \| 58.6 \| 28.6 \| 71.6 \| 1366 \| 51.8 \| 84.4 \| 63.8 \| 55.6 \| 74.0 \| 63.8 \| 48.1 \| 60.1 \| 25.6 \| 55.1 \| \| C-Abstractor \| 2232 \| 826% \| 53.7 \| 23.7 \| 70.9 \| 1209 \| 48.7 \| 82.8 \| 58.0 \| 50.3 \| 68.2 \| 48.6 \| 48.0 \| 53.4 \| 23.9 \| 50.6 \| \| 2D Pool \| 2232 \| 826% \| 57.5 \| 28.7 \| 71.5 \| 1426 \| 53.1 \| 84.0 \| 64.2 \| 58.7 \| 74.3 \| 64.2 \| 50.0 \| 66.5 \| 34.3 \| 56.1 \| \| 1D Pool \| 2232 \| 826% \| 58.3 \| 29.2 \| 71.4 \| 1434 \| 53.6 \| 83.8 \| 64.8 \| 58.6 \| 74.5 \| 65.0 \| 49.1 \| 66.8 \| 35.0 \| 56.3 \| It is worthy mentioning that Honeybee focus on the design of projector outside LLM while our focus is the visual context compression within LLM. Therefore, our method has the potential to accelerate the inference and training of Honeybee (i.e., keep Honeybee's 2D Conv on projector before LLM, and add our 1D pool compressor on visual tokens within LLM). We will include these new results in our revision. For question 1, we thank the reviewer for the comments. We add experiment with 2D pooling in the same Table above. As results, 1D pooling perform slightly better than 2D pooling (+0.2\%) under the same compression ratio. This could be due to the visual tokens being processed in a 1D manner within several LLM layers. Besides, we opted for 1D pooling because it allows for more adaptable compression ratio. For instance, using a 1D pool with a stride can reduce the number of tokens by 2x, whereas a 2D pool with a stride of 2 reduces the token count by 4x. We will include these new results in our revision. For question 2, we appreciate the reviewer's feedback. K is set to 2 with stride set to 8 in Table 2, resulting in a compression ratio of 556\% to approximate [41]’s 514\% (fixed CR) for a fair comparison of compressor. For question 3, we thank the reviewer for the comments. The previous validated four benchmarks are considered representative because they assess diverse capabilities: GQA evaluates visual understanding, SQA focuses on scientific knowledge across 26 topics, MME tests perception and cognition through 14 subtasks, and MM-Vet measures ensembling skills in recognition, OCR, math, and spatial awareness (encompassing other benchmarks like VQAv2, COCO, and TextVQA). In Table 1 of GLOBAL RESPONSE PDF, we have included all 13 benchmarks in our ablation now and evaluated based on the average performance across these benchmarks. The evaluated 13 benchmarks include not only VQA but also multimodal benchmarks. We achieve consistent conclusions with our original paper for both the ablation study and final results. We will include these new results in our revision. We sincerely hope our rebuttal addresses your concerns, and we look forward to your feedback. Your response will motivate us to refine our paper into a more solid version. Kind regards, Authors of Paper 2615	null	null	Rebuttal 1: Rebuttal: We are grateful to the reviewers for their feedback. Reviewer dJy4 commends the paper for presenting "The paper presents a compelling study. The idea is clean and simple". Reviewer ENde notes "the paper addresses an underexplored area in MLLMs", and Reviewer PsS7 appreciates the thorough empirical studies conducted. Additionally, Reviewer PsS7 suggests that the work could provide a direction for future research, highlighting its potential to influence further advancements. We thanks the reviewers for the constructive comments. To address them, we provide a global response PDF (attached to this post) with more benchmark results (dJy4, ENde), scalability to large 13B model (ENde, PsS7), and additional visual analysis (ENde). We also address each reviewer's questions with new results on a point-by-point basis. We sincerely hope our rebuttal addresses your concerns, and we look forward to your feedback. Your response will motivate us to refine our paper into a more solid version. Kind regards, Authors of Paper 2615 Pdf: /pdf/54bc6dddcc1930c2cc739f1fbd7501a9085143d8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MambaTalk: Efficient Holistic Gesture Synthesis with Selective State Space Models	Accept (poster)	Summary: This paper explores the application of state space models (SSMs) to co-speech gesture generation. The authors identify the computational challenges and jittering issues associated with the direct application of SSMs to gesture synthesis. To address these, they propose a two-stage modeling strategy with discrete motion priors and hybrid fusion modules. The first stage involves learning discrete holistic gesture priors with multiple VQVAEs, and the second stage refines latent space representations using local and global scans. Experiments demonstrate that MambaTalk outperforms state-of-the-art models in generating natural, rhythmic, and contextually appropriate gestures. Strengths: 1. The paper is well-structured and clearly written. 2. It seems that MambaTalk is the first to explore the potential of the selective scan mechanism for co-speech gesture synthesis. 3. The methodology is thoroughly validated through extensive experiments, including both subjective and objective evaluations. Weaknesses: 1. Limited novelty. Integrating the local and global scans seems not novel. More discussion on the comparison between MambaTalk and baseline methods is needed. 2. Limited datasets for method evaluation. Only the BEATX-standard dataset is adopted to evaluate the performance of the method. How about the zero-shot quantitative comparison on other datasets, such as the BEAT dataset or the RAVDESS dataset, which take emotion into account? Technical Quality: 4 Clarity: 4 Questions for Authors: See the "Weakness" section. Confidence: 2 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: See the "Weaknesses" section. If possible, I suggest adding more visualization results and videos to the supplementary materials. Additionally, please note that the final score for this study is not solely determined by the peer reviewers' discussion. If the authors can address my main concerns, I would be willing to raise the score. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: A1: Comparison with baseline methods Thanks for your advice. Comparing with the baseline methods, we not only propose a method that applies a selective scan mechanism on co-speech gesture synthesis with local and global scans (a novel method not found in previous work), but we also consider the different motion patterns of different parts of the human body. We found that directly applying Mamba would cause serious shaking problems. Therefore, we leverage VQVAE and learnable queries to incorporate motion priors. Additionally, the direct application of Mamba also has the issue that the limb movements of different body parts tend to be average. Therefore, we integrate attention and selective scanning mechanisms into our framework to model spatial and temporal relationships. We will refine our statements and make our framework clearer (e.g., adding pseudo-code) in our revised version. Comparing to baselines, we have an advantage on all metrics, especially on BC (16.72%), MSE (36.35%), and LVD (13.99%). Our method also has efficiency in training and inference time (Appendix A.2). The training time for a single epoch of MambaTalk is only 42 seconds. In comparison, CaMN and EMAGE require 493 seconds and 83 seconds per epoch, respectively, which translates to 1173% and 212% of MambaTalk's epoch time. When considering the total time, MambaTalk also boasts a faster convergence rate, with our method requiring only 100 epochs. \| Method \| Time per Epoch (s) \| Epochs \| \|--------------------------\|------------------\|-----------------\| \| CaMN \| 493(1173%) \| 120 \| \| EMAGE \| 83(212%) \| 400 \| \| Ours \| 42(100%) \| 100 \| A2: Method evaluation Thanks for your advice. Conducting a zero-shot comparison between the BEAT dataset, which uses BVH to represent gestures, and RAVDESS, represented by videos, would be challenging due to the use of SMPLX in BEATX. Therefore, we retrain our method using the BEAT dataset to consider emotions and verify the broad advantages of our approach for upper body movements. To incorporate emotions, we utilize embedding to convert them into style features and employ adaptive layer normalization to integrate them into the latent space of our framework. Due to time constraints, we only conducted preliminary experiments. The results are shown in the table. Our method demonstrates a significant improvement compared to the baseline, which also validates the effectiveness of our approach. We will conduct more comprehensive experiments and include additional details in the revised version. \| Method \| FGD$\downarrow$ \| SRGR$\uparrow$ \| BeatAlign$\uparrow$ \| \|--------------------------\|------------------\|-----------------\|--------\| \| CaMN (baseline) \| 123.7 \| 0.239 \| 0.783 \| \| Ours \| 51.3 \| 0.256 \| 0.852 \| A3: More visualization results Thanks for your kind advice. We will include more visualization results and videos in the supplementary materials in our revised version. --- Rebuttal Comment 1.1: Comment: Thanks to the author for providing the rebuttal. The initial score remains the same after the rebuttal. --- Reply to Comment 1.1.1: Comment: We would like to express our sincere appreciation for your careful review and the time you have dedicated to evaluating our manuscript. We are grateful for the opportunity to provide a rebuttal and address the points you raised. We respect your decision and the thorough consideration you have given to our manuscript. We are pleased that our work has been recognized with a "weak accept" rating, which we interpret as a positive endorsement of our research. We would like to assure you that we have taken your feedback seriously and have made every effort to improve the manuscript. Should there be any further opportunities to refine our work or address any additional concerns, we are more than willing to do so. Once again, thank you for your valuable feedback.	Summary: This study explores the use of state space models (SSMs) to enhance gesture synthesis, addressing challenges such as diverse movement dynamics and unnatural jittering in generated gestures. Through a two-stage modeling approach and the introduction of MambaTalk with hybrid fusion modules, the study demonstrates superior performance compared to existing models in subjective and objective experiments. Strengths: 1. This work is the first to explore the potential of the selective scan mechanism for co-speech gesture synthesis, achieving a diverse and realistic range of facial and gesture animations. 2. The writing of this work is fine, there are no obvious typos. 3. The workflow pipeline figure is clear to understand. Weaknesses: My concerns and suggestions of this manuscript are listed as blew: 1. The motivation of this work sounds unclear and looks a little incremental. As claimed in the abstract section 'the high computational complexity of these techniques limits the application in reality'. This statement is not clear. Does computational complexity mean more model parameters? or training time? or inference time? or more GPU memory cost? The authors didn't have a clear statement. 2. Moreover, in the abstract, 'which stem primarily from the diverse movement dynamics of various body parts.' are actually the general difficulties of the co-speech gesture generation task, not the specific one of how to employ the SSMs to this task. 3. The introduction section is not clear. The authors pay much more attention to the related works. Not the motivation and the high-level technical contributions of their work. This led me to feel that the entire work lacked technological innovation after reading the Introduction. 4. As for the methods, it is just directly imply the SSMs to this work. I cannot see any design on how to effectively solve the problem of 'computational complexity'. Actually, the technical contribution is poor, and the overall pipeline is very similar to the previous work EMAGE[1]. 5. Could the authors explain why they only experimented on the BEATX dataset? As far as I know, TED[2] and TED-expressive[3] are also two commonly used co-speech gesture datasets. 6. The author did not conduct experiments to demonstrate how to effectively reduce the computational complexity by using SSMs, which directly led to the unclear motivation of this work. [1] Liu, H., Zhu, Z., Becherini, G., Peng, Y., Su, M., Zhou, Y., ... & Black, M. J. (2023). Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374. [2] Yoon, Y., Cha, B., Lee, J. H., Jang, M., Lee, J., Kim, J., & Lee, G. (2020). Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6), 1-16. [3] Liu, X., Wu, Q., Zhou, H., Xu, Y., Qian, R., Lin, X., ... & Zhou, B. (2022). Learning hierarchical cross-modal association for co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10462-10472). Technical Quality: 2 Clarity: 2 Questions for Authors: please refer to Weaknesses Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: please refer to Weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: A1: Computational efficiency Thanks for your question. Our computational efficiency is mainly reflected in the inference time, which has been analyzed in section A.2 of the appendix. We leverage the linear computational complexity of Mamba and the sequence compression capability of VQVAE within our framework, which helps in reducing computational complexity. Additionally, our method holds a significant advantage in training time. The training time for a single epoch of MambaTalk is only 42 seconds. In comparison, CaMN and EMAGE require 493 seconds and 83 seconds per epoch, respectively, which amounts to 1173% and 212% of MambaTalk's epoch time. When considering the total time, MambaTalk also demonstrates a faster convergence rate, as our method only requires 100 epochs. \| Method \| Time per Epoch (s) \| Epochs \| \|--------------------------\|------------------\|-----------------\| \| CaMN \| 493(1173%) \| 120 \| \| EMAGE \| 83(212%) \| 400 \| \| Ours \| 42(100%) \| 100 \| A2: Writing improvement Thanks for your valuable comments. The goal of our work is also to use Mamba to address the difficulties of the co-speech gesture generation task. We will refine our statement in the abstract. For the introduction section, we will simplify the description of related work and move this part to the related work section to make space for our motivation and the high-level technical contributions of our work (as illustrated in A3: Comparison with EMAGE). A3: Comparison with EMAGE Our work does not simply involve implementing Mamba for gesture synthesis. We found that directly applying Mamba would cause serious shaking problems. Therefore, we incorporated motion priors using VQVAEs and individual learnable queries for different parts of the body. Our solution differs from EMAGE, which involves extracting motion cues from masked body joints. At the same time, the direct application of Mamba also has the issue that the limb movements of different body parts tending to be average. Therefore, we refine the design of spatial and temporal modeling in latent spaces by proposing a local-to-global modeling strategy and incorporating them with attention and selective scan mechanism into the design of our framework. Our work is also the first framework based on SSMs designed for co-speech gesture synthesis. We will refine our statements and make our framework clearer (e.g., adding pseudo-code) in our revised version. For comparison with the transformer-based method EMAGE[1], our SSM-based method also has an advantage over EMAGE in all metrics, especially in terms of BC (16.72%), MSE (36.35%), and LVD (13.99%) metrics. Our work is not a disruptive innovation. However, in addition to using SSMs instead of transformers, we have also developed numerous adaptive designs for SSMs to make it work. A4: Reasons for choosing BEATX and the generalizable benefits of our method Thanks for your question. We experiment on the BEATX datasets since our work focuses on holistic gesture synthesis. However, most current datasets only include movements of specific body parts, not the entire body. For example, the TED[2] and TED-expressive[3] datasets that you mentioned only focus on upper body gesture synthesis. TED includes 10 upper body joints, while TED-expressive includes 13 upper body joints and 30 finger joints. Moreover, TED and TED-expressive datasets are based on the 3D pose estimator ExPose for extracting gestures, which contain some errors in the ground truth. Therefore, we conducted experiments using another motion capture dataset called BEAT to validate the generalizable benefits of our method. This dataset contains data on upper body and hand movements to validate the generalizable benefits of our method for specific parts of body movements. Due to time constraints, we only conducted preliminary experiments. The results are shown in the table. Our method demonstrates a significant improvement compared to the baseline, which also validates the effectiveness of our approach. We will conduct more comprehensive experiments and include them in the revised version. \| Method \| FGD$\downarrow$ \| SRGR$\uparrow$ \| BeatAlign$\uparrow$ \| \|--------------------------\|------------------\|-----------------\|--------\| \| CaMN (baseline) \| 123.7 \| 0.239 \| 0.783 \| \| Ours \| 51.3 \| 0.256 \| 0.852 \| --- Rebuttal Comment 1.1: Title: Response to authors Comment: Dear authors, Thanks for your efforts and responses to my questions. Although most of my concerns are addressed by the authors, I still have some main concerns about the motivation and writing of this work. Therefore, I raise my rating from 3 to 4. Thank you. Best regards --- Reply to Comment 1.1.1: Comment: We are grateful for your engagement with our rebuttal and for the opportunity to address your concerns. We understand that despite our efforts to respond to your initial questions, there are still aspects of our work that have not fully met your expectations. Your feedback is invaluable to us, and we are committed to enhancing the quality of our research. We acknowledge that our motivation and the clarity of our writing may not have been as compelling as they should be. To address these issues, we will take the following steps in our revised paper: - Clarification of Motivation: We will revisit the introduction and conclusion sections to more explicitly articulate the significance and novelty of our research. We have included additional context to underscore the importance of our work within the broader scope of co-speech gesture synthesis. For the motivation of long-term and real-time motion synthesis, this approach is critical for human-computer interactive scenarios, where real-time performance is crucial, and some conversations require extended periods of speech. In this context, we integrated VQVAE with Mamba in our framework, effectively addressing the shortcoming of diffusion models (e.g., slow inference speed) through VQVAE's compression expression ability and Mamba's linear computational complexity. - Enhanced Writing Quality: We will undertake a thorough review of the paper to ensure that the writing is clear, concise, and engaging. We will also seek the assistance of professional editors to refine our language and presentation. - Addressing Specific Concerns: We have carefully considered each of your points and provided targeted feedback. We believe these changes have strengthened the overall coherence and persuasiveness of our arguments. If you have any questions, please feel free to raise them. We hope that these revisions have adequately addressed your concerns and have brought our manuscript closer to acceptance. We are open to further suggestions and are willing to make additional revisions as needed to meet the standards of NeurIPS. Thank you once again for your constructive feedback.	Summary: The paper explores the selective scan mechanism for gesture generation. First, it trains VQVAE to reconstruct faces and body parts using discrete latent space. Then, It uses local and global scanning mechanisms to improve the latent representations of various body parts for the purpose of gesture generation. Strengths: 1) The paper explores Mamba for gesture generation. 2) The objective function includes features related to gesture acceleration and velocity, which are important to capture. 3) The paper contains a user study where even the ground truths are evaluated. I think that is the right direction for gesture-related studies. Weaknesses: 1) The related work contains many acronyms not introduced or known to a wide community (e.g., VQVAEs, HiPPO, and LSSL). 2) How the authors used Mamba in their work is unclear. The explanation in section 3.1 is vague. 3) Audio feature extractor is not strong. The authors should have explored stronger speech representations that can be used to improve speech prosody. 4) The technical contributions of the paper might be limited to implementing Mamba architecture for gesture analysis. Technical Quality: 3 Clarity: 2 Questions for Authors: 1) Can you please dedicate section 3.1 to how you used Selective State Spaces for gesture generation? The explanation given is very generic. The reader can not follow this generic description without prior knowledge of Mamba. 2) How are the motion priors (the codebook) initialized? Also, how often is this codebook updated? 3) Why are the results reported in "EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling" differs from what you report in Table 1. FGD in the EMAG paper is 5.512, while in your paper, they are reported as 5.423. If you are using two different backbone models (and hence features), then how fair is this comparison if you fine-tune your models with the best parameters? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: One limitation related to gesture diversity across speakers and cultures is not addressed. The mentioned limitations are mainly related to technical details, which I think are already being addressed using Transformer and Diffusion based models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: A1: Related work and Preliminaries section We apologize for any inconvenience caused to your reading due to the generic explanation. We will provide a more detailed explanation for related work (e.g., VQVAE, HiPPO, LSSL) in section 3.1, which acts as preliminaries of our work. For the use of Mamba, we mainly discuss its application in Section 3.3, where we combine Mamba with attention mechanisms and learnable queries to model motion sequences using a local-to-global scanning strategy. In the revised version, we will supplement more details, including pseudo-code, to provide a clearer explanation of our method. A2: Codebook Thanks for your questions. We utilize uniform initialization for the codebook, with a numerical distribution range of [-1/codebook_size, 1/codebook_size), and conduct the first training stage on the training dataset. The codebook is solely updated during the first stage, and in the second stage of training for the speech-to-gesture mapping, the codebook remains frozen. A3: Audio representation Thanks for your valuable comments. Enhancing the audio representation can potentially improve performance, but it might also result in unfair comparisons with previous methods. In our revised version, we will include additional experiments using various audio feature extractors (e.g., wav2vec2, Whisper). A4: Technical contribution Our work did not simply involve implementing Mamba for gesture synthesis. We found that directly applying Mamba would cause serious shaking issues. Therefore, we explored incorporating motion priors using VQVAEs and individual learnable queries for different parts of the body. Additionally, the direct application of Mamba also has the issue that the limb movements of different body parts tend to be average. Therefore, we refined the design of spatial and temporal modeling in latent spaces by proposing a local-to-global modeling strategy and incorporating them with attention and selective scan mechanism into the design of our framework. We will refine our statements and make our framework clearer (e.g., adding pseudo-code) in our revised version. A5: Results of EMAGE For the difference in results, we compared the v3 version released by EMAGE on arXiv, which the authors of EMAGE claimed to be the CVPR 2024 camera-ready version. We checked again and found that they later updated multiple versions of arXiv. The latest version, v5, does indeed report 5.512. We will incorporate their updated results in our revised version. To ensure fairness in the comparison, we maintained consistency with them in the experimental setting. We have an advantage over EMAGE in all metrics, with significant performance improvements in BC, MSE, and LVD metrics, demonstrating enhancements of 16.72%, 36.35%, and 13.99%. A6: Limitation section Thanks for your valuable comments. We will add this discussion to the limitations section. --- Rebuttal Comment 1.1: Comment: Thank you very much for the clarification and the response, which is much appreciated, especially the point related to the comparison of the result. I have raised my ratings from 5 to 6. --- Reply to Comment 1.1.1: Comment: We are deeply grateful for your thoughtful reconsideration of our manuscript and the increase in your rating (weak accept). Your feedback is instrumental in helping us refine our work, and we are pleased that our clarifications and responses have been well-received. Thank you once again for your constructive feedback and for the opportunity to improve our manuscript based on your valuable insights.	Summary: This paper focuses on the problem of co-speech gesture generation, particularly aiming to address the challenges of jittery movements, long motion sequences, and holistic gesture generation (including both face and body movements). To tackle these issues, the authors present a new method that combines diffusion models with state space models and incorporates both local and global scans. The proposed method has been evaluated on the BeatX dataset. Strengths: This paper introduces novel methodological aspects for co-speech feature generation. For instance, combining diffusion models with state space models and local and global scan approaches has not been explored before, to the best of my knowledge. The experimental results clearly demonstrate the effectiveness of the proposed approach. Moreover, the paper presents both quantitative results and qualitative results through user studies, providing a complete and strong evaluation. Additionally, the proposed approach is faster compared to the state-of-the-art methods, as presented in the appendix. However, there are some weaknesses that need to be addressed. Please see below for detailed comments. Weaknesses: Major Comment: The main weakness of the paper is its limited evaluation. While the paper excels in generating facial, body, or holistic gestures, evaluating on only one dataset is substandard. To be competitive, the authors should consider additional standard benchmarks such as those used in the GENEA challenge [1]. There is also a long list of papers using diffusion models for co-speech gesture synthesis [1], none of which are mentioned in the paper. While I appreciate the authors' work and acknowledge the scarcity of datasets with both face and body motions, the proposed approach handles face and body separately. Therefore, at least body motion generation performance could be compared on existing datasets to demonstrate generalisable benefits. Another weakness is the insufficient details regarding long-term motion generation, which is claimed as a key contribution. The paper lacks sufficient details on the length of the sequences considered in training and evaluation. Moreover, there is no comparison to support this claim. Minor Comment: The lower part of Figure 2 is helpful for understanding the workflow. However, including pseudo-code would be beneficial to increase the reproducibility of this method. [1] https://genea-workshop.github.io/2023/challenge/ Technical Quality: 3 Clarity: 3 Questions for Authors: Why is only one dataset considered? Please explain the rationale behind evaluating the method on only one dataset. How are these results generalisable and competitive given this limitation? Generalizability and Competitiveness: How do you ensure that the results obtained from the BeatX dataset can be generalized to other scenarios? Discuss any potential limitations in terms of generalisability and how your method addresses them. Training and Evaluation Sequence Lengths: Please elaborate on the length of the motion sequences considered during training and evaluation. Impact on Co-Speech Gesture Generation Performance: Discuss how the length of motion sequences impacts the performance of co-speech gesture generation. What are the implications of using different sequence lengths, and how does your method handle long-term motion generation effectively? Including comparisons or experiments that highlight these aspects would be beneficial. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have satisfactorily discussed the limitations and broader impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: A1: Reasons for choosing BEATX Thanks for your kind advice. Our work focuses on holistic co-speech gesture generation. We chose to use this dataset because it includes global movements and the smplx sequences of the entire body (e.g., face, upper body, lower body, and hands). However, most current datasets (such as TED, TED-X, BEAT, GENEA) only include movements of specific parts of the body (e.g., upper body), not the entire body. A2: Generalisable benefits of our method To further validate the generalizable benefits of our method on upper body and hand movements, we used the BEAT dataset, which includes upper body and hand movements. Due to time constraints, we only conducted preliminary experiment. The results are shown in the table. Our method demonstrates a significant improvement compared to the CaMN (baseline), which also validates the effectiveness of our approach. More comprehensive experiments with detailed information will be included in the revised version. \| Method \| FGD$\downarrow$ \| SRGR$\uparrow$ \| BeatAlign$\uparrow$ \| \|--------------------------\|------------------\|-----------------\|--------\| \| CaMN (baseline) \| 123.7 \| 0.239 \| 0.783 \| \| Ours \| 51.3 \| 0.256 \| 0.852 \| A3: Related diffusion-based work Thanks for pointing out these wonderful works. For the diffusion-based model, we have compared our method with DiffuseStyleGesture in Table 1 and Table 4, which is a similar method to DiffuseStyleGesture+ (Reproducibility Award of GENEA workshop). Compare to DiffuseStyleGesture, DiffuseStyleGesture+ primarily considers the text modality as an additional input and utilizes channel concatenation to merge the text feature with the audio feature. Another leading solution (Diffusion-based co-speech gesture generation using joint text and audio representation) also incorporates the text modality as an additional input and employs contrastive learning to enhance the features. We will cite and further analyze the differences using all of these methods in our revised version. A4: The length of the motion sequences considered during training and evaluation Thanks for your professional advice. However, what we want to declare in our paper is that our method can generate long sequences with low latency (with analysis in Appendix A.2). For the length of the motion sequences considered during training and evaluation, we adopt a segmented modeling strategy. This strategy involves dividing the target sequence into multiple segments, each 64 frames in length, for processing. Therefore, our method can effectively handle long-term motion generation. The length of the motion sequence does not have a significant impact on performance because almost all segmented sequences share the same length and are processed in a similar manner. A5: Pseudo-code Thanks for your advice. We will include pseudo-code in our revised version. We will also make our code open-sourced if the paper get accepted. --- Rebuttal Comment 1.1: Title: reply: Rebuttal by Authors Comment: Thank you for the additional information. After reviewing all the feedback and responses, I have decided to maintain my initial score. While the paper introduces some novel methodological aspects, the primary concerns remain: the lack of experimental results and the need for further clarification, particularly regarding the motivation for long-term and real-time motion synthesis. --- Reply to Comment 1.1.1: Comment: We would like to extend our sincere gratitude for the time and effort you have invested in reviewing our rebuttal. We appreciate the detailed feedback and understand your concerns regarding the lack of experimental results and the need for further clarification on the motivation for long-term and real-time motion synthesis. For experimental results, we have conducted comprehensive experiments on holistic co-speech gesture synthesis (the main focus of our method), including quantitative results, qualitative results, and efficiency analysis in our initial version. In response to your feedback, we have incorporated supplementary experiment results on upper-body co-speech gesture synthesis in our rebuttal. The enhancement in our method is substantial, demonstrating improved overall performance compared to the baseline. Future experiments will focus on visualization. For the motivation of long-term and real-time motion synthesis, we expand our discussion on the motivation behind the need for long-term and real-time motion synthesis. This approach is critical for human-computer interactive scenarios, where real-time performance is crucial, and some conversations require extended periods of speech. In this context, we integrated VQVAE with Mamba in our framework, effectively addressing the shortcoming of diffusion models (e.g., slow inference speed) through VQVAE's compression capability and Mamba's linear computational complexity. As demonstrated in Appendix A.2, our method can generate sequences with extremely low latency, which is beneficial for the application of co-speech gestures. We look forward to your further feedback and are grateful for the opportunity to refine our work based on your valuable insights. Thank you once again for your consideration.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Transformers on Markov data: Constant depth suffices	Accept (poster)	Summary: This paper attempts to provide a possible explanation the capability of transformer architecture for efficient next token prediction of (stationary) $k$th order Markov data. The main result is a constructive proof that a $3$-layer transformer with a single head per layer can emulate conditional $k$ grams, and it necessarily uses the non-linearities in the transformer architecture like layer-normalization. The paper also provides lower bounds on representational powers of the transformers for Markov data. Strengths: (1) The results presented in the paper serve as a crucial step towards understanding the prediction capabilities of transformers for $k$th order Markov processes with finite state spaces (of small to moderate sizes). (2) The paper provides a novel explanation of the (possible) way in which a layer-normalization (LN) might be used in modelling $k$-grams. To my knowledge, the non-linearities are discarded in typical prior work(s). (3) The lower bound is provided for single layer transformer (Theorem 5) is a strong evidence signifying of importance of multiple layers. (4) The paper is well-written and easy to follow. Weaknesses: (1) In spite of solving a general and interesting problem of generative modelling using Markov structure, the evaluation is very limited. For example, I believe that the paper does not have empirical validation of $k$-gram modelling by the transformer when the size of the state space $S$ is large. (2) The paper does not address or comment how the training data can affect the performance during testing. This step of training might be very significant when the size $S$ of the state space is high. It would be nice to see how the mechanism works when $S$ is of order of tens or a couple hundreds (my understanding is that Figure 2 is only for binary data, $S=2$). Typo in Architecture 1: $\textbf{x}_i^{(1)} = \texttt{Emb}(x_n)$ instead of $\textbf{x}_n^{(1)} = \texttt{Emb}(x_n)$ Technical Quality: 2 Clarity: 3 Questions for Authors: (1) I believe that the state space is binary in your evaluations $S=2$. Can you please confirm this? (2) In Table 3 of Appendix F, the embedding dimension is mentioned to be chosen as per grid search over \{16, 32, 64\}. Should this be not the same as $d=6S+3$ as in the proof of Theorem 4 in Appendix C.3? In this case, if $S=2$, then $d=15$ is enforced. Can you provide the explanation for grid search? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: (1) Although the result sheds light on the abilities of a transformer architecture in modelling conditional $k$-grams, this is only optimal for time-homogeneous stationary Markov data. Hence, the paper does not directly have a huge impact towards understanding how transformers perform on general stochastic sequences (like language tokens, or frames in a video). (2) The constructions in the paper work only for small state spaces (tens to a few hundreds), as the embedding dimension used in the proofs scales with its size. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and constructive criticism. We have fixed the typo pointed out. Below we address the main weaknesses discussed in the review: ### [W1,W2] Experiments for and learning dynamics under larger state spaces? In the attached rebuttal pdf, we ran experiments for Markov models on state spaces of size $\|S\| = 5$. Here too, we show that the main phenomenon discussed in the paper ($2/3$-layer $1$-head transformers learn $k$-th order Markov processes for $k > 1$) holds true. In particular, we show that a $2$-layer $1$-head transformer learns order-$3$ Markov processes. We were unable to run this experiment for $k=4$ keeping the state space size fixed since the model longer to converge than our computational allowance would permit. When the state space becomes even larger, say on the order of tens or a couple of hundreds, it becomes much harder to train the models on reasonably large values of $k$ within the computational budget in our availability. ___ ### [Q1] Binary state space You are correct to note that we run our experiments on binary data (to be able to scale to the largest possible values of $k$ on a computational budget). ### [Q2] Precise dependency of embedding dimension The experiments we run are more in line with the case of the standard transformer architecture, which corresponds to Theorem 4, where the embedding dimension scales as $O(S)$ (where the constant is implicit). In the common rebuttal, we discuss how to theoretically reduce the dependency on the embedding dimension at the cost of a small increase in the loss incurred by the model. On a separate note, there are usually a few reasons to expect that there would be some differences in theoretical and empirical findings especially when it comes to the precise dependency on quantities like the embedding dimension, etc. Empirically, transformers are trained from a random initialization and in a way where the model may be more / less efficient at realizing certain mechanisms compared to a theoretical construction. We would argue that it is in fact promising that the embedding dimension is of a similar order as what is predicted in our theorems, as it is often the case that the theoretical and empirical findings are off by several orders of magnitude. ### [L1] Time inhomogeneous processes This is a great question, and something we hope to study as the next step. Time homogeneity of the data model (such as with $k$-th order Markov processes) implies that it is possible to write down the estimate of the next-symbol conditional distribution using an unweighted empirical estimate. When the data process becomes time inhomogeneous, the statistical estimators would look more akin to weighted conditional empirical estimates (under assumptions on the nature of time inhomogeneity). The constructions we consider in the paper do have the ability to translate to these kinds of settings as well. However, to understand the limits of how transformers do this is a great direction for future research, and something we hope to explore going forward. It is plausible that transformers need more depth to be able to capture these kinds of processes, depending on the nature of the time inhomogeneity. ### [L2] Constructions only apply for small state spaces In commercial (for which metadata is available) and larger open source models, the vocabulary size has usually been in the range 30K-250K. However, the embedding dimensions have been of the scale of 1000-20,000, which is one order of magnitude smaller. It is worth pointing out that our constructions, while optimized for the dependency on the depth and the number of heads, can also be improved on the depth dependency. In the common rebuttal, we discuss how to improve the dependency of the embedding dimension from $O(S)$ to $O(\log (S)/\epsilon^2)$ where $\epsilon$ is a parameter which now captures an approximation error. Thus, even when the state space is larger, for practical ranges of $\epsilon$, the dimensionality implied by this result should be much smaller than $O(S)$. --- Rebuttal Comment 1.1: Title: Discussion period ending soon: call for response Comment: Dear Reviewer TFSx, We sincerely appreciate the time you have taken to provide valuable feedback for our work. As we are getting closer to the end of the discussion period, could you let us know if our responses above have adequately addressed your concerns? We remain at your disposal for any further questions. If you agree that our responses to your reviews have addressed the concerns you listed, we kindly ask that you consider whether raising your score would more accurately reflect your updated evaluation of our paper. Thank you again for your time and thoughtful comments! Sincerely, The Authors --- Rebuttal Comment 1.2: Comment: Thank you for the comments and clarifications. I shall retain my score as the applicability of such a theoretical statement is rather limited.	Summary: This paper studies the learning process and representational power of transformers on the data sequence generated by k-th order Markov processes (or k-gram). Theoretically, this paper proved that (1) an attention-only transformer with $O(\log k)$ layers with one head for each layer can represent the conditional (in-context) $k$-gram. (2) Enhanced with MLP and layer normalization a 3-layer transformer can express the conditional $k$-gram task. They also complement their results with a lower bound for attention-only transformers/a 1-layer transformer with MLP and layer-norm with some assumptions. Detailed empirical validations are conducted to corroborate the construction (1). Strengths: 1. This paper improved the construction using $k$-head, 2-layer in previous papers on the in-context $k$-gram data by constructing a $\log k$-depth, 1-head transformer. The proof technique is based on the binary-tree-like aggregation procedure of the previous $k$ tokens information, which greatly improves the memory cost of the transformer parameters. Empirically, the construction can be partially validated by experiments showing that a 3-layer transformer can learn an in-context 8-gram. Also, a lower bound result is included (with some empirically validated assumption) to make this bound tight. 2. The construction of a 3-layer transformer with the MLP and LayerNorm is quite novel. The technique is based on the unique ternary representation of integers. The ternary representation is to maintain the uniqueness of $v_i$ after normalizing the embedding vector. The role of the non-linearity in the transformer in this $k$-gram mechanism can be crucial for improving the representation power of the model, and this result can serve as an implication. Weaknesses: The two main construction results improve the previous expressivity results. However, the second theorem (Theorem 4) still lacks empirical evidence showing that the constructed solution is the minimizer that Adam/GD converges to. The previous experiments only show that $\log k$-depth transformer can somehow be learned. It is important to figure out whether the solution is only for construction or it can be actually learned in some way. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Is it possible to add a set of experiments for the 3-layer transformer with LayerNorm and MLP trained on the k-gram model when $k>8$? Or showing that it is hard to obtain from optimization? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive comments, and constructive criticism. Below we address the main weakness pointed out, as well as the question asking about experiments for $k > 8$. ### [W1] Evidence that ADAM/GD converges to theoretical construction This is a great question. In Fig. 3 of the rebuttal pdf attached, we plot the attention patterns learnt and observe some sort of approximately exponential decay. The theoretical constructions also match this kind of exponential decay (as seen from the attention pattern in eq. (50) in the proof of Theorem 4). While this is promising, there is certainly more to explore about how and what the transformer learns. In particular, it is interesting to see how these attention patterns evolve over the course of learning before converging at these kinds of exponential patterns. This may reveal some more about how transformers learn, and get us one step closer to a theoretical understanding of the learning dynamics of transformers when exposed to Markovian data. ### [Q1] Experiments for $k > 8$? This was also pointed out by other reviewers, and is indeed an important point. We would be happy to include something which provides more clarification around what happens when $k=8$, however in the form we currently consider, it is not within our computational budget to scale this to $k=9,10$ or anything higher. A longer discussion as to why this kind of computational “phase transition” occurs is provided in the common rebuttal above. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed clarifications. I will maintain my score.	Summary: This paper studies the ability of transformers to learn $k^{th}$ order Markov chains. They first conduct experiments showing that transformers with 2 layers and 1 head can learn Markov chains of up to order $k=4$. Similarly, with 3 layers, they can learn Markov chains of order $k=8$. Based on these observations, they show theoretical results about the representation power. Specifically, they present novel constructions to show that attention-only transformers with 1 head can learn $k^{th}$ order Markov chains with $\log_2(k+1)$ layers. On the other hand, 2-layer attention-only transformers require $k$ heads. These results show that increasing the depth is much more beneficial than increasing the number of heads. Next, the authors show that for the full transformer model, a constant depth of 3 suffices to learn $k^{th}$ order Markov chains with an embedding dimension of the order of the vocabulary size. This result reveals the benefit of non-linearities arising from layer normalization. They also present a lower bound on the depth requirement for 1-head attention-only transformers under some assumptions on the attention pattern learned by the model. Strengths: - Overall, the paper is well-written and easy to follow. The authors present nice intuitions about the constructions and the theoretical results. - The results in the paper showcase the benefit of depth over the number of attention heads in transformers, which contributes to our understanding of transformers. The paper also provides insights into the benefit of non-linearities that arise from layer normalization in transformers, which is interesting, and helps compare attention-only transformers with standard transformers. Weaknesses: - The main weakness is that some of the statements made by the authors about their results/observations contradicting prior work lack clarity and are not supported by a thorough discussion of the potential reasons for the differences. I think these statements should be accompanied by more context. Please see the Questions section for queries/suggestions regarding this. - There are some minor typos and grammatical errors that should be corrected. Please see the next section for details. Technical Quality: 3 Clarity: 3 Questions for Authors: - The statement in lines 28-31 needs further clarification: - It’s not clear what range of $k$ is considered in prior works. It seems that the observations are not exactly in contradiction to prior work, since a) 2-layer 1-head transformers can’t seem to be able to learn Markov chains of order $k>4$, and b) the result in Table 1 and Theorem 2 states that for order $k$, the number of heads for 2 layers is $k$, at least for attention-only transformers. Table 1 is missing a column for the number of heads needed for standard transformers with 2 layers. - Has prior work considered training transformers with 3 layers? How does the statement after (ii) contradict prior observations? - Regarding the experiments, particularly Fig. 2: - In Fig. 2(a), the test loss gap seems a bit high. Can the authors include results for $k=1$ in this plot for a comparison? - Can the authors share results when training the 3-layer 1-head transformer on Markov chains of higher order $k>8$? Is there a reason why higher-order Markov chains were not considered for the experiment in Fig. 2(b)? I am curious if there is a gap between the theoretical result about the representation power (which holds for $k>8$), and the training dynamics. - Other suggestions: - In the abstract, it would be good to emphasize the word ‘constant’ in line 14. - I suggest using $V$ instead of $S$ for the vocabulary size. - It would be good to include some discussion on whether similar construction techniques (for Theorem 3) have been used in the literature. - I suggest including some discussion on related work on the role of softmax attention, the benefit of layer normalization techniques, and the benefit of depth in neural networks. - Typos/grammatical errors: - Missing citation in line 23. - Extra ‘studies’ in line 27. Extra ‘them’ in line 132. Extra ‘how’ in line 272. - Please check the phrasing in lines 220-222. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments. ### [W1] Thorough discussion for why contradictions arise in past work As discussed in the common rebuttal, the main reason for the differences in observations compared to [1] and [2] is the fact that the models were not trained for sufficiently long to observe the increase in test loss. Please refer to Fig. 2 in the attached document where we plot this explicitly on a random $5$-state order-$3$ Markov chain. In the beginning phase of the training (iterations 1-200), the test-loss curves plateau around the bigram performance, but as we continue training, the loss continues decaying slowly. We believe that this difference only appears because the model was not trained for sufficiently long, keeping all other variables unchanged. We will include this plot in the paper. ___ ### [Q1] Range of $k$ considered in previous work The contradiction we mention in the $2$-layer case is empirical. Prior work such as [1],[2] train $2$-layer $k$-head models on $k$-th order Markov chains for $k=1,2,3,4$. However, we are lead to believe that the $2$-layer $1$-head model was not trained for sufficiently many iterations to learn $k=2,3,4$. Fig. 2 in the attached material showcases this difference when a $2$-layer $1$-head transformer is trained on $k$-th order Markov chains on $5$ states for $k=3$. The initial phase (first ~100 iterations), the model stagnates in performance at the level of the best bigram model. When we continue training, the loss of the model starts decreasing further and breaks past this bigram phase to learn higher order conditional $k$-gram models. Finally, it’s worth noting: while we do not have a column for the $2$ layer standard transformer in Table $1$, we believe that the main $3$-layer construction can also work for $2$-layer transformers with $2$ heads in the first layer. This is because in the $3$-layer construction we consider, the first two layers of the model essentially do not interact with each other (and operate on different parts of the embedding space), they can be implemented across two different heads of the first layer. We will add this into the paper to clarify that there exists a contradiction even theoretically. Formally: For $2$-layer transformers with $2$ heads, prior work argues that these models can learn up to $k \le 2$. Our work shows that these models can represent much higher values of $k$ theoretically (via an extension of Theorem $4$). ### [Q2] Clarifications regarding Fig. 2 We have included the plot for $k=1$ in Fig. 1 of the attached document in the paper (this experiment trains $k$-th order Markov chains on $k$-head $2$-layer transformers). We will include this plot in the paper as well. There is a gap in the test loss which decreases when we move to $3$-layers and $1$-head. The reasons for this small but non-negligible test loss gap are unclear - while our theoretical construction works for the $3$ layer case, it may be possible that with $2$ layers and $1$ head transformers are unable to exactly represent the conditional $k$-gram model. It may also be possible that ADAM/GD are unable to find this solution, even if it is exactly representable. Understanding the limits of how $2$-layer $1$-head models learn is an interesting question to look into. On a separate note, It is worth mentioning that the best achievable test loss will increase as we keep the sequence length fixed and increase $k$. In the learning setup we consider, the training and test data do not come from the same Markov process, since the transformer is trained on data from randomly chosen Markov processes and tested on data from another randomly chosen Markov process. So the transformer has to do some form of in-context learning. When the sequence length is fixed and $k$ increases, the in-context learning problem becomes harder. The model has to learn $S^k$ conditional distributions (for each possible prefix of $k$ states). So as $k$ grows, the model has less data for each of these distributions which makes the estimation task harder. This is a major reason for why the $3$-layer transformer’s loss increases a little bit when dealing with $k=8$. We were computationally bottlenecked to train on sequence lengths of up to $500$ and at this scale, the model has just about enough data to learn each of the $256$ conditional distributions. Fixing the sequence length as $500$, the models should not be expected to have small test loss as $k$ grows to $9$ and beyond. The reason we do not train on $k > 8$ is discussed in the common rebuttal. The sequence length to consider would have to be at least around $1000$ or longer for $k=9$ and $2000$ for $k=10$. The model also takes longer to train, and we were unable to train for long enough for the model to reach convergence. In the future, we hope to be able to test out the results in this paper on longer sequence lengths (thereby allowing larger values of $k$) while keeping the depth of the model and the number of heads fixed. ### [Q3] Additional suggestions We shall plan to incorporate the notational changes and fix the typos as suggested by the reviewer. With the additional page in the final version, we shall also include a discussion on related work on the role of softmax attention and depth which were present in existing work. Finally, we will also include a discussion on related works which consider techniques similar to the composition technique we considered in Theorem $3$.   ___ ### References [1] Makkuva, Ashok Vardhan, et al. "Attention with markov: A framework for principled analysis of transformers via markov chains." arXiv preprint arXiv:2402.04161. [2] Nichani, E., Damian, A. and Lee, J.D., 2024. How transformers learn causal structure with gradient descent. arXiv preprint arXiv:2402.14735. [3] Edelman, Benjamin L., et al. "The evolution of statistical induction heads: In-context learning markov chains." arXiv preprint arXiv:2402.11004. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal and additional figures. The clarifications are helpful and I am happy to maintain my score.	Summary: This paper investigates the representation capability of transformer with different number of layers or heads when learning k-th order Markov Process. They provide theoretical results demonstrating that attention-only transformers with O(log2(k)) layers can represent the in-context conditional k-th order Markov Process. This conclusion is supported by empirical results and is novel compared with previous training dynamics results. Adding LayerNorm, they prove that standard transformers with just three single-head layers can represent arbitrary order Markov processes. The paper also presents lower bounds on the size of transformers needed to represent certain attention patterns. Strengths: 1. The writting is good and easy to follow. And the proof is relatively solid. Most of the intuition is clear, like using hierarchical way to construct the intermediate logits which include information of multiple tokens. 2. The main contribution of this paper is that the authors prove there exists some transformers which have fewer head and logarithm-limited layers that can learn the determined k-th Markov Process. And these results are supported by the empirical experiments, and it's more related to the real-world case. (Like we don't need too many heads) Weaknesses: 1. This work analyzes the representation capability of transformer of learning Markov Process, however, there is no clear statement of error bounds in the key theorem like Theorem 4 (although there are some helpful discussions like Remark 2). And this may related to some questions like Questions-2. 2. The intuition of the proof of Theorem 5 is still not very clear, maybe adding more examples will be helpful, like how LayerNorm works. 3. Like mentioned in limitation parts, it's more focused on representation capability rather than training dynamics. The latter one may be more difficult for analyzing. More questions in Questions-1. 4. In Figure 2, the authors should better add k=8 and k=16 to further support the conclusion. 5. There is a citation typo in Line 23. 6. Maybe you can explicitly say that learn the conditional k-gram model is equivalent to letting transformer generate the same logits as that model (i.e., Eq.1), otherwise the reviewer maybe confused by the conclusion in the main theorems. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. This work is more related to representation capability of transformer, rather than training dynamics, which are the focus of related works [6], [7] mentioned in the papers (Line 114 - 133), so is it possible that the construction of k heads is more theoretically friendly than the that in your paper for analyzing dynamics like gradient flow? And what's the possible way for training dynamics under your structure? 2. It seems that you still need $\Omega(k)$ bit precision for learning $k$-th order Markov Process, but in real-world case we don't need such a high precision. Is this necessary? and is it required by the constant layers transformer setting and not required by the $O(log(k))$ attention-only setting? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors have mentioned their limitations in the end of paper, like they just focus on representation capability rather than training dynamics, which maybe the futural direction. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed comments. Below we address the main questions and weaknesses: ### [W1] Error bounds We are happy to include a longer discussion of error bounds in the paper. To expand, suppose all the weights in the transformer model are upper bounded by $1$. When the transformer is implemented with a precision of $m$ bits, the new attention weights $\widehat{\text{att}}_{n,i}$ satisfy, $$ \frac{1+e^{-\varepsilon}}{1+e^{\varepsilon}} \le \frac{\widehat{\text{att}}\_{n,i}}{\text{att}_{n,i}} \le \frac{1+e^{\varepsilon}}{1+e^{-\varepsilon}}; \quad \text{where, } \varepsilon \triangleq d \cdot 2^{-m+1} $$ This uses the fact that by truncating vectors to $m$ bits, the approximation error in the inner product of the inner product of two $d$-dimensional vectors. i.e. $\varepsilon$, is at most $d \cdot 2^{-m+1}$. In particular, simplifying assuming $\varepsilon$ is small, we get a multiplicative error of $[ 1 - \Omega (\varepsilon), 1 + \Omega (\varepsilon)]$. A similar analysis works for layer-normalization and results in a similar error scaling. Thus, within a single attention layer, we compute vectors which are within $[1-\varepsilon', 1+\varepsilon']$ where $\varepsilon' = T \cdot 2^{-m-\log(d)}$. Iterating across $L$ layers, the error incurred scales as a multiplicative $1 \pm 2^{-m-L\log(dT)}$. Thus, when the bit-complexity scales as $m \gg L \log (dT)$, the approximation error is a multiplicative polynomially small constant in $T$. There is only one caveat in case of the constant depth transformer, the weights of the transformer in each of the $3$ layers are upper bounded by $O(2^{k})$, rather than $1$. Thus, the bit precision must now scale as $k + O(\log(ST))$. We chose to de-emphasize these details in the paper to avoid unnecessarily complicated theorem statements, and to avoid detracting from the (arguably more interesting) other phenomena discussed in the paper. ### [W2] Intuition behind Theorem 5 We have rewritten this section to make the intuitions more clear. We have added a new “proof sketch” section, clarified what the $k$-th order induction head does in this context, and emphasized how layernorm allows the transformer realize these $k$-th order induction heads. In short, layernorm allows changing attention from $\text{att}\_{n,i} \propto \exp (\langle k_i, q_n \rangle)$ to instead look like $\text{att}\_{n,i} \propto \exp ( - \\| \hat{k}\_i - \hat{q}\_n \\|\_2^2)$ where $\hat{k}\_i$ and $\hat{q}\_n$ are the $L_2$-normalized key and query vectors. This is nice, because if we can realize the key vectors $k_i = \sum_{j=1}^k 2^j e_{x_{i-j}}$ and the query vectors as $q_n = \sum_{j=0}^{k-1} 2^j e_{x_{n-j}}$, the attention is maximized if and only if $\hat{k}_i = \hat{q}_n$, which we can prove occurs only when $x\_{i-1} = x\_n, \cdots, x\_{i-k} = x\_{n-k+1}$ (using the fact that the binary representation of a number is unique). This realizes a $k$-th order induction head. ### [W3] Representation vs Training dynamics In this paper, we focus on the representation capacity, which itself turns out to be a complicated phenomenon. On the optimization side, there are indeed many questions open. Unfortunately, even in the simplest settings (1 layer transformers), the analysis of gradient descent on transformers trained on Markovian data is incredibly complex. The rigorous analysis in [Makkuva et al] is over 70 pages long for this case. Extending to higher depth, is incredibly non-trivial. [Lee et al] discusses this case, however, their analysis is only amenable under very strong assumptions i.e., the “reduced model” of transformers. Even with these assumptions, the analysis is over 60 pages long. We hope to convince the reviewer that while training dynamics is indeed an interesting and timely question, it certainly requires a dedicated effort to be able to analyze, and would constitute a separate work of its own. ### [W4] Experiments on higher values of $k$ Discussed in the common rebuttal. ___ ### [Q1] Learning dynamics At a high level, yes, we believe that the k-head mechanism discussed in prior works may be more amenable for analysis. When the depth is constant, the transformer is forced to be more creative in the ways it learns the $k$-th order chain, and this leads to more intricate mechanisms. However, unless simplifying assumptions are imposed, we believe that understanding learning dynamics falls into the realm where current tools in optimization theory are not strong enough to be able to answer. ### [Q2] $\Omega(k)$ bit precision for learning $k$-th order Markov processes One may consider an extension of $k$-th order Markov processes to the case where the conditional distribution depends on a sparse subset (of size $p$) of the previous symbols observed. The conditional distribution has a low degree of dependency on the past, even though it is not the immediately preceding $p$ symbols. While these kinds of processes are special case of $k$-th order Markov processes, modeling them as such may necessitate a very large value of $k \gg p$, since the there may be dependencies on symbols which appeared long ago. Our constant depth constructions work even in this setting, with the bit-complexity scaling with $p$ and not with $k \gg p$. This analysis shows that transformers don’t require high bit-precision as long as the degree of dependency of the conditional distribution on the past is small. In practice, this is often the case. On a separate note, our constant depth constructions require $\Omega(k)$ bit complexity, but the embeddings are now $O(S)$ dimensional, rather than being $\Theta(Sk)$ dimensional in the attention-only setting. Thus, the total number of bits in each embedding vector is the same in both cases. We believe that it is possible to reduce the bit complexity of the constant depth construction at the cost of making the embeddings $\Theta (Sk)$ dimensional. --- Rebuttal Comment 1.1: Title: Discussion period ending soon: call for response Comment: Dear Reviewer YGsY, We sincerely appreciate the time you have taken to provide valuable feedback for our work. As we are getting closer to the end of the discussion period, could you let us know if our responses above have adequately addressed your concerns? We remain at your disposal for any further questions. If you agree that our responses to your reviews have addressed the concerns you listed, we kindly ask that you consider whether raising your score would more accurately reflect your updated evaluation of our paper. Thank you again for your time and thoughtful comments! Sincerely, The Authors	Rebuttal 1: Rebuttal: ## Common Rebuttal We thank all the reviewers for taking the time to go through our paper and suggest constructive criticism. Please find attached a pdf containing additional plots. Below we address some common points raised by multiple reviewers. ### Long training reveals contradictions with prior work We point the reviewers to Fig. 2 in the rebuttal pdf. Here, we clearly observe the stages of learning for a $2$-layer transformer. Initially (in the first ~200 iterations) the model stagnates and this is around the loss of the best bigram model (which is what previous work claimed these models could learn). As we continue training, we see that the loss starts decaying but over a much slower time-scale. The model indeed requires around 20K iterations to get close to the optimal loss. The closest related phenomenon to why the model performance appears to improves very slowly is that of grokking: zooming into any interval, it may appear that the model has converged, however this only appears to be the case as the parameters of the model may be in a region where the loss landscape is very flat. In the rebuttal pdf, we also point the reviewers to Fig. 1. The authors of [1] also present similar results, but it appears that their experiments requires more samples / optimization cycles than in our paper. Here, we observe that $4$-head transformers with $2$ layers need around 3K iterations to learn $k$-th order Markov chains for $k=4$. In contrast, according to Fig. 2 in the main paper, $2$-layer transformers with $1$ head trained on the same data processes require around 7K iterations. ### Running our experiments on higher values of $k$ Running our experiments in Fig. 2 on higher values of $k$ is certainly important. However, there is a computational tradeoff which we were not able to navigate when $k$ grows to be larger than $8$. We were computationally bottlenecked to train transformers on sequence lengths of up to around $500 \approx 2^9$ for the number of iterations required for the test-loss to begin to converge. At this scale, the model has “just” enough data to be approximately able to estimate order-8 Markov chains in-context. Note that the model has to approximately estimate $2^8 \approx 250$ conditional distributions to learn order-8 Markov chains. At this scale, if were to increase $k$, keeping the sequence length fixed, the model would stop performing well. This is the scale at which no in-context estimator could work. Information theoretically, there is just not enough data in a test-sequence to be able to estimate the next sample distribution, even approximately. In order to test our results, say for $k=10$, we would have to evaluate on transformers of sequence length $\approx 2000$ (4x blowup). While the forward/backward passes naively takes 16x longer, the transformer also seems to require many more iterations to optimize the loss to low error. These reasons make it prohibitive to evaluate on larger values of $k$. ### Precise dependency on embedding dimension, especially when $S$ is large In practice, the embedding dimension of transformer models like GPT-2/3 and LLama-2/3, and plenty of other open source models (for which data is available), the embedding dimension falls roughly in the range (1000-20,000). This is an order of magnitude smaller than the vocabulary size of these models, which is often in the range (30,000-200,000). It should be possible to improve our results in the dependency on the embedding dimension, which we didn’t fully optimize for (in comparison with the number of heads and depth). At a high level, the reason the embedding dimension in our Theorems and ___ scales as $O(S)$ is because the transformer stores information about the statistics of each symbol along orthogonal components of the embedding vector. The orthogonality of these components prevents information about the statistics of any one symbol from affecting those of another symbol. It remains to be formally verified, but we believe that all of our constructions, as well as the mechanisms empirically learnt by transformers are robust to errors which may appear even when information about the statistics of each symbol does not appear in exactly orthogonal components of the embedding vector, but in approximately orthogonal components. By allowing for approximate orthogonality, the transformer can be much more efficient about its use of the embedding space: the Johnson Lindenstrauss theorem states that there are $S$ approximately orthogonal (pairwise inner products $\in [-\epsilon,\epsilon]$) vectors in $\mathbb{R}^p$ even when $p = O(\log (S)/\epsilon^2)$ is much smaller than $S$. In principle, it suffices to choose the embedding dimensions to scale roughly as $\log (S)/\epsilon^2$ to have an approximate version of Theorem $4$, paying a error scaling with $\epsilon$. Finally, it’s worth mentioning that for practical ranges of $\epsilon$ and $S$, this number ($\log(S)/\epsilon^2$) can indeed be much smaller than what is predicted by what is in the current version of Theorem $4$. We are happy to include this discussion in the paper to provide more context about the embedding dimension.   ___ ### References [1] Edelman, Benjamin L., et al. "The evolution of statistical induction heads: In-context learning markov chains." arXiv preprint arXiv:2402.11004. Pdf: /pdf/5b951e64aa8ae491d3323e9e0172c2fe763eed9b.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper studies the representation capacity of transformers in in-context learning of order-$k$ Markov chains. First, the authors theoretically show that $O(\log (k))$ layers are sufficient to represent $k$-th order induction heads in attention-only transformers. The paper also demonstrates the benefit of non-linearities, such as layer-norm, by showing that a slightly modified transformer (modified from the original architecture) with constant depth is also sufficient to represent the in-context conditional distribution. Strengths: Given the prevalence of in-context learning, I believe understanding the ICL of Markov chains is a very important problem, not only because of the Markovian nature of language but also the simplicity and control it provides in a synthetic setup. This is a good work that builds on recent research on ICL of Markov chains by showing that in a slightly modified transformer, scaling of heads with the order of the chain ($k$) is not necessary to represent $k$-th order induction heads. The lower bound (albeit contingent on the $k$-th order induction head assumption) also provides useful insights. I really enjoyed reading the paper; it was very clearly written and easy to follow. Generally speaking, all the ideas are laid out very clearly, with intuitive explanations that follow before and after the theorems. Weaknesses: 1. It would be beneficial to compare single head results with multiple heads (e.g., in Figure 2a, $2$ heads for $k=2$ and $4$ heads for $k=4$) as a sanity check, given previous works [1] that highlight the multi-head requirement for learning order-$k$ induction heads. I mention this because, to me, it seems like in Figure 2a, the loss isn’t approaching zero unlike in Figure 2b. 2. The work mentions "long training" to observe contradictory results (2-layer single head tfs being able to learn up to order-4 MCs in-context). How long are you training compared to [1]? How significant is this effect, and what impact do you think it has in terms of the underlying mechanism being picked up? I think this needs to be discussed further. [1] Benjamin L. Edelman, Ezra Edelman, Surbhi Goel, Eran Malach, and Nikolaos Tsilivis. The evolution of statistical induction heads: In-context learning markov chains, 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What do you think is the reason 2-layer, 1-head transformers can learn up to order-$4$ MC (if they do, see Weakness 1)? What mechanism do you think is inherently being picked up? Is it learning order-$k$ induction heads or some alternate mechanism? I ask because if there is an alternate mechanism, it could be useful for the lower bound (Thm. 6) in establishing if "the lower bound representing $k$-th order induction heads implies an unconditional lower bound." 2. Regarding Theorem 4, even though the embedding dimension requirement is reduced compared to attention-only transformers (Thm. 2, Thm. 3), it still feels quite large compared to what we see in practice. Any comments on this? How tight do you think this is? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have discussed the limitations separately in the Appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the comments and questions about the paper. Below we address the main questions and weaknesses raised: ### [W1] Comparing single vs. multi-head In the attached material, please refer to Fig. 1, which we will add into the paper in the subsequent version (as the new Fig. 1). This plot discusses how the multi-head transformer learns, while the number of iterations to converge to optimality is of the same order, the test-loss these models achieves is nearly $0$. In comparison, the single head $2$-layer model has non-zero but small test-loss, while the single head $3$-layer model has even smaller test loss. At this scale, we did not distinguish between the small loss of the transformer vs. negligible loss incurred by the multi-head setup. It is indeed plausible that with better choices of hyperparameters (optimizing over embedding dimension, for instance), we may see the delta become even smaller. In our evaluation setup, the transformer does not know what the distribution of samples at test-time are (only that they are Markovian). Thus, any model which achieves low test-loss has to resort to something like an in-context estimate. We were computationally bottlenecked to train transformers on sequence lengths of up to around $500 \approx 2^9$ for the number of iterations required for the test-loss to begin to converge. At this scale, the model has “just” enough data to be approximately able to estimate order-8 Markov chains in context (i.e., having to approximately estimate $2^8 \approx 250$ conditional distributions on $\{ 0,1\}$). At this scale, if were to increase $k$ the model would stop performing well, and more importantly, no in-context estimator could work at this scale. Information theoretically, there is just not enough data in a test-sequence to be able to estimate the next sample distribution, even approximately. TLDR; When the sequence length is around $500$, $k=8$ is the point where we would start to see degrading performance (for any model, and not just a transformer) by virtue of approaching the regime where we don’t have enough samples to estimate conditional probabilities from the test-sequence accurately. ### [W2] Long training to observe contradictory results This is discussed in the common rebuttal above. ___ ### [Q1] Why do 2-layer 1-head transformers learn order-4 Markov chains? This is a great question. While our construction shows that the standard $3$-layer transformers with $1$ head are able to represent $k$-th order induction heads, it may be possible that at an even smaller scale (say $2$-layer with $1$ head) that transformers are still able to represent the in-context conditional $k$-gram with some amount of error. It is also possible that at an even smaller scale, (say $2$-layer $1$-head transformers for sufficiently large $k$, or if the embedding dimension is reduced even further) that induction heads are no longer representable at all. Understanding this regime presents new avenues of research: here the transformer may still achieve low loss, while “provably” not being able to use the induction head mechanism to achieve this loss. It’s worth pointing out the case of $1$-layer $1$-head transformers which were studied in [1]. The authors show that transformers are still able to learn Markov chains, but under this extreme size constraint, learning does not happen in a transductive way. At this scale, the transformer does not do in-context learning, rather learns the parameters of the Markov chain from which data is generated directly. This is a slight departure from our setting, since learning the parameters is infeasible when the training and test data come from different distributions (such as two different randomly sampled Markov chains). ### [Q2] Precise dependency on embedding dimension This is discussed in the common rebuttal above.   ___ ### References [1] Makkuva, Ashok Vardhan, et al. "Attention with markov: A framework for principled analysis of transformers via markov chains." arXiv preprint arXiv:2402.04161 (2024). --- Rebuttal Comment 1.1: Comment: Thank you for the detailed responses and additional figures. I have read other reviews and responses as well, and I am happy to retain my rating.	null	null	null	null	null	null
Language models scale reliably with over-training and on downstream tasks	Reject	Summary: While existing scaling law studies look at compute-optimal pretraining, this paper considers scaling laws in the context of both pretraining and downstream performance. They perform scaling experiments and find that performance is predictable even in overtraining, and average downstream performance is also predictable. Strengths: I think this is a solid paper that attempts to answer an important question. While I’m concerned about the lack of novelty (see the weaknesses), I overall lean towards accepting the paper. I think that the fact the paper reproduces findings from other papers using a different methodology is a good sign that the overall results are correct, and is valuable information (e.g. [Owen 2024](https://arxiv.org/pdf/2401.04757) also finds that average downstream performance is more predictable than for individual downstream tasks). To me, this is the primary contribution of the paper, and an important one as such. Weaknesses: To my mind, the largest weakness of this paper is the lack of novelty. In particular, my understanding is that the most important findings are that (1) performance is predictable in overtraining, and (2) average downstream performance is predictable. I’m not sure why (1) should be surprising – doesn’t a parametric scaling law, such as method 3 from the Chinchilla paper, also give the ability to predict loss when overtraining? I think the authors could improve the motivation for this consideration by providing a back of the envelope calculation: for instance, I plugged in the model size and dataset size for Gopher 280B into the Chinchilla scaling law and got a predicted test ppl of ~7.3 on MassiveText. However, Gopher actually had a (validation) perplexity of ~8.1, this constitutes a relative error of around 10% – substantially larger than the relative errors obtained by the authors of this paper. If the authors can provide an argument of this sort I'd find that helpful. I thought (2) was a more interesting claim, but I’ve seen this analyzed in Owen 2024 (https://arxiv.org/pdf/2401.04757), albeit with a different methodology. As such, I felt that the core results of the paper weren’t very novel. However, I’d be happy to update my assessment if the authors can provide evidence that my understanding is incorrect. One interesting point that the authors mentioned is that performance on individual tasks is less predictable. But this is only mentioned in passing, and I felt that it could be expanded upon a fair bit. What are the implications of this observation? Are there any patterns for which individual tasks are or aren’t predictable? I’m slightly concerned about data leakage being an issue for the downstream tasks, given that the training data (The Pile and RedPajama) covers a wide swath of the internet, and some of the downstream benchmarks have been criticized for data leakage or having label errors. Minor comment: The last paragraph of section 5 is a bit confusing: “There has been a rise in over-trained models [113, 114] and accompanying massive datasets [112, 82, 104, 3]. For example, Chinchilla 70B [45] is trained with a token multiplier of 20, while LLaMA-2 7B [114] uses a token multiplier of 290.” This makes it sound a bit like Chinchilla is overtrained, which I don’t think the authors are trying to say, so I’d suggest something like the following instead: “For example, while Chinchilla 70B is trained compute-optimally with a token multiplier of 20, LLaMA-2 7B…” Technical Quality: 3 Clarity: 3 Questions for Authors: - Doesn’t a standard parametric scaling law fit, such as method 3 from the Chinchilla paper, also allow one to predict loss when overtraining? - How do the results from this paper differ from those in this other paper, which also looks at downstream performance? https://arxiv.org/pdf/2401.04757 - If I plug in the coefficients from the Chinchilla scaling law into equation 8, with $\alpha = 0.34$ and $\beta = 0.28$, I find that the predicted value for $\eta$ is $\frac{\alpha \beta}{\alpha + \beta} \approx 0.154$. In comparison, the values of $\eta$ in table 6 are generally around 0.25. What’s the cause of this difference, and how do you know? - Did the authors consider alternative functional forms for the downstream performance scaling law, besides the negative exponential? - Did the authors check for data leakage? How? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: I felt that the authors did a good job describing some of the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the attention to our work! Please see below for responses to your review. We are happy to provide more clarification or results should it be helpful! Over-training novelty. Thank you for pointing out that Chinchilla Approach 3 implies that over-trained model behavior is predictable. We agree, which is why we framed Equation (4) as a reparameterization of Chinchilla Approach 3 rather than as a novel scaling law (L101-109). While Approach 3 implies reliable scaling in the over-trained regime, this is not empirically verified in the Chinchilla paper. Even though the equation implies a phenomenon, it may not be empirically true. Hence, we argue that our over-training results are empirically novel and hence valuable to the community. Furthermore, we consider validation loss, while Chinchilla considers only training loss, and we explicitly measure relative error between predictions and ground truth to actually quantify how good scaling fits are. These features are missing in the original Chinchilla paper; however, we feel they are important for setting scaling research on solid footing. Downstream error prediction novelty. Thank you for bringing up the Owen 2024 paper, which we were not aware of at the time of our submission. We added the reference to our main related work section to contextualize that others find a relationship between compute and downstream average error. Our main innovation relative to the Owen study is again empirical. The models considered in the Owen study are not standardized: different architectures, training codebases, optimization schemes, and training datasets––to name a few. Each of these factors introduces confounders. In contrast, we create a standardized, open-source setting, which controls these factors. Comparison to Gopher 280B. Thank you for pointing out that the Gopher 280B model does not seem to follow Chinchilla scaling laws. Here we note that Gopher 280B was trained on 300B tokens, which amounts to a token multiplier of $M=\sim1.1$, which is far from the $M=\sim20$ that the Chinchilla team finds is compute optimal on MassiveText. Given that Gopher 280B is under-trained, we might expect its scaling behavior to be less predictable, as also observed in our under-training experiments at $M=5$ (see L252, Appx. Figure 9). Predictability on individual tasks. Thank you for mentioning that individual tasks being hard to predict is interesting! We agree and will expend in L236-249. Particularly, we believe that this observation motivates future work on understanding interactions between training sets and downstream eval predictability. Looking at Table 2, it appears that predictably is influenced by training distributions. For example, training on RedPajama allows for predicting relative error for the 7B run on ARC-Easy at $\sim 5\\%$; however, the prediction error is much higher for C4 and RefinedWeb trained models at $>26\\%$. While influence functions [1, 2] are one promising avenue; we believe there is much more work to be done here. Also, looking at the ablations over downstream eval choices in Appendix Figure 8 suggests that generally adding more evals trends towards better predictability. Dataset leakage concerns. Thank you for bringing this up, dataset leakage is indeed an important problem to consider and can be hard to mitigate in web-scale data. We used standard, open-source datasets as is and did not conduct additional decontamination past what was done in the original releases. However, there has been recent evidence suggesting that contamination, even when it does exist, may not be catastrophic. For example the DataComp-LM project finds, in their Section 4.6, that when explicitly decontaminating against MMLU, performance is comparable (51.8 without decontamination and 52.7 with decontamination). We also note that evaluation in NLP is an active area of research and has many open problems, which are not the focus of our study. We hope that as the science of evaluation in NLP advances, researchers will revisit our scaling testbed, critique its shortcomings, and iterate on methodology. Related work wording. Thank you for mentioning that it is currently unclear that the Chinchilla 70B model is not over-trained. We have applied your suggestion to make this clear. Discrepancies in scaling exponents with Chinchilla. Thank you for catching this! After reviewing this discrepancy, we realized that we printed values of $\alpha$ not $\eta$. As mentioned in L108, $\eta = \alpha / 2$. Hence our values for $\eta$ are actually 0.121, 0.136, and 0.127, more closely matching the Chinchilla value. Thanks again for your diligence looking through the Appendix, we have fixed the table. Exponential decay relating downstream performance and loss. We did not extensively explore alternatives to exponential decay and rather considered empirical phenomena to inform our choice. L131-132 provides the intuition that the functional form should be bounded from above by $\epsilon$, which represents model performance close to random chance error. The error is naturally bounded from below as loss approaches the irreducible loss $E$. --- Additional references. [1] Pang Wei Koh and Percy Liang. Understanding Black-box Predictions via Influence Functions. ICML, 2017. https://arxiv.org/abs/1703.04730. [2] Roger Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, Evan Hubinger, Kamile Lukosiute, Karina Nguyen, Nicholas Joseph, Sam McCandlish, Jared Kaplan, Samuel R. Bowman. Studying Large Language Model Generalization with Influence Functions. arXiv, 2023. https://arxiv.org/abs/2308.03296. --- Rebuttal Comment 1.1: Comment: Thanks for the responses to my questions! I believe this addresses my main concerns. I also think I was being a bit harsh on the novelty point, because even though there's a lot of overlap in results the methodology is quite different and probably more reliable than Owen 2024. As such, I'll increase my score. --- Reply to Comment 1.1.1: Comment: Thanks! We appreciate you reconsidering your score post-rebuttal!	Summary: The authors propose a scaling law for the “Chinchilla over-trained” regime where models are trained on many more tokens (in this paper, up to 30x) than Chinchilla-optimal. They motivate a scaling law relating pre-training compute and “over-training” to validation loss. They empirically demonstrate that the proposed scaling law accurately predicts the validation loss of a 1.4B 32x over-trained model and a 6.9B Chinchilla-optimal model. They then study a simple scaling law relating perplexity to downstream benchmark error. They select a subset of 17 benchmarks for which a 154M parameter models performs 10% above random chance accuracy, and show show that average downstream error of the 1.4B and 6.9B models is predictable. Strengths: The setting considered by the paper is very relevant at the moment, as major model releases in the past year fall precisely in the “Chinchilla over-trained” regime. The work is also novel, as I am not aware of prior work that proposes and empirically validates scaling laws tailored for the Chinchilla over-trained regime. The proposed scaling law in the over-trained regime is well-motivated both by prior scaling laws and by further empirical observations in the over-trained regime (Figure 2). While the models trained (5-8 1e21 FLOPs) are at least two orders of magnitude smaller than the latest open models (e.g., Gemma 2, Llama 3, Qwen 2), the experiments are of a large enough scale for a proof of concept. The paper is clear and it gives sufficient details on the experimental set-up. Weaknesses: My main concerns are twofold: authors do not compare with the standard scaling laws of Kaplan et al. (fitted on their own testbed), and it is unclear how authors choose the models used to fit their scaling laws (Table 1). Authors do not compare their over-trained scaling law with that proposed by Kaplan et al. The authors could fit this scaling law as in Hoffman et al. , Section 3.3, without any additional model training. Specifically, how well can the standard Kaplan et al. law predict the validation loss of the 1.4B over-trained model, when fitted on the model testbed with N < 1B described in Section 3.2? Regarding the claim that validation loss (resp accuracy) is predictable with 300x (resp. 20x) less compute, I find this misleading, since authors train and evaluate a reasonably large model testbed, but only report the compute required for the 5 (resp. 6) models that they ultimately choose for the fit. The authors do not discuss how this “train”/“test” split was chosen. Clearly, the train/test split should be chosen before seeing the evaluation results for any of the models, rather than including models until the fit seems "good enough", or choosing the smallest subset for which the fit is “good enough”. Otherwise, both the claim of 300x/20x compute as well as Figure 5 are misleading. Similarly, Figure 1 would be misleading, and it should include all models with N < 1.4B. The authors consider token multiplier M <= 640, however current models are even more overtrained. If am not mistaken, for Llama 3 8B, M ~= 2000. Demonstrating the validity of the proposed scaling laws for the amount of over-training of current models would have been ideal, even at smaller model scales. Lastly, the proposed scaling laws are validated at substantially lower compute scales than current models with publicly available weights. It would have been ideal to at least see results for over-trained 7B models. I understand that the experiments presented in the paper already require a substantially amount of compute, and more closely matching the compute scales of recent models would be unfeasible for most research labs — therefore I am not taking this point into consideration when scoring the paper. Technical Quality: 2 Clarity: 3 Questions for Authors: * How well can the standard Kaplan et al. law predict the validation loss of the 1.4B over-trained model, when fitted on the model testbed with N < 1B described in Section 3.2? * How were the models in Table 1 chosen? Were they chosen prior to the other models being evaluated? Was it apparent that including only one over-trained model, namely 11M, 320M, would suffice? * Figure 18 shows that HellaSwag and ARC-Easy scores are very correlated with validation loss. However, Table 2 shows very large individual top-1 error for HellaSwag and ARC-Easy. How can this be? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Limitations are adequately discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Comparison to Kaplan et al. Thank you for mentioning this important prior work. Unfortunately, the methodology in Kaplan et al., which utilizes early-stopped models and a different learning rate schedule, is not compatible with our scaling testbed, which is similar to the more contemporary Hoffmann et al. setup. Additionally, we argue that Kaplan et al. is not a practically relevant point of comparison given recent work (Porian et al., 2024 [1]) identifying methodological weaknesses in the Kaplan et al. scaling study. Specifically, Porian et al. find errors in the way Kaplan et al. discount last layer compute and set warmup for small models. We chose to base our scaling study on the more recent Chinchilla paper given its wide adoption (e.g., Muenninghoff et al., 2023) and its intentionality towards correcting methodological oversights in Kaplan et al. Clarification on experimental setup choices: model configurations in Table 1 and train/test splits. Thank you for bringing this up. We detail our approach to selecting model configurations in Section 3.2 and Figure 4. In brief, we train models in a large grid search on a mixture of Pile and RedPajama data. We attempt to predict the validation loss on OpenLM validation data (the OpenLM codebase, recent arxiv papers, and news articles) for larger 1.4B and 7B chinchilla optimal runs. At this stage we do not consider 1) over-training, 2) C4 eval which is used for test loss evals, 3) C4, RedPajama, or RefinedWeb non-mixed datasets, or 4) downstream task performance. Hence our grid search happens in a validation environment, which we are confident is not indexed on our test setup. As for choosing values of $M$, we wanted to test extrapolation to a $N=1.4B, M=640$ over-trained run. Hence, we chose a $M=320$ datapoint and happened to do so at the smallest parameter scale to save compute. To better understand potential human bias that went into this decision, we plot relative error vs. many different potential configurations in Appx. Tables 14-16. The takeaway is that while our eventual choice showed favorable trade-offs between compute and predictive power, there are ways to create more accurate scaling laws with privileged knowledge of test metrics. Clarification on reported compute savings for scaling laws. Thank you for the opportunity to clarify our 20x or 300x compute savings for our scaling laws relative to our largest runs. While hyperparameter searches can be expensive, our team bore the upfront cost to find and open-source reliable configurations, which the community can now use for future scaling studies. Hence our claim is that, using our final configurations, one should be able to predict large runs with the aforementioned compute multiples, which we feel is justified. Over-training in Llama3 8B. Thanks for bringing up the Llama3 8B release. We note that this release happened after the NeurIPS deadline and that our token multiplier range up to $M=640$ is reasonable for models at the time we submitted. Also Llama3 8B is an outlier in that many popular models are less over-trained (e.g., Mistral 7B). Additionally, it is unclear if, at 15T tokens, Llama3 8B was trained for a single epoch. Practically, training a 1.4B parameter model for $M=\sim 2000$ worth of tokens is prohibitive in our setting as 1) this is more compute than we have access to unfortunately and 2) this requires 2.8T tokens for a single-epoch run, which is larger than public datasets at the time of our training runs. Many open-source models are larger than 7B parameters. Thank you for pointing this out and also recognizing we were not able to train Llama-sized models given compute limitations. We agree that verifying scaling trends is valuable for larger models. However, we do not feel that this weakness in our manuscript is a critical flaw for a couple of reasons: 1. Increasingly, teams are pushing the capabilities of models with under 7B parameters. For instance, Phi-2 (2.7B parameters) and Gemma 2B (2B parameters) are both performant models. 2. One of the main motivations of over-training is to have a model with fewer parameters to save on inference costs (L88-91). Hence, we argue it makes sense to study over-training in a low parameter regime to see how far small models can be pushed with over-training. This being said, we attempt to address the spirit of your concern by predicting the validation loss of Llama2 7B and 13B models, under the assumption that these models were trained on datasets similar to RefinedWeb. To run this experiment, we re-tokenize RefinedWeb with the Llama2 tokenizer and re-train our small-scale models from Table 1. As we see in the attached pdf (Figure A), with this assumption, our over-trained scaling laws predict accurately both models’ performance. Note, we must make the aforementioned data assumption as scaling laws are fit on a suite of models trained and evaluated on standardized distributions. To truly have a clean experiment, we would need access to the Llama2 training data and internal details. Nevertheless, we feel this experiment suggests over-trained performance can be predictable for larger runs. Top-1 error on HellaSwag and ARC-Easy in Table 1 vs. Figure 18. Thanks for bringing this up. Table 2 and Figure 18 show different things. Table 2 shows predictions based on only the six configurations from Table 1. Figure 18, in contrast show the empirical trend with all 104 models we trained. One takeaway here is that it should be possible to achieve reliable downstream prediction on individual tasks with increased compute investment. However, this represents diminishing returns as the scaling law investment approaches the cost of actually training the large scale run. --- New references. [1] Tomer Porian, Mitchell Wortsman, Jenia Jitsev, Ludwig Schmidt, Yair Carmon. Resolving Discrepancies in Compute-Optimal Scaling of Language Models. arXiv, 2024. https://arxiv.org/abs/2406.19146. --- Rebuttal 2: Comment: Thank you for your response. My two main concerns were not effectively addressed. Having read the other reviewers’ reviews, I’ll increase my score, since I agree that the paper does show that language models scale reliably with over-training. Comparison to Kaplan et al.: What I mean is to consider how well the “standard” functional form used by prior work, which you write in Equation 3, can predict the over-training regime. Without this comparison, it is simply not possible to judge whether the proposed reparametrization in Equation 4 is a valuable contribution in itself or not. Why should practitioners fit Equation 4 and not Equation 3? Note that Equation 3 is the power law of Kaplan (Equation 4.1 in the Kaplan paper) + the irreducible loss term. I included in my review “The authors could fit this scaling law as in Hoffman et al. , Section 3.3, without any additional model training”. Reviewer CG6A echoes a similar weakness “doesn’t a parametric scaling law, such as method 3 from the Chinchilla paper, also give the ability to predict loss when overtraining?”. I fail to understand why you cannot provide this comparison. You do not need to train any additional models, you have the (N, D, L) triplets. It is just a matter of fitting Equation 3. Reported compute savings: I am not referring to the compute spent on the grid search for hyper-parameter tuning. I am referring to the compute required to train the scaling test bed itself, bar the N=1.4B and N=6.9B models which are the ones you ultimately aim to extrapolate to. “Hence, we chose a datapoint M=320 and happened to do so at the smallest parameter scale to save compute”. This is not convincing. A single M=320 datapoint happened to be good enough for the fit. But you did not happen to train a single M=320 model, you trained many models with M > 20. The concern is in choosing the fit/predict split after the models were already trained and evaluated. Otherwise, the stated compute savings and Figure 5 altogether can be misleading. > we plot relative error vs. many different potential configurations in Appx. Tables 14-16 These plots precisely illustrate my point. Looking at the C4 eval plots, the blue stars are in the Pareto frontier of relative error vs FLOPs spent. The 300x/20x multipliers are not representative of what one would typically obtain, they represent a “best” case scenario, potentially indicative of cherry-picking regarding the fit/predict split. Over-training in Llama3 8B and larger models: I maintain my initial positive assessment that the experiments are of a large enough scale for a proof of concept. --- Rebuttal 3: Comment: We apologize for not addressing your concerns; we unfortunately misunderstood your original messages! This said, we appreciate you attending to the other reviews and still reconsidering your score! We aim to address your outstanding concerns below. Comparison to Kaplan et al. There is no reason we cannot provide this comparison. We unfortunately misunderstood your original comment, but please see the table below for the requested results! Note, Equations (3) and (4) in the paper are the same (i.e., both have four free parameters, with the relation given in L108). Concretely, "we reparameterize Equation (3) in terms of compute $C = 6ND$ and a token multiplier $M = D/N$ [to] get, [Equation (4)]" (L107). We include Equation (4) to provide the reader with intuition about what should happen when one over-trains (L110-114). We are also careful not to claim Equation (4) as a novel scaling law. For example, in the introduction: "We explain our observations by reparameterizing existing scaling laws in relation to the amount of over-training" (L42-43). Our main contributions with respect to over-training are empirical (as detailed in our response to reviewer CG6A). \| Scaling Form \| Model \| Relative Error on C4 eval loss\| \|--------------\|-------\|--------------------------------\| \| $L(N,D) = AN^{-\alpha} + BD^{-\beta}$ (Kaplan et al. like) \| open_lm_1b (M=640.0) \| 16.1281% \| \| \| open_lm_7b (M=20.0) \| 17.2352% \| \| $L(C, M) = E + (aM^{\eta} + bM^{-\eta}) C^{-\eta}$ (our reparameterization of Hoffmann et al.) \| open_lm_1b (M=640.0) \| 0.7103% \| \| \| open_lm_7b (M=20.0) \| 0.7320% \| Reported compute savings. Sorry that we misunderstood your comment! We understand your skepticism about our 300x/20x savings claim and ultimately your concern about cherry-picking. We reiterate, using the configurations in Table 1, a practitioner should expect to predict runs similar to our N=1.4B and N=6.9B runs with the claimed compute savings. We will refine the wording to make sure this claim is more precise and clear. With respect to the entire scaling testbed, we do feel the need to clarify that the results are not cherry-picked. For the sake of transparency, we detail key moments in our development timeline, specifically for the over-training portion of the project. We hope this will provide clarity and help address your very legitimate concerns. 1. Chinchilla Approach 3 hints that scaling should be reliable; however, they do not directly measure (in terms of relative error), how good scaling predictions are emperically. Hence, we started our investigation by trying to reproduce Chinchilla. This resulted in a large grid search, as we did not understand how to set hyperparameters to get reliable scaling trends. The grid search ultimately culminated in Section 3.2 of our writeup. Note, at this stage, we had not touched any of our downstream evaluations or our main training distributions. In fact, at this stage we did not even have a project idea! We were, however, largely able to reproduce Chinchilla Approach 3 and observed reliable scaling for compute-optimal models. 2. At this stage, we were looking for promising research directions. Motivated by many open-weight releases at the time, we thought it would be nice to understand what happens empirically when one over-trains models. We hypothesized that we should see “parallel lines” in the log-log plot based on the functional form of Approach 3. Hence, we chose a subset of our entire grid search configurations (as detailed in Section 3.2), tokenized C4, and trained these configurations for various token multipliers (i.e., number of tokens). This set of training runs eventually contributed to our scaling testbed. We did in fact see parallel lines and then set out to understand if this phenomenon could be reliably extrapolated with some of our smaller-scale runs. 3. At first, we thought we could predict over-trained behavior with only Chinchilla optimal models. However, after playing around with equations and parameterizing with token multipliers, we realized that to extrapolate in M, we would have to fit to at least 1 data point where $M \neq 20$ (i.e., at least one non-compute-optimal run). We knew that we would have enough compute to over-train our largest run at $M=640, N=1.4$B. From our scaling testbed we chose the smallest config ($N=11$M) with the largest token multiplier ($M=320$) such that we could still probe for extrapolation in $N, M$. The other models in Table 1 are, somewhat arbitrarily, Chinchilla optimal. 4. We constructed Appx. Tables 14-16 to ablate our decisions. Our meta-analysis here is that the combination of our compute constraints and intuition that small models (e.g., over-trained 11M models) could give reasonable signal were reasonable.	Summary: This paper investigates the power laws (scaling laws) of neural language models, particularly from the perspective of over-training and the relationship between validation loss (perplexity) and NLP downstream tasks. The authors define over-training as the situation where runs consume large amounts of computational resources, and they introduce a token multiplier, M. It is computed by D / N, where D is the number of training data tokens and N is the number of parameters. Through various model training setups and three different training corpora, the authors demonstrate that the validation loss can be computed and predicted using an equation that includes M. They also introduce another equation that illustrates the relationship between validation loss and downstream task error. Strengths: The main contribution of this paper is the exploration of the scaling laws in language models concerning over-training and NLP downstream tasks. These results, including equations and practical outcomes, are beneficial for researchers and engineers developing large language models. As highlighted by the authors, these insights are valuable for researchers in their future work. Weaknesses: The authors mentioned several limitations and future work in the paper. I agree with them, and especially the ‘scaling up’ part is the primary concern of this paper. The model sizes range from 0.011B to 6.9B, but open-source models are larger than these sizes - for instance, Llama 2 starts at 7B, and Llama 3 starts at 8B [1]. Furthermore, model size is crucial for techniques such as CoT [2]. I hope to hear the authors' opinions on this concern. Technical Quality: 3 Clarity: 3 Questions for Authors: - M could be related to overfitting if N is large but D is not sufficient. The authors mentioned that when M equals 5 (under-training), the scaling becomes unreliable. I assume this is because the models are not well trained—underfitting. Could you explain and discuss the relationship between over-training and overfitting? - (Minor) It is hard to distinguish between M=320 and M=640 in the figures. Could you change the colors to make them more distinguishable? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the attention to our work! Please see below for responses to your review. We are happy to provide more clarification or results should it be helpful! Many open source models are larger than 7B parameters. Thank you for pointing this out––we agree that verifying scaling trends is valuable for larger models. However, we do not feel that this weakness in our manuscript is a critical flaw for a couple of reasons: 1) Increasingly, teams are pushing the capabilities of models with under 7B parameters. For instance, Phi-2 (2.7B parameters) and Gemma 2B (2B parameters) are both performant models. 2) One of the main motivations of over-training is to have a model with fewer parameters to save on inference costs (L88-91). Hence, we argue it makes sense to study over-training in a low parameter regime to see how far small models can be pushed with over-training. This being said, we attempt to address the spirit of your concern by predicting the validation loss of Llama2 7B and 13B models, under the assumption that these models were trained on datasets similar to RefinedWeb. To run this experiment, we re-tokenize RefinedWeb with the Llama2 tokenizer and re-train our small-scale models from Table 1. As we see in the attached pdf (Figure A), with this assumption, our over-trained scaling laws predict accurately both models’ performance. Note, we must make the aforementioned data assumption as scaling laws are fit on a suite of models trained and evaluated on standardized distributions. To truly have a clean experiment, we would need access to the Llama2 training data and internal details. Nevertheless, we feel this experiment suggests over-trained performance can be predictable for larger runs. Over-training vs. overfitting. Thanks for the opportunity to clarify. While over-training and overfitting are related concepts, they have distinct definitions. Given a Chinchilla optimal allocation of tokens to parameters, over-training refers to training on disproportionately more tokens than parameters (L88-91). Critically, validation loss can still go down with more over-training, as seen in Figure 2. Hence, over-trained models are not necessarily overfit to the data (i.e., overfit models necessarily experience increasing validation loss). Also, in our single epoch training setting, there is no data repetition, so over-fitting is not expected to happen. We agree that the terminology can be confusing, especially given the ubiquity of the term overfitting in machine learning. We will add this explanation to the paper for clarity. Underfitting and unreliable scaling. Our empirical observation is that under-trained models (trained on less tokens than is Chinchilla optimal) appear to scale unreliably. We cannot, however, make a stronger claim that all underfit models scale unreliably. For example, in Figure 2, we see that the Chinchilla optimal models ($M=20$) actually underfit the data, as training for $M>20$ decreases validation loss. However, the models at $M=20$ do scale reliably, following a power-law with irreducible error. Improved color contrast for Figures. We agree that the contrast could be improved for clarity, thanks for pointing this out. We plan to update Figures 1 and 2, which now use the `plasma` matplotlib color pallet. Please see the attached pdf, Figure B, for a sample. --- Rebuttal Comment 1.1: Comment: Thank you for your responses and the revised figures. Your response have addressed my concerns and clarified my questions. Good luck! --- Reply to Comment 1.1.1: Comment: Thanks! Given that we've addressed the concerns, we were wondering if you might reconsider our score. We understand this decision is entirely at your discretion and are thankful for your attention regardless!	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their attention to our work, constructive comments, and positive feedback. Specifically, we are grateful for reviewers highlighting our empirical efforts and strengths of our methodology (RC8r, CG6A). We also appreciate their mentioning the relevance of our scaling study for model development (7fVK, RC8r) especially in light of contemporary research releases where model over-training is commonplace. We also appreciate the reviewers pointing out room for improvement. In the comments below we address all concerns raised by reviewers, including additional empirical evidence when warranted. We are happy to continue the discussion during the rebuttal period and provide additional clarification and experiments as needed! Again, we thank all reviewers for helping us improve our work! (Please see below for a pdf containing rebuttal Figures A and B mentioned in the comments below) Pdf: /pdf/58750808484c3fbf522e6c05a8aa64e9dacc3daf.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Supervised Kernel Thinning	Accept (poster)	Summary: This paper applies the kernel thinning approach to non-parametric regression, with two kinds of estimators (NW and KRR) considered. This approach speeds up the computational efficiency by using carefully chosen coresets for approximation. Theoretical results on the approximation error are established. Numerical experiments include synthetic data and the well-known California housing dataset. Strengths: 1. The paper is well organized and written, with enough background and clear motivation 2. Solid theoretical results are developed for the proposed method 3. Certain experimental results are given to justify the proposed method Weaknesses: 1. Despite the well developed theory, I am not fully convinced yet by the significance/contribution of this method - I am willing to stand corrected if more comparisons can be made with other existing methods for efficient non-parametric regression. 2. The application are only to toy examples or standard small datasets - e.g. for the Californian housing dataset with n = 20,640, doing non-parametric regression without thinning also seems fine. 3. In the experiments, the baseline is chosen to be NW / KRR with naive thinning - some more competitive, state-of-the-art baselines could have been chosen. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In the beginning of Sec 3, the NW and KRR estimators are restricted to kernels taking the form of (4) and (6). What are the specific reasons to make these restrictions? And to see how restrictive/general this is, could you give some examples of commonly used kernels that fall under this category? 2. The theory makes the sub-Gaussian tail decay assumption on the kernel, which would exclude many commonly used kernels such as the Matern kernel. How would the theory work out, and how would the rates change if this assumption is generalized to sub-Weibull tail decay, or say a bound on the Orlicz-norm? 3. The method (or at least the theory) is developed for one-dimensional response data. How could this be generalized to multi-dimensional response data, where issues with computational efficiency becomes more significant? 4. In the discussion section, it is mentioned that the rates on the excess risk could be sub-optimal. Do you have a sense of what rate could be minimax optimal? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weaknesses and questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the time you’ve spent reviewing our work and for your thoughtful feedback. We now address each of your questions and concerns in turn. $\blacktriangleright$ $\textbf{Comparisons to existing methods for efficient non-parametric regression}$ - _"The application are only to toy examples or standard small datasets..."_ - _"In the experiments, the baseline is chosen to be NW / KRR with naive thinning..."_ Thank you for this suggestion. - We have added an experiment on the SUSY dataset ($d=18, N=5\times 10^6$) from Diaz et al. (2023). We sample $4\times 10^6$ points to use for training. (Full KRR is now infeasible as we can only store 3k-4k columns at a time, e.g., on a machine with 100 GB RAM). - We include the RPCholesky preconditioning method from Diaz et al. (2023) and the FALKON method from Rudi et al. (2017) as SOTA baselines. As a sanity check, we also include Conjugate Gradient method (without preconditioning). - In Fig. 2 of the attached PDF, we have added RPCholesky-NW (where $S_{NW}$ are the pivot points from RPCholesky) and Restricted KRR with RPCholesky for landmark selection (see box plots in red). While it outperforms the KT method in the KRR setting, RPCholesky is less suitable for speeding up the NW estimator as shown in Fig. 2a of the attached PDF. $\newline$ $\blacktriangleright$ $\textbf{Kernel assumptions}$ - _"In the beginning of Sec 3..."_ - _"The theory makes the sub-Gaussian tail decay assumption on the kernel..."_ Thank you for raising this question. - Our theory in fact permits a large class of base kernels. This strength is borrowed partially from the Kernel Thinning results of Dwivedi & Mackey (2021). - In our revision, we will use $\mathbf{k}$ exclusively to denote the base kernel and use $\mathbf{k} _{ALG}$ to denote the kernel for thinning. Given a valid base kernel $\mathbf{k}$, we can construct $\mathbf{k} _{NW}$ and $\mathbf{k} _{RR}$ following (4) and (6), respectively. - For NW, we only showed the results when the kernel is compactly supported or admits sub-Gaussian tails (for ease of presentation); remarkably, the same results hold when the kernel has sub-exponential tails or poly-tails as well. In fact, $\mathbf{k}$ can be any radial kernel $\mathbf{k}(x_1,x_2) := \kappa(\lVert x_1-x_2\rVert_2/h)$ such that $\kappa(u)$ is a bounded, L-Lipschitz function and decays (at any rate). This includes kernels with sub-Weibull tail decay. The kernel density estimation guarantees from Dwivedi & Mackey (2021), which we use in our analysis, indeed use an Orlicz bound. - For KRR, $\mathbf{k}$ can be any kernel such that the log-covering number satisfies the PolyGrowth (e.g., finitely-times differentiable kernels) or LogGrowth (e.g., analytic kernels, finite-rank kernels) property (see Assump. 4). For KRR, our results do degrade with increasingly less smooth kernels. $\newline$ $\blacktriangleright$ $\textbf{Multi-dimensional response data}$ - _"The method (or at least the theory) is developed for one-dimensional response data..."_ The reviewer’s comment on multi-dimensional response data has direct and interesting applications for speeding up multivariate response regression problems. - Roughly speaking, KT with input kernel $\mathbf{k}_{ALG}$ provides a coreset that well approximates the averages of functions in its RKHS $\mathcal H$. - Dwivedi & Mackey (2022), Thm. 4 showed that thinning with $\mathbf{k} _{ALG} := \mathbf{k} _1 + \mathbf{k} _2$ produces a coreset that simultaneously well approximates averages for any $ f\in \mathcal{H}(\mathbf{k} _1)$ or $f\in \mathcal{H}(\mathbf{k} _2)$ (aka the "aggregate kernel trick"). - We use the aggregate kernel trick to tackle multi-variate regression via one-dimensional kernel thinning on an aggregated kernel. - To be precise, let $f = (f _1, …, f _m)$ be multi-dimensional response function, and for all $j = 1, …, m$, $f _j : R^d \to R$ are coordinate functions in some RKHS $\mathcal{H}( \mathbf{k} _j)$ generated by some base kernel $\mathbf{k}(j)$. - Now define the aggregated kernel $\mathbf{k} _{ALG}((x_1,y_1),(x_2,y_2)) := \sum _{i=1}^{m} w _j \mathbf{k}(j) ((x _1,y _{1,j}),(x _2,y _{2,j}))$ (for NW and use $\mathbf{k} _{RR}(j)$ instead of $\mathbf{k} _{NW}(j)$ for KRR), where weights $w _j$ reflect prior beliefs. - Then $\mathbf{k} _{ALG}$-kernel thinned core-sets are shared across all coordinate functions, that also balances information across all coordinates. - Notably, aggregate kernel tricks have been used in Generalized Kernel Thinning by Dwivedi & Mackey (2022) for improving integration error guarantees and Compress-then-Test by Domingo-Enrich et al. (2023) for increasing power of kernel-based hypothesis tests. Here we use it for speeding multi-variate non-parameteric regression. $\newline$ $\blacktriangleright$ $\textbf{Minimax optimal rates}$ - _"In the discussion section, it is mentioned that the rates on the excess risk could be sub-optimal..."_ We will include the expanded discussion in the revision: - The following is classically known (see Tsybakov (2009)): - When $f^\star$ belongs to a Hölder class with smoothness $\beta > 0$, the minimax optimal rate is $\mathcal{O}(n^{-\frac{2\beta }{2\beta + d}} )$. - When $f^\star$ belongs to a Sobolev space with smoothness $\nu=1/\alpha$ (for $\alpha\in (0,2)$), the minimax optimal rate is $\mathcal{O}(n^{-\frac{2}{2+\alpha}})$. - As shown by Dwivedi & Mackey (2021), KT produces a coreset that achieve the minimax rates (as if all $n$ points were used) for a large class of kernel density estimation tasks—e.g., employing any radial kernel $\mathbf{k}(x_1,x_2) = \kappa( \lVert x_1-x_2\rVert _2 / h)$ with bounded, L-Lipschitz $\kappa: \mathbb{R} \to \mathbb{R}$. However, our current proof techniques do not yield a minimax rate for supervised learning task. We believe a modified proof technique should yield the same rate as obtained by $n$ points for some class of kernels. --- Rebuttal Comment 1.1: Title: Further response Comment: We thank Reviewer QHUc for your comments, which we believe have been addressed in our response. Please let us know if you have any other questions that we can address.⬤	Summary: The authors speed up two non-parametric regression estimators, Nadaraya-Watson (NW) and Kernel Ridge Regression (KRR), by using a kernel thinning (KT) technique to compress the input data in a way that preserves important statistical properties. They include a theoretical analysis proving that that KT-based regression estimators are computational more efficient and also exhibit improved statistical efficiency over i.i.d. subsampling of the training data. They also include empirical validation of their method on simulated and real-world data. Strengths: Good presentation with a solid analytical analysis providing theoretical guarantees. Overall well-organized paper. The proposed method of combining KT with the estimators is novel (to my knowledge) and shows potential. Weaknesses: The empirical analysis is limited. The shortcomings on the real-world data are not adequately addressed (see also limitations). Technical Quality: 3 Clarity: 3 Questions for Authors: I have no particular questions. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The comparison with other large-scale kernel methods is limited to RPCholesky; a more comprehensive comparison and overview would be nice. Performance/utility of the proposed method on real-world data remains unclear. The results obtained on real-world dataset in section 4.2 are not convincing, and the authors acknowledge the problem of kernel mis-specification but do not elaborate or provide any resolution. Btw, there is an incomplete sentence in lines 283-284. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the time you’ve spent reviewing our work and for your thoughtful feedback. We now address each of your questions and concerns in turn. $\blacktriangleright$ $\textbf{Comparison with large-scale kernel methods}$ - _"The comparison with other large-scale kernel methods is limited to RPCholesky"_ Thank you for this suggestion. We have added new results in the global response pdf. - It includes an experiment on the SUSY dataset ($d=18, N=5\times 10^6$) from Diaz et al. (2023). We sample $4\times 10^6$ points to use for training. (Full KRR is now infeasible as we can only store 3k-4k columns at a time, e.g., on a machine with 100 GB RAM). - We include the _RPCholesky preconditioning method_ from Diaz et al. (2023) and the _FALKON method_ from Rudi et al. (2017) as SOTA baselines. As a sanity check, we also include the _Conjugate Gradient method_ (without preconditioning). - In Fig. 2 of the attached PDF, we have added RPCholesky-NW (where $S_{NW}$ are the pivot points from RPCholesky) and Restricted KRR with RPCholesky for landmark selection (see box plots in red). While it outperforms the KT method in the KRR setting, RPCholesky is less suitable for speeding up the NW estimator as shown in Fig. 2a of the attached PDF and is outperformed by KT-NW. $\newline$ $\blacktriangleright$ $\textbf{Kernel mis-specificiation}$ - _"the authors acknowledge the problem of kernel mis-specification but do not elaborate or provide any resolution"_ We will elaborate our discussion in the revision. - The supervised KT guarantees apply when $f^\star$ lies in the RKHS $\mathcal{H}(\mathbf{k})$, where $\mathbf{k}$ is the base kernel. - In practice, choosing a good kernel $\mathbf{k}$ is indeed a challenge _common to all prior work_. - Our framework is friendly to recent developments in kernel selection to handle this problem: - Dwivedi & Mackey (2022), Cor. 1 provide integration-error guarantees for KT when $f^\star \notin \mathcal{H}(\mathbf{k}). - Moreover, there are recent results on finding the best kernel (in testing; Dominigo-Enrich et al. (2023), Sec. 4.2). - Radhakrishnan et al. (2024) introduce Recursive Feature Machines, which use a parameterized kernel $\mathbf{k} _{M}(x_1,x_2) := \exp(-\frac{(x_1-x_2)^\top M (x_1-x_2)}{2\sigma^2})$, and propose an efficient method to learn the matrix parameter $M$ via the average gradient outer product estimator. An exciting future direction would be to combine these parameterized (or "learned") kernels with our proposed KT methods for non-parametric regression. - Formally combining these results, potentially with the recursive feature machines of Radhakrishnan et al. (2024) to build a practical strategy with theoretical guarantees is an exciting future direction. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response! Best of luck to the authors.	Summary: After the rebuttal, I have updated my score from 3 to 5. ---- This paper provides a meta-algorithm based on Kernel Thinning for non-parametric regression, in particular, the Nadaraya-Watson (NW) regression, and the Kernel Ridge Regression (KRR). The idea is to run NW or KRR on a thinned coreset by KT. The core of the method is to choose a suitable kernel for which the KT assumptions are satisfied. The paper introduces to meta-kernels that takes a base kernel of features x and produces the kernel over (x,y) suitable for regression tasks. The paper provides the theoretical guarantees for kernel thinned NW and kernel thinned KRR. In empirical study, the paper includes simulation results and real data results which demonstrate the proposed method's strength over full computation and standard thnnning in accuracy and time; however, the SOTA RPCholesky method seems to have a better empirical performance than kernel thinned KRR. Strengths: - The paper in general is easy to read - The idea to introduce kernel thinning to non-parametric regression methods is novel. - Provides theoretical guarantees of the method under suitable assumptions. Weaknesses: - The empirical advantage of the proposed kernel thinned KRR over RPCholesky is not clear. In fact, from figure 5, RPCholesky has slighly higher training time compared to kernel thinned KRR, however, a much lower test MSE. - Some technical background and discussion can be made more clear. See questions. - Writing can be improved. See questions. Technical Quality: 2 Clarity: 2 Questions for Authors: - There is not an algorithmic description for Kernel Thinning algorithm. - Eq 4 and 6: why do you choose the kernel in this way? In particular, can you comment on the part $k(x_1,x_2) y_1, y_2$. Does the performance of the kernel rely on the relationship of $y$ and $x$, and if so, how? Also, in eq 6, is this a typo $k^2(x_1,x_2)^2$? - Notation suggestion: instead of $S_{KT}$ for KT-NW, could use $S_{NW}$ to match the use of $K_RR$ in KT-KRR part. - Sec 4.1: are the kernels listed in 1.2.3 used as the base kernel to construct the kernel for thinning, or they are the kernel for thinning? - Fig 2: Right panel: why not include RPCholesky? - Fig 2: why si the full not performing the best? - Fig 3 and Fig 4: why not use the function in (11)? - Fig 3 and Fig 4: what is the unit of y axis? Are they all in the same unit? - Sec 4.2: what is the conclusion from Fig 5? - Sec 4.3: line 283-284: "In this example (and ....)" incomplete sentence. - It'd be good to include Fig 5 in the main text instead of appendix. Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: - KT-KRR does not demonstrate an empirical advantage over RPCholesky. Missing a discussion on this point. - Missing a discussion on potential failure cases of the method, though there is a discussion on kernel mis-specification and potential way to deal with it. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the time you’ve spent reviewing our work and for your thoughtful feedback. We address each of your concerns below. $\blacktriangleright$ $\textbf{Comparison with RPCholesky}$ - _"The empirical advantage..."_ - _"KT-KRR does not demonstrate..."_ Thank you for raising these points. The main focus of this work was to build a stepping stone towards bridging two literatures: distribution compression and non-parametric regression, by designing algorithms and developing theory. Tuning to optimize empirical performance, e.g., by combining the pre-conditioning step of RPCholesky with KT, is a great future direction. As another sanity check, we have added an experiment on the SUSY dataset ($d=18, N=5\times 10^6$) from Diaz et al. (2023). KT-KRR's (22%) performance lies between RPCholesky (20%) and ST-KRR (22.7%). Thinning 4 million points with our method took only 1.7 seconds on a single CPU core, with further speed-ups to be gained from parallelizing on a GPU. - _"Fig 2: Right panel: why not include RPCholesky?"_ In Fig. 2b of the attached PDF, we have added Restricted KRR with RPCholesky for landmark selection. In Fig. 2a of the attached PDF, we have added RPCholesky-NW, where $S_{NW}$ are the pivot points from RPCholesky. $\newline$ $\blacktriangleright$ $\textbf{Choice of kernel}$ - _"Eq 4 and 6: why do you choose the kernel in this way?..."_ - _"Sec 4.1: ... the kernels listed in 1.2.3 used... are the kernel for thinning?"_ This is a great question central to the novelty of our work while bridging the two literatures. Directly applying generic kernels would not work in theory and our experiments confirm that. Overall the choice we make are informed by the following points: - Roughly speaking, KT with input kernel $\mathbf{k}_{ALG}$ provides a coreset that well approximates the averages of functions in its RKHS $\mathcal H$. - Dwivedi & Mackey (2022), Thm. 4 showed that thinning with $\mathbf{k} _{ALG} := \mathbf{k} _1 + \mathbf{k} _2$ produces a coreset that simultaneously well approximates averages for any $ f\in \mathcal{H}(\mathbf{k} _1)$ or $f\in \mathcal{H}(\mathbf{k} _2)$ (aka the "aggregate kernel trick"). - The NW estimator is a ratio of averages: an average of $f(x)=k(x, x_0)$ in the denominator and an average of $f _{num}(x,y)=\langle y,1 \rangle k(x,x_0)$ for the numerator. - $f _{num}$ lies in the RKHS corresponding to the kernel $\mathbf{k} _{num}((x_1,y_1),(x_2,y_2)) := \mathbf{k}(x_1,x_2) y_1 y_2$, so thinning with this kernel provides good approximation of the numerator. - Applying aggregate kernel trick to $\mathbf{k} _1 = \mathbf{k}$ and $\mathbf{k} _2= \mathbf{k} _{num}$ yields the $\mathbf{k} _{NW}$ from (4). - For KRR, we approximate the training loss, which is an average of $\ell_f(x,y) = (f(x)-y)^2 = f^2(x) -2 f(x)y + y^2$. - Assuming $f\in \mathcal{H}(\mathbf{k})$, $\ell_f$ lies in the RKHS of $\mathbf{k}^2(x_1,x_2) + y_1 y_2 \mathbf{k}(x_1,x_2) + y_1^2 y_2^2$, where the last term can be dropped (see Sec. 3 of _Compressed Empirical Measures_ by Grünewälder (2022)), yielding (11) (thanks for pointing out the typo). - Remarkably, our ablation studies in Fig. 2 of the original manuscript and Fig. 2 of the attached PDF confirm this theory. - Thinning with $\mathbf{k} _{ALG}((x_1,y_1),(x_2,y_2)) = \mathbf{k}(x_1,x_2)$ (orange; Ablation #2 in Sec. 4.1) performs poorly because it only exploits the $x$ information. - Thinning with $\mathbf{k} _{ALG}((x_1,y_1),(x_2,y_2)) = \mathbf{k}(x_1\oplus y_1,x_2\oplus y_2)$ (i.e., ) (brown; Ablation #1 in Sec. 4.1)—despite incorporating information from both $x$ and $y$ by concatenating the two—also yields poor empirical performance. - Thinning with $\mathbf{k} _{ALG} = \mathbf{k} _{NW}$ and $\mathbf{k} _{ALG} = \mathbf{k} _{RR}$ perform the best for the NW and KRR settings, respectively. $\newline$ $\blacktriangleright$ $\textbf{Improvements to notation and writing}$ - _"Notation ... instead of for KT-NW, could use...KT-KRR part."_ We will use $S_{NW}$ to denote the coreset for the thinned NW estimator and $S_{RR}$ to denote the coreset for the thinned KRR estimator. We will use $\mathbf{k}$ exclusively to denote the base kernel and use $\mathbf{k} _{ALG}$ to denote the kernel for thinning. - _"There is not an algorithmic description for Kernel Thinning algorithm."_ We apologize for this oversight. We will add a paragraph in the revised paper and an algorithm with details in the appendix. - _"Fig 3 and Fig 4: why not use the function in (11)?"_ - _"Fig 3 and Fig 4: what is the unit of y axis?..."_ In the attached PDF, we have reproduced Fig. 3 and 4 with (11) and improved the formatting. The y-axis is in seconds and we use log-scale to make the difference between ST, KT, and RPCholesky clearer. - _"Sec 4.2: what is the conclusion from Fig 5?"_ We will revise the discussion for Fig. 5 as follows: On the California Housing dataset, KT-KRR lies between ST-KRR and RPCholesky in terms of test MSE. In terms of train time, KT-KRR is quite a bit faster than RPCholesky (0.0153 s vs 0.3237 s). $\newline$ $\blacktriangleright$ $\textbf{Discussion on potential failure cases}$ - _"Missing a discussion on potential failure cases of the method..."_ We will expand our discussion: - The supervised KT guarantees apply when t $f^\star$ lies in the RKHS $\mathcal{H}(\mathbf{k})$, where $\mathbf{k}$ is the base kernel. In practice, choosing $\mathbf{k}$ is indeed a challenge _common to all prior work_. - Dwivedi & Mackey (2022), Cor. 1 provide integration-error guarantees for KT when $f^\star \notin \mathcal{H}(\mathbf{k}). Moreover, there are recent results on finding the best kernel (in testing; Dominigo-Enrich et al. (2023), Sec. 4.2). - Combining these results, potentially with the recursive feature machines of Radhakrishnan et al. (2024) (also see our response to Rev PSUT) to build a practical strategy is an exciting future direction. --- Rebuttal Comment 1.1: Title: followup questions Comment: I would like to thank the authors for their detailed reply. Most of my concerns are addressed and I have some followup questions: - In Figure 2b in the attached pdf, why is RPCholesky not included at n=2^14? It is because it is not feasible to run RPCholesky at that point? - Can you also address this question: Fig 2: why is the full not performing the best? --- Rebuttal 2: Comment: 1. Thank you for the catch. RPCholesky-KRR achieves an estimated test loss of 0.9980 when $n=2^{14}$ (see Table B below)—close to the Bayes optimal loss of 1. We have estimated the population test loss by sampling 10,000 points from our data generating distribution (as a proxy for integrating), so the estimated test loss can actually be less than 1 with some probability. When we use log-scaling on the y-axis, negative values of excess risk become undefined, so they don't appear on the plot. Note that Full-NW and Full-KRR for $n=2^{14}$ also face this issue. We will make sure to add a note about this in the revision. 2. In Fig. 2, Full-NW and Full-KRR are the black box plots (always the first in each group when reading from left to right). Full is deterministic, so the box plots appear as lines. _Note that Full-NW and Full-KRR indeed perform the best across all $n$._ For additional clarity, we have attached tables of the test loss (mean ± standard deviation across 100 trials). (Subtract each value by 1 to get the excess risk values used in Fig. 2.) Table A: Test loss for NW estimators in Figure 2a of the attached PDF \| \| $n=2^{8}$ \| $n=2^{10}$ \| $n=2^{12}$ \| $n=2^{14}$ \| \|:------------------------------------------------------\|:----------------\|:----------------\|:----------------\|:----------------\| \| Full \| 1.1423 ± 0.0000 \| 1.0494 ± 0.0000 \| 1.0189 ± 0.0000 \| 0.9997 ± 0.0000 \| \| RPCholesky \| 3.0737 ± 0.2548 \| 2.8583 ± 0.3150 \| 2.0772 ± 0.2920 \| 1.4720 ± 0.1026 \| \| ST \| 3.1181 ± 0.1813 \| 2.8844 ± 0.3063 \| 2.4720 ± 0.3058 \| 1.7632 ± 0.2364 \| \| KT w/ $\mathbf{k}_{\mathrm{RR}}((x_1,y_1),(x_2,y_2))$ \| 2.9203 ± 0.4216 \| 2.2230 ± 0.4160 \| 1.3952 ± 0.0955 \| 1.1403 ± 0.0274 \| \| KT w/ $\mathbf{k}(x_1,x_2)$ \| 3.1040 ± 0.1928 \| 2.4149 ± 0.2585 \| 1.6458 ± 0.1386 \| 1.3679 ± 0.0705 \| \| KT w/ $\mathbf{k}((x_1\oplus x_2), (y_1\oplus y_2))$ \| 3.0814 ± 0.1648 \| 2.6633 ± 0.3603 \| 1.7260 ± 0.1791 \| 1.4282 ± 0.0825 \| \| KT w/ $\mathbf{k}_{\mathrm{NW}}((x_1,y_1),(x_2,y_2))$ \| 2.9838 ± 0.2056 \| 2.2423 ± 0.3736 \| 1.3799 ± 0.0808 \| 1.1363 ± 0.0267 $\newline$ Table B: Test loss for KRR estimators in Figure 2b of the attached PDF \| \| $n=2^{8}$ \| $n=2^{10}$ \| $n=2^{12}$ \| $n=2^{14}$ \| \|:------------------------------------------------------\|:----------------\|:----------------\|:----------------\|:----------------\| \| Full \| 1.3750 ± 0.0000 \| 1.0584 ± 0.0000 \| 1.0121 ± 0.0000 \| 0.9981 ± 0.0000 \| \| RPCholesky \| 2.7112 ± 0.3636 \| 1.7153 ± 0.3839 \| 1.0112 ± 0.0029 \| 0.9980 ± 0.0000 \| \| ST \| 2.9433 ± 0.3129 \| 2.5855 ± 0.3570 \| 2.1328 ± 0.3374 \| 1.5634 ± 0.1437 \| \| KT w/ $\mathbf{k}_{\mathrm{NW}}((x_1,y_1),(x_2,y_2))$ \| 2.3068 ± 0.4381 \| 1.9520 ± 0.2783 \| 1.4859 ± 0.1146 \| 1.2394 ± 0.0538 \| \| KT w/ $\mathbf{k}(x_1,x_2)$ \| 2.8103 ± 0.1677 \| 2.4297 ± 0.2559 \| 1.7077 ± 0.1433 \| 1.4974 ± 0.0885 \| \| KT w/ $\mathbf{k}((x_1\oplus x_2), (y_1\oplus y_2))$ \| 3.0312 ± 0.2431 \| 2.6189 ± 0.3281 \| 2.1086 ± 0.3178 \| 1.4590 ± 0.1206 \| \| KT w/ $\mathbf{k}_{\mathrm{RR}}((x_1,y_1),(x_2,y_2))$ \| 2.0422 ± 0.3406 \| 1.8856 ± 0.2575 \| 1.4336 ± 0.1142 \| 1.2316 ± 0.0499 \| --- Rebuttal Comment 2.1: Comment: Thank you for your clarification. - I don't get why the loss of 0.9980 becomes undefined when y is in log2 scale? - Additionally, can you provide some reasoning between the performance comparison of Cholesky and the proposed method? In table A, cholesky performs better than the proposed for most of cases, but the trend is reversed in table B. --- Rebuttal 3: Title: Response to further comments Comment: Thank you for your follow-up questions. 1. The values in Table A and B are test loss while the values in Figure 2 are excess risk. Hence a test loss of 0.9980 becomes an excess risk = test loss - 1 = -0.0020, which is not a valid value on log scale; hence the missing marker for RPCholesky in Figure 2. 2. Since the Tables are test loss, a lower value is better. So the opposite trend to what you wrote is true. In particular, in Table A, RPCholesky performs worse than our proposed method. On the other hand, RPCholesky is tends to do better than our proposed method In Table B. Overall the main takeaways from Table A/B (and Figure 2) is that our Kernel Thinning based estimator - _improves upon standard thinning for both NW and KRR estimators across all coreset sizes_ - _improves upon RPCholesky for NW estimator across all coreset sizes_ - _improves upon RPCholesy for small coreset sizes for KRR, and remains competitive (but a bit worse) in large coreset sizes for KRR._ RPCholesky’s superior performance over KT for KRR is not too surprising since RPCholesky a spectral method tuned to perform better in such settings, while NW estimator does not benefit immediately from an improved spectral approximation. We remark that these results are consistent with our original motivation of designing a unified framework for speeding up general non-parametric regression estimators using a single recipe.	null	null	Rebuttal 1: Rebuttal: We thank all the reviewers for their helpful and detailed feedback on our work. We summarize our additions as follows: $\blacktriangleright$ $\textbf{New experiment on SUSY dataset}$ - In Table 1 of the attached PDF, we have added results on the SUSY dataset ($d=18,N=5\times 10^6$). We use $4\times 10^6$ points for training and the remaining $10^6$ points for testing. - We compare our method KT-KRR ($\sigma=10,\lambda=10^{-1}$) to several large-scale kernel methods, namely RPCholesky preconditioning, FALKON, and Conjugate Gradient (all with $\sigma=10,\lambda=10^{-3}$). The parameters are chosen with cross-validation. - We show that KT-KRR achieves test MSE between ST-KRR and RPCholesky with training time almost half that of RPCholesky. These findings are consistent with our results on the California Housing dataset (Fig. 5 in the original manuscript). $\newline$ $\blacktriangleright$ $\textbf{Addtitional comparisons with RPCholesky}$ - In Fig. 2a of the attached PDF, we have added a new baseline: RPCholesky-NW, where $S_{NW}$ are the pivot points from RPCholesky (see box plots in red). - KT-NW outperforms RPCholesky-NW in terms of both test MSE and train times. While it provides a good low-rank approximation of the kernel matrix, RPCholesky is not designed to preserve averages. Moreover, RPCholesky requires $\mathcal{O}(n^2)$ time to produce a coreset of size $\sqrt{n}$, whereas KT-NW requires $\mathcal{O}(n \log^3 n)$ time. This difference is reflected in our wall-clock timings (see Fig. 3a of the attached PDF). - In Fig. 2b of the attached PDF, we have added a new baseline: Restricted KRR using RPCholesky for landmark selection (see box plots in red). - Consistent with our real-world experiments, RPCholesky-KRR outperforms KT-KRR in terms of test MSE (Fig. 2b of the attached PDF), but with slower train times (Fig. 3b of the attached PDF). $\newline$ $\blacktriangleright$ $\textbf{Insight into the choice of kernels}$ We expand the discussion of the following points: - Choice of base kernels $\mathbf{k}$: - Our theory for KT-NW permits any bounded, Lipschitz radial kernel with tails that decay at any rate—e.g., compact, sub-Gaussian, sub-exponential, and poly tails. - Our theory for KT-KRR permits any kernel satisfying PolyGrowth or LogGrowth log-covering number—e.g., finitely-times differentiable, analytic, finite-rank kernels. - Construction of supervised kernels: We explain the form of $\mathbf{k} _{NW}$ and $\mathbf{k} _{RR}$ and validate these design choices with our ablation studies (see Fig. 2 of attached PDF and Fig. 2 of original manuscript). - Kernel mis-specification: We describe several potential approaches for finding the right kernel and for deriving guarantees when the regression function $f^\star$ lies outside the RKHS $\mathcal{H}(\mathbf{k})$. - Multi-dimensional response data: Our theory in fact applies effortlessly to the setting when $f^\star$ is multi-dimensional (e.g., for multi-class classification tasks). The resulting supervised kernel looks like the sum of one-dimensional supervised kernels (or equivalently, replacing $y_1 y_2$ with $\langle \mathbf{y} _1,\mathbf{y} _2\rangle$ in $\mathbf{k} _{NW}$ and $\mathbf{k} _{RR}$ ). - Minimax rates: We discuss the minimax rates for each regression problem and although our rates our sub-optimal, we believe we this is a limitation in the analysis and that minimax rates can be achieved for some class of kernels. $\newline$ We now respond to the specific comments and questions from each review. Pdf: /pdf/dff8d30b378f0740a8cfc0f524075933781a527b.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
LION: Linear Group RNN for 3D Object Detection in Point Clouds	Accept (poster)	Summary: This paper proposes to leverage linear networks such as RWKV and Mamba to capture long-range dependencies in LiDAR-based outdoor 3D object detection, leading to relatively larger group sizes of the voxel partition. The proposed techniques include voxel merging/expanding and voxel generation. Experiments are conducted on the Waymo Open dataset and nuScenes dataset, achieving state-of-the-art performance. Strengths: - This paper is an early attempt to utilize linear RNN for outdoor LiDAR-based 3D object detection. - LION achieves state-of-the-art performance on the mainstream datasets. - LION can be built upon multiple linear RCNNs such as mamba, RetNet, and RWKV, showcasing the universality of the proposed framework. Weaknesses: - Some claims are obscure and not well supported by evidence, such as: "However, effectively applying linear group RNN to 3D object detection in highly sparse point clouds is not trivial due to its limitation in handling spatial modeling." Why is it not trivial? What is the limitation? What is spatial modeling? - The proposed techniques are not novel. The spatial descriptor is a common spconv-based module. Voxel merging and expanding are trivial. The voxel generation is new but also similar to some methods, such as the "virtual voxel" in "fsdv2: Improving fully sparse 3d object detection with virtual voxels". It would be better to conduct a discussion. - The ablation for larger group size is not sufficient. The authors should conduct experiments with different group sizes to showcase the workings of group sizes since it is a main claim. - Since the most significant advantage of utilizing linear RCNN is efficiency, the authors are encouraged to conduct a more detailed runtime evaluation to reveal the latency of each component and how the designs affect the efficiency. Technical Quality: 3 Clarity: 3 Questions for Authors: - It seems like the multiple diffused voxels can occupy a single position. How do you handle this situation? - How many new voxels will be generated? Does the voxel generation reduce the efficiency? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: Some claims are obscure and not well supported by evidence ... Before feeding voxel features into linear RNN, we need to flatten 3D voxel features into 1D sequence features. Unlike the common 3D sparse convolution operation that directly deals with 3D voxel features in 3D space, the linear RNN only processes 3D voxel in 1D space, bringing limitations in handling spatial modeling. Here, the limitation is that the linear RNN is a sequence model, which is less effective in perceiving 3D spatial information (e.g. two adjacent voxels). Spatial modeling means capturing the local 3D geometric information for each voxel. W2: The proposed techniques are not novel. - Although the 3D spatial descriptor consists of a common spconv-based module, it is crucial to help our linear RNN-based network to better capture 3D local spatial information. Therefore, the combination of 3D spatial descriptor and linear RNN operators could make each voxel feature perceive the local spatial information and long-range relation, which is important to improve detection performance. - Voxel merging and expanding operations are applied to reduce the computation cost without harming the detection performance. But we do not claim these operations are our contributions (Refer to L63-71 of the main paper). - For voxel generation, we make the first attempt to leverage the auto-regressive properties of linear RNNs (refer to Table 5 in the main paper) to generate new voxel features for 3D detection. This is different from the mentioned method FSDv2. Specifically, FSDv2 first votes center points and converts these points to virtual voxels. For virtual voxel features, FSDv2 aggregates these point features by MLP. Therefore, we could categorize it as a KNN manner for voxel diffusion (Refer to the experiments in Table 5 of the main paper). Finally, we will add this discussion in the revised paper. W3: Ablation for larger group size. Thanks. We provide the ablation studies of different group sizes in the following table. Here, we set a minimum group size of 256 for all our four LION blocks (Baseline: [256, 256, 256, 256]). We could observe that the manners with larger group size (i.e, II, III, IV, V) bring consistent performance improvement over the baseline (I). However, we carefully find that there is a drop in performance from IV to V by further enlarging the group size. This might lead to less effective retention of important information in excessively long sequences due to the limited memory capacity of linear RNNs. \|#\|Group Size\|Vehicle\|Pedestrian\|Cyclist\|mAP/mAPH (L2)\| \|:---\|:---\|:---:\|:---:\|:---:\|:---:\| \|I\|[256, 256, 256, 256]\|65.6/65.2\|72.3/65.0\|68.3/67.2\|68.8/65.8\| \|II\|[1024, 512, 256, 256]\|66.9/66.5\|74.9/69.6\|70.8/69.8\|70.9/68.6\| \|III\|[2048, 1024, 512, 256]\|66.7/66.3\|74.9/69.7\|72.2/71.2\|71.3/69.1\| \|IV\|[4096, 2048, 1024, 512]\|67.0/66.6\|75.4/70.2\|71.9/71.0\|71.4/69.3\| \|V\|[8192, 4096, 2048, 1024]\|66.5/66.1\|74.6/69.5\|71.6/70.6\|70.9/68.7\| W4: More detailed runtime evaluation. We provide the detailed latency of each part of our LION-Mamba. Considering that some linear RNN-based operators (Mamba, Retnet, or RWKV) are usually applied to replace Transformer in modeling long sequences thanks to their efficiency. Therefore, we sufficiently compare the latency with transformer-based backbone (DSVT) to illustrate the efficiency of our LION. Here, we evaluate the latency on one NVIDIA GeForce RTX 3090 with a batch size of 1. Due to the quadratic complexity of the transformer, DSVT adopts a small group size of 48 in their paper. When we increase the group size of DSVT to 256, the latency obviously increases and the corresponding memory is even unacceptable (produces about 20G GPU memory for inference and fails to train this model due to OOM). In contrast, benefiting from the high efficiency of the linear RNN in modeling long sequences, our LION could adopt a larger group size (4096) for feature interaction while maintaining acceptable latency (146.2 ms) and low GPU memory (about 3G during the inference). \|Component\|Voxel Extraction (ms)\|3D Backbone (ms)\|BEV Backbone (ms)\|Detection Head (ms)\|Total Latency (ms)\|mAPH (L2)\| \|:---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \|DSVT(official paper)\|7.5\|82.8\|28.6\|17.8\|136.7\|72.1\| \|DSVT(256 group)\|7.3\|164.4\|28.6\|17.8\|218.1\|OOM\| \|LION-Mamba\|4.3\|97.1\|27.1\|17.7\|146.2\|73.2\| \|LION-Mamba-L\|5.6\|136.0\|32.2\|17.5\|191.3\|74.0\| Furthermore, we provide a detailed ablation study on the Waymo validation set with 20% training data to analyze the effect of each component on the efficiency and performance. \|3D Spatial Descriptor\|Voxel Generation\|Latency (ms)\|mAPH (L2)\| \|:---:\|:---:\|:---:\|:---:\| \| \| \|123.2\|65.8\| \|√\| \|131.3\|68.6\| \|√\|√\|146.2\|69.3\| Q1: It seems like the multiple diffused voxels can occupy a single position. We apologize for the unclear presentation. We merge the diffused voxels and raw voxels by summing these voxels at the same position in the voxel merging operation to deduplicate voxels. We will make it clearer in the revised version. Q2: How many new voxels will be generated? Does the voxel generation reduce the efficiency? Actually, the number of new generated voxels depends on the number of input voxels and the diffusion ratio. Moreover, the voxel generation does reduce the efficiency of the whole network. To better illustrate the effect of voxel generation, we provide the ablation study on the important hyper-parameter of diffusion ratio in the following table. We find that a larger diffusion ratio brings more latency but better performance. In this paper, we set the diffusion ratio as 0.2 to trade off the performance and latency. \|ratio\|Latency (ms)\|Vehicle\|Pedestrian\|Cyclist\|mAP/mAPH (L2)\| \|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \|0\|131.3\|66.5/66.1\|74.8/69.6\|70.9/70.0\|70.8/68.6\| \|0.1\|141.0\|66.9/66.5\|75.0/69.8\|71.5/70.5\|71.1/68.9\| \|0.2\|146.2\|67.0/66.6\|75.4/70.2\|71.9/71.0\|71.4/69.3\| \|0.5\|158.9\|67.2/66.8\|75.3/70.0\|72.1/71.1\|71.5/69.3\| --- Rebuttal Comment 1.1: Comment: Dear Reviewer, Thank you very much for your valuable reviews and comments. We have carefully addressed the concerns you raised and hope that our responses satisfactorily resolve them. If you have any questions or need additional clarification after reading our rebuttal, please do not hesitate to let us know during the discussion period. Thank you once again for your time and consideration.	Summary: This paper proposes a linear Group RNN-based backbone for 3D object detection tasks. It can achieve a larger window size compared to previous transformer-based methods. A 3D spatial feature descriptor is also introduced to capture 3D spatial information. Furthermore, to address the sparsity of point clouds, this paper leverages the auto-regressive property of RNNs to generate voxels for distinguishing foreground features. Experiments on the large-scale nuscenes and waymo datasets validate the effectiveness, as well as the generalization across different linear compositional RNN operators such as Mamba, RWKV, and RetNet. Strengths: 1.While transformer-based backbone networks have demonstrated superior performance, their quadratic complexity has limited their application scenarios. This paper explores the potential of linear group RNNs as feature extraction backbones for 3D detection tasks, and presents state-of-the-art performance on large-scale datasets, which is interesting. 2. The proposed method has been shown to be effective across multiple linear group RNN operators such as Mamba, RWKV, and RetNet, demonstrating the generalization ability. 3. The paper is well-written, with a clear expression of the motivation and is easy to read. Weaknesses: 1.In this paper, the author has transformed the irregular point cloud into regular voxel representation. However, L172-175，the author claim max or average-pooling operations is not suitable to achieve downsampling or upsampling operations. This is conflicted. Furthmore, I think the motivation provided for the voxel generation approach is not sufficient, making it difficult to accept the motivation behind the proposed voxel generation method. 2.The author should report the test set results on the waymo dataset of the experimental results. In addition, the authors should also report the results on the multi-frame setting on the waymo dataset to prove the effectiveness of the method in the multi-frame setting. I want to the know the test set results on waymo dataset and multi-frame setting results. 3.The authors should report the inference time of LION so that we have a clearer understanding of the latency and resource consumption of the proposed method. 4.When will release the code for the community? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. L124, I’m not clear about the Window Partition. As far as I understand, non-empty voxels are extracted line by line into the window along the X or Y axes. Every time a sequence of 4096 lengths is filled, the remaining non-empty voxels are continued to be placed in the next window. So, how do you deal with it if the final window is not enough for 4096 voxels. Regarding the division and description of windows, it should be clearer. The current description is difficult for me to understand. Other questions pleae see the weakness. Will raise up the score after solving my concerns. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author claim the corresponding limitation in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1-1: In this paper, the author has transformed the irregular point cloud into regular voxel representation. However, L172-175, the author claim max or average-pooling operations is not suitable to achieve downsampling or upsampling operations. This is conflicted. Sorry for this misleading! Since the distribution of voxels in 3D space is sparse, the regular max or average pooling operations on a dense 2D image are not suitable for performing voxel downsampling or upsampling. W1-2: Furthmore, I think the motivation provided for the voxel generation approach is not sufficient, making it difficult to accept the motivation behind the proposed voxel generation method. Our motivation of voxel generation has two folds: 1) Voxel generation can densify some key voxel features to enhance the feature representation in highly sparse point clouds; 2) Voxel generation can mitigate the information loss from voxel merging operation, which is an effective operation to reduce the computation cost in our LION. We will make them clearer in the revised version. W2: The author should report the test set results on the waymo dataset of the experimental results. In addition, the authors should also report the results on the multi-frame setting on the waymo dataset to prove the effectiveness of the method in the multi-frame setting. I want to the know the test set results on waymo dataset and multi-frame setting results. Thanks for your nice suggestion. We provide the results with 3 frames on the val set in the following table. It could be found that our LION-Mamba-L even outperforms DSVT with 2.2 mAPH/L2, which effectively illustrates the superiority of our LION. \| Method \| Frames \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH (L2) \| \|:------------\|:-------:\|:-----------:\|:------------:\|:---------:\|:------------:\| \| SST \| 3 \| 68.5/68.1 \| 75.1/70.9 \| -/- \| -/- \| \| DSVT \| 3 \| 73.6/73.2 \| 78.2/75.4 \| 77.2/76.4 \| 76.3/75.0 \| \| LION-Mamba-L \| 3 \| 73.9/73.5 \| 81.0/78.3 \| 80.7/79.8 \| 78.5/77.2 \| Furthermore, for submitting Waymo test benchmark, it is a common trick to unify the training and validation datasets to train the model for better performance. Therefore, we need to reorganize our dataset and re-train our model for the test set. Considering the limited time, we train our LION-Mamba-L with 3 frames for only 12 epochs to save the training time. Our LION-Mamba-L achieves state-of-the-art performance on Waymo test dataset, whose corresponding results are as follows (We also provide the screenshot for results on the Waymo official website with 3 frames in the uploaded global PDF file): \| Method \| Epochs \| Frames \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH (L2) \| \|:-------------\|:------:\|:-------:\|:-----------:\|:-----------:\|:---------:\|:-----------------:\| \| CenterPoint++ \| 30 \| 3 \| 75.5/75.1 \| 75.1/72.4 \| 72.0/71.0 \| 74.2/72.8 \| \| PillarNeXt \| 36 \| 3 \| 76.2/75.8 \| 78.8/76.0 \| 71.6/70.6 \| 75.5/74.1 \| \| LION-Mamba-L \| 12 \| 3 \| 77.2/76.9 \| 82.0/79.3 \| 76.8/75.9 \| 78.7/77.4 \| W3: The authors should report the inference time of LION so that we have a clearer understanding of the latency and resource consumption of the proposed method. Thanks for your suggestion. We provide the detailed latency of our LION-Mamba and LION-Mamba-L inference time in the following table. We evaluate the latency on one NVIDIA GeForce RTX 3090 with a batch size of 1. For a more detailed discussion of the Transformer-based method DSVT in terms of latency and performance, please refer to W4 of Reviewer 9fJS. \| Method \| Voxel Extraction (ms)\| 3D Backbone (ms)\| BEV Backbone (ms)\| Detection Head (ms) \| Total Latency (ms) \| mAPH (L2) \| \|:-------------\|:----------------:\|:-----------:\|:------------:\|:--------------:\|:------------------:\|:-------:\| \| LION-Mamba \| 4.3 \| 97.1 \| 27.1 \| 17.7 \| 146.2 \| 73.2 \| \| LION-Mamba-L \| 5.6 \| 136.0 \| 32.2 \| 17.5 \| 191.3 \| 74.0 \| W4: When will release the code for the community? We will release all codes and models by September 30, 2024. Q1: L124, I’m not clear about the Window Partition. As far as I understand, non-empty voxels are extracted line by line into the window along the X or Y axes. Every time a sequence of 4096 lengths is filled, the remaining non-empty voxels are continued to be placed in the next window. So, how do you deal with it if the final window is not enough for 4096 voxels. Regarding the division and description of windows, it should be clearer. The current description is difficult for me to understand. We apologize for the unclear presentation. For the final window that is not enough for 4096 voxels, we repeat these remained voxels in this window up to 4096 voxels. We will make it clearer in the revised version. --- Rebuttal 2: Comment: Thanks for the author's rebuttal. My concerns have been solved and I raise up my score. Release code will be great for the community. --- Rebuttal Comment 2.1: Comment: Sincerely thank you for your valuable comments again! We will open source all the codes for this community.	Summary: This paper targets the problem of long-range feature interactions for point cloud detection. It proposes a window-based 3D backbone based on linear group RNN and sparse convolution. In contrast to existing Transformers methods, this work increases the group size by leveraging the linear complexity of recent Mamba and RWKV. The paper also introduces the 3D spatial feature descriptor to capture 3D local spatial information and 3D voxel generation strategy to address the sparsity of point clouds. Experiments demonstrate the efficacy of this proposed method. Strengths: $\bullet$ The problem studied in this paper is important, as the long-range relationship is critical in point cloud detection. $\bullet$ LION-Mamba achieves strong performance with low GFLOPs on widely used outdoor datasets, including the Waymo Open Dataset and nuScene. $\bullet$ The proposed 3D spatial feature descriptor and voxel generation are simple and easy to follow. Weaknesses: $\bullet$ [Novelty] My main concern is the overall limited technical contribution. The paper does not show significant differences from previous works. Additionally, it lacks a thorough discussion comparing it with existing approaches. $\qquad$(1) Model Structure. The proposed LION block uses the same encoder-decoder structure as SED block in HEDNet [1]. Furthermore, the 3D spatial feature descriptor is identical to SSR block, and the voxel merging and expanding operations are similar to RS conv. Besides, the LION layer has the same structure with DSVT block. This paper appears to merely integrate the DSVT block into SED block and replace Transformers with linear RNNs. $\qquad$(2) Window Partitions. The equal-size window partition along X/Y-axis has been widely adopted in voxel-based detectors, such as Flatformer and DSVT. $\qquad$(3) Voxel Generation: The way to “distinguishing foreground voxels without supervision” has been widely used in point cloud detection, such as SPSS-Conv [2] and VoxelNeXt [3]. $\bullet$ [Latency] While there is some analysis of computation cost (GFLOPs), it lacks a comparison of latency with state-of-the-art algorithms, such as HEDNet and DSVT. For 3D object detection in outdoor scenarios like autonomous driving, the real-time application makes the latency more critical. Besides, in Figure 1, the comparison with LION-Mamba-L is missing. $\bullet$ In line 42-44, the authors claim LION can support thousands of voxel features to establish long-range relationships. However, in Line 153-157, this paper proposes the spatial information loss issue and need additional local sparse convolution to address. In my point of view, a sequence of thousands based on window partitions must include the voxel and all its neighbors. This inconsistency raises concerns about the robustness and effectiveness of LION. $\bullet$ The linear RNN, such as Mamba, is a unidirectional model. Is it reasonable to use two single-layer RNNs in a LION Layer? Will the bidirectional Mamba, such as Vision Mamba, enhance performance? $\bullet$ [Motivation] The X/Y-axis window partition is proposed to address the feature interactions problem with group size limitation. The motivation to use X/Y-axis window partition is not clear, given the large group size. $\bullet$ [Motivation] The motivation for choosing only four different offsets in voxel generation with auto-regressive property is unclear. A voxel has many neighbors in 3D space, so why did the author choose only these four offsets? Further explanation is needed. [1] HEDNet: A hierarchical encoder-decoder network for 3d object detection in point clouds. NIPS’23 [2] Spatial Pruned Sparse Convolution for Efficient 3D Object Detection. NIPS’22 [3] VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking. CVPR’23 Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the weaknesses section. Additionally, $\bullet$ Please provide a comparison with HEDNet in Waymo, including latency, accuracy, and parameters. $\bullet$ It would be better to illustrate visualization results to support the claim LION can model long-range dependencies. $\bullet$ What is the performance of replacing the LION Layer with DSVT block (use the setting in the original paper) in the LION Block? Moreover, please provide a comparison of the latency between the two variants. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Limitations have been included Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your careful review and valuable suggestions. Here, we further address our contributions in this paper: 1) Linear RNN-based 3D detection framework: support kinds of linear RNNs (e.g., Mamba, RWKV, RetNet) to allow long-range feature interaction. 2) 3D spatial feature descriptor: compensate for the lack of linear RNNs in capturing local 3D spatial information. 3) Voxel generation: utilize auto-regressive property of the linear RNN for voxel diffusion and obtain a more discriminative feature representation in highly sparse point clouds. [Novelty] 1) Model Structure. - The encoder-decoder structure is not our contribution and we adopt it to reduce the computational costs while maintaining superior performance. We provide the experiments in the following table. \|Method\|Latency (ms)\|Vehicle\|Pedestrian\|Cyclist\|mAP/mAPH (L2)\| \|:---\|:---:\|:---:\|:---:\|:---:\|:---:\| \|LION-Mamba(w/o encoder-decoder structure)\|180.1\|66.8/66.4\|75.1/70.2\|72.1/71.1\|71.3/69.2\| \|LION-Mamba\|146.2\|67.0/66.6\|75.4/70.2\|71.9/71.0\|71.4/69.3\| - Although the 3D spatial descriptor consists of a common and simple spconv-based module, it is crucial to help our linear RNN-based network to better capture 3D local spatial information. Therefore, the combination of 3D spatial descriptor and linear RNN operators could make each voxel feature perceive the local spatial information and long-range relation, which is important to improve detection performance ((Refer to the Table 3 and 4 in the main paper)). - The structure of LION layer is a widely used structure (e.g., Swin Transformer, FlatFormer, and DSVT). We keep consistent with DSVT for convenience. 2) Window Partition. We follow the previous work FlatFormer (Refer to L130-133 in the main paper). Besides, we do not claim that the equal-size window partition is not our contribution. 3) Voxel Generation. Thanks for your valuable feedback. We would like to emphasize that our primary contribution in voxel generation lies in the first attempt of the auto-regressive capacity of the linear RNN. The part of “distinguishing foreground voxels without supervision" only serves it. We appreciate your point regarding the distinction of foreground voxels without supervision, and we will revise this part in the final version. [Latency] Thanks. We provide the comparison of different methods. We evaluate the latency on single 3090 with a batch size of 1. LION is an early attempt to adopt linear RNN to 3D object detection, and the running time needs to be optimized compared with the highly-optimized operator (e.g., sparse conv). We will improve the speed in future work for the real-time application. Please refer to the limitation (L323-325) in the main paper. For Figure 1 in the main paper, we will revise it in the final version. For a more detailed discussion of the Transformer-based method DSVT, please refer to W4 of Reviewer 9fJS. \|Method\|mAPH (L2)\|Latency (ms)\|params (M)\| \|:---\|:---:\|:---:\|:---:\| \|HEDNet\|73.4\|74.8\|11.9\| \|DSVT\|72.1\|136.7\|8.7\| \|LION-Mamba\|73.2\|146.2\|8.6\| \|LION-Mamba-L\|74.0\|191.3\|16.1\| [Motivation] 1) Motivation to use X/Y-axis window partition. For linear RNN models, different sequence arrangements lead to varying feature interactions. More arrangements can improve feature richness, so we use X/Y-axis window partitioning to generate diverse sequences. 2) Why did the author choose only these four offsets? Setting more different offsets usually results in producing more voxels, which brings more computation costs for the following LION layers. To reduce computation cost, we only consider the offset of each voxel for diffusion in BEV space and ignore the offset along the Z axis. Here, we simply choose these four offsets along the two diagonals in BEV space. [Weakness] 1) A sequence of thousands based on window partitions must include the voxel and all its neighbors ... Although a sequence with thousands of voxels usually contains the voxel features of their neighbors, two adjacent voxels might be far apart in this sequence (please refer to Figure 4 in the main paper) since the linear RNN extract features along the sequence. Therefore, we adopt an additional local sparse convolution (or 3D spatial feature descriptor) to address this problem. 2) Bidirectional modeling. Sorry for missing this detail! In this paper, all linear RNNs (Mamba, RWKV and RetNet) adopt the bidirectional manner for better feature interaction. We will add the details in the revised version. [Question] 1) Comparison with HEDNet Thanks. We provide the comparison with HEDNet in [Latency]. 2) Visualization of long-range dependencies. Thanks. We provide the visualization for long-range dependencies in Figure 1 in our uploaded PDF file. 3) Performance of replacing the LION Layer with DSVT block. We replace the LION layer with DSVT block to validate the effectiveness of LION, as shown in the following table. * represents the results without our 3D spatial feature descriptor and voxel generation. All models are trained with 20% data and 12 epochs. We report the official results of DSVT (I). We can observe that the performance of integrating DSVT into LION (II and IV) is lower than LION (III and V). Finally, the manners with our 3D spatial feature descriptor and voxel generation (IV and V) produce much better performance than the methods (II and III), which effectively demonstrates the importance of our proposed components. \|#\|Method\|Latency (ms)\|Vehicle\|Pedestrian\|Cyclist\|mAP/mAPH (L2)\| \|:---\|:---\|:---:\|:---:\|:---:\|:---:\|:---:\| \|I\|DSVT\|136.7\|67.2/66.8\|72.5/66.4\|70.1/69.1\|69.9/67.4\| \|II\|LION-DSVT\|122.7\|64.8/64.3\|71.0/63.6\|67.4/66.2\|67.7/64.7\| \|III\|LION-Mamba\|123.2\|66.2/65.7\|73.7/67.2\|68.7/67.6\|69.5/66.9\| \|IV\|LION-DSVT\|157.7\|66.1/65.7\|74.4/68.8\|70.7/69.8\|70.4/68.1\| \|V\|LION-Mamba\|146.2\|67.0/66.6\|75.4/70.2\|71.9/71.0\|71.4/69.3\| --- Rebuttal Comment 1.1: Comment: Dear Reviewer, Thank you very much for your valuable reviews and comments. We have carefully addressed the concerns you raised and hope that our responses satisfactorily resolve them. If you have any questions or need additional clarification after reading our rebuttal, please do not hesitate to let us know during the discussion period. Thank you once again for your time and consideration. --- Rebuttal Comment 1.2: Comment: Thanks for your detailed reply. I appreciate the clarifications, which have addressed some of my initial concerns. However, after careful consideration of your responses and the manuscript, I still have significant reservations about this paper. $\bullet$ Motivation need to be improved. Following previous Transformer-based framework, using linear RCNN in 3D detection is straightforward and simple matter. The support for linear RNN is hardly inspiring for 3D detection community. $\bullet$ The contributions and novelty are limited to this field. In response, techniques like the encoder-decoder (HEDNet), the feature descriptor (SubmConv), window partition (FlatFormer) are already widely used in 3D detection. Moreover, increasing voxel density in 3D space has proven useful in prior work. $\bullet$ Low efficiency. Lion shares the same structures as HEDNet, yet its latency is nearly three times higher. The increased computation complexity offers minimal performance gains (only +0.6 L2 mAPH). I have serious concerns about using linear RNN for outdoor 3D detection that demand real-time performance. Moreover, the authors have not addressed this critical issue but instead added voxel generation, which further increases the computation burden. I suggest the authors focus on improving the efficiency of applying linear RCNN in 3D detection, rather than merely applying various types. This would be more valuable to the community. Given these considerations, I maintain my original rating. --- Reply to Comment 1.2.1: Comment: Thank you for your patient and detailed comments. We will further clarify these comments. 1. Motivation need to be improved. - Linear RNN is an important operator for supporting larger group size for 3D object detection, since the long-range relationship is critical in point cloud detection. - In fact, directly applying linear RNN (Note that we keep the same structure except for our proposed 3D spatial feature descriptor and voxel generation.) can only achieve a poor performance (66.9 mAPH/L2). When adopting the proposed 3D spatial feature descriptor and voxel generation, there is a 2.4 mAPH/L2 improvement, which demonstrates the effectiveness of our contribution. * represents the results without our 3D spatial feature descriptor and voxel generation. \| # \| Method \| Latency(ms) \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH(L2) \| \| :--- \| :---------: \| :---------: \| :-------: \| :--------: \| :-------: \| :----------: \| \| I \| LION-Mamba* \| 123.2 \| 66.2/65.7 \| 73.7/67.2 \| 68.7/67.6 \| 69.5/66.9 \| \| II \| LION-Mamba \| 146.2 \| 67.0/66.6 \| 75.4/70.2 \| 71.9/71.0 \| 71.4/69.3 \| 2. The contributions and novelty are limited to this field. - The encoder-decoder and window partition are not our contirbution. - Although the 3D spatial descriptor is a simple SubmConv, it is crucial for our linear RNN-based network to better capture 3D local spatial information, which is very important for how to make good use of linear RNN. - Increasing voxel density in 3D space is a common problem. We would like to emphasize that our primary contribution in voxel generation lies in the first attempt of the auto-regressive capacity of the linear RNN and achieves better performance compared with other methods. 3. Performance and Efficiency Performance: The performance of Waymo is relatively saturated. We think 0.6 mAPH/L2 is a relatively large improvement. We provide the performance on the nuScenes in the main paper. Besides, we additionally provide the results on Argoverse V2 dataset in this response. - nuScenes: LION can significantly outperform the HEDNet with 1.9 NDS. Besides, the latency gap between LION and HEDNet has further narrowed compared with the results on Waymo. \| Method \| Latency (ms) \| NDS \| mAP \| \| :--------- \| :----------\| :------------- \| :------------- \| \| HEDNet \| 162.5 \| 72.0 \| 67.7 \| \| LION-Mamba \| 183.8 \| 73.9 (+1.9) \| 69.8 (+2.1) \| - Argoverse V2: LION can significantly outperform the HEDNet with 4.4 mAP, leading to a new SOTA result. Besides, LION is faster than HEDNet in the large-range scenario (200m × 200m), which demonstrates the effectiveness of our LION for processing large group size point clouds. We will add the result in the revised version. \| Method \| Latency (ms) \| mAP \| \| :--------- \| :----------\| :-------------\| \| HEDNet \| 192.3 \| 37.1 \| \| LION-Mamba \| 186.6 \| 41.5 (+4.4) \| Efficiency: LION is an early attempt to adopt linear RNN to 3D object detection, which achieves SOTA performance on Waymo, nuScenes, and Argoverse V2. The running time needs engineering optimization compared with the highly-optimized operator (e.g., sparse conv). We will improve the speed in future work for the real-time application.	Summary: This paper presents the LION block, a neural component that builds up a backbone to extract 3D features with linear group RNNs for 3D object detection. The authors introduce a 3D spatial feature descriptor to extract point features, and a novel auto-regressive voxel generation method to density the foreground feature in sparse point clouds. Strengths: 1. The presentation is clear and easy to follow. 2. The proposed method surpasses previous state-of-the-art methods. Weaknesses: 1. Figure 3 (a) is confusing. Since the 3D Spatial Feature Descriptors are neural layers with learnable parameters, it would be better to represent these layers with blocks like “LION Layer”. 2. It would be beneficial to discuss the rationale for positioning the 3D Spatial Feature Descriptor after the Linear Group RNN, as placing it beforehand could potentially better preserve spatial information. 3. An ablation study on Voxel Merging and Expanding would help quantify their contributions to the framework's performance. Technical Quality: 3 Clarity: 3 Questions for Authors: Since objects-of-interest in autonomous driving is relatively small comparing to the whole scene, why long-range relation ($K=4096, 2048, …$) is important (line 133 - 135) in this scenario? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations are properly discussed in the main paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1: Figure 3 (a) is confusing. Since the 3D Spatial Feature Descriptors are neural layers with learnable parameters, it would be better to represent these layers with blocks like “LION Layer”. Thanks for your nice suggestion! We will revise "3D Spatial Feature Descriptors" as "3D Spatial Description Layer" to make it clearer in the final version. W2: It would be beneficial to discuss the rationale for positioning the 3D Spatial Feature Descriptor (3D SFD) after the Linear Group RNN, as placing it beforehand could potentially better preserve spatial information. Thanks! We provide the corresponding results of different placements as following table. For Placement 1, we place the 3D SFD after voxel merging. For Placement 2, we place 3D SFD before the voxel merging. We agree with your explanation that placing 3D Spatial Feature Descriptor beforehand could potentially better preserve spatial information. Finally, we will add this discussion of the rationale for positioning the 3D Spatial Feature Descriptor in the revised version. \| Method \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH (L2) \| \|:------------\|:---------:\|:----------:\| :-------: \|:-------------:\| \| Baseline \| 66.4/66.0 \| 73.5/67.4 \| 70.4/69.3 \| 70.1/67.6 \| \| Placement 1 \| 66.5/66.1 \| 74.8/69.1 \| 71.1/70.2 \| 70.1/68.6 \| \| Placement 2 \| 67.0/66.6 \| 75.4/70.2 \| 71.9/71.0 \| 71.4/69.3 \| W3: An ablation study on Voxel Merging and Expanding would help quantify their contributions to the framework's performance. Thanks! Note that removing the voxel merging and voxel expanding operations means that the input voxels of the linear group RNN are not processed by any downsampling or upsampling operations, which results in some additional computational cost. In the following table, we provide the experiment of our method with/without the voxel merging and voxel expanding operations. It can be observed that our method with adopting the voxel merging and voxel expanding operations could effectively reduce the computational costs while maintaining superior performance. \| Method \| Latency (ms) \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH (L2) \| \| :------------------------------------------\| :---------: \| :-------: \| :--------: \| :-------: \| :----------: \| \| LION-Mamba (w/o Voxel Merging and Expanding) \| 180.1 \| 66.8/66.4 \| 75.1/70.1 \| 72.1/71.1 \| 71.3/69.2 \| \| LION-Mamba \| 146.2 \| 67.0/66.6 \| 75.4/70.2 \| 71.9/71.0 \| 71.4/69.3 \| Q1: Since objects-of-interest in autonomous driving is relatively small comparing to the whole scene, why long-range relation (K=4096, 2048, ...) is important (line 133 - 135) in this scenario? Good question! In fact, capturing the long-range relationship is helpful to obtain richer context information for better understanding the whole scene. Therefore, it is not only important for detecting large objects, but also for detecting small objects. In this paper, we build the long-range relationship by the linear group RNN operations. We provide the ablation studies of different group sizes in the following table. Here, we set a minimum group size of 256 for all our four LION blocks (Baseline: [256, 256, 256, 256]). We could observe that the manners with larger group size (i.e, II, III, IV, V) bring consistent performance improvement over the baseline (I). However, we carefully find that there is a drop in performance from IV to V by further enlarging the group size. This might lead to less effective retention of important information in excessively long sequences due to the limited memory capacity of linear RNNs. Finally, to better understand the long-range relationship, we provide the visualization in Figure 1 (Please refer to our uploaded PDF file). We will add these discussion and experiments in the final version. \| # \| Group Size \| Vehicle \| Pedestrian \| Cyclist \| mAP/mAPH (L2) \| \|:----\|:------------------------\|:-------: \| :--------: \| :-------: \| :----------: \| \| I \| [256, 256, 256, 256] \| 65.6/65.2 \| 72.3/65.0 \| 68.3/67.2 \| 68.8/65.8 \| \| II \| [1024, 512, 256, 256] \| 66.9/66.5 \| 74.9/69.6 \| 70.8/69.8 \| 70.9/68.6 \| \| III \| [2048, 1024, 512, 256] \| 66.7/66.3 \| 74.9/69.7 \| 72.2/71.2 \| 71.3/69.1 \| \| IV \| [4096, 2048, 1024, 512] (Ours) \| 67.0/66.6 \| 75.4/70.2 \| 71.9/71.0 \| 71.4/69.3 \| \| V \| [8192, 4096, 2048, 1024] \| 66.5/66.1 \| 74.6/69.5 \| 71.6/70.6 \| 70.9/68.7 \|	Rebuttal 1: Rebuttal: We are grateful for the valuable suggestion and feedback of all reviewers, which will greatly improve the quality of our paper. We will carefully revise our paper according to your suggestions. To Reviewer ehGN and wvps: We provide the visualization of long-range dependencies in the uploaded PDF file. To Reviewer 5dKF: We provide the anonymous screenshot for testing results on the Waymo official website in the uploaded PDF file. Pdf: /pdf/6a5d24cb1a3fccd33a66cccc33e3114c14856457.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification	Accept (poster)	Summary: The paper presents an adaptive, mixed-precision quantization method for compressing KV cache in LLMs. It proposes a channel-separable tokenwise quantization scheme to establish a robust quantization baseline, reducing the memory overhead of quantization parameters. A saliency metric, based on normalized attention scores, is employed to identify salient tokens accurately. This approach preserves essential information while aggressively quantizing less salient data, allowing for higher compression ratios with minimal impact on model accuracy. The method can be integrated with FlashAttention with an efficient approximation of the proposed metric. Experimental results validate the superiority of ZipCache on speed and accuracy over existing methods. Strengths: 1. The idea of preserving salient tokens while compressing less critical tokens is both intuitive and effective, with the proposed accurate saliency metric being crucial for achieving this. 2. The paper presents a novel quantization scheme tailored for channel-wise outliers in LLMs, providing a notable reduction in memory overhead compared to group-wise quantization schemes. 3. The paper demonstrates strong experimental results on GSM8k and HumanEval, convincingly showcasing the benefits of adaptive KV cache compression. 4. The integration with FlashAttention greatly enhances overall generation speed. 5. The paper is well-organized. Weaknesses: 1. Evaluating ZipCache on other generation or comprehending tasks will make the findings more robust. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Have you considered using only the attention scores from the most recent token to determine token saliency? What impact might this have on the overall performance? 2. What does saliency mean in Figure 3c? What is the probability of each token being selected as a salient token, as mentioned in the caption of Figure 3? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks to the reviewer for the valuable comments. Q1: Evaluating ZipCache on other generation or comprehending tasks. As shown in Table C in the rebuttal PDF, we evaluate the performance of ZipCache on LongBench. The results show that ZipCache outperforms the previous state-of-the-art method, KIVI. Due to the limited time slot of the rebuttal, we will conduct experiments on more models and add the results to the revised version. Q2: Using only the most recent token to determine token saliency. As shown below, determining token saliency with only the most recent token leads to a 3.2% accuracy drop compared to ZipCache. Table: The effect of different saliency metric on GSM8k with CoT prompts. Here, "H/L" denotes the bit-width for salient tokens (high-precision) and regular tokens (low-precision), respectively. The compression ratio is calculated with an average input length of $l=840$. \| Model \| Metric \| Bit-width (H/L) \| Saliency Ratio \| Compression Ratio \| Acc. (%) \| \|-------------\|----------------------------\|-----------------\|----------------\|-------------------\|----------\| \| Mistral-7B \| FP16 \| 16/16 \| 100% \| 1$\times$ \| 41.62 \| \| Mistral-7B \| Last Token's Attention Scores \| 4/2 \| 60.0% \| 4.98$\times$ \| 38.04 \| \| Mistral-7B \| Normalized Attention Scores \| 4/2 \| 60.0% \| 4.98$\times$ \| 41.24 \| Q3: What does saliency mean in Figure 3c? There are many attention layers and heads in a model. For each token, this value is derived by counting the number of times this token is selected as a salient token among all attention heads, then normalizing this count by the total number of attention heads. --- Rebuttal Comment 1.1: Title: Review for rebuttal Comment: Thanks for the efforts from the authors. My concerns are addressed in the rebuttal. I think this paper is well-motivated and has its merit to the community. It demonstrates strong performance in the experiments. Therefore, I raise my score after rebuttal. --- Reply to Comment 1.1.1: Title: Appreciation for Your Valuable Feedback Comment: Dear Reviewer E63Z, Thank you for your feedback. We truly appreciate your careful consideration of our responses. Best regards, Authors of #5969.	Summary: The paper introduces ZipCache, an adaptive KV cache compression method for LLMs by accurately identifying salient tokens. It first presents a channel-separable tokenwise quantization scheme that reduces the overhead of quantization parameters. Next, it proposes a metric for identifying salient tokens based on normalized attention scores. By efficiently approximating the saliency metric, the method integrates seamlessly with fast attention mechanisms like FlashAttention. The efficacy of ZipCache is demonstrated through extensive experiments, showing superior performance in terms of compression ratio, speed, and minimal accuracy loss compared to existing methods. Strengths: 1.The introduction of normalized attention scores as a metric for token saliency is significant and promising for adaptive KV cache compression. It can accurately identify salient tokens, marking a substantial improvement over prior methods. 2.The approximated saliency metric can be integrated with FlashAttention, making it practical for real-world applications. 3.The idea of the channel-separable tokenwise quantization approach is novel and well-motivated, reducing memory overhead compared to groupwise counterparts. 4.The experimental results are promising, clearly demonstrating the efficacy of ZipCache in terms of compression ratio, speed, and minimal accuracy loss. 5.The paper is clearly written and easy to follow. Weaknesses: 1.The experiments should encompass a broader diversity of tasks to comprehensively assess the method's effectiveness, such as LongBench. 2.The font size in Figures 5 and 6 is too small, making them difficult to read. It is recommended to increase the font size for better clarity and accessibility. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: See weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks to the reviewer for the valuable comments. Q1: Evaluate ZipCache over LongBench. As shown in Table C in the rebuttal PDF, we evaluate the performance of ZipCache on LongBench. The results show that ZipCache outperforms the previous state-of-the-art method, KIVI. Due to the limited time slot of the rebuttal, we will conduct experiments on more models and add the results to the revised version. Q2: The font size in Figures 5 and 6 is too small. Thanks for your valuable comment. We will revise it in the final version. --- Rebuttal Comment 1.1: Comment: Thank the authors' detailed response. After reviewing all comments and responses, my concerns are addressed well. The proposed method is novel and effective, and I believe this paper is ready to be published. So I would like to raise my score to 8. --- Reply to Comment 1.1.1: Title: Appreciation for Your Valuable Feedback Comment: Dear Reviewer S3gk, Thank you for your feedback. We truly appreciate your careful consideration of our responses. Best regards, Authors of #5969.	Summary: This paper proposes a post training quantization framework named ZipCache for quantizing the key and value cache of LLMs. The authors introduce a channel-separable token-wise quantization scheme, which consumes less memory than the group quantization in terms of the quantization parameters. They also select saliency tokens based on a normalized attention matrix to avoid the accumulated bias. Additionally, the authors introduce an approximation method to apply it to FlashAttention. Strengths: - Overall, the writing of the paper is clear and easy to follow. Especially, the figures are drawn well. - By proposing the ZipCache, this work promotes the PTQ quality of LLMs on three chanllenging datasets. - The methods are well-integrated into popular framwork like FlashAttention. Weaknesses: - The idea of channel-separable quantization is not novel. It is a common practice to smooth the outliers of activations in the channel dimension, e.g., SmoothQuant [1], OmniQuant [2]. - As shown in Table 3 (a), the values on the diagonal of the matrix are relatively large, while the values at other positions are relatively small. Based on this circumstance, if the mean operation is performed for none-zero positions, earlier tokens may have smaller $p_i$ for they have more tiny values counted. Therefore, it may cause a prefer to the latest tokens. I think this is a issue for the proposed technique in Section 4.2, which needs more analysis. - Ablation studies are not provided. - As stated in Line 268 in the paper, new tokens remain un-quantized unless they reach 100. As far as I know, keeping part of tokens in full precision can improve the accuracy [3, 4]. Therefore, it is better to distinguish the contributions of this implementation and the proposed techniques. - I think the value of 'H/L' should be carefully considered. This is because both KIVI and ZipCache keep new tokens in FP16 at first and then perform quantization. Since the 'H' value of KIVI is 16, why it becomes 4 for ZipCache? - The equation (5) is not accurate for not subtracting the zero-point 'z' in the de-quantization. [1]. SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. [2]. Omniquant: Omnidirectionally calibrated quantization for large language models. [3]. KIVI: ATuning-Free Asymmetric 2bit Quantization for KV Cache. [4]. SKVQ: Sliding-window Key and Value Cache Quantization for Large Language Models. Technical Quality: 3 Clarity: 3 Questions for Authors: - The authors have compared the channel-separable quantization with many dynamic methods, but lacks comparison of accuracy with static methods, e.g., WKVQuant [1]. It is better to further explore whether the $c_i$ in equation (6) can outperform well-trained parameters, since they consumes the same memory space. - How does the calculation of $c_i$ in equation (6) effect the inference latency? It is better to compare the channel-separable quantization with group quantization and static methods in terms of the inference latency. [1]. WKVQuant: Quantizing weight and key/value cache for large language models gains more. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: This paper contains the describtion of limitations. For suggestions, please refer to the Weakness and Questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks to the reviewer for the valuable comments. Q1: The idea of channel-separable quantization is not novel. Please refer to General Response Q1. Q2: Performing the mean operation may cause a prefer to the latest tokens. Indeed, there might be a prefer to the latest tokens since the values on the diagonal of the matrix are relatively large. However, the overall token saliency is adaptively determined in our method. As shown in Figure A in the rebuttal PDF, we present the token saliency with a 100-line retrieval sample as input. Besides the latest tokens, the system prompts in the front and related context in the middle are also assigned high saliency by our method. Moreover, Table B in the rebuttal PDF shows that our method achieves superior performance compared to using accumulated attention scores or a fixed prefer to the latest tokens. Q3: Ablation studies are not provided. As referred to Table 1 and Table 2 in the paper, we have conducted ablation experiments to demonstrate the effectiveness of channel-separable quantization scheme and the efficient approximation of the saliency metric, respectively. We further conduct ablation studies on our saliency metric, as shown in Table B in the rebuttal PDF. It demonstrates the superiority of our saliency metric over using accumulated attention scores or consistently prioritizing the latest tokens. Q4: Remaining new tokens un-quantized can improve accuracy. During the prefill phase, KIVI [i] maintains full-precision KV caches for recent tokens to ensure accuracy, while our approach adaptively quantizes all KV caches. During the decoding phase, we quantize all KV caches with mixed-precision every 100 tokens, while KIVI adopts a sliding window and always keeps KV caches for latest tokens (128 tokens by default) in full-precision. This implies that KIVI always retains more KV cache at full precision compared to our method. Moreover, the streaming strategy we adopted aligns with GEAR [ii] and is aimed at enhancing decoding speed. Therefore, the comparison with KIVI and GEAR fairly demonstrates the efficacy of our method. Q5: The value of `H/L' should be carefully considered. As referred to Q4, during the prefill phase, we quantize all KV caches, with the bit-width for salient tokens set to 4. This approach contrasts with KIVI's method of maintaining full-precision (FP16) for the recent tokens. Q6: The equation (5) is not accurate. Thanks for your valuable comment. Equation (5) has been revised as follows: $\hat{\mathbf{x}}=\mathcal{Q}_U(\mathbf{x},k)=(\mathrm{clip}(\lfloor \frac{\mathbf{x}}{s}\rceil +z, 0, 2^{k}-1) - z) \cdot s.$ Q7: Comparisons between channel-separable quantization and static methods like WKVQuant. Due to different settings, ZipCache and WKVQuant [iii] are not directly comparable. WKVQuant quantizes model weights and KV cache together, optimizing parameters through cross-block reconstruction regularization and a gradient descent algorithm. In contrast, our method focuses solely on KV cache compression and is entirely training-free. Q8: Compare the channel-separable quantization with group quantization and static methods in terms of the inference latency. Inference latency comparisons among channel-separable quantization, groupwise quantization, and static quantization methods are detailed in Table A of the rebuttal PDF. Using static quantization can slightly reduce latency (2556.03 ms vs. 2584.01 ms). Conversely, groupwise quantization increases inference speed (2664.05 ms vs. 2584.01 ms) and reduces the compression rate (3.81$\times$ vs. 4.43$\times$) due to massive quantization overhead. [i] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache. [ii] GEAR: An Efficient KV Cache Compression Recipe for Near-Lossless Generative Inference of LLM. [iii] WKVQuant: Quantizing Weight and Key/Value Cache for Large Language Models Gains More. --- Rebuttal Comment 1.1: Title: Review for Rebuttal Comment: Thanks to the authors for the detailed response. I understand the motivation of the channel-seperable token-wise quantization, as well as the effectiveness of the token saliency metric after the rebutall. The integration of this kind of methods to flash-attention is also a valuable contribution. Therefore, considering the answers from Q1 to Q5, I decide to raise my score to 5. Although, I think the description of the Section 4.1 can be improved, especially from the aspect of accuracy and latency. As claimed to be a more advanced baseline, it should process better generalization ability compared to group quantization and well-trained static parameters, rather than solely performing well on a single dataset. Additionally, I think the answers to Q6 and Q8 should be included into the main text. --- Rebuttal 2: Title: Follow-Up on Rebuttal Comment: Dear Reviewer Nc8R, We greatly appreciate the time and effort in reviewing our work. We have carefully considered your comments and suggestions and made significant revisions to address the concerns you raised. We are eager to ensure that our paper meets the high standards of our respected reviewers. Please don’t hesitate to let us know if there is any additional feedback you might have at this stage. Best regards, Authors of #5969. --- Rebuttal 3: Title: Follow-Up on Rebuttal Comment: Dear Reviewer Nc8R, Thank you for dedicating your time to reviewing our paper. As the discussion period deadline is approaching, we kindly invite any further comments or concerns you might have. Your feedback has been immensely valuable to us for refining the paper. Best regards, Authors of #5969. --- Rebuttal 4: Title: Appreciation for Your Valuable Feedback Comment: Dear Reviewer Nc8R, Thank you for your feedback. We truly appreciate your careful consideration of our responses and will carefully revise the paper based on your and other reviewers' suggestions. Best regards, Authors of #5969.	Summary: The paper introduces a channel-separable quantization scheme that decouples the quantization along channel and token dimensions. This method significantly reduces the quantization overhead without compromising performance. To accurately recognize salient tokens, the paper introduces a new token saliency metric based on normalized attention scores, which alleviates the bias towards earlier tokens that accumulate more values. The authors have demonstrated on different LLM models with different tasks. Strengths: The idea of channel separable token saliency quantization seems to work better than standard groupwise quantization. The results are strong and beats the SoTA! The demonstration with flashattention integration is useful Weaknesses: The paper needs to be proof read by a native English speaker, to improve its readability. The contribution is incremental, and the idea of group quantization (different for different K and V) are not new as already noted by the authors. The idea of speeding up through flashattention and flashdecoding while the baselines not leveraging that is a bit of unfair comparison. Both GEAR and KIVI can exploit similar speed up benefit, thus Fig 1 comparison and later in table speed comparison is not fair. The details of system level implementation is missing, it will be good to help the reviewer provide how the speed up can be demonstrated on a single GPU inference system. Technical Quality: 3 Clarity: 2 Questions for Authors: refer to weakness. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The system benefits are not clear or thorough for the reviewer to clearly appreciate. Also more fair system eval is needed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks to the reviewer for the valuable comments. Q1: The writing can be more native. Thanks for your valuable comment. We will carefully proofread the paper in the final version. Q2: The idea of group quantization (different for different K and V) is not new. There might be some misunderstandings regarding groupwise quantization. Groupwise quantization is a scheme built upon tokenwise quantization, further processing outlier channels in distinct groups, as illustrated in Figure 2(c) of the paper. Additionally, applying different quantization schemes for K and V respectively (channelwise quantization for K and channel-separable tokenwise quantization for V) is not the key contribution of our paper. In terms of the quantization scheme, our main contribution lies in the introduction of channel-separable quantization to the KV cache compression, significantly reducing the overhead of quantization parameters without sacrificing performance, as detailed in General Response Q1. Q3: The contribution is incremental. Please refer to General Response Q1 for the detailed contribution of our channel-separable quantization scheme and to General Response Q2 for an elaboration on the overall contribution of our paper. Q4: The speed comparison is not fair. Thank you for your valuable comments. Firstly, we highlight that we are the pioneering work that enables adaptive KV cache compression to be compatible with FlashAttention, greatly enhancing the generation speed compared to previous approaches like H2O [i] and MiKV [ii]. Moreover, the reported speed results of GEAR [iii] and KIVI [iv] were obtained with their official implementations, which did not integrate FlashAttention at the time of our paper submission. To ensure a fair comparison, we implement GEAR and KIVI with FlashAttention integration, and the results are shown in Table A in the rebuttal PDF. Notably, the latency of ZipCache is lower than that of GEAR, which can be attributed to our efficient quantization scheme, whereas GEAR has a high overhead due to outlier extraction. Compared to KIVI, ZipCache's latency is slightly higher (2584.01 ms vs. 2482.26 ms), but ZipCache achieves a higher compression ratio and better performance. This difference is due to KIVI's fixed compression strategy, while we adaptively compress the KV cache based on the saliency. These results will be revised in the final version. Q5: The details of system level implementation are missing. The detailed processes of ZipCache for both prefill and decoding phases are summarized in Algorithms 2 and 3 in the Appendix of our paper. Currently, we implement ZipCache based on the Hugging Face Transformers library, with specific modifications to the KV cache module. Our method is also orthogonal to other system level frameworks such as vLLM [v]. It should be noted that we utilize FlashAttention to maximize computational efficiency for both the prefill and decoding phases, eliminating the need to customize additional matrix multiplication kernels. We will release the source code upon acceptance. [i] H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models. [ii] No Token Left Behind: Reliable KV Cache Compression via Importance-Aware Mixed Precision Quantization. [iii] GEAR: An Efficient KV Cache Compression Recipe for Near-Lossless Generative Inference of LLM. [iv] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache. [v] Efficient Memory Management for Large Language Model Serving with PagedAttention. --- Rebuttal Comment 1.1: Title: Response. Comment: Thanks to the authors for the new results with attached pdf. Particularly, I appreciate the fair pre-fill stage comparison with all quantization schemes leveraging flashattention. 1. While I appreciate the author's effort, I think further clarifications are needed. In specific, the KV compression can be useful for longer context based evaluation, the current manuscript including that in the one in rebuttal pdf. The LLaMA2 model has pretraining context length of ~4k which is not sufficient enough for the long context demonstration. Beyond long context evaluation the benefits of KV cache compression is not that usefully demonstrable. And the main manuscript demonstrates results with around 900 and 200 tokens on average as said by the authors. 2. The idea of normalized attention score is not new either, so I believe their is a bit of overclaim here. Please refer to this paper, that proposed something similar, namely MAS [1]. 3. Why the generation efficiency is compared with only MiKV, instead of other SoTA quantization methods like KIVI? It is understandable that MiKV falls short as it cant leverage flash-decoing/attention, this is again unfair comparison. 4. It is also not clear if the probe tokens are selected randomly how it will still help manage perform flashattention or flashdecoding efficiently. Also, during decode phase do you implement flashattention or flashdecoding? please provide more details. 5. Overall, I believe the paper is built on top of the main findings of probe tokens, however the details in that front is yet to be sufficient. Additionally, how the selection of probe tokens help guide the bit width assignment of salient tokens if a token does not have any representative in the randomly selected probe token set? 6. The algorithm of decode stage seems ambiguous, or not fully informative. For example, the new tokens getting added on top of previous KV, is it the hybrid H/L quantized KV or or some high precision KV? As we may need to compute attention and softmax etc at high precision to have numerical stability. This needs clarification. 7. It looks like the probe tokens during decode phase are kept at higher precision, did the author include that in their total compression scheme? Also, the probe token computation and memory overhead is not clearly discussed in the paper's results. This should impact accuracy as well as throughput compared to schemes like KIVI. [1] On the Efficacy of Eviction Policy for Key-Value Constrained Generative Language Model Inference, 2024. --- Rebuttal 2: Title: Follow-Up on Rebuttal Comment: Dear Reviewer ozSB, We greatly appreciate the time and effort in reviewing our work. We have carefully considered your comments and suggestions and made significant revisions to address the concerns you raised. We are eager to ensure that our paper meets the high standards of our respected reviewers. Please don’t hesitate to let us know if there is any additional feedback you might have at this stage. Best regards, Authors of #5969. --- Rebuttal 3: Title: Follow-Up on Rebuttal Comment: Dear Reviewer ozSB, Thank you for dedicating your time to reviewing our paper. As the discussion period deadline is approaching, we kindly invite any further comments or concerns you might have. Your feedback has been immensely valuable to us for refining the paper. Best regards, Authors of #5969. --- Rebuttal 4: Title: Response to additional questions (Part 1) Comment: Thanks to the reviewer for the valuable comments. Q1: The KV compression can be useful for longer context based evaluation. Firstly, in alignment with prior literature [1-4], we have evaluated our method on three challenging and widely-recognized benchmarks, including the long context Line Retrieval task. As highlighted in Table 3, our approach consistently surpasses previous state-of-the-art methods, underscoring its effectiveness. Secondly, as mentioned in lines 36-38 of our paper, during batch inference, the KV cache with an input length of approximately 4K is already a significant bottleneck in the storage system. For example, the KV cache can occupy 1.2TB of memory space with a batch size of 64 and an input length of 4096. Moreover, to further address your point, we have also evaluated the performance of ZipCache on LongBench using the longchat-7b-v1.5-32k model, as shown below. Due to the limited time slot of the rebuttal, we will conduct experiments on more long-context tasks and add the results to the revised version. Table: Performance comparisons on LongBench with LongChat-v1.5-7B-32k. The saliency ratio is 60% for ZipCache. \| Model \| Method \| TREC↑ \| SAMSum↑ \| LCC↑ \| RepoBench-P↑ \| \|------------------------\|---------\|-------\|--------\|-------\|-------------\| \| LongChat-v1.5-7B-32k\| FP16 \| 66.06 \| 41.19 \| 52.96 \| 56.8 \| \| \| KIVI-2 \| 66.0 \| 40.57 \| 47.99 \| 52.6 \| \| \| ZipCache\| 66.0 \| 40.64 \| 51.25 \| 52.87 \| Q2: Similar idea of normalized attention scores has been proposed. Thank you for sharing this literature [5] and we will include it in the references in the revised version. However, we would like to emphasize that our proposed saliency metric can be applied universally to all tokens without exception, whereas the approach in [5] excludes certain tokens from the eviction scope based on their standard deviation. Additionally, our work provides a comprehensive analysis of the limitations of using accumulated attention scores as a saliency metric, as discussed in lines 200-210 and illustrated in Figure 3 of the paper. Q3: The generation efficiency is compared with only MiKV and the comparison is unfair. As referred to Figure 1 in the paper, Table A in the Appendix and Table A in the rebuttal PDF, we have compared the generation efficiency with H2O, GEAR and KIVI. Moreover, the previous adaptive KV cache compression methods [1-2] are not compatible with FlashAttention, which is a major drawback of them. By contrast, our method integrates seamlessly with FlashAttention, enhancing generation speed. Q4: How do probe tokens help perform flashattention or flashdecoding efficiently? Firstly, as referred to lines 230-232 and Equation (9) of the paper, we select a small set of probe tokens and compute attention scores using their queries, which allows us to approximate the saliency for all tokens. Once we have determined the token saliency, we can efficiently compute the attention output using FlashAttention. For a more detailed explanation of our method, please refer to Algorithm 2 and Algorithm 3 in the Appendix. Q5: Do you implement flashattention or flashdecoding during decoding phase? Yes. As mentioned in lines 267-268, 304-307, and detailed in Algorithm 3 in the Appendix, we implement FlashDecoding during the decoding phase. Similar to the prefill phase, we explicitly compute attention scores for a small set of probe tokens to approximate the saliency of all tokens, enabling the majority of tokens to be computed using FlashDecoding. Q6: How does the selection of probe tokens help guide the bit width assignment of salient tokens if a token does not have any representative in the probe token set. As mentioned in lines 254-255 of the paper, our approach to selecting probe tokens involves a hybrid strategy, where 5% of the tokens are the latest ones, and 5% are randomly selected. All previous tokens are attended when computing attention scores for the latest token. Q7: The algorithm of decode stage seems ambiguous. For example, what is the precision of the KV cache for new tokens. As referred to lines 267-268 of the paper, we implement a streaming strategy during the decoding phase, where the new KV cache is quantized every 100 new tokens generated. This approach is consistent with the strategy used in GEAR [3] and is designed to enhance decoding speed. Consequently, before reaching the 100-token threshold, the new KV cache is maintained in full precision. After 100 new tokens are generated, all KV cache will be quantized based on their estimated saliency. Q8: Attention and softmax need to be computed at high precision to have numerical stability. Indeed. The KV cache will be dequantized to full-precision when calculating attention. This is consistent with previous KV cache quantization work [2-4]. --- Rebuttal 5: Title: Response to additional questions (Part 2) Comment: Q9: The probe tokens during decoding phase are kept at higher precision. This is not the case. All KV caches, including those for probe tokens, are quantized after every 100 new tokens are generated. As referred to Q5, during the decoding phase, the primary difference between probe tokens and other tokens is that we use the queries of probe tokens to compute attention scores explicitly, while other tokens are processed using FlashDecoding. Q10: Keeping probe tokens at higher precision will impact accuracy as well as throughput. Please refer to Q9. Probe tokens are not kept in higher precision. Q11: What is the computation and memory overhead for probe tokens? As shown below, computing attention scores with the queries of probe tokens introduces limited computation and memory overhead. Table: Computation and memory overhead for probe tokens. Here, "ZipCache w/o Probe Tokens" denotes the token saliency is randomly generated rather than approximated with probe tokens. Data is collected by serving LLaMA3-8B on a NVIDIA A100 GPU with a batch size of 8 and sequence length of 3072. \| Model \| Method \| Prefill-phase Latency (ms) \| Max GPU Memory (MB)\| \|------------------------\|---------\|-------\|--------\| \| LLaMA3-8B \| ZipCache w/o Probe Tokens\| 2503.06 \| 34990 \| \| \| ZipCache w/ Probe Tokens\| 2584.01 \| 34992 \| [1] H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models. [2] No Token Left Behind: Reliable KV Cache Compression via Importance-Aware Mixed Precision Quantization. [3] GEAR: An Efficient KV Cache Compression Recipe for Near-Lossless Generative Inference of LLM. [4] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache. [5] On the Efficacy of Eviction Policy for Key-Value Constrained Generative Language Model Inference. --- Rebuttal Comment 5.1: Title: Final remark Comment: Thanks to authors for their comprehensive response! Overall majority of my concerns seems to be addressed by the authors. However, regarding Q3, I would recommend the authors to change their wordings as it is not true that majority of the quantization can not support flashattention. Flashattention is an orthgonal method that can be merged with most of the quantization schemes. Please tone down this claim about others don't supporting it. For Q4, if so, then please explicitly mention in L307 of the manuscript that flashattention during prefill and decoding during decode, despite this being intuitive (as flashattn does not provide much benefit during decode phase). I thus increase my score. --- Reply to Comment 5.1.1: Title: Appreciation for Your Valuable Feedback Comment: Dear Reviewer ozSB, Thank you for your feedback. We truly appreciate your careful consideration of our responses and will carefully revise the paper based on your and other reviewers' suggestions. Best regards, Authors of #5969.	Rebuttal 1: Rebuttal: We thank all reviewers for their valuable feedback. Overall, our work has been well recognized as - "It is novel and well-motivated" (Reviewers S3gk and E63Z) - "It is well-integrated into FlashAttention" (All Reviewers) - "It is clear and easy to follow" (Reviewers Nc8R, S3gk, and E63Z) - "It demonstrates strong performance" (Reviewers ozSB, S3gk, and E63Z). We have summarized and addressed the main concerns as follows: Q1: The proposed channel-separable quantization is not novel. We highlight a significant challenge in KV cache groupwise quantization: the number of quantization parameters grows linearly with the product of sequence length and hidden dimension, as referred to Table 1 in the paper, which greatly impacts the KV cache compression ratio. This crucial issue is overlooked in previous literature and motivates us to disentangle quantization along channel and token dimensions to reduce quantization overhead. Therefore, our motivation is fundamentally different from model quantization such as SmoothQuant [i] and OmniQuant [ii], which focus on migrating quantization difficulties between activations/weights or query/key states before matrix multiplications. As referred to Table 1 in the paper, our scheme significantly reduces the overhead of quantization parameters and achieves superior performance compared to groupwise quantization. Q2: The contribution of our paper. Our contributions are summarized as follows: 1) We propose a channel-separable quantization scheme to reduce the overhead of quantization parameters. It brings a higher compression ratio and achieves superior performance compared to groupwise quantization. 2) We introduce an accurate metric to identify salient tokens and adaptively quantize all KV caches based on their saliency, thereby improving the overall compression ratio. 3) ZipCache is the pioneering work that enables adaptive KV cache compression to be compatible with FlashAttention, significantly enhancing the generation speed and practicality. Overall, as referred to Table 3 in the paper, ZipCache achieves the highest KV cache compression ratio and the highest accuracy compared to predecessor methods, demonstrating the efficacy and contribution of our method. Q3: Additional experiments on LongBench. As shown in Table C in the rebuttal PDF, we evaluate the performance of ZipCache on LongBench. The results show that ZipCache outperforms the previous state-of-the-art method, KIVI [iii]. Due to the limited time slot of the rebuttal, we will conduct experiments on more models and add the results to the revised version. [i] SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. [ii] Omniquant: Omnidirectionally Calibrated Quantization for Large Language Models. [iii] KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache. Pdf: /pdf/7644c938474db91f1d748cf24f6a1d44cd4b0026.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Perceiving Longer Sequences With Bi-Directional Cross-Attention Transformers	Accept (poster)	Summary: The paper proposes Bi-Directional Cross-Attention Transformers, a new architecture that aims to reduce the computational complexity of self-attention in transformers. The authors claim linear complexity with respect to sequence lengths. This is achieved by replacing the query embeddings in self-attention with a fixed length sequence of learned embeddings. The authors perform ablations and evaluations on ImageNet 1k to demonstrate the effectiveness of the new model. Strengths: The paper addresses the efficiency of Vision Transformers. With the increasing size of current models and the ongoing trend that larger models perform better, it is of great importance to the research community to improve the quality / size trade-off. I really appreciate the quality of the presentation of the paper. The proposed method is clearly presented and the paper very well written. The paper includes the right amount of technical depth to follow. The proposed method is a creative solution to a long standing issue. Weaknesses: The main limitation of the paper is the limited evaluation. Specifically, the paper focuses its evaluation on Transformers with short sequence lengths. This is a bit disappointing, since the main benefit of the proposed method is with respect to sequence length and its impact on the complexity of self-attention. I wish there were experiments with 10k+ sequence lengths. Along similar lines, the paper focuses mainly on the visual task of image classification, which tends to use relatively short sequences. It would be great to see experiments in the language domain, where much longer sequences are commonly observed. Technical Quality: 3 Clarity: 3 Questions for Authors: Long sequence length is mainly an issue in language models with large sequence lengths. Could you add a discussion on the, whether the proposed model could also work for language models. Specifically, does the proposed model work with causal attention, and if yes how? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper at hand has a sufficient discussion of limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your helpful suggestions, and address your points in the following. [P1] - Focus on 'short' sequences: We would like to point out that the arguably most-common sequence lengths in vision tasks like classification are around 197 tokens (224x224 w/ patch size 16x16 + cls token). → Our experiments significantly go beyond this: classification up to 9216 (384/p4), a sequence 47x as long. → Our results further show that this processing at a more fine-grained level can help to significantly boost performance (Table 2) -- something that 'conventional' (vision) Transformer models are unable to do at this given computational budget; (also see the 28% reduction in FLOPs and 8.4x higher throughput in Sec 3.5/Table 3 for lengths 4k-8k on language-like tasks.) While trying 10k+ would be interesting, we believe that the presented results in terms of performance improvement and efficiency do constitute a valuable contribution that we hope sparks further research into the quality / size trade-off of models (as you also outlined in your review). --- [P2] - Vision = short sequences; Long sequences mainly in language: While we agree that longer context lengths have been a popular topic in NLP, we would like to highlight that it is actually just as important in other domains like vision/general perception. → For applications like autonomous driving or flying, high resolution can be crucial to detect objects in one's path and avoid critical incidences! Note that even an 'older' HD1080 image results in sequences of 8100 tokens (p16) or 32400 (p8). → The current 'short-sequence' nature of images is mainly a decision made by the community during the definition of popular datasets or specific tasks; which we expect to change in the future, given recent developments in perception/capturing methods (e.g. 4K, 8K...) Our work demonstrates that higher resolution is beneficial even in classification tasks (Table 2, bottom;) as well as segmentation (Table A6), and we hope our paper provides valuable insights and an architecture that enables future research into more specialized applications in these and other domains. --- [P3] - Language & causal masking: - Language in general: BiXT can be seen an encoder-based approach (similar to BERT-like models), and we expect it to therefore be applicable to similar tasks that require understanding and modelling of the whole sequence (e.g. full sentences) -- which is what we demonstrate to a certain extent in Section 3.5 / Table 3 on the two LRA tasks. - Causal masking: As BiXT circumvents the expensive token self-attention via the bi-directional mechanism, causal masking in the sense it is used for decoder-only methods on generative language tasks is not directly applicable to our architecture; when simply masking cross-attention, information of later tokens would be able to 'leak' to earlier ones via the latent refinement. One possibility to enable causal reasoning in this setup could be to assign groups of tokens to specific latents by masking the bi-directional cross-attention accordingly, combined with causal masking on the latent self-attention -- so that later groups can see earlier ones, but not inversely. (This would, however, reduce the processing resolution of the latent/token interaction to certain groups during training); → Given that the focus of this work has been mainly on perception tasks centered around encoding, we have not run experiments in this direction, an therefore cannot make a confident prediction how well it would perform. We thank you for pointing this out as we see it as an interesting possibility for future work building on BiXT, and we will add a discussion of this to the paper's appendix (extended) as well as the limitation section. --- --- We hope our answers addressed all your questions and concerns. If you have any further queries, please do not hesitate to reach out. --- Rebuttal Comment 1.1: Title: Final review Comment: I would like to thank the authors for their response to both my questions and the questions by the fellow reviewers. After going over the other reviews and considering all the answers, I still believe that the contributions and novelty of the paper are sufficient to pass the bar for acceptance. --- Rebuttal 2: Title: Thank you for the feedback and appreciation of our work Comment: We would like to thank you again for your valuable feedback and for supporting our work!	Summary: The paper presents a novel Transformer architecture called BiXT (Bi-directional Cross-Attention Transformers) that efficiently processes longer sequences like point clouds, text, or images while maintaining competitive performance across various tasks. The BiXT model is inspired by the Perceiver architecture but replaces iterative attention with a bi-directional cross-attention module. This module allows simultaneous attention between input tokens and latent variables, leveraging an attention symmetry between the two. BiXT scales linearly in terms of computational cost and memory consumption, making it suitable for processing longer sequences. The BiXT model achieves competitive performance on tasks such as point cloud part segmentation, semantic image segmentation, image classification, hierarchical sequence modeling, and document retrieval. Strengths: 1. The linear scaling of computational cost with input size is a significant advantage, allowing the model to handle larger datasets and longer sequences more effectively than traditional Transformers. 2. BiXT can incorporate modality-specific components in a plug-and-play fashion, improving results while maintaining generality across different input modalities. Weaknesses: In Figure 1, why do the main improvements come from replacing iterative with sequential, rather than the proposed bi-directional structure? Also, why the FLOPs of bi-directional structure is larger than the sequential one? Technical Quality: 3 Clarity: 3 Questions for Authors: See the weakness Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your helpful review and address your questions in the following: [Q1] - Improvement over iterative method: As we outline in Section 3.2 and in more detail in Appendix A.4, a big improvement in performance comes due to 'unblocking' the bottleneck that exists in iterative attention methods like Perceiver. Moving from an iterative to either sequential or bi-directional approach significantly extends the effective working memory of the method as tokens are refined alongside the latents. The important advantage of our bidirectional approach over the sequential one is its increased efficiency: BiXT's bi-directional cross-attention only requires 4 instead of 6 projection matrices (2x [R,V] vs. 2x [Q,K,V]) and BiXT only computes the most-expensive attention matrix once instead of twice. As stated in lines 245-247: Contrasting the two CA-based approaches with identical numbers of layers (‘d12’) demonstrates the clear advantage of our proposed bi-directional CA, that achieved similar results but requires: - ~7% fewer FLOPs, - ~15% less memory, and - ~5% fewer parameters. --- [Q2] - FLOPs Table 1(a): In Table 1 (a), the architectural 'depth' (i.e. number of layers) is provided in parentheses behind the model name: e.g. BiXT (d12) means a 12-layer BiXT model. Note that if we compare the two models of the same depth, this would be lines 2 and 3 of the "Cross-Attn" part: - Bi-dir: 1.68 GFLOPS, 7.86M Memory, 15.12M param - Seq.: 1.81 GFLOPS, 8.54M Memory, 15.94M param → This leads to the savings in FLOPs, memory and parameters introduced by our more efficient bi-directional cross-attention stated in the previous answer (also see lines 245-247 of the paper). → This then allows us to add one additional layer (i.e. bi-dir d12 vs. seq d11) while having comparable FLOPs (1.68G vs 1.66G) and still less memory (7.86M vs 8.44M), enabling BiXT to consistently outperform the sequential approach across all our experiments while still being 7-8\% more memory efficient. (lines 248/249) We thank you for pointing out that the indication of model depth might not be clear enough, and we will make sure to additionally detail this in the Table's caption. --- --- We hope our answers have clarified all your questions. If you have any further ones, please let us know and we are happy to answer them. --- Rebuttal Comment 1.1: Title: Response to Authors Comment: I thank the authors for your positive responses, which addressed my concerns. I believe it is a good work to be accepted, while the specific rating needs further discussion with the AC and other reviewers. --- Reply to Comment 1.1.1: Title: Thank you for the feedback and your support Comment: We are happy to hear that we have addressed your concerns, and would like to thank you again for your valuable feedback and the support of our work!	Summary: This research paper presents an enhancement to the Perceiver architecture in terms of accuracy and efficiency. The key innovation is a bidirectional cross-attention module designed to iteratively stack query-to-token and token-to-query cross-attention modules, revealing a symmetry between these two attention mechanisms. Consequently, a novel bidirectional transformer architecture is introduced, which scales linearly with the number of input tokens, efficiently handling general modal input data. This replacement reduces computational costs by approximately one-third compared to iterative cross-attention stacking, while achieving higher accuracy. The improved method achieves an impressive 81.9% accuracy for classification tasks on ImageNet-1K with only 4.7G FLOPs and 5M parameters, which require only a fraction of the FLOPS compared to the original Perceiver. The paper also includes extensive experiments on more generalized input modalities, underscoring the versatility and effectiveness of the proposed enhancements to the Perceiver architecture. Strengths: - The proposed Bi-Directional Cross-Attention Transformer is novel. The mechanism of bi-directional cross-attention is analogue to processing semantics (‘what’) and location (‘where’), which makes the paper easy to follow. - The experiments on image classification, point cloud classification, and semantic segmentation show that BiXT achieves a good trade-off between accuracy and efficiency. - It is appreciated that the scaling trends are explored regarding the number of latents and dimensions. Weaknesses: - The presentation can be improved. First, the comparison between the Perceiver-IO series and BiXT should be presented in visualizations as BiXT is claimed to improve Perceiver architecture. Second, the architectural configuration of BiXT should be also visualized or specified in the table. There are many variants of BiXT in Tables 1 and 2, which makes the reviewer confused. - FLOPs may not reflect the speed of model inference. Cloud the authors provide comparisons of throughputs (images/second) of different models in Table 2. - Lack of analysis for the mechanism behind bi-direction cross-attention. The authors claim that bi-direction cross-attention is used to refine ‘what’ and ‘where. However, there is sufficient analysis and experiments to support this. Technical Quality: 3 Clarity: 2 Questions for Authors: See Weakness. Overall, this paper proposes a new transformer architecture by stacking multiple bi-direction cross-attention blocks. The experiments are sufficient. It would be better to provide more in-depth analyses. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for the helpful suggestions and address your points individually in the following. [P1] - Presentation & architectural details: - Visuals: We have included into the document attached to the global response: 1) Transition from iterative to sequential and then bidirectional attention 2) Side-by-side comparisons of BiXT's bi-directional & Perceiver's iterative attention block details → We will add this to the revised version of the paper (main or appendix, space permitting) - Configs Tab. 1 & 2: - The configuration of BiXT in Table 1 (a) is indicated in parentheses behind the model name ('d11' or 'd12'), and all experiments in Table 2 use a 12-layer architecture (d12). - As stated in lines 206/207, all of these models use 64 latent vectors, embedding dimension 192 and 6 heads, and are exclusively composed of the blocks visualised in Figure 2 (without optional token refinement). - The variants BiXT/16, /8 and /4 are all the same architecture, with the only change being the patch-size of the tokenizer. - Due to space constraints, we moved the specification of the point clouds experiments in Table 1(b) to Appendix D1. → We thank you for pointing this out, and will add a clear reference to the specification into the main paper. --- [P2] - Empirical Throughput Due to space constraints, we decided to move our analyses regarding empirical throughput to Appendix B.2, where we provide further insights regarding complexity and throughput for different sequence lengths, and contrast two different BiXT variants to three recent Vision Transformer models across token sequences from 196 (224 w/ p16) to 9216 (384 w/ p4). → The results outline BiXT's advantage of being able to efficiently scale up sequence length (i.e. image size or processing resolution in this case). We are naturally happy to include additional architectures into this comparison if you think it further elevates our work. Please also note the empirical throughput results we report in Section 3.5 on the long-sequence tasks, where our throughput is up to 8.4x faster on long sequences compared to a conventional Transformer model. --- [P3] - 'What' and 'where' As you point out, we use the analogy of 'what' and 'where' mainly to motivate how the two branches of the bi-directional architecture can be interpreted. This conceptual decomposition of the data applies to a range of tasks, especially when perceiving a scene composed of various objects (provided as 2D images, 3D point clouds, ..); but can equally be used for 1D sequences (e.g. groups of words in sentences, or hierarchical structures like in Table 3). While there is indeed no proof or guarantee that the two branches will always satisfy this concept for any type of input data -- and we use this analogy to ease interpretation and understanding for the reader (as you point out in strengths) -- there are some indications by the empirical results we obtained that support this interpretation: 1) Visualization of our image classification task show that all latents generally attend to regions that we humans perceive as belonging to 'one entity', like the building or flag. We provide additional visualizations of the bi-directional attention for all latents and different layers in the appendix in Figs A2-A4. 2) For image segmentation, we present results where we predict a local mask directly from each token with a linear layer -- which requires each token to represent the information of the particular local region it represents, i.e. 'where' things are. 3) The empirical performance of our methods obtained across tasks, which is based on this analogy: Instance-based tasks use the information in the latents ('what'), whereas dense tasks like segmentation use the tokens to provide region-specific output ('where') -- allowing BiXT to obtain competitive results. If you have the feeling we have overstated this aspect of our work, we are naturally happy explicitly highlight that this might not apply to all cases and should rather be used as an analogy that is empirically supported for select tasks. --- --- We hope our answers have helped to address your concerns. Please do not hesitate to reach out if there are any remaining unclear points or questions.	null	null	Rebuttal 1: Rebuttal: Dear reviewers and AC, We want to genuinely thank you for your valuable time and effort spent reviewing our manuscript, and are grateful for the detailed and constructive remarks that have helped us to further improve the quality of our paper. We individually address each reviewer's comments as direct rebuttal to their respective review. As requested by reviewer pu66, we have attached additional visualizations outlining the conceptual differences between iterative, sequential and bi-directional approach, as well as a detailed side-by-side comparison of the internal components of the iterative and bi-directional attention blocks. It is of course possible that we might have misinterpreted a comment or question, in which case we would cordially ask the reviewers to point this out to us so we can clarify any remaining points as promptly as possible. Thank you very much! The Authors Pdf: /pdf/b4aa99d69a36d96798270ee1a781125c118ef5ac.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Algorithmic progress in language models	Accept (poster)	Summary: The authors of this paper examine the performance improvements of language models over the past decade, and investigate how much of it can be attributed to algorithmic improvements of language models. Strengths: I note here that, given that this paper proposes a method to evaluate the historical progress of language models, I believe that the primary goal for the paper would be to provide interesting insights for the field and outline current open questions. As such, I will structure the rest of my review with that in mind. - The topic of language models is more central than ever to the broad NeurIPS community, and I think that analysis on the historical progress of the field is of great interest. - The authors obtain valuable insight on the cause of language model improvements over the years. I find their conclusion that data/compute scaling has been a major driving force for improvement over the past few years interesting, if somewhat expected. - The result on the importance of the transformer is also interesting, and provides a nice retrospective justification of the broad adoption of this architecture. - The analysis of data from past works is also sound. Weaknesses: - The most crucial weakness of this paper is that, despite the analysis of past trends, it does not provide clear insights or suggestions on where the field should direct its efforts, moving on. The authors acknowledge this limitation, but it seems to me that such discussion Is crucial, in order for this paper to be of interest to the community. - It is also not clear to me how the insights provided by the paper could extrapolate in the future. While there has been a lot of effort and improvements gained in performance from data scaling, this is not something that can reliably keep on - at some point, sources of data and compute limitations catch up. As such, it is not clear to me how the insight that the effective compute for language models doubles every set period of time can be useful for many years down the line. - On more minor note, I think some points can be improved for clarity of presentation: - The captions of Figures 1a and 1b are joined together - some spacing between them is required. - It would be better if the authors clarified the statistical models used in the main paper, rather than in the appendix. Overall, my key concern for the paper is that it does not provide enough insight for future directions. Technical Quality: 3 Clarity: 3 Questions for Authors: I would be grateful if the authors could expand on the points I raised above, regarding the insight for future directions. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have adequately addressed the limitations of their work. Regarding negative societal impact, I do not foresee any arising from this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the feedback – we address your questions and concerns below. We fully agree with you that there is substantial uncertainty about how our results could extrapolate into the future. Indeed, we have mentioned this point in the limitations section, and pointed out that our core focus has always been to estimate historical rather than future rates of progress. That said, although future extrapolations are not within the primary scope of our paper, our results still represent a substantial step up compared to the prior status quo. For one, it provides a rate of progress that can be extrapolated in the first place, and it paves a clear path forward for future work. E.g. We agree with your point that compute limitations might be relevant for understanding future algorithmic progress, and this provides a clear next step of trying to quantify the significance and timing of this bottleneck. In fact, we feel that this strongly relates to your point about providing insight for future directions. Our work is most strongly directed towards people interested in understanding trends and drivers of progress in ML. Thus while it is of general interest for ML practitioners, by far the most important future directions pertain to future research on ML trends, such as in understanding future compute/data bottlenecks. We also mention that further research could extend the analysis to specific benchmarks or different domains, or consider the impact of individual algorithmic innovations by performing ML experiments. As we alluded to in the related work section, there has been relatively little work studying these important high-level questions about progress in ML and we believe our work points strongly towards additional work in this area. As such, we completely agree with you that the primary goal of the paper should be to provide interesting insights and outline future directions. We believe that our work achieves these two criteria most directly for people interested in studying ML trends/progress, as per the previous discussion. As a final point, thanks for the suggestions regarding improving the paper’s presentation – we will incorporate these changes when we next update the paper. --- Rebuttal Comment 1.1: Title: Thank you for your comments. Comment: I would like to thank you for your response to my review. I understand better now how the findings of the paper can relate to future work. I believe that further highlighting the above points in the main paper will increase the paper's impact. Given that this was my main concern, I am raising my score.	Summary: This paper investigates the rate of algorithmic progress on the task of language modeling, using a dataset of over 200 LM evaluations on WikiText and Penn Treebank between 2012-2023. The authors fit an augmented scaling law to the data and show that the models are requiring 2x less compute roughly every eight months -- a rate which supersedes Moore's law. Further, the authors find that more recent performance gains have been primarily due to compute scaling and that the contribution of the transformer architecture is roughly equivalent to two years of algorithmic progress in the field. Strengths: 1. This paper presents an interesting take on quantifying the progress on the task of language modeling using an analysis of data collected from over 200 model evaluations in the past 10-11 years. 2. The authors clearly specific the assumptions made in conducting the analysis, which makes the approach quite readable. Weaknesses: 1. There a number of assumptions made to elicit a quantification of algorithmic progress, and the nature of the task itself necessitates those assumptions be made. However, this also means that the resulting analysis is brittle and as such the technical contributions of the work don't rise to the level of a solid contribution. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Can you give examples of the ways in which the core results would change with different scaling law formulations are used? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The limitations were adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the feedback on our paper. We agree that we make several assumptions in our work quantifying algorithmic progress, and that this introduces uncertainty into our conclusions. However, we do not believe that these assumptions undermine the core results of our paper. We have performed extensive robustness checks for this purpose, e.g. we consider different ways of managing autocorrelation in Appendix I, different ways of estimating doubling times in Appendix G, etc. In each case, our robustness checks provide grounding for our core empirical results. If you do not believe that these robustness checks address your concern, could you please specify why not, and what assumptions you believe need to be addressed? To address your question about scaling laws in particular, our work considers a range of different model specifications, as outlined in Tables 10 and 11. These models are varied across different degrees of freedom, such as algorithmic progress impacting scaling exponents, progress that is specific to each benchmark, etc. We illustrate the variation across these degrees of freedom in figure 1b. Furthermore, in appendix C.1, we compare our model parameter estimates to existing work on scaling laws (e.g. Hoffmann et al 2022 and Kaplan et al 2020). We find that our estimated scaling exponents are consistent with theirs within the time range over which the majority of our data lies. In appendix H, we analyze the impact of including an irreducible loss term in the scaling law formulation, and compare this with our core model. Again, our findings are consistent with our core results. Overall we believe that having consistent results across all these changes provides strong evidence in favor of our core results.	Summary: This paper presents an analysis of the relative contribution of algorithmic progress to overall LM performance gains over the window of 2012-2023. The authors evaluate a large number of potential equation variants for modeling algorithmic progress using leave-one-out CV. By making use of defined notions of effective data/parameters/computer, the authors estimate that 2x less compute is needed to acheive the same performance roughly every 8 months. The authors find that roughly 2/3 of the scaling progress is due to the scaling of physical compute resources, with the remainder being attributed to algorithmic progress. The singular contribution of the Transformer architectures is individually analyzed. Thorough analysis of the techniques applied and their limitations are presented. Strengths: The paper is clearly written and well-presented. The depth and quality of the anlaysis are exemplary. The methodology of analysis is not highly original, but its application to algorithmic progress is a novel and useful contribution to the community. The limitations of this kind of analysis are well-discussed, which is a useful contribution in its own right. Weaknesses: Nits: Figure 2: label the y axes, consider making especially the right plot a bit larger/more readable, the text is very small Fig 1: fix the spacing between the (a) and (b) captions, they are nearly overlapping Technical Quality: 4 Clarity: 4 Questions for Authors: None. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The limitations are thoroughly discussed in the paper. No relevant missing potential negative social impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	null	Summary: This paper breaks down the driving force behind language models into two factors: scaling up the amount of compute used to train the models and algorithmic innovations. A statistical model, similar to the scaling law, is built and fitted to analyze the contributions of these two factors. The paper claims that models with the same performance level require half the compute every eight months, reflecting algorithmic improvement. The authors also find that the majority of scaling progress is due to the scaling of physical compute resources rather than algorithmic progress. Strengths: The paper presents a very interesting and innovative approach to quantifying the algorithmic progress of language models. By covering over 200 language models from the past decade, the empirical foundation for the conclusions drawn is solid. Additionally, the author has performed extensive robustness checks to ensure the validity of their core results. Weaknesses: While the paper makes several assumptions to quantify algorithmic progress (including the extra coefficients in the augmented scaling law), these assumptions and the many degrees of freedom undermine the robustness of the proposed model and the conclusions drawn. I have doubts about whether the statistical model built can inform future progress in language modeling. Minor issue: The caption of Figure 1 seems to be incorrect. It should address Figures 1a and 1b instead of 3a and 3b. Technical Quality: 4 Clarity: 2 Questions for Authors: None Confidence: 3 Soundness: 4 Presentation: 2 Contribution: 3 Limitations: I do not see any potential negative social impact of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Title: Is there a mix up? Comment: Dear reviewer, is this the review for the write paper? The authors and the area chair suspect that this review is for a different paper. Could you kindly update your review? Thanks, Your Area Chair --- Rebuttal Comment 1.1: Title: Appologise for mix up Comment: Yes, I mistakenly pasted the wrong review. I sincerely apologize to the authors and the area chair for the confusion and inconvenience caused. I have updated my review.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Scalable and Effective Arithmetic Tree Generation for Adder and Multiplier Designs	Accept (spotlight)	Summary: This paper presents a new method for designing arithmetic modules by modeling tasks as single-player tree generation games, using reinforcement learning techniques. This approach combines prefix and compressor tree modules to find optimal multipliers. Experiments show that the developed 128-bit adders and multipliers outperform the latest designs, significantly reducing delay and size. The method enhances computational efficiency and hardware size within hours and integrates seamlessly into advanced technology, offering valuable insights for future hardware design. Strengths: The strengths of the paper are listed below: The paper introduces an innovative method for designing arithmetic modules that outperforms traditional human design techniques. The paper is well-presented, offering a clear explanation of its methodology. It improves results against other method PrefixRL Weaknesses: The weaknesses of the paper are listed below: the paper innovation limited to application of RL and Tree Search to a circuit design problem, which are studied in other previous work such as: Mirhoseini, Azalia, et al. "A graph placement methodology for fast chip design." Nature 594.7862 (2021): 207-212. Technical Quality: 3 Clarity: 3 Questions for Authors: A few question and notes are listed below: It seems the main additions of this work are two-level Retrieval and MCTS compared to PrefixRL. Are there other inovation that distingush this work vs PrefixRL? How one would extend to other arithmetic operations, such as exponentiation? Seems a weakness to the method is that one needs to design the "game" for addition operators. Maybe There are work that design envs that author could try: Ma, Yecheng Jason, et al. "Eureka: Human-level reward design via coding large language models." arXiv preprint arXiv:2310.12931 (2023). Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations and, if applicable, potential negative societal impact of their work Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your positive comments. Weakness 1 [Innovation] - Circuit design is a broad field encompassing various tasks. Our work focuses on arithmetic unit design (adders and multipliers), which differs significantly from graph placement in search space and evaluation metrics. Graph placement involves optimizing the positions of circuit modules with graph structures, while arithmetic unit design deals with the structures of prefix and compressor trees, resulting in vastly different combinatorial space and optimization criteria. - Moreover, our approach includes several innovations: - When designing adders, we use the MCTS framework to balance exploring new design possibilities and refining existing ones, ensuring a comprehensive and efficient search across the design space. - For the compressor tree search in MultGame, the search performance is enhanced by redefining actions and states, and building the design from scratch, providing greater search flexibility. - The optimization process is also optimized by two-level retrieval and pruning strategies, enabling the efficient scaling of the design's bit-width (128-bit adder and 64-bit multiplier). - Our co-design framework allows simultaneous optimization of prefix and compressor trees, further enhancing the performance of designed multipliers. Question 1 [Difference with PrefixRL] Both our approach and PrefixRL [17] apply an RL framework to circuit design. However, our method incorporates several critical enhancements that we would like to clarify: - Advancement in the RL (state, action, and reward) formulation: - State: In PrefixRL, prefix trees are represented using a binary matrix, which is solely applicable to the prefix tree structure. In contrast, our model does not directly represent the prefix tree structure. Instead, each node, representing a prefix tree, is evaluated by $W(s)$, a scoring function designed to effectively balance exploration and exploitation. - Action: PrefixRL’s actions include adding and deleting cells. However, adding cells does not improve the theoretical metrics (level/size) of the adder it represents. Therefore, we have strategically omitted actions that add cells when optimizing these metrics. This refinement significantly boosts search efficiency, especially in the case of a 128-bit adder, by focusing on actions that directly contribute to performance improvement. - Reward: The reward in PrefixRL is calculated based on the performance improvement at each step. Our method, however, utilizes the performance value of the final node reached during the simulation phase of MCTS. This shift reduces the computational overhead from multiple time-consuming simulations and provides a more direct assessment of performance outcomes. - Advancement in Task Setting and Method Design: - Task Setting: PrefixRL only addresses adder design, whereas we work on adder and multiplier designs. - Method Design: We have introduced a novel framework for the co-optimization of the multiplier’s compress tree and prefix tree. This integrated approach significantly enhances the overall effectiveness of multiplier design. Question 2 [Extend to Exponentiation Operation] Exponentiation operations, whether employing the naive method or the method of exponentiation by squaring, involve numerous multiplication operations. Thus, our existing multiplier design approach is inherently well-suited to optimize these procedures. By harnessing the repeated multiplication principle intrinsic to both naive and binary exponentiation techniques, we can extend our RL framework, initially developed for multipliers, to enhance the design and efficiency of exponentiation units. This extension also allows us to further improve our model to accommodate the iterative nature of exponentiation, ensuring that our optimization effectively addresses the distinctive characteristics and performance demands of these hardware modules. Question 3 [Need for Designing Game] We understand your concern regarding the need to design the "game" for our work. Indeed, the majority of problems solved using RL necessitate the explicit definition of states, actions, and rewards, much like designing a game. The benefit of this approach is that a well-defined structure can significantly enhance the agent's ability to optimize the target efficiently and effectively. Regarding the work "Eureka", we fully agree that it represents an excellent contribution to the field. We will discuss Eureka in our revised paper and consider applying its reward design methodologies to arithmetic unit design in our future work. We believe that leveraging Eureka's approach can further improve our algorithm's efficiency and performance. --- Rebuttal Comment 1.1: Title: rebuttal acknowledgment Comment: thank you for clarifications and addressing the questions --- Reply to Comment 1.1.1: Comment: Thank you very much for the update. We appreciate your time and effort in helping us improve our paper.	Summary: The paper introduces a novel approach to optimizing arithmetic module designs, particularly adders and multipliers. By modeling design tasks as single-player tree generation games and employing RL techniques (MCTS and PPO), the authors effectively explore the design space to uncover superior structures. Their method significantly enhances computational efficiency and reduces hardware size, outperforming existing techniques in both aspects. This work demonstrates the effective application of RL in hardware design, presenting an innovative tree-generation strategy for improved design optimization. Strengths: Originality: 1. Few competing works address adder-multiplier co-design automation using RL or machine learning techniques. This paper stands out as an interesting and pioneering effort in optimizing such designs with RL. 1. The introduction of tree-generation games and the application of RL techniques to optimize arithmetic module designs are novel and innovative. Quality: 1. Based on the experimental results, this paper has discovered several optimal 128-bit adder architectures, a noteworthy achievement given the vast search space explored. 1. The experiments are thorough, validated across different technology nodes and using both open-source and commercial tools, confirming the method's effectiveness. Clarity: 1. The paper is well-organized, featuring a clear introduction, detailed methodology, and systematic presentation of results. This structure facilitates understanding the complex concepts introduced. 1. The introduction video is a nice touch and helps in understanding the proposed method. Significance: 1. The authors have made the research open-source and documented most of the experimental details comprehensively, which significantly enhances the reproducibility of the results. 1. The improvements of the proposed method over existing techniques in terms of delay and area are significant and demonstrate the potential of the proposed approach in optimizing hardware design. Weaknesses: 1. This study primarily focuses on optimizing adders and multipliers, with limited exploration of other modules within the hardware system. 1. While the proposed methods demonstrate notable improvements for 128-bit adders and 64-bit multipliers, their scalability to even larger bit widths needs further investigation. Minors: 1. I suggest the authors consistently use either the abbreviation 'RL' or the full term 'reinforcement learning' throughout the paper. 1. In Equation 3, $\sum_{i=0}^{n}$ should be $\sum_{i=1}^{N}$. 1. Lines 226-227: '... from step 0 to step T' should be '... from step 1 to step T'. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Could you provide detailed information on the design flow? Moreover, how are the floorplan and placement parameters set in OpenROAD? 1. Could you elaborate on how your method might be extended or adapted to optimize other hardware units? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your positive comments. Weakness 1 [Focus on Adders and Multipliers] - While our current results focus on adders and multipliers, these operations are among the most time-consuming on GPUs, making their optimization highly significant. - We would like to clarify that our work can also be extended to other arithmetic modules. Taking exponentiation as an example, exponentiation operations require multiple multiplication operations. By leveraging the underlying principle of repeated multiplication in exponentiation, we can apply our RL framework, initially designed for optimizing adders and multipliers, to enhance the design and efficiency of exponentiation units. This extension would involve tuning our model to accommodate the iterative nature of exponentiation, ensuring that the optimization captures the unique characteristics and performance requirements of these hardware modules. Weakness 2 [Scalability] We understand your concern regarding the scalability of our methods to larger bit widths. Currently, most computer systems predominantly use integer widths within the 128-bit range (with 64-bit multipliers producing 128-bit results), covering most practical applications. Moreover, we have employed pruning and two-level retrieval techniques to effectively explore and discover superior 128-bit prefix adder structures. Additionally, through optimized synthesis, we have scaled our multiplier designs from 16-bit to 64-bit, demonstrating significant improvements over the RL-MUL approach. Thus, our approach addresses current practical requirements and showcases the potential for scalability and enhanced performance through advanced optimization techniques. Minors Thank you for pointing out. We will correct them in the revised version. Question 1 [Details about the synthesis] We have provided detailed information on the timing aspects of the design flow in the appendix of our paper. Specifically, Fig. 12 and Fig. 13 contain comprehensive script codes that outline our approach to both logical and physical synthesis. For Fig. 12, we have maintained consistency with the RL-MUL methodology, while for Fig. 13, we utilized the default physical synthesis flow parameters provided by OpenROAD. These figures should provide the detailed information you have requested regarding the timing aspects and the specific parameters used in OpenROAD. Question 2 [Extension to Other Hardware Units] Our method can also be extended to exponentiation units. See Weakness 1. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I would like to keep my score (strong accept). --- Reply to Comment 1.1.1: Title: Thank you for the update Comment: We are very grateful for your positive comments once again. We also truly appreciate the time and effort you have put into reviewing our work.	Summary: This paper discusses the application of reinforcement learning to optimize the design of arithmetic circuits, specifically adders, and multipliers. Two single-player tree generation games, AddGame and MultGame, are designed to formulate adder and multiplier design problems. AddGame re-designs the search method, following a similar state space and action space as PrefixRL. MultGame optimizes the compressor tree from scratch instead of starting from existing solutions. Results show they outperform PrefixRL and RL-MUL. Strengths: 1. The paper re-designs the search techniques for adders and multipliers. Experimental results demonstrate the effectiveness of the proposed methods. The findings are significant for hardware design. 2. Main concerns are addressed which include the performance of designed adders and multipliers and the time cost of the searching process. 3. The writing is clear and well-structured. Weaknesses: 1. The paper would benefit from a more detailed introduction of the states and actions for both AddGame and MultGame in the main part. It was confusing to understand how a compressor tree is represented until I referred to Appendix A.4. 2. Some details are missing. For example, it is unclear whether the results shown in Figure 6 are synthesized using Nangate45 technology. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The paper states, “The compressor tree is built from scratch instead of starting from existing solutions for more design flexibility.” Does building from scratch indeed offer more design flexibility while maintaining performance? 2. Is the objective of Table 7 to minimize delay? Can you present a result with a trade-off objective? 3. Please provide the definition of accuracy mentioned in Appendix B.6. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations and societal impacts are discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your positive comments. Weakness 1 [Add Detailed Introduction in Main Text] Thank you for the suggestion. In our revised paper, we will include the introduction of the states and actions for both AddGame and MultGame, and the tree representations in the Appendix, to the main text. Weakness 2 [Technology Used in Fig. 6] The Fig. 6 results are based on Nangate45 PDK (45 nm). We will add this information to the caption of Fig. 6 in our revised version. Question 1 [Benefits of Building from Scratch] Building from scratch theoretically allows for the exploration of a broader range of compressor tree structures. All compressor trees can be viewed as being composed of basic compressor units, and building from scratch begins with adding these fundamental compressors. This approach grants greater freedom compared to modifying an existing compressor tree. This is because modifying existing compressor trees limits the exploration of new structures due to constraints such as the number of iterations. Additionally, the number of compressors in a compressor tree is finite, which means the steps in our building-from-scratch method are also limited. Consequently, optimization performance can be maintained within an acceptable range. As shown in Table 8, the optimization time remains within a reasonable timeframe. This demonstrates that building from scratch offers more design flexibility while maintaining performance. Question 2 [Objective in Table 7] The designs in Table 7 are obtained by performing trade-off optimization with open-source tools and then directly testing these optimized circuits with commercial tools. The increase in area can be attributed to the differences between the open-source and commercial tools used. Due to variations in their synthesis methodologies, the results from open-source tools might not achieve optimal performance when used with commercial tools. However, our approach still demonstrates significant advantages in terms of delay. We will clarify this in the new version of our paper. Question 3 [Accuracy Definition in Table 9] As we introduced in the two-level retrieval, we defined a fast flow and a full flow. The fast flow is quick but may not be accurate without routing, whereas the full flow takes longer but provides precise delay and area outputs. We assume that the delay and area predicted by the fast flow are $ d' $ and $ a' $, respectively, while the delay and area achieved from the full flow are $ d $ and $ a $. The accuracy reported in Table 9 is defined as $ \frac{d'}{d} $ and $ \frac{a'}{a} $. Experimental results show that the fast flow tends to slightly underestimate delay, while its area predictions are completely accurate.	Summary: This paper aims to leverage reinforcement learning (RL) techniques for automatic arithmetic circuit design. The authors propose to cast the design tasks as single-player tree generation games, and leverage reinforcement learning techniques to optimize these arithmetic tree structures. For adder circuits, the proposed approach discovers designs of 128-bit adders that achieve Pareto optimality in theoretical metrics. Moreover, the proposed approach significantly outperforms previous state-of-the-art RL-based approaches for adders and multipliers design. Strengths: 1. The paper is well-written and logically sound. 2. The paper explores the application of RL methods to arithmetic circuit design, presenting a unique intersection of interest for RL community and AI chip development. 3. The authors introduce a novel approach by modeling arithmetic module design tasks as single-player tree generation games, namely AddGame and MultGame, leveraging the well-established RL capabilities for complex decision-making tasks. 4. The method exhibits high flexibility and scalability, making it applicable to both 7nm technology and higher-bit units. These characteristics are crucial for practical hardware design and industrial applications. 5. Experiments show that the proposed approach discovers designs of 128-bit adders that achieve Pareto optimality in theoretical metrics, which is an impressive result. Weaknesses: The paper employs 45nm and 7nm PDKs. They are open-source and academically oriented, but not commercial-grade industry PDKs. This choice might introduce variations in the designs when transitioning from theoretical models to real-world industrial applications. Thus, the authors are encouraged to test their approach on real industry PDKs to validate the practicality in future work. A minor grammatical correction is needed on line 114, where “it derived” should be revised to “it is derived.” Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why the authors do not test on real industry PDKs? 2. Does the proposed method have any limitations when applied to real industry PDKs? 3. In Fig. 6, why do various RL approaches produce a range of designs, whereas traditional designs like BK and Sklansky adders result in only a single design? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N.A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your positive feedback. Weakness 1 [PDK Selection] Although the 45nm and 7nm PDKs employed in our study are open-source and academically oriented, they are designed to closely mimic corresponding industry nodes and are widely used for benchmarking purposes within the academic community. This ensures complete reproducibility of our results, which is a critical aspect of rigorous scientific research. For example, both PrefixRL and RL-MUL utilize the open-source Nangate45 PDK. In contrast, commercial-grade industry PDKs are proprietary and restrict the sharing of detailed process information. While we acknowledge the importance of validating our approach with industry PDKs in future work to confirm its practicality in real-world applications, open-source PDKs provide a robust foundation for demonstrating our methodology's feasibility and effectiveness within this paper's scope. Weakness 2 [Grammatical Error] A good catch. We will correct this in our revised version. Question 1 [Usage of Industry PDK] The PDKs used in our study are designed to mimic corresponding industry nodes and are widely adopted for benchmarking purposes. They help enhance reproducibility, while industry PDKs, being proprietary, may present some challenges with reproducibility. Question 2 [Applicability to Real Industry PDKs] Our proposed method can be directly applied to real industry PDKs. The only requirement is to modify the library files corresponding to the parameters during synthesis. Question 3 [Design Number Variability] Traditional designs like Brent-Kung and Sklansky adders have fixed prefix structures, resulting in a single, consistent design each time they are implemented. In contrast, RL methods explore a wide design space during a single execution, generating various tree structures and diverse design outcomes. This flexibility allows RL methods to produce multiple designs within one optimization process. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response. I will keep my current score. --- Reply to Comment 1.1.1: Title: Thanks for the reply Comment: Thank you for your positive feedback and the time spent reviewing our work.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper presents a reinforcement-learning-based approach to optimizing arithmetic units, specifically adders and multipliers, to enhance computational efficiency and reduce area consumption. The key idea is to frame the design as tree generation games. The evaluation shows that the proposed method can generate 128-bit adders with up to 26% reduced delay and 30% smaller size, and multipliers achieving up to 49% faster speeds and 45% size reductions compared to existing techniques. Strengths: - The paper is in good writing style. Figures are well plotted. - The use of reinforcement learning and tree generation games for Adder and Multiplier generation is novel and interesting. - The evaluation setup is clear and the comparison between different methods is comprehensive. The trade-off between the area and delay of the resulted ALU designs is well demonstrated. - The reported enhancements in speed and size for adders and multipliers with large bit width are substantial. Weaknesses: - The improvement of PPO-based method becomes smaller as the bit width becomes lower. The paper claims that multipliers and adders are more important for large models in AI applications. However, the modern large models are typically running with 16 bits or even lower bit width. The applicability of the proposed method in modern applications remain questionable. - The evaluation is mostly performed on integer data types. It is unclear how proposed method works on the adders and multipliers of floating-point data type. - While the results are promising, the specific modeling seems to only work for adders and multipliers. The generalization of the proposed approach on other basic arithmetic units are unclear. Technical Quality: 2 Clarity: 3 Questions for Authors: Please refer to the weakness. Confidence: 2 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your constructive comments. Weakness 1 [Application in Modern LLM] - We appreciate this observation regarding the diminishing improvements of our PPO-based method as the bit width decreases. Indeed, the paper highlights the significance of multipliers and adders for large models, which traditionally have used higher bit widths. However, it is essential to note that while modern AI models often utilize 16 bits or even lower for efficiency, the computational demands and precision requirements can vary significantly depending on the application and deployment scenario. Recent studies [1] have shown that the performance of LLAMA 3 significantly degrades with low-bit quantization. This indicates that our method for high-bit precision has substantial potential for improving the computational speed of future large models and high-performance computing tasks. - In addition to theoretical improvements, we are actively collaborating with industry partners to validate and refine our approach in real-world scenarios, ensuring that our method remains relevant and effective for contemporary AI applications, irrespective of their bit-width requirements. [1] How Good Are Low-bit Quantized LLAMA3 Models? An Empirical Study Weakness 2 [About Floating-Point Data] - Thank you for raising the concern regarding the performance of our method on floating-point v.s. integer data types. It is important to clarify that at the fundamental level, the arithmetic operations of addition and multiplication share core similarities between floating-point and integer data types. For example, in floating-point multiplication, the exponents are added (integer addition), and the significands are multiplied (integer multiplication). This indicates that the core computational workload lies in integer arithmetic. Thus, while our current evaluation is on integer data types, the principles and optimizations easily apply to floating-point arithmetic. - Although our initial evaluations focused on integer operations due to their straightforward implementation and common usage in specific applications, the core optimizations are inherently applicable to floating-point arithmetic as well. Future work will explicitly focus on floating-point evaluations to demonstrate this applicability and to fine-tune our method for any floating-point-specific optimizations that may be required. Weakness 3 [Significance in Addition and Multiplication. Extension to Other Arithmetic Units] - While our current results focus on adders and multipliers, these operations are among the most time-consuming on GPUs, making their optimization highly significant. - We would like to clarify that our work can also be extended to other arithmetic modules. Taking exponentiation as an example, exponentiation operations require multiple multiplication operations. By leveraging the underlying principle of repeated multiplication in exponentiation, we can apply our RL framework, initially designed for optimizing adders and multipliers, to enhance the design and efficiency of exponentiation units. This extension would involve tuning our model to accommodate the iterative nature of exponentiation, ensuring that the optimization captures the unique characteristics and performance requirements of these hardware modules.	Summary: A reinforcement learning-based method is proposed to design adders and multipliers within a framework of single-player tree generation games, named AddGame and MultGame. The method leads to the discovery of superior designs for 128-bit adders and multipliers, achieving significant improvements in delay and size compared to previous methods. Strengths: 1. The results look quite promising, especially on large designs. Weaknesses: 1. The novelty is limited, as the RL environment (state, action, reward) is pretty much similar to previous work [17], and MCTS is also not something new in a game-playing problem. 2. Some additional data points in the results might not look reasonable. See questions below. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What's the specific representation of adder and multiplier in the proposed method? Is it similar to previous works (e.g., a matrix in [17])? 2. Did the obtained design go through an equivalence-checking process to ensure the functionality correctness? If so, the details should be provided. 3. Figure 6 presents that the proposed method achieves promising results. However, it is observed that the Sklansky adder achieves even better delay than the Kogge-stone, which looks questionable and degrades the convincingness of the experiment flow. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your constructive feedback. Weakness 1 [Distinction and Improvement over PrefixRL [17]] Both our approach and PrefixRL [17] apply an RL framework to circuit design. However, our method incorporates several critical enhancements that we would like to clarify: - Advancement in the RL (state, action, and reward) formulation: - State: In PrefixRL, prefix trees are represented using a binary matrix, which is solely applicable to the prefix tree structure. In contrast, our model does not directly represent the prefix tree structure. Instead, each node, representing a prefix tree, is evaluated by $W(s)$, a scoring function designed to effectively balance exploration and exploitation. - Action: PrefixRL’s actions include adding and deleting cells. However, adding cells does not improve the theoretical metrics (level/size) of the adder it represents. Therefore, we have strategically omitted actions that add cells when optimizing these metrics. This refinement significantly boosts search efficiency, especially in the case of a 128-bit adder, by focusing on actions that directly contribute to performance improvement. - Reward: The reward in PrefixRL is calculated based on the performance improvement at each step. Our method, however, utilizes the performance value of the final node reached during the simulation phase of MCTS. This shift reduces the computational overhead from multiple time-consuming simulations and provides a more direct assessment of performance outcomes. - Advancement in Implementation: - Two-level Retrieval: We use the fast synthesis flow to retrieve potential designs, and then use full synthesis flow to get the final performance results, improving the efficiency. - Advancement in Task Setting and Method Design: - Task Setting: PrefixRL only addresses adder design, whereas we work on adder and multiplier designs. - Method Design: We have introduced a novel framework for the co-optimization of the multiplier’s compress tree and prefix tree. This integrated approach significantly enhances the overall effectiveness of multiplier design. Weakness 2 [Data Points] See the response in Question 3. Question 1 [State Representation] Specific representation of adder and multiplier: - Adder: We represent the prefix tree structure based on the score $ W(s) $ at the nodes of the search tree, eliminating the need to explicitly design a matrix representation model in PrefixRL [17] and making it more generalizable. - Multiplier: We select high-level meta-features, such as the maximum estimated delay value of bits, different from the matrix representation in counterparts (e.g. RL-MUL). Utilizing these higher-level features enhances the model's ability to learn the construction of multipliers with detailed information, thereby improving design outcomes. Difference with PrefixRL [17]: Please refer to our response in [Weakness 1] for the advantages of our approach over PrefixRL. Question 2 [Did the obtained design go through an equivalence-checking process to ensure the functionality correctness?] Sure, we did it. Specifically, we used corresponding testbenches for the generated Verilog code, which includes hundreds of sets of addition or multiplication operations. By observing the outputs, we verified the correctness of the computations. Question 3 [Sklansky v.s. Koggle-Stone] We agree with the reviewer that, in general, the Kogge-Stone adder generally has lower latency than the Sklansky adder due to its highly parallel, low fanout, and balanced prefix computation structure. As observed in our paper (Figure 6, middle subfigure), the Kogge-Stone adder exhibits lower latency compared to the Sklansky adder. However, in some specific experimental conditions, the Sklansky adder can also exhibit lower latency, as observed in our experiments (Figure 6, left subfigure) and also reported in the literature [1, 2, 3, 4]. The specific reasons for this observation are the following: - Impact by the Interconnection Complexity: In real-world implementations, the Kogge-Stone adder's structure results in very complex interconnections. Typically, with abundant hardware resources, this interconnection complexity can be effectively managed. However, when actual hardware resources are limited, this complexity can lead to significant wiring congestion and increased actual delay. In contrast, the Sklansky adder, with its reduced interconnection, offers simpler wiring. This simplicity decreases the likelihood of congestion and potentially enhances overall performance under resource limitations. - Impact by Synthesis Tools: Delay is influenced not only by the differences in the prefix tree structures (in Kogge-Stone and Sklansky adders) but also by the subsequent synthesis processes. Typically, the high fan-out in the Sklansky adder tends to increase its delay. However, modern synthesis tools (e.g., Yosys, OpenROAD) are equipped with advanced optimization techniques that can effectively manage high fan-out by balancing the load, buffering critical paths, and optimizing the placement of logic elements. These tools can significantly reduce the adverse impact of high fan-out on delay, ensuring that the Sklansky adder performs efficiently. To further validate our experimental results, we also consulted an expert in adder design, who believes that it is entirely acceptable for the Sklansky adder to have a slightly lower delay than the Kogge-Stone adder after physical synthesis. [1] Performance comparison among various parallel prefix adder topologies [2] PrefixRL: Optimization of parallel prefix circuits using deep reinforcement learning [3] Implementation of 64 Bit Arithmetic Adders [4] An Efficient Design and Performance Analysis of Novel 8 Bit Modified Wallace Multiplier Using Sklansky Adder in Comparison with Kogge-Stone Adder (KSA) --- Rebuttal Comment 1.1: Title: rebuttal acknowledgment Comment: Thanks for the clarification. Most of the concerns are addressed. A few more comments: 1. If the logic equivalence checking is performed, please add the description in future versions. 2. Regarding the clarification on Kogge-stone vs. Sklansky, if it is due to interconnection, the authors may inspect the final layout and/or tool settings to verify. The score is raised accordingly. --- Reply to Comment 1.1.1: Title: Thank you for the update Comment: Thank you very much for the update. 1. We will add the checking description in our revised version. 2. Thank you for the suggestion. Our codebase, available to the research community, includes detailed settings that facilitate such investigations. We will further inspect the final layout and tool settings, and report them in revision. We appreciate your time and effort in helping us improve our paper.	null	null	null	null
Verifiably Robust Conformal Prediction	Accept (poster)	Summary: The authors propose a novel method to recover coverage guarantees for conformal predictions in the presence of adversarial attacks. Unlike previously proposed approaches, the authors directly leverage verifiable methods for NN to compute prediction scores. Through empirical tests, the authors prove the benefit of the proposed approach against attacks bounded by different norms. Strengths: 1. Empirical results support the benefit of the proposed approach against vanilla CP, RSCP+ 2. The proposed approach seems to be robust across different values of the adversarial perturbation 3. The approach is well-described and the coverage guarantees are supported by theoretical arguments Weaknesses: 1. In the paper, the authors focus on PDG. However, the literature offers more advanced and sophisticated attacks. It would be beneficial to assess the robustness of the proposed approach against other type of attacks. 2. Since it is not specified in the paper, I assume the ML model under attack has not been trained using adversarial training. The reader might wonder whether the benefit of the proposed approach remains valid in the presence of an adversarially trained model. Concerning the latter case, will vanilla CP still violate the coverage guarantees? 3. Figure 2 is hard to read. Which curve represents the proposed approach, and which one represents the benchmarks? Concerning Figure 2a: Why does the coverage go to one as the magnitude of the adversarial perturbation goes to one? Perhaps it is because all the methods converge to a trivial prediction set. However, this should be clearly explained in the text. Concerning Figure 2b: Why does the coverage decrease as a function of the number of samples? Technical Quality: 3 Clarity: 3 Questions for Authors: See Weaknesses Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are discussed Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > 1. In the paper, the authors focus on PDG. However, the literature offers more advanced and sophisticated attacks. It would be beneficial to assess the robustness of the proposed approach against other type of attacks. With regard to the chosen attack methods, we evaluate PGD in the case of classification tasks and FGSM for the regression task. We chose PGD as it remains a popular choice within existing literature on robust CP approaches, and so as to be consistent and draw fair comparisons against the RSCP/+ methods. In theory, however, our approach is agnostic to the adversarial attack algorithm used to perturb the inputs, as it ensures robustness to any $\ell^p$ bounded adversarial attacks. > 2. Since it is not specified in the paper, I assume the ML model under attack has not been trained using adversarial training. The reader might wonder whether the benefit of the proposed approach remains valid in the presence of an adversarially trained model. Concerning the latter case, will vanilla CP still violate the coverage guarantees? We appreciate that we should have been clearer about the adversarial training aspect, and your assumption is correct in that we do not use adversarial training. Whilst an adversarially trained model is more reliable on adversarial inputs, using one in combination with vanilla CP doesn't guarantee robust coverage. For this to hold, we would require the adversarially trained model to guarantee prediction invariance for every input and the corresponding $\epsilon$-bounded region around it, which is not the case. Our VRCP method would provide valid prediction regions instead. We will clarify this point further in the paper. > 3. Figure 2 is hard to read. Which curve represents the proposed approach, and which one represents the benchmarks? Concerning Figure 2a: Why does the coverage go to one as the magnitude of the adversarial perturbation goes to one? Perhaps it is because all the methods converge to a trivial prediction set. However, this should be clearly explained in the text. Concerning Figure 2b: Why does the coverage decrease as a function of the number of samples? Perhaps confusingly, the legend for all the plots can be seen in the 3rd figure due to an oversight in figure placement. We will update the manuscript to include the legend more clearly below the figures. As the reviewer pointed out, in Figure 2a, coverage and prediction set size increase for larger epsilon due to the robust methods including more classes within their prediction sets to account for the stronger potential perturbation. This trend converges to trivial prediction sets and a coverage of 1. Note that in our experiments with $\epsilon=0.05$, VRCP-I has an empirical coverage of $1$ but does not produce trivial prediction sets (its average set size is $8.52$ against a maximum of $10$). In Figure 2b, we observe that increasing the number of Monte-Carlo samples used in RSCP+ methods improves their prediction set efficiency. The higher the number of samples used for randomised smoothing, the lower the effect is of the Hoeffding/Bernstein bound required to correct for sampling error. We will add both of these discussions to the main text to improve clarity. --- Rebuttal Comment 1.1: Comment: I would like to thank the authors to take the time to write the rebuttal to address my concerns. Although I appreciate the paper motivation, I think that the contribution is limited. I will keep my previous score.	Summary: The paper provides a verifiably robust conformal prediction method via neural network verification. It considers two paradigms for considering the perturbation either at the calibration stage or the inference stage. They also consider the validity of method for both classification and regression. Strengths: 1. The paper is well written. I believe people with other background can also understand it well. 2. The methods work for me. It is clear that NN verification provides a lower/upper bound of the logits, which are transformed to certified bound for conformity scores. Weaknesses: 1. Novelty: I understand that it is a new angle that combines NN verification and conformal prediction to provide a certifiably robust conformal prediction framework, but I wonder technically, can we do more about the combination. For example, can we have new conformity score adapted to NN verification that can provide tighter bound? Or at least empirically, we can analyze what types of conformity scores are more suitable for randomized certification? What are more suitable for deterministic certification (NN verification). 2. What is the intuition that the method can outperform RSCP+ in Table 1? Since usually randomized smoothing provides better certification than NN verification, then coming to the conformal prediction context, what makes the difference? 3. Missing related work: Kang, Mintong, et al. COLEP: Certifiably Robust Learning-Reasoning Conformal Prediction via Probabilistic Circuits. ICLR 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the weakness part. Overall, I do not have many concerns about the paper's soundness. If the authors can provide more insights into comparisons between randomized certification and deterministic certification in robust conformal prediction, it would be beneficial to the community. Basically, the authors need more support of why NN verification-based conformal prediction is worthy besides it can provide a general $\ell_p$ norm guarantee. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: discussed in Sec 6.1 Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > 1. Novelty: I understand that it is a new angle that combines NN verification and conformal prediction to provide a certifiably robust conformal prediction framework, but I wonder technically, can we do more about the combination. For example, can we have new conformity score adapted to NN verification that can provide tighter bound? Or at least empirically, we can analyze what types of conformity scores are more suitable for randomized certification? What are more suitable for deterministic certification (NN verification). We address the question regarding verification-friendly non-conformity score functions in Part 1 of the Global Rebuttal. > 2. What is the intuition that the method can outperform RSCP+ in Table 1? Since usually randomized smoothing provides better certification than NN verification, then coming to the conformal prediction context, what makes the difference? We address the question regarding the intuition behind VRCP’s improved empirical performance to RSCP+ in Part 2 of the Global Rebuttal. > 3. Missing related work: Kang, Mintong, et al. COLEP: Certifiably Robust Learning-Reasoning Conformal Prediction via Probabilistic Circuits. ICLR 2024. We will add this to the related work of the updated manuscript. We understand that this paper introduces a learning-reasoning framework (COLEP) that integrates the original RSCP (Gendler et al. 2021) to provide adversarially robust conformal prediction sets by using auxiliary models and probabilistic circuits. This work is complementary to ours because VRCP can be directly applied to COLEP in place of RSCP. > Please refer to the weakness part. Overall, I do not have many concerns about the paper's soundness. If the authors can provide more insights into comparisons between randomized certification and deterministic certification in robust conformal prediction, it would be beneficial to the community. Basically, the authors need more support of why NN verification-based conformal prediction is worthy besides it can provide a general $\ell_p$ norm guarantee. We addressed this question in detail above. We would also like to note that VRCP provides more benefits as listed below: - Other existing robust CP approaches introduce additional dependencies, such as the need for a hold-out set or a large number of samples (used for randomized smoothing) - VRCP supports regression tasks without requiring any modifications, unlike the existing approaches - VRCP supports arbitrary $\ell^p$ norms --- Rebuttal Comment 1.1: Title: Thanks for the rebuttal Comment: Thank the authors for the rebuttal. Overall, I think this paper is technically sound and the first combination work of NN verification and conformal certification. I will maintain a borderline because the motivation for using it instead of RSCP is not very strong besides the arbitrary $\ell_p$ norm. Although, RSCP requires many samples, this work is also not efficient due to NN verification. I also do not think RSCP can not be tailored for regression tasks.	Summary: This submission proposes VRCP, a framework for verifiably robust conformal prediction under Lp-bounded adversarial attacks. The framework integrates existing bound propagation tools for verification of conformal predictions. Two variants, VRCP-I that triggers bound computation at inference time and VRCP-C that triggers bound computation at calibration time, are proposed. VRCP-I has inference overhead but the set size is smaller. Experiments on CIFAR10, CIFAR100, and a regression task demonstrate the effectiveness. Strengths: 1. The submission introduces a feasible method to turn well-studied neural network verifiers into conformal prediction verifiers. The method is sound in theory and is also verified empirically. This is a novel contribution to the field. 2. Compared to existing verifiably robust conformal prediction methods, VRCP has superior empirical performance especially when compared to randomized smoothing, demonstrating its practical value. 3. Presentation is generally great and easy to follow. Weaknesses: 1. The framework is relatively straightforward. The foundation is some probability relaxations that can be intuitively derived. I would not view this intuitiveness as a weakness. However, it would be great if the authors could discuss some extensions and optimizations, e.g., proof sharing for VRCP-C, alternative but more verification-friendly score function design, etc. 2. Some in-depth discussion could benefit the submission. Concretely, why does VRCP work with different Lp norms? Does the benefit come from the existing verifier's flexibility? Does the bound tightness differ with different Lp norms? Why does the method surpass RSCP? 3. The experimental setup is not very clear: What is the model's normal accuracy? Could you provide more background on the setup of regression experiments especially the physical meaning of the reward bound and nominal performance? The setting is not very common in the literature. Minor: Line 129-130: the lengthy sentence seems to be lacking a period or comma. Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Yes, authors discuss the limitations in the last paragraph of the main text. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > 1. The framework is relatively straightforward. The foundation is some probability relaxations that can be intuitively derived. I would not view this intuitiveness as a weakness. However, it would be great if the authors could discuss some extensions and optimizations, e.g., proof sharing for VRCP-C, alternative but more verification-friendly score function design, etc. As you correctly mention, our methods rely on the output bounds computed by the NN verifier. Thus, the same features of NN architectures that are friendly for verification will be beneficial to our methods. For example, using transformations and activation functions with low Lipschitz constants would result in tighter bounds with linear bound propagation approaches. We address the question regarding verification-friendly non-conformity score functions in (1) of the global review. > 2. Some in-depth discussion could benefit the submission. Concretely, why does VRCP work with different Lp norms? Does the benefit come from the existing verifier's flexibility? Does the bound tightness differ with different Lp norms? Why does the method surpass RSCP? Regarding the first two questions, the existing verification methods indeed grant VRCP's ability to extend to other $\ell^p$ norms. These methods theoretically provide verification for perturbations bounded by any $\ell^p$ norm, although in practice, often only the most common $\ell^p$ norms are implemented ($\ell^1, \ell^2$ and $\ell^{\infty}$). The bounds' tightness differs with different $\ell^p$ norms because, for the same epsilon bound, the epsilon-bounded $\ell^1$ ball is strictly smaller (in volume) than the corresponding $\ell^2$ ball, which is in turn smaller than the $\ell^{\infty}$ ball. Hence, for larger values of $p$, we will see larger input regions and looser output bounds. We address the question regarding the intuition behind VRCP’s improved empirical performance to RSCP+ in (2) of the global review. > 3. The experimental setup is not very clear: What is the model's normal accuracy? Could you provide more background on the setup of regression experiments especially the physical meaning of the reward bound and nominal performance? The setting is not very common in the literature. For the classification experiments, the test accuracies of the models are: \|Accuracy/Model\|CIFAR10\|CIFAR100\|TinyImageNet\| \|---\|---\|---\|---\| \|Top-5 Test\|98.27%\|82.87%\|55.72%\| \|Top-1 Test\|76.52%\|55.73%\|29.64%\| It should be noted that the accuracy of the model has no effect on VRCP's validity and only affects the efficiency of the prediction sets (more accurate models, tighter prediction regions). For the regression experiments, the train and test losses of the models are: \|Loss/Environment\|Adversary\|Spread\|Push\| \|---\|---\|---\|---\| \|Train\|0.066\|0.075\|0.075\| \|Test\|0.051\|0.053\|0.068\| In terms of the reward bound, we have scaled the total cumulative reward between the range [0, 1] and thus the prediction intervals are also taken over this range. We will clarify this in the updated paper. --- Rebuttal Comment 1.1: Comment: Thanks for the response! Most of my concerns are resolved. So I maintain my score. > Does the bound tightness differ with different Lp norms? Sorry the question may not be clear enough. I was asking whether the gap between verifier's score bound and actual score minimum / maximum for perturbed inputs can become larger/smaller among different Lp norms. --- Reply to Comment 1.1.1: Comment: Thank you for your response. > Sorry the question may not be clear enough. I was asking whether the gap between verifier's score bound and actual score minimum / maximum for perturbed inputs can become larger/smaller among different Lp norms. Yes, it differs between different $\ell_p$ norms. For example, $\ell_\infty$ results in larger gaps for the same attack budget $\epsilon$. This is because the $\ell_\infty$ ball contains other $\ell_p$ balls. However, it should be noted that verifiers tend to over-approximate non-linear norms (e.g., $\ell_2, \ell_3, \dots$) more than linear norms (e.g., $\ell_1$ and $\ell_\infty$). This means for non-linear norms, we may get larger gaps than linear norms. We hope this answers your questions.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their useful comments. Here we respond to the common feedback amongst all the reviews. --- ## Part 1. Verification-Friendly Non-Conformity Score Functions We agree that investigating verification-friendly score functions is a great idea for future work. We use $1-f_y(x)$ as the score function, as we find this to be, for classification tasks, the most popular choice across the literature, including existing randomised smoothing approaches. This function happens to be verification-friendly because it is completely linear and does not introduce further over-approximations. Score functions that introduce over-approximation (e.g., where there are multiple expressions to bound) would make our approach more conservative and possibly favour randomised smoothing approaches. --- ## Part 2. The intuition behind VRCP's Improved Empirical Performance to RSCP+ In the RSCP+ approach, the robustness guarantee is dependent on a number of factors that affect performance, listed below: - The size of the Lipschitz constant $\sigma/\epsilon$. - The size of the Hoeffding/Bernstein bounds used to correct for sampling error when estimating the mean of their smoothed scores. - The representativity of the holdout set w.r.t. the rest of the calibration distribution (when using PTT). Both VRCP-C and VRCP-I methods perform agnostically of the aforementioned factors and thus exhibit improved efficiency where RSCP/+ falls short.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
CLODE: Continuous Exposure Learning for Low-Light Image Enhancement using Neural ODEs	Reject	Summary: This paper formulates the higher-order curve estimation problem as a NODE problem, enabling effective and accurate solutions with standard ODE solvers. Strengths: This paper is well-written and structurally organised. Weaknesses: Reference formats are not consistent. Technical Quality: 3 Clarity: 3 Questions for Authors: Why choose E=0.6 in Eq.(11)? The setting of hyper-parameters in Eq.(17) should be mentioned. What's the challenge of applying ODE to continuous exposure learning? The background of the proposed method in Figure 4 is quite different from GT, why? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: As described above Flag For Ethics Review: ['Ethics review needed: Research involving human subjects'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are glad to hear that you found the paper to be well-written and structurally organized. We have diligently examined your comments and concerns as a reviewer, and have prepared responses addressing the raised concerns. W1: Reference format - Thanks to your thorough suggestion, we believe we can make the reference format consistent. We promise to make the revision in the final version. Q1: Hyper-parameters - The exposure level parameter $E$ in Eq.(11) guides the network to train the resulting final brightness of the output image. - For the exposure level parameter, *we followed the prior settings of previous curve-adjustment based methods ([6, 9]).* To ensure fairness and eliminate any effects from changes in $E$, we consistently used the same value. - *For the setting of hyper-parameters in Eq.(17), is detailed in L459-461 of the Appendix.* As we provided in the Appendix A.1, the weights for each loss function are set to balance the scale of losses. To reiterate, the weights for the loss function $w_{col}$, $w_{param}$, $w_{spa}$, $w_{exp}$ and $w_{noise}$ are set to 20, 200, 1, 10 and 1 respectively. Q2: Challenge of applying ODE to continuous exposure learning. There were three main challenges in applying ODE to continuous exposure learning: inference speed, tolerance setting, and time consumption in training. [Inference speed] - First challenge is inference speed as written in L330-332. Since Neural ODE requires iterative processes to find the optimal solution, can result in somewhat slower inference times. To tackle this problem, we provide CLODE-S which is a compact version composed of a 2-layer network (R-fig. 1(b)). Although CLODE-S takes 0.0004M parameters and takes 0.005 second for image inference, CLODE-S shows promising performance in Table 5. Improving the inference speed of CLODE is our future research goal. - One possibility is to apply RectifiedFlow [1] as mentioned in UK8A W2. RectifiedFlow[1] transforms the solution paths of Neural ODEs into straight lines, enabling faster estimation of the Neural ODE system. We can first assume that the optimal solution found by CLODE is the expected optimal solution and then apply CLODE specifically to [1]. - Given the potential to apply such cutting-edge Flow Matching methods, we believe that CLODE is highly promising and can achieve fast inference speeds. [Tolerance setting]* - In addition, setting the tolerance is crucial in the NODE system because the ODE solver determines the state is optimal and has to be terminated, by the error of the current state within the allowable error rate. The error rate is defined by the following formula: $$\Gamma_t = atol + rtol \times \text{norm}(Err_t)$$ - Here, $Err_t$ is the current state error, and atol and rtol represent the absolute and relative tolerance, respectively. - If $Err_t > \Gamma_t$, the ODE solver adjusts the step size to minimize the error. - If $Err_t \leq \Gamma_t$, the current image is considered the optimal solution state, and the process terminates. - Since $\Gamma_t$ is calculated by atol and rtol, setting these tolerance value is important. If atol and rtol are too small, training may fail, and if they are too large, effective brightness enhancement may not be achieved. As mentioned in the Appendix. L472, CLODE empirically sets atol and rtol to $1e^{-5}$. *For additional details about CLODE, please kindly refer to our global rebuttal G1.* [Time consumption in training] - Finally, as NODE involves simulation-based training, the model takes longer to train as it grows in size. To overcome this obstacle, we designed the architecture of CLODE to be as compact as possible while maximizing low-light image enhancement performance. For a better understanding of "simulation-based learning", we provide a schematic representation of CLODE (dopri5) in actual image enhancement at the bottom of R-Fig.2. Q3: Figure 4 - We understand Reviewer Qjwa's curiosity about Fig.4. We can explain the reason why the background of Fig.4 is quite different from the ground truth in two aspects. - One reason is the unsupervised methodology of CLODE. Since training is done without ground-truth images, CLODE improves the input images using only the given information from itself. The other reason comes from the lack of information in the input images. In some overexposed images in the SICE dataset, the overexposed regions do not contain the same details as the ground truth images. - In particular, the lack of information in the input image causes the same problem with supervised methods too, as shown in Fig.4. CLODE’s enhancement results may not match the ground truth perfectly, but despite this, our method performs better than other unsupervised methods and competes well with supervised methods. *To address issues with over-exposed images, we will need to use a generative model, which we plan to explore in future research.* Thank you very much for taking the time to review our work. If there are any additional questions or points, we would be delighted to address them. --- >References [1*] Liu, Xingchao, and Chengyue Gong. "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow." In ICLR, 2023 --- Rebuttal Comment 1.1: Comment: Thanks for your response. I have no more questions. --- Reply to Comment 1.1.1: Comment: We appreciate Reviewer Qjwa's time and constructive comments and discussions. We are also thankful for the acknowledgment of the effectiveness of our method. In the final version, we will include the discussions.	Summary: This paper mainly addressed the problem of insufficient data for low-light enhancement. Specifically, it proposed CLODE , which employs Neural Ordinary Differential Equations to learn the continuous dynamics of the latent image for the first time. The experiments demonstrate the CLODE performs better than other unsupervised learning methods. Strengths: + This is the first attempt to formulate the higher-order curve estimation problem as a NODE problem. + CLODE can offer user controllability, enabling users to manually adjust exposure. Weaknesses: - Details of User Controllable Design. Despite the better result with use control, detail of the users is missing. For example, the number of volunteers, and whether they are banned from the ground truth image before they adjust the output image. Also, involving human feedback bring much more time in the inference stage. - In Sec. 3.3 Inference Process, the relationship between the output image IT and noise-free image is questioned. Each iteration includes a noise removal module, yet the output image still contains some noise, contradicting the expectation of a noise-free result in the model. - Experimental Setup: The experimental setting described in [1] seems more suitable for unsupervised methods. Using only a single dataset for training in this study does not adequately reflect the advantages of the proposed method. A specific analysis comparing and justifying the differences in experimental setups is necessary. - Model Iteration Selection in "CLODE+" (Table 2): The manual operation required to select the iteration step raises concerns. How is this value determined to ensure suitable results? This approach appears more suited to image retouching tasks than enhancement. - Concern about the fair comparison with previous methods. This paper uses 5 different losses. I wonder whether only part of them is used in previous methods, are the proposed method align with previous methods? For example, some Retinex-based method does not explicitly consider the impact of noise, and they do not have Noise Removal process. Does CLODE still outperform other methods without noise removal? More ablation experiments are needed for thorough explanation. - Effectiveness of Noise Removal Module: In the first toy scene in Figure 4, as well as Figures 7 and 8, there is noticeable noise residue and some degree of color distortion, which casts doubt on the effectiveness of CLODE and its noise removal module for low-light enhancement. - More explanation of the superiority of CLODE. Can author provide clearer explanation of the mechanism? For example, in Figure 9 of Supp material, is the better results comes from the more iterations, or more iterations at the early stage, where the estimation is harder? [1] Learning a Simple Low-light Image Enhancer from Paired Low-light Instances Technical Quality: 3 Clarity: 2 Questions for Authors: - In tab. 3, can the continuous method be considered to have an early stop mechanism? While discrete methods adjust in every iteration as discrete value, the continuous method adjusts as a continuous value. Further analysis and additional visual results comparing these methods across iterations are anticipated. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for recognizing our contribution and pointing out constructive concerns. We have made efforts to address the reviewer’s concerns as follows. Q1, W7: More explanation of CLODE Due to character limitation we placed explanation of this question in global rebuttal. *Please refer to global rebuttal G1 to address reviewer's concern.* We have prepared explanations along with additional analysis (R-fig. 2). In brief, CLODE includes an early stop mechanism. W1, W4: User Control - *First, we would like to cautiously mention that CLODE, even without user-control, is already state-of-the-art.* - In Table 1 and 2, CLODE+ was determined based on the average values of images selected by two experts in the field. - To address Reviewer Z6DS’s curiosity, we conducted an extra user study on the LOL dataset with 21 participants who had no prior knowledge of the ground-truth images. The results are in R-Table 6. From the optimal state provided by CLODE, five images were generated for the participants by shifting the time steps by -0.5, -1, +0.5, and +1, respectively. - The results of the user study fall between the values for CLODE and CLODE+. *Since CLODE already produces high-quality images through optimal solutions and empirically achieves more appealing images near the optimal state, finding user-preferred images is not challenging. Furthermore, the reason for the lower results in the user study is due to exceptionally dark reference images, even though the results from CLODE+ look better. Therefore, there is no significant difference in terms of visual quality (PI). [Image retouching]* - CLODE leverages NODE to offer user controllability as a kind of ‘free bonus feature', and this can be understood in terms of image retouching as Reviewer Z6DS's comment. However, the ability to perform image retouching through unsupervised learning is a clear advantage that makes CLODE more practical for use in diverse environments. We would also like to assert that our method demonstrates superior performance compared to various methods without user control. W2: Noise Problem - *Before the explanation, we would like to clarify that $I_T$ contains noise.* - In Section 3.2.1, the input to the ODE function is $I_t$, which is passed through the Noise Removal module to obtain the denoised image $\tilde{I_t}$ as shown in Eq.(7). The denoised $\tilde{I_t}$ is used as input to the Curve Parameter Estimation module and is provided by Eq.(8). The reason for using the denoised $\tilde{I_t}$ as an input of Curve Parameter Estimation module is to obtain a fine-grained curve parameter map $\mathcal{A}_t$. The final output for the ODE function is expressed as $\mathcal{A}_t \otimes I_t \otimes (1-I_t)$ (Eq.(9)), where we utilize $I_t$ rather than the denoised image $\tilde{I_t}$. The reason for not using $\tilde{I_t}$ is that it is difficult to preserve the details of the input image when image enhancement is performed by repeatedly denoising the image. Thus, the mentioned $I_T$ is the result of the improvement using the fine-grained curve parameter map, and since no direct denoising is performed in the iteration process, $I_T$ contains some noise. Lastly, to obtain a noise-free image $\tilde{I}_T$, we apply the Noise Removal module to $I_T$ as described in L196-197. W3: Experimental step - The reason for using a single training dataset for each task is *to ensure a fair comparison with previous methods that include both supervised and unsupervised approaches. - Since most supervised methods’ official weight parameters are trained on a single dataset, we adopted the same approach. - We agree that training with diverse datasets is beneficial and provide results from training on all LOL and SICE datasets in R-Table 7. The overall performance is comparable to that achieved with a single training dataset. W5: Losses* - The four losses, excluding $\mathcal{L}_{noise}$, were used in the same way as in the previous curve-adjustment methods. We acknowledge that some models do not account for noise, and although we mentioned the case without the noise module in Table 4, we kindly present a comparison again in R-Table 3. As the image becomes brighter, noise becomes more amplified, leading to a slightly lower SSIM. However, *our method still outperforms other methods. W6: Noise Module, Color Casts* - Our noise removal module consists of very few parameters (Model size: 0.085MB, 22,275). While this may result in lower performance compared to existing denoising models, it effectively learns during the image enhancement process in CLODE. - *For more explanation of the denoiser, we recommend referring to global rebuttal G2.* Additionally, compared to low-light enhancement models that include noise removal ([10, 13]), the unsupervised denoising performance is competitive. (R-Table 3) - CLODE enhances the image based on the color statistics of the input image in an unsupervised manner, which can lead to the occurrence of color casts. To elaborate, while curve adjustment methods preserve the details of the input image and enhance it to a naturalness, the color loss follows the Gray-World hypothesis (L224), leading to these issues. We use the same color constancy loss as in previous curve-adjustment method [6]. Nevertheless, in comparison to existing methods, CLODE exhibits superior performance in terms of naturalness image quality metrics and color matching histogram loss in R-Table 4. - For further explanation on color casts, please refer to our response to 9pkv’s W1, Q1. We apologize for the lack of detailed explanation due to space limitations but assure you that we will address any additional questions thoroughly during the discussion period.	Summary: This manuscript introduces CLODE, which learns low-light image enhancement using neural ordinary differential equations (NODE). The key innovation lies in formulating the higher-order curve estimation problem as a NODE problem. Experimental results show that the proposed approach outperforms state-of-the-art unsupervised counterparts across several benchmarks. Strengths: 1. The paper is easy to follow. 2. Using neural ordinary differential equations to address the iterative curve-adjustment update process shows better performance. Weaknesses: 1. The novelty is limited, and the technical contribution is incremental. Apart from formulating the curve estimation as a NODE problem, the paper lacks innovation，which is the main reason why I gave this paper a lower score. 2. More strong supervised baselines should be included for reference. Comparing only a few relatively weak baselines can lead to a misunderstanding of the current gap between supervised and unsupervised methods. 3. Additionally, the authors should report some perceptual metrics for better comparison. 4. The writing and the presentation need improvement. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to `Weakness'. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have discussed the limitations of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing a thoughtful review. For enhancing our paper, we have diligently reviewed the weaknesses and questions raised regarding our paper and have prepared additional experiments and answers. W1: Novelty - We cautiously wish to assert the novelty of our approach. In contrast to previous curve-adjustment methods that use discrete updates for gradual image enhancement, CLODE addresses the limitations of existing methods by reformulating them into NODE, which facilitates solving for the optimal solution in continuous space. - In addition, we would like to refer to Reviewer UK8A's mention, our motivation is strong and solid, and we address existing shortcomings in a straightforward and effective manner. - The proposed method shows optimal training results by reorganizing the curve-adjustment equation to NODE, which optimally handles the various exposure images in the inference time, different from a limitation of the existing curve-adjustment method. In addition, we designed a suitable network consisting of Noise Removal module and Curve Parameter Estimation module for this purpose (R-fig. (a)), as shown in Section 3.2.1 ODE function, and showed excellent performance compared with the existing unsupervised method. - Furthermore, by offering a user controllability that utilizes the features of NODE as shown in Fig.3 in the main paper, the potential of the model has been enhanced. User controllability features are among the most effective and practical aspects of applying NODE, as they allow for customized brightness outcomes to suit individual preferences. - In the context of unsupervised methods in low-light image enhancement, we firmly believe that the first attempt of transporting discrete curve-adjustment method problem to continuous space by neural ordinary differential equations, designing adequate architecture for the NODE method, and providing high visual performance, user-controllability, constitute meaningful contributions. Accordingly, we kindly ask for a reconsideration of the novelty of our work. W2: References (strong supervised methods) - **In the main paper, Retinexformer is referred, a transformer-based state-of-the-art method and the second-best performer in the NTIRE 2024 low-light challenge [1\], as a strong supervised comparison method*. - We are thankful for Reviewer 9uXV’s concern and we are going to add additional strong supervised diffusion-based baselines GSAD** [2] and PyDiff* [3]. Even when compared to diffusion-based supervised methods, CLODE demonstrates competitive performance. We will include performances of two methods in our final revised version. Diffusion-based method\|LSRW/LOL (PSNR)\|LSRW/LOL (PSNR mean GT) :---\|:---:\|:---: GSAD\|17.37/23.01\|19.51/27.60 PyDiff\|17.00/20.49\|20.11/26.99 CLODE\|17.28/19.61\|19.61/23.16 W3: Perceptual metrics* - In response to this question, we provided perceptual metrics (NIQE, BRISQUE, PI, Entropy) in Table 2 (SICE dataset) and Table 7 (LOL and LSRW datasets) in the Appendix. - Additionally, following the reviewer’s suggestion, we provide LPIPS (Learned Perceptual Image Patch Similarity) [4] performance results. CLODE exhibits superior average LPIPS performance compared to other methods.* \|Dataset\|URetinexNet\|RetinexFormer\|SCI\|RUAS\|ZeroDCE\|NightEnhancement\|PairLIE\|CLODE :---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---: LSRW\|0.308\|0.315\|0.398\|0.469\|0.317\|0.583\|0.342\|0.331 LOL\|0.121\|0.131\|0.358\|0.270\|0.335\|0.241\|0.248\|0.263 SICE\|0.264\|0.263\|0.486\|0.608\|0.239\|0.360\|0.305\|0.235 MSEC\|0.393\|0.362\|0.396\|0.668\|0.329\|0.462\|0.431\|0.223 Average\|0.272\|0.268\|0.410\|0.504\|0.305\|0.412\|0.332\|0.263 W4: The writing and the presentation need improvement. - Lastly, as Reviewer 9uXV advised, we are willing to improve writing and presentation of the paper based on comments. In the final revision, we will incorporate the reviewer’s suggestions and additional results and improve the quality of the manuscript. --- > Reference [1] Liu, Xiaoning, et al. "NTIRE 2024 challenge on low light image enhancement: Methods and results." In CVPRW, 2024. [2] Hou, Jinhui, et al. "Global structure-aware diffusion process for low-light image enhancement." In NeurIPS, 2023. [3] Zhou, Dewei, Zongxin Yang, and Yi Yang. "Pyramid diffusion models for low-light image enhancement." In IJCAI, 2023. [4] Zhang, Richard, et al. "The unreasonable effectiveness of deep features as a perceptual metric." In CVPR, 2018.	Summary: This paper This paper proposes an ODE-based method to tackle low-light image enhancement problems. The motivation of the paper is inspired by the observation that the conventional discrete iterative approaches set fixed update steps. It does not only miss the optimal solution and also does not guarantee the convergence. Hence, the proposed method takes the iterative curve-adjustment approach and formulates them into solving neural ordinary differential equations. This method is used to work with unsupervised learning to estimate the higher order curve parameters to reconstruct image structure details. Comprehensive experiments demonstrate that the proposed method outperforms the baseline methods on LOL and SICE benchmarking datasets. Strengths: Strengths 1. This paper proposes a novel method that integrates the neural networks into an ODE optimization framework. The neural network is playing an adaptive set of updatable parameters. 2. Comprehensive experiments show that the proposed method outperforms the baseline methods in the task of low-light image enhancement. 3. The motivation of this paper is strong and solid. It is inspired by the drawbacks of the existing methods and tackle the problems directly in the proposed method. 4. This paper addresses the limitation of the proposed method. Weaknesses: Weakness 1. Based on the visual comparison in Figure 4, the proposed method tends to produce over-exposed areas for highlight regions. 2. The processing speed of the proposed method is one of the limitations. Technical Quality: 3 Clarity: 3 Questions for Authors: This paper is a novel low-light image enhancement work with solid equation derivation to support the its objective. The authors are expected to address the questions raised in the weakness sections in the rebuttal period. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitation is included in the main manuscript. The processing speed (inference speed) of the proposed method is slow compared to dedicated supervised DL methods. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are glad to hear that you found the paper to be strong and solid. We have diligently examined your comments and concerns as a reviewer, and have prepared responses addressing the raised concerns. W1: Highlight region - We understand the reviewer’s concern in Fig. 4. As a reminder, CLODE is trained in an unsupervised manner without ground truth images and loss functions are aimed to enhance regions of the image to the desired exposure values (e.g., $E$=0.6) while preserving spatial and color constancy based on the input image statistics. - Although we employ spatial consistency loss and color constancy loss, because our method relies solely on the input image statistics, and the inferred color may appear grayish in enhanced images due to the Gray-World hypothesis (L224). This is a significant issue that needs to be addressed in unsupervised methods. We have spent a lot of effort searching for an appropriate color loss to apply to unsupervised methods, but we could not find a more suitable loss for environments that use only a single input image. - However, we would like to emphasize that our model is more robust to over-exposed images compared to other models. The third row in Fig.4 compares the results of each model for a given over-exposed image, and our model is the most robust to the over-exposure situation. The tonal adjustment result is closer to the ground truth than the other methods. - Over-exposed images partially contain pixel values outside the color range, which do not provide sufficient information for image enhancement. As can be seen in Fig.4, this issue is also present in supervised methods. However, we would like to underline that CLODE can provide visually reasonable results even for over-exposed images. Please also have a look at Fig.15 in the Appendix for additional results. Our method shows visually impressive results on various exposure conditions. Additionally, resolving issues with over-exposed images will require the use of a generative model, which we plan to explore in future research. W2: Processing speed - In Sec. 5 Limitations, as Reviewer UK8A concerned, we reported less inference speed compared to existing methods. As Reviewer UK8A mentioned, the inference speed of CLODE is comparable to that of RetinexFormer. However, in Table 5 and L332, we also reported CLODE-S which is a compact version of the proposed method. CLODE-S consists of 0.0004M parameters and takes 0.005 second to infer, and its architectural details are shown in R-Fig.1(b). In addition, we remain a distillation approach for shortening inference time as a future work. - Another possibility is to apply RectifiedFlow [1] in an unsupervised manner, which transforms the solution paths of Neural ODEs into straight lines, facilitating faster estimation of the Neural ODE system. We can initially assume that the optimal solution found by CLODE aligns with the expected optimal solution in [1], and then apply CLODE specifically to [1]. - Given the potential to implement advanced Flow Matching methods [1, 2], we believe CLODE holds great promise and could achieve rapid inference speeds. --- > References [1] Liu, Xingchao, and Chengyue Gong. "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow." In ICLR, 2023 [2*] Tong, A., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Fatras, K., ... & Bengio, Y. "Improving and generalizing flow-based generative models with minibatch optimal transport." In Transactions on Machine Learning Research, 2024 --- Rebuttal Comment 1.1: Title: Comment after rebuttal Comment: Thank you for the authors' feedback. The weakness concern is addressed in the rebuttal. --- Reply to Comment 1.1.1: Comment: Thank you for your response. It was a pleasure to write the rebuttal, and we appreciate your continued support of our work. Once again, thank you for your efforts, and we will incorporate the feedback into the final version.	Rebuttal 1: Rebuttal: To the reviewers. First and foremost, we appreciate the reviewers' efforts. We have prepared responses to all comments, along with figures and tables in the attached PDF that show additional experiments to enhance our explanations. Below are brief explanations for each figure and table: There are 4 figures and 7 tables in the attached PDF. R-Figure: * 1. UK8A W2, Qjwa Q2, 9pkv W2,Q2 * (a): Architectural details of two modules in CLODE. * (b): Architectural overall of CLODE-S which consists of a 2-layered network. * 2. Z6DS Q1, W7: Depicting inference trajectories of (a1) to (e1) and CLODE in Table 3 of the main paper. * 3. 9pkv Q3: The plot of loss values of under exposed and normal exposed input images according to time steps. * 4. 9pkv W4, Q4: Examples of employing existing denoising modules. R-Table: * 1. 9pkv W3, Q4: Results of existing denoising methods. * 2. 9pkv W3, Q4: Results of denoising ablations. * 3. Z6DS W5: Comparison results without using $L_{noise}$. * 4. 9pkv W1, Q1: Results of Color matching histogram loss * 5. 9pkv Q3: Results of step statistics * 6. Z6DS W1, W4: User study of CLODE+ * 7. Z6DS W3: Results of CLODE trained on LOL + SICE. --- Additionally, we have prepared global rebuttals for Z6DS Q1, W7 and 9pkv W3, Q4, including *G1. Explanation of CLODE* and *G2. Weakness of Denoiser. G1. Explanation of CLODE* - As mentioned in [1], ODE reformulation provides benefits such as continuous space estimation, memory efficiency, and accurate problem-solving with ODE solvers. We’ve attached R-Fig.2* for additional analysis. - The top of R-Fig.2 shows discrete trajectories of models (a1) to (e1) from Table 3, while the bottom displays CLODE trajectories with Euler and dopri5 solvers. (Top) Discrete methods (a1) to (e1) enhance images but don’t achieve optimal exposure. (Bottom) CLODE (dopri5) provides more realistic image enhancement in continuous space. - CLODE (dopri5) uses an *early stop mechanism. It tracks error at each state, terminating when the error is within allowable error rate. For dopri5, $k$-order solutions ($k$=5) are used to calculate error ($Γ_t$) as follows: $$Γ_t = atol + rtol × norm(\|O_t^k - O_t^{k-1}\|), (Eq.(29))$$ - where the $k$-order solution at time $t$ is denoted as $O_t^k$ and the $(k-1)$-order solution is denoted as $O_t^{k-1}$. - If $\|O_t^k - O_t^{k-1}\| > Γ_t$ the step size is re-adjusted. If it’s within $Γ_t$, the solution is deemed optimal, and the process terminates. - ODE solvers are designed to find optimal solutions through iterative steps. R-Fig.2* shows that discrete methods can’t guarantee optimal solutions, which led us to develop the NODE method for continuous ODE problems. Thus, improvements are due more to NODE reformulation than to iteration count. Table 3 shows that NODE outperforms simple discrete repetition. For example, using the Euler method in 30 steps achieves better performance than method (e1). - As shown in Fig.9 and R-Fig.2, many steps are found initially because adaptive solvers like dopri5 need initial step estimation with higher-order solutions to ensure accuracy. - Dopri5 uses higher-order solutions to ensure the accuracy of the optimal solution, as seen in Eq (29), requiring at least 6 evaluations per state. Thus, it uses short intervals initially to store evaluations. We chose dopri5 as CLODE’s default solver for its stability and reliability across platforms like MATLAB. G2. Weakness of Denoiser [Effectiveness of denoiser] - Since NODEs do simulation-based training, as the denoiser (Noise Removal module) gets heavier, it consumes more time for training. Therefore, we employ a lightweight 3-layer network (0.085MB) as the denoiser in CLODE. - Noise removal module has a fewer number of parameters, but accomplishes an important task. The module learns to denoise the image at each step (Eq.(7)), thus learning jointly with the image enhancement process and helping to predict the fine-grained curve parameter maps at each step. - We train denoising during the image enhancement process to effectively train our noise module. - To demonstrate this effect, we compared three different scenarios (Pre-denoising, CLODE, Post-denoising). For clarity, "Pre-denoising" trains the denoiser only on the input image $I_0$, while "Post-denoising" trains the denoiser only on the enhanced image $I_T$. - CLODE shows outperforming results, which implies that CLODE is the best scenario among them. (R-Table 2) - The reason for these results is low-light images have low pixel values, which provide insufficient information for denoising. After enhancement, the original noise becomes entangled with the image content. Thus, we believe that continuous denoising is crucial for effective low-light correction because noise is amplified with continuous exposure enhancement. [Using existing methods] - We could find performance improvements by utilizing existing denosiers to get $\tilde{I}_T$. - For experiments, we adopted DnCNN [2] and Restormer [3] as a denoiser. In quantitative terms, there is 0.66 gain in SSIM by [3] and 0.1dB gain in PSNR by [2] and also in R-Table 1, we achieve visually outstanding results. (R-Fig. 4) - Thanks to 9pkv's suggestion, we have demonstrated that using the existing denoiser or increasing the size of the module can also be beneficial. --- > References [1] Chen, Ricky TQ, et al. "Neural ordinary differential equations." In NeurIPS, 2018. [2] Zhang, Kai, et al. "Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising." In IEEE transactions on image processing, 2017. [3*] Zamir, Syed Waqas, et al. "Restormer: Efficient transformer for high-resolution image restoration." In CVPR, 2022. Pdf: /pdf/1342a40163d606b2647d31aaf59271d96584c1fe.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper proposes a Neural ODE method for curve-adjustment-based low light image enhancement methods to achieve better results which are often sub-optimal for fixed discrete step methods. Specifically, the proposed method reformulates the curve-adjustment-based from the discrete version into the ODE problem by introducing a continuous state. An ODE solver is adopted for the optimization to find the optimal step for the enhancement. Additionally, a simple denosier and a curve parameter estimation module are proposed for noise removal and parameter estimation, respectively. Extensive experiments are conducted to show the effectiveness of the proposed method. Strengths: 1. Turn the discrete curve-adjustment method into a NODE problem, benefited from the optimization to search for the optimal step. 2. User control support during inference is good for the application of the proposed method. 3. The proposed method seems to have good performance over other competitors. Weaknesses: 1. The proposed method faces color casts, which is obvious in almost all qualitative results, even with a color constraint in loss functions. 2. The proposed method proposes to denoiser and curve parameter estimator in the NODE framework, however, generalize the method to existing curve-ajustment-based method seems to be a more attractive solution. 3. The denoiser seems to be weak since there is so much noise left for the qualitative results. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What leads to the color cast, even with color constraints applied. And it seems that the steps increases, the color shifts become stronger. Comparisons with user-controlled result are also evident for the color shifts. An evaluation on color consistency is necessary for this issue. 2. Is there any potential to generalize such a method to previous methods, like turn previous methods into NODE and find optimal steps for them? 3. In Table 3 in the main paper and Figure 9 in the supp., the comparisons between discrete methods and the proposed method show good advantage of the proposed method over previous methods. But what is the statistics for a dataset or some test examples, like the most common optimal steps for inextreme cases? It may help to get interesting findings. 4. Why is the denoiser so weak? Can it be improved by use existing modules? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, it is discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for thoroughly reviewing our paper. We appreciate your feedback. To address the concerns you raised, we are providing several experimental results and our perspectives. W1, Q1: Color cast [What leads to the color cast?] - CLODE enhances the image based on the color statistics of the input image in an unsupervised manner, which can lead to color casts. To elaborate, while curve adjustment methods preserve image details and enhance naturalness, color loss following the Gray-World hypothesis (L224) can lead to color cast issues as the exposure level changes. - CLODE follows the same color constancy loss Eq. (12) as previous zero-reference based method [6] for color constancy. Nevertheless, compared to existing methods, *CLODE using NODE scenario exhibits superior performance in terms of naturalness image quality metrics and color matching histogram loss. (R-Table 4)* [Color-matching histogram loss function] - To address Reviewer 9pkv’s color casts concern more precisely, we prepare additional comparison results on *color-matching histogram loss function* in HistoGAN [1] which is designed for controlling color of input image in [1] by matching color histograms of target image and input image. We utilized *color-matching histogram loss* to measure color casting degrees from ground-truth images to output images. - In the R-Table 4, we compare non-reference metrics and color-matching histogram loss results with existing unsupervised methods. The color-matching histogram loss of CLODE+ and CLODE ranked first and second-best in SICE dataset and first and third-best in LOL dataset. This indicates that CLODE exhibits fewer color casts compared to existing unsupervised methods, despite achieving high-quality enhancement results. W2, Q2: Generalizing the method - As suggested, CLODE can be generalized to existing curve adjustment methods. While it is possible to apply NODE to existing architectures, accurate curve estimation is crucial for high quality. Therefore, we developed a new compact and efficient architecture that can effectively apply on NODE and estimate the fine curves as depicted in R-Fig.1 (a). - Regarding the conversion of previous methods to NODE, our experiments with the existing network [6] demonstrated that while applying NODE is feasible, it is not effective. ([6]_large* is a version of the network from [6] with increased parameters for performance improvement.) Method\|#params (M)\|PSNR/SSIM :---\|:---:\|:---: [6]+NODE\|0.0794\|17.17/0.571 [6]_large+NODE\|0.3593\|19.26/0.637 CLODE\|0.2167\|19.61/0.718 - We can also set the maximum allowable steps for CLODE, but instead of setting a fixed optimal step, using optimal steps is effective for diverse image datasets. (Please refer to R-Fig.2 and global rebuttal G1 for further explanation of CLODE). - Thus, we developed our own network (CLODE) using an adaptive solver with optimal steps that must vary for each image. - For additional explanations regarding the optimal step, please refer to the following Q3. Q3: Optimal step statistics* - We believe this concern is very constructive. As previously mentioned in L465, the maximum allowable step for the ODE solver is empirically set to 30, considering speed. The ODE solver terminates early if it finds the optimal solution within the maximum steps. - Furthermore, in order to provide the statistical analysis proposed by 9pkv, the average number of steps for the ODE solver was calculated across the SICE, BSD100, DIV2K, and LOL datasets. The SICE dataset comprises five to seven images per sample, with exposure levels ranging from under-exposed to over-exposed. Additionally, BSD100 and DIV2K were used to provide additional statistics for normal-exposed conditions. The results based on the exposure conditions are in R-Table 5. - The number of calculation steps increases in the order of under-exposed, over-exposed, and normal-exposed images. This is because improving a normal-exposed image is considered a stiff problem. For normal-exposed images, where minimal improvement is required, the dynamics are more stiff than for other images. Specifically, for dynamically stiff ODE problems, the step size taken by the solver is forced down to a small level even in a region where the solution curve is smooth, and these decreased step sizes may require more evaluation steps. - For a better understanding, we plotted R-Fig.3. In the inference time, CLODE aims to find the optimal solution by minimizing the loss functions, so in R-Fig.3 the y-axis represents the non-reference loss value, the x-axis represents time, and each point indicates a step. - The dopri5 solver we primarily use is a nonstiff solver. Although a stiff solver (e.g., ode15s) could be employed to shorten inference steps for normal-exposed images, it is not efficient for improving under- or over-exposed images that are close to nonstiff problems. Thanks to the quality suggestion, future research may explore ode solver algorithms that are dynamically adapted based on input image conditions. W3, Q4: Weakness of denoiser - First of all, as mentioned in Q4, our method can be improved by adopting existing denoising modules. However by considering the inference speed issue of the model, we designed a compact size of denoiser. For 9pkv’s worry of the effectiveness of denoiser, we cautiously want to assert Table 4 in the main paper. From comparison of case (2c) and (2d), we could find the Noise Removal module (denoiser) raises PSNR 0.94dB and SSIM 0.141. *To address the concern clearly, we prepared experiments on diverse aspects (+using existing denoiser) in global rebuttal. Please refer G2 in global rebuttal. --- > Reference [1] Mahmoud Afifi, Marcus A. Brubaker, and Michael S. Brown. "HistoGAN: Controlling Colors of GAN-Generated and Real Images via Color Histograms." In CVPR, 2021. --- Rebuttal 2: Comment: Thank you for your feedback. My concerns are well addressed, and I will raise my rating to borderline accept. --- Rebuttal Comment 2.1: Comment: Thank you for your response and constructive questions. We will include rebuttal content in the final vision. Once again, thank you for taking the time and raising our score.	null	null	null	null	null	null
Unconditional stability of a recurrent neural circuit implementing divisive normalization	Accept (poster)	Summary: Stability is a critical notion in the understanding of dynamical systems as well as for learning them. This paper studies the stability of a biologically plausible recurrent cortical model, ORGaNICs, that implements divisive normalization. More precisely: - it demonstrates the local stability of two specific subclasses of the model. - it shows that the model performs competitively on standard machine learning tasks. Strengths: Understanding how properties of biological neural networks lead to better learning properties is an important area of research and this paper nicely contributes to it by highlighting the potential importance of divisive normalization in stabilizing neural dynamics and facilitating training. The paper is overall well-written and nicely connected to both theoretical and empirical neuroscience results. The theoretical stability analysis is non-trivial and provides important insights into the model. The empirical results are solid. Weaknesses: The main weakness of the paper is that it might not be fully suited to the NeurIPS community in its current form: - the introduction of the dynamics lacks some intuition on what the dynamics achieves. Equation 2 partially does so, but is likely not sufficient to fully convey the intuition. If possible, connecting to existing RNN architectures would help. Introducing the Lypanuov energy earlier could also help. - many notions used in Section 4 might deserve a better introduction as NeurIPS is not a dynamical systems conference. This holds for "Lyapunov diagonally stable", "Z-matrix", "M-matrix" and, to some extent, to "stability" and "indirect method of Lyapunov". To make their results more accessible, the authors may want to provide intuitive definitions of these terms and why they are important. Additionally, the connections made to the machine learning literature are sometimes imprecise: - in L43, it is written that stability helps mitigate the vanishing gradient problem, which is inaccurate (e.g. a system with very fast time constants is stable but suffers from vanishing gradients). On top of that, stable recurrent networks having slow time constants can still suffer from some form of exploding gradients [Zucchet and Orvieto, 2024](https://arxiv.org/abs/2405.21064). - in L61, the authors write that divisive normalization generalizes batch and layer normalization. While there are definitely some links, this is not true: divisive normalization affects the recurrent dynamics whereas both batch and layer normalizations are usually applied after the recurrent layer. In particular, those normalization schemes do not affect the stability of the system. - in L349, divisive normalization is directly compared to the form of attention used in Transformers. Transformers normalize over the time axis, which is highly implausible and different from the type of attention mentioned here. This point deserves to be made more precise. Finally, there are some additional links to existing ML literature that would nicely fit within the paper: - the kind of neural network studied in Section 6.1 is known as [Deep equilibrium models, DEQ](https://arxiv.org/abs/1909.01377). One of the main difficulties that come with training such models is to [keep the dynamics stable](https://arxiv.org/abs/2106.14342) throughout learning. The studied architecture may be less prone to these behaviors: adding the DEQ baseline to the experiments and monitoring the stability of the dynamics during training would be interesting additions to Section 6.1. - in Section 6.2, the baseline architectures are rather "old". The state-of-the-art networks on tasks requiring modeling long-range dependencies are state-space models (e.g. [S4](https://arxiv.org/abs/2111.00396) or [LRU](https://arxiv.org/abs/2303.06349)). Additionally, sequential models are known to be extremely sensitive to initializations on tasks such as sMNIST, see [Amos et al. 2023](https://arxiv.org/abs/2310.02980). It is therefore important to describe the initialization scheme of the recurrence in the main text, as it is likely a main driver of the performance. I hope these remarks can help the authors to improve their paper. Technical Quality: 4 Clarity: 3 Questions for Authors: - The model considered has the same number of E and I neurons, which is, to the best of my knowledge, far from being the case in the brain. Is the system still stable when the number of I neurons is much smaller than the number of E neurons? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Limitations are properly addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: - “the introduction of the dynamics … why they are important”. Answer: We thank the reviewer for the insightful feedback. We will add more intuition about the dynamics in the model description and compare ORGaNICs to LSTM, an RNN architecture similar to ORGaNICs. We will also provide intuitive definitions of the terms used in the stability proof. - “in L43, it is written that stability … Zucchet and Orvieto, 2024”. Answer: We appreciate the referee's observation. While stability indeed addresses the issue of exploding gradients, excessive stability—where the real parts of the eigenvalues of the Jacobians are significantly smaller than 0—can lead to a lossy system prone to vanishing gradients. ORGaNICs effectively mitigates both issues: Exploding gradients: through its inherent stability. Vanishing gradients: by processing information across various timescales (and via the long effective time constant when the membrane time constants are fixed) while maintaining stability, resulting in a blend of lossy and non-lossy neurons. The efficacy of ORGaNICs in mitigating vanishing gradients is demonstrated by its competitive performance against architectures specifically designed to address this issue, such as LSTMs. We will provide a more detailed explanation of this in our revised manuscript. Moreover, we posit that the built-in normalization feature of ORGaNICs may alleviate the "curse of memory" described by Zucchet and Orvieto (2024), as normalization is proposed as a potential solution to this problem. - “in L61, … stability of the system”. Answer: Divisive normalization (DN) was introduced as a model of steady-state response of neurons, functioning as a static nonlinearity similar to batch and layer normalization. However, in the brain, there are no static nonlinearities; it is proposed that DN is achieved via a recurrent circuit. ORGaNICs is such a recurrent circuit designed so that the responses of neurons at steady state follow the DN equation, with stability being an emergent property of the circuit. The DN equation indeed generalizes batch and layer normalization, as shown in the work by Ren, M., Liao, R., Urtasun, R., Sinz, F. H., & Zemel, R. S. (2016). - “in L349, … made more precise”. Answer: The key similarity is that both operate via changing the input gain, which also models a wide range of results from neurophysiological and psychophysical experiments. In particular, it has been shown that normalization explains experimental measurements of temporal attention (Denison, Carrasco, Heeger, Nature Human Behavior, 2021), i.e., over the time axis, analogous to transformers. - “the kind of neural network … additions to Section 6.1”. Answer: We thank the reviewer for pointing this out. We checked if the models dynamics are stable throughout training and found that this is indeed true. Please see Fig.2 of the attached pdf. - “in Section 6.2, … of the performance”. Answer: We are aware that state-space models perform better at these tasks and we will mention this (along with citations) in the paper, but our goal is to compare ORGaNICs with other RNNs designed with the property of less-dissipation and stability to mitigate the problem of vanishing and exploding gradients. We thank the reviewer for pointing out the sensitivity of initialization on sequential modeling tasks. We have added that detail (uniform random initialization) in the revised manuscript. Questions: - “The model … E neurons?” Answer: Experimental data suggest that the ratio of E and I neurons is 4:1. Our results are also applicable for a different E/I ratio than 1. --- Rebuttal Comment 1.1: Comment: I acknowledge the rebuttal and keep my score as it is.	Summary: This work is based on ORGaNICs, a particular type of RNN architecture that implements divisive normalization (from neuroscience). The authors explore whether ORGaNICs are stable enough to be meaningfully (and stably) trained by gradient descent. Due to their stability, which the authors have proven theoretically for specific cases and empirically in general, the authors claim that ORGaNICs can learn long-term dependencies and solve vanishing and exploding gradients found in other neurodynamical models. Strengths: Trained RNNs in neuroscience have always had a standard vanilla architecture or have minimal modifications. This work introduces the possibility of training a more complex model motivated by biology while having no stability issues. Weaknesses: Subjectively, I believe that the scope of this work is too limited for the audience at NeurIPS, but I will not consider this in my final score. The general theme of my objective issues with this work is the slight overselling of this work: 1. Fundamentally, ORGaNIC is an architecture. Song et al. [2016] implemented Dale's law by manipulating the recurrent weight matrix, which is a method applicable to any architecture. Soo et al. [2024] trained vanilla RNNs on long-term dependencies using skip connections through time, which once again is a method applicable to any architecture. So the claim that these "models do not accurately reflect cortical dynamics" in lines 105-106 do not make sense to me. To illustrate my point, can the authors comment on whether there are any issues if Soo et al. [2024] is applied to ORGaNICs (beyond the fact that it is not needed)? 2. This work seems to be strongly driving the point that divisive normalization is a key criteria for biological plausibility. There are a variety of biological properties that the brain exhibits. For example, there are 22 different properties of V1 identified in Marques et al. [2020], of which specific versions of DN make up some of them. It is not convincing to select a model based on one particular effect found in V1. Marques et al. [2020]: https://www.biorxiv.org/content/10.1101/2021.03.01.433495v2 3. The authors claim that ORGaNICs is stable and therefore solves exploding and vanishing gradients, which does not make sense. The way the authors used the word "stable" here is in the language of dynamical systems. "Stability" in the context of exploding and vanishing gradients has a different meaning. The fact that the model can learn sequential MNIST (gradient stability) does not mean that it is due to its (dynamical) stability. A completely (dynamically) stable architecture can still have vanishing gradients. Lines 115-116 suggest to me that the authors are somewhat aware of this but not completely. Technical Quality: 3 Clarity: 4 Questions for Authors: ORGaNIC begins with the word "oscillatory". Can the authors comment on whether the network must always be oscillatory? I see that the analysis looks into stable and unstable fixed points, which means that the answer is no, which makes the name confusing to me. Likewise, can the authors elaborate on the "gate" in ORGaNIC and how it is the "biophysically plausible extension" (line 86) of gating in machine learning? Specifically, why it is biophysically plausible, and why it is an extension. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 1 Limitations: In line with my issue of overselling this work, there really is not any honest discussion of genuine limitations in the discussion section, which the authors claim they have done so in the checklist. Discretization problems (lines 350-354) and using single-layers (line 355) are not specific to this work at all. I would encourage the authors to talk about genuine limitations. For example, I suspect but cannot prove, that there would be some increased computational complexity compared to vanilla RNNs, since having multiple differential equation for a single model means that more things need to be done in a single time step. This work, focusing on a biologically plausible model, shows the model performing machine learning experiments (in contrast, in Heeger and Mackey [2019] the model is performing neuroscience experiments). Also, while this model implements divisive normalization, as I mentioned above there are many biological phenomena that the brain exhibits. It would be unreasonable to ask the authors to state what ORGaNICs cannot do, but at least I would like to ask the authors if ORGaNICs can do everything that SSNs (which they cited) can do? These are the limitations that I believe readers would like to see. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: - “Fundamentally, ORGaNIC … one particular effect found in V1”. Answer: DN is observed in numerous cortical areas beyond V1 and across different species (Carandini & Heeger, 2012). It explains a wide range of experimental phenomena in various neural systems and cognitive processes (citations in lines 49-53). DN can also be shown to be the appropriate nonlinearity for removing dependencies in sensory inputs, resulting in a set of neural responses that are statistically independent (Simoncelli and Olshausen, 2001). Consequently, DN is not just an effect observed in V1, but a computational principle (Carandini & Heeger, 2012) that can be derived from a statistical optimization criterion. It is thus unsurprising that DN has been shown to dramatically improve the performance of ML models in several applications (e.g., Balleé, Laparra, Simoncelli, 2015; Balleé, Laparra, Simoncelli, 2017). Moreover, batch and layer normalization (of which DN is a generalization, see Ren, M., Liao, R., Urtasun, R., Sinz, F. H., & Zemel, R. S. (2016)) are critically important architectural motifs in modern deep learning architectures. There is thus overwhelming evidence that DN is an essential feature of both natural and artificial neural systems. In this paper we demonstrate that DN can be integrated into the dynamics of an RNN, leading to a provably stable circuit for which an interpretable normative principle (Lyapunov function) can be derived. This work does not just propose an architecture, but provides 1) unprecedented insight into the impact of normalization on the dynamics of RNNs, through an interpretable normative principle; 2) establishes an important precedent for how the incorporation of neurobiological principles can drive advances in ML. - “The authors claim … aware of this but not completely”. Answer: In RNN architectures derived from discretized ODEs, such as ORGaNICs, the stability of the dynamical system is intrinsically linked to the vanishing and exploding gradients (VG and EG) problem. Mathematical analysis and details can be found in, Haber and Ruthotto (2017); Chang, Chen, Haber, and Chi (2019); and Erichson et al. (2020). ORGaNICs effectively addresses both problems: EG: Mitigated through the architecture's inherent stability. VG: Addressed by processing information across various timescales (and via the long effective time constant when the membrane time constants are fixed), resulting in a blend of lossy and non-lossy neurons. The effectiveness of ORGaNICs in tackling VG is evidenced by its competitive performance against architectures specifically designed to address this issue, such as LSTMs. Questions: - “ORGaNIC begins … name confusing to me”. Answer: ORGaNICs is oscillatory in the sense that it can be mapped to a damped harmonic “oscillator” as we show in the paper. This means that for the right parameters and inputs, stochastically driven ORGaNICs exhibit a peak in the power spectrum, which is consistent with the LFP (Local Field Potential) oscillatory activity observed in neural recordings, even though there are no sustained oscillations. - “Likewise, can the authors elaborate … it is an extension”. Answer: Gating in ORgaNICs is performed by modulating the input gain ($\mathbf{b}$) and the recurrent gain ($\mathbf{1} - \mathbf{a}^+$). Possible mechanisms of how the responses of these neurons may be computed by loops through higher visual cortical areas and/or thalamocortical loops are discussed in the Discussion/Mechanisms of Heeger, D. J., & Zemlianova, K. O. (2020) and its SI Appendix. The gate in ORGaNICs is an extension of the gates in LSTMs/GRUs because the recurrent gain/gate in ORGaNICs (unlike LSTMs/GRUs) is a particular nonlinear function of the output responses/activation, designed to achieve normalization. Limitations: - “In line with my issue … in a single time step”. Answer: The computational cost of ORGaNICs is similar to LSTMs since like LSTMs we have three extra sets of variables ($\mathbf{a}, \mathbf{b}, \mathbf{b}_0$) which are not directly used for prediction. Therefore, ORGaNICs is more computationally expensive than vanilla RNNs. We thank the reviewer for pointing this fact out. We have included this fact in the discussion. However, we do not consider this a limitation, as the increased computational cost results in improved stability, trainability, and interpretability. - “This work, … neuroscience experiments)”. Answer: Since training biologically plausible neurodynamical models using backpropagation is a challenging task (Soo, W., & Lengyel, M. (2022)), we demonstrate the trainability of ORGaNICs by naive backpropagation and backpropagation through time, without gradient clipping/scaling, on well studied ML tasks. Training ORGaNICs on neuroscience experiments will be done in future work. - “Also, while this model implements … SSNs (which they cited) can do?” Answer: ORGaNICs has been shown to model a wide range of neural phenomena, including sustained activity, sequential activity, motor preparation, and motor control (Heeger, D. J., & Mackey, W. E. (2019)), as well as simulate the dynamics of V1 activity (Heeger, D. J., & Zemlianova, K. O. (2020); S. Rawat, D.J. Heeger, and S. Martiniani. Cosyne Abstracts 2024) and the emergence of communication subspaces (S. Rawat, D.J. Heeger, and S. Martiniani. Cosyne Abstracts 2023). To address the reviewer’s question, ORGaNICs can do everything that an LSTM can do (while being stable, easily trainable, and interpretable). While a direct comparison of ORGaNICs and SSN across many tasks has not been done, on the one task we consider (static MNIST, following Soo, W., & Lengyel, M. (2022)) ORGaNICs outperforms SSN on the first try (i.e., without hyperparameter optimization). Consistent with our claims, ORGaNICs also performs comparably to LSTMs on sequential tasks such as sequential (and permuted) MNIST. --- Rebuttal Comment 1.1: Title: Early response Comment: I thank the authors for the thorough response. I will reply again in a few days after carefully reading them. But I need to post this early response first to give the authors a chance to reply. In my original review, I mentioned the point about the lack of any honest discussion of limitations, and then suggested three possible points for the authors to discuss as a limitation. Perhaps the authors felt the need to defend them as if they were criticisms, which is why I am writing this post to clarify: those points are not criticisms and I am truly encouraging them to give limitations of their work just like in any complete piece of research. Basically I am saying that in the "limitations" part of the response, I still do not see any limitations. --- Reply to Comment 1.1.1: Title: Discussion of limitations Comment: In addition to the general limitations (discretization problem and using single-layers), we also discussed other limitations throughout the text. For clarity, we will summarize them again in the Discussion in a new version of the manuscript. These are, - Limitation #1: We pointed out very clearly that we cannot rigorously prove stability for the case of a general recurrent weight matrix. So our most general results rely on empirical evidence obtained through extensive numerical tests presented in the original manuscript, as well as additional results presented in response to the reviewers (see Fig.1 of the attached pdf for a demonstration of stability from random initial conditions). - Limitation #2: We are not yet taking advantage of the modulators to control the effective time constant and instead are learning the intrinsic time constants. This potential limitation was pointed out by reviewer NGE1 who asked for further validation of the model with fixed time constants. We welcomed the suggestion and performed further numerical tests in which we found that ORGaNICs achieves good performance even when we fix the intrinsic time constants (as shown in Table 1 of the attached pdf). - Limitation #3: In the current work, the weight matrices $ \mathbf{W_{by}}$, $ \mathbf{W_{ba}}$, $ \mathbf{W_{b_0 y}}$ and $\mathbf{W_{b_0 a}}$ are $n \times n$. This leads to the number of parameters increasing faster than other RNNs with hidden state size, as is evident from Table 2. This was pointed out by reviewer 2CUH. We will stress this limitation further in the updated version of the work. In future work, we plan to test the performance with compact and/or convolutional weights so that the # of parameters does not increase markedly with the hidden state size. To the extent that we tested the model for the purpose of demonstrating stability, and trainability by SGD without gradient clipping/scaling in a standard ML task, these are the only limitations that we encountered. We will add a section to the Discussion to summarize them clearly. Based on our current understanding and results, we have no evidence (theoretical or experimental) of additional limitations. In future work, we will address the challenges noted above, and benchmark the model extensively on additional ML tasks. If the reviewer has specific questions that we have not thought of, we would be happy to investigate them and perform additional numerical tests as we did in response to other reviewers. --- Rebuttal 2: Title: Good response Comment: I thank the authors for the follow up response. I am glad that authors finally understand the position I am coming from, so I will continue with additional comments. My main point, since the very first review I posted here, is that the tone of the work is a little overblown, citing the following examples: 1. criticizing method papers that are not specific to any architecture to promote your own model (point 1 of my weaknesses that remains unaddressed) 2. absolutely judging every model based on whether it can or cannot do DN (point 2 of my weaknesses that the authors doubled down on in the reply, highlighting the importance of DN) which leads me to feel like the paper is not coming from a genuine standpoint of trying to contribute to academia, but rather trying to sell itself. Also, it is appalling that the authors can claim "we have no evidence (theoretical or experimental) of additional limitation" when V1 has so many interesting phenomena (see bolded below of biological traits that ORGaNIC does not have). Again, from my purely sincere point of view for this work to read like a good paper, might I suggest a more contributing narrative that fosters collaboration and improvement as a field instead of relying on criticism: - the vanilla RNN architecture was adapted to be biologically realistic by [Song et al. 2016] by incoporating Dale's Law. [Soo et al. 2024] developed a method for such RNNs to learn long-term dependencies. ORGaNIC is a model that is already built on biological principles, and can learn long-term dependencies intrinsically, therefore not needing any of those results. (those papers provided methods, not models to be compared with ORGaNIC) - DN is a phenomenon found in the brain, and ORGaNIC is able to express this effect. The biological visual system also expresses other phenomena, such as surround suppression, adaptation, attention-based mechanisms, foveal and peripheral vision, retinotopic mapping, columnar structure, binocular and monocular effects, unique processing of color and plasticity. Other models have been built to study those traits, but here we focus on DN as it is one of the more interesting effects that is believed to stem purely from dynamical effects. (Edit: To be clear, these are examples of how to write a collaborative narrative. I am in no way forcing the authors to include anything in the paper.) Once again, I am glad that the authors are now giving genuine limitations, and I hope that they can continue to accept my constructive criticism on the points above. I will reply again right before the deadline to make my final decision. --- Rebuttal Comment 2.1: Title: Thank you Comment: As the discussion period is coming to an end, I realize that this back and forth might have required too much time and effort from the authors, so I will post my concluding remarks now, giving the authors full benefit of the doubt in outstanding issues. Overall, the main rebuttal has addressed many doubts that I have, and I will raise my score by 1. This work represents progress for previously handcrafted models in neuroscience to be trained and actually perform tasks, which is a meaningful contribution to neuroscience. --- Reply to Comment 2.1.1: Title: Updating statements and summary of key results Comment: As we noted in our original response, the model’s limitations had already been stated in the original manuscript. We will summarize them in the Discussion. The paper investigates a neurobiologically plausible RNN model that achieves divisive normalization (DN) through recurrent excitation. This model has been shown to recapitulate a broad range of neurophysiological observations, but the application of this theory to neuroscience is not the focus of the current paper. The core contributions can be summarized at a lay level as follows: - Unconditional Stability: ORGaNICs is shown to be unconditionally stable under mild constraints (this is highly nontrivial). We develop mathematical machinery to prove this in a couple of limiting cases, by expressing ORGaNICs as a mechanical (specifically, gyroscopically stabilized) system – to do this we even need to prove new results on mechanical systems. - Implications for ML: This stability allows ORGaNICs to be trained using naïve BPTT without gradient clipping/scaling, performing comparably to LSTMs. The point is not simply that ORGaNICs is at least as good as LSTMs, but that imposing DN dynamically makes training more efficient and robust. Moreover, ORGaNICs trained by naïve BPTT fares well when compared to SSN trained by a DNG (the comparison is to SSN, not to DNG). - Dynamic Normalization: Unlike standard RNNs, where normalization is imposed a-posteriori to the output layer, ORGaNICs implements normalization dynamically: this is the first time this idea is tested in ML and analyzed theoretically. We believe that this is an important contribution. In contrast layer/batch normalization does not affect the stability of the system. - Theoretical Insights: The connection to mechanical systems enables deriving a Lyapunov function, offering a normative principle for understanding the model's dynamics. Much of the reviewer’s concerns are focused on the language used in the Related Work section which we will rephrase as follows (adapting the reviewer’s first suggestion): (Song et al., 2016) incorporated Dale’s law into the vanilla RNN architecture, which was successfully trained across a variety of cognitive tasks. Building on this, (Soo et al., 2024) developed a technique for such RNNs to learn long-term dependencies by using skip connections through time. ORGaNICs is a model that is already built on biological principles, and can learn long-term dependencies intrinsically, therefore it does not require the method used by (Soo et al., 2024). Regarding the remaining concerns. The reviewer raises concerns about modeling the full range of V1 phenomena, but V1 is not the focus of this paper. There is no mention of V1 or the visual system anywhere (except when discussing the history of DN). Many of the phenomena mentioned by the reviewer have been or can be incorporated into ORGaNICs (see paragraph below). Application of the ORGaNICs theory to model V1 and several other neural systems will be done in future work. Regardless, we welcome the reviewer’s suggestion and we will add the following text: DN has been proposed as a canonical neural computation (Carandini & Heeger, 2012) and is linked to many well-documented physiological (Brouwer & Heeger, 2011; Cavanaughet et al., 2002) and psychophysical (Xing & Heeger, 2000; Petrov et al., 2005) phenomena. DN models diverse neural processes: adaptation (Wainwright et al., 2002; Westrick et al., 2016), attention (Reynolds & Heeger, 2009)), automatic gain control (Heeger, Simoncelli & Movshon, 1996), decorrelation and statistical whitening (Lyu, & Simoncelli, 2009). Since ORGaNICs’ response follows the DN equation at steady-state, it already incorporates this wide variety of neural phenomena. ORGaNICs have been shown to capture some of the dynamics of neural activity (Heeger & Zemlianova 2020; Rawat et. al., Cosyne 2024). Additional phenomena not explained by DN (Duong et. al., NeurIPS 2024) can in principle be integrated into the model. Regardless, more work needs to be done, of course, to explain the full range of neurophysiological phenomena. In this paper, however, we focus on the effects of DN on the dynamical stability of ORGaNICs. When we state that “we have no evidence (theoretical or experimental) of additional limitations” we are referring specifically to ORGaNICs as an instance of RNNs for ML (because that is the focus of this paper), not to limitations of the theory as a neurobiological circuit model (which we intend to publish separately). We have no evidence that there is something that ORGaNICs couldn’t do when compared to an LSTM. Finally, we assure that the paper has been written with the most genuine intentions, purely driven by curiosity. In fact, the core contributions are conceptual and technical (in the form of theorems), rather than ad-hoc benchmarks. We are grateful for the reviewer’s constructive comments which we believe will help clarify the key contributions of the paper.	Summary: This paper studies the stability properties of a model of cortical circuits which was introduced in 2019 (and I wasn't yet aware of): the ORGaNIC model by Heeger & Mackey. This LSTM-like model uses a simple set of differential equations that unify several phenomena observed in cortex, including normalization (Carandini & Heeger, 2012), and in some sense map LSTMs onto plausible cortical circuitry. This paper presents two main theoretical results: (i) that the model is locally stable around its unique fixed point when the recurrent matrix is the identity, and that (ii) existence and uniqueness of a stable fixed point can be shown for a general 2D model (really just one "principal neuron", and a gate variable) under certain conditions on the parameters. The paper wraps up with a few training experiments on (regular/sequential/permuted) MNIST classification, showing that ORGaNICs perform well. Strengths: This paper is very strong on a technical level. The maths are sound as far as I could tell from going through the paper carefully (mostly main text; took only a cursory look at the appendices) but not re-deriving the equations myself. The proof of local stability in ORGaNICs when $W_r = I$ is elegant and actually rather sophisticated; perhaps some of the most useful collaterals of this proof are (i) the explicit derivation of the eigenvalues of the Jacobian when additionally all normalization weights are equal to the same (positive) value, and (ii) the energy function of Eq 13 that provides further insight into the dynamics of ORGaNICs. Both of these add to the interpretability of this model class. The iterative fixed-point-finding algorithm (Algorithm 1) seems novel and useful, too. Weaknesses: Overall this is a highly esoteric paper for which I have difficulty assessing potential impact (hence my high uncertainty rating). The theoretical results are impressively detailed but seem fairly limited in scope (how are the W_r = I and 2D cases relevant for either ML or neuroscience?). I am also not super convinced by the utility of the broader conjecture on stability at the end of section 5. The experiments are fairly limited in scope and breadth, too (e.g. no mention of hyperparameter tuning for the SSN comparison). The abstract claims that it is “thanks to its intrinsic stability property” that ORGaNICs perform comparably to LSTMs; while this seems like a sensible hypothesis I don't think the paper really shows that, yet I believe the paper would be much stronger if that was shown to be true. In this respect, I wonder in what way the comparison to SSNs (e.g. Soo & Lengyel) helps make this case; is it because SSNs also perform a form of normalization at steady-state, much like ORGaNICs, but have otherwise no stability guarantees? Is it even true that SSN training is brittle / often fails because the network loses stability, thus giving rise to a stronger-than-usual tradeoff between training stability and learning rate magnitude? How could the author rule out that there might be another fundamental difference between these 2 models that has little to do with stability and yet underlies the performance difference? After all, no formal stability guarantees exist for standard LSTMs as far as I am aware, and that doesn't prevent them from training very well on most tasks (was gradient clipping even necessary for the LSTMs in those particular experiments?). In summary, I suppose the paper lacks a couple of convincing ablations to make the point that stability guarantees increase training performance. On clarity of exposition: this is a pretty hard paper to follow, it's very dense and doesn't really offer the hierarchy of exposition that a reader would want to see -- the theoretical results are presented in a very flat way, with a permanent back and forth between details and bigger picture that I found hard to navigate. Technical Quality: 3 Clarity: 2 Questions for Authors: - The abstract (and main text) says you trained your models using "simple backpropagation through time" but appendix I says you trained all models using Adam -- in principle there is no contradiction between these two claims, as backprop is also involved in Adam, but I wonder why you wrote "simple backpropagation" (which could easily be misunderstood as "simple gradient descent", i.e. not Adam). What's non-simple backpropagation? - The formatting of references is annoying :D please use parentheses to clearly separate references from the main text. - typo on l104: "dynamics-neural growth" → "dynamics-neutral growth" Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Some technical limitations are stated in the Discussion section. EDIT: following the rebuttal, I am raising my score to a 5. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: - “Overall this is a … uncertainty rating)”. Answer: We argue that the potential impact is high. LSTMs and GRUs have had a huge impact on ML/AI applications even though they are not always stable and hence require ad hoc techniques for training (e.g., gradient clipping/scaling). Furthermore, although normalization has been shown to improve the training and performance of CNNs, there has been no principled way of adding normalization to RNNs (putting a normalization layer between RNN layers is not the same as normalizing the activations within each RNN layer). ORGaNICs overcomes these limitations, with comparable capabilities to LSTMs and GRUs, with built-in recurrent normalization (motivated by neurobiology) that, as we show, is sufficient to guarantee dynamical stability. Moreover, we note also that normalization is an essential architectural motif in modern ML, and that divisive normalization has been shown to generalize batch and layer normalization, as shown in the work by Ren, M., et. al., (2016). As such, the ability to integrate normalization in the dynamics of an RNN, and to analytically demonstrate the impact of normalization through the derivation of an interpretable normative principle (Lyapunov function) is far from esoteric and very much needed in the theory of deep learning. Finally, our work establishes an important precedent for how the incorporation of neurobiological principles can drive advances in ML. - “The theoretical … end of section 5”. Answer: Considering the cases of $\mathbf{W}_r = \mathbf{I}$ and the two-dimensional model, we explore two different limits of arbitrary $\mathbf{W}_r$. The first limit involves relaxing constraints on all parameters in a high-dimensional model while assuming $\mathbf{W}_r = \mathbf{I}$. The second limit involves relaxing constraints on all parameters, including $w_r$, for a 2D model. These two limits provide a foundation for understanding the empirical results regarding stability in models with arbitrary recurrence, which is intractable with the approach used in the paper. There is also a rich history of studying neural mean field theories (Wilson, H. R., & Cowan, J. D. (1972); Kraynyukova, N., & Tchumatchenko, T. (2018)) which yield two-dimensional E-I circuits used for modeling the cortex. The conjecture is crucially important because we want to learn an arbitrary recurrent weight matrix during ML and neuroscience tasks. We have empirical evidence in support of this conjecture (see Fig.4,5 and Fig.1,2 of the attached pdf). We also use this conjecture in the paper itself (for the static input task). - “The experiments … SSN comparison)”. Answer: We thank the reviewer for the suggestions regarding additional tasks. To satisfy the reviewer, in a subsequent version of the paper, we will include ORGaNICs’ performance on the parity task for comparison. However, we stress that the primary focus of this paper is to demonstrate the stability, trainability, and interpretability of ORGaNICs compared to other RNNs. There was no hyperparameter tuning for any of the tasks – ORGaNICs simply outperforms SSNs on the first attempt at the task we considered. - “The abstract claims… to be true”. Answer: By that statement, we mean to convey that ORGaNICs mitigates the problem of exploding gradients: through its inherent stability; and vanishing gradients: by processing information across various timescales while maintaining stability, resulting in a blend of lossy and non-lossy neurons. This leads to a trainable RNN circuit, with a performance comparable to LSTMs without the need for specialized techniques like gradient clipping/scaling. We have revised this sentence for better clarity. - “In this respect, ... training performance”. Answer: Like SSN, ORGaNICs is a neurodynamical model originally designed to explain cortical activity. The comparison with SSN is simply meant to demonstrate that just like SSN, ORGaNICs can be successfully trained on a static MNIST task. So we are not trying to make the point that ORGaNICs performs better due to its stability guarantees, but rather that ORGaNICs can be trained by naive BPTT and perform comparably to SSN trained by the non-trivial method of dynamics-neutral growth (DNG). The result of the comparison is that ORGaNICs performs slightly better (despite no hyperparameter tuning) while being much easier to train than SSN. If we were to speculate on the performance gap between ORGaNICs and SSN, our best guess would be that DNG constrains the range of possible models (viz. parameters) thus reducing the expressivity of SSN. It is beyond the scope of this paper to identify why SSN trained by DNG does not do as well, but what is certain is that it is much harder to train SSN than ORGaNICs. The main takeaway is that ORGaNICs can be trained naively because it is stable, and it performs as well or better than alternative neurodynamical models (e.g., SSN) trained by sophisticated techniques. Regarding the LSTMs, they were trained using gradient clipping in Arjovsky, M., Shah, A., & Bengio, Y. (2016). - “On clarity of exposition: … navigate”. Answer: We thank the reviewer for the suggestion. We will revise the introduction of the paper to provide a clearer roadmap of the main results of the paper and how they are organized. Questions: - “The abstract … backpropagation?” Answer: We mean to convey the fact that we train ORGaNICs by naive backpropagation (not simple gradient descent), meaning that 1) we do not use gradient clipping/scaling that is typically adopted during BPTT when training RNNs such as LSTMs; 2) we do not need to resort to specialized techniques such as the sophisticated DNG method used to train SNN, an alternative neurodynamical model. We have resolved this ambiguity by revising the statement. - “The formatting … text”. Answer: We thank the reviewer for the suggestion, we have updated the references using the suggested format. --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal. I can see the potential impact of this work a bit more clearly now, and will raise my score to a 5, also reflecting a comparative rating taking into account the strength of the field across all my review assignments. Regarding clarity of exposition, I would encourage the authors to do more than just “revis[ing] the introduction of the paper to provide a clearer roadmap of the main results of the paper” -- my concern is that the main text itself is hard going, with the authors often going off on a tangent, and including a plethora of very technical details where the reader would instead like to be given a clear synthesis. Also: > There was no hyperparameter tuning for any of the tasks – ORGaNICs simply outperforms SSNs on the first attempt at the task we considered. Surely the comparison to SSN is meaningless unless you (at least) hyper-tune the SSN results; otherwise, how do we know if you weren't just lucky here? In general though, I think it's a good idea to hyper-optimize even one's own method, as this saves a lot of time downstream to people who would like to compare to your method. For example, if your answer to my concern above is that you, in fact, did not run your own SSN simulations but directly imported the (hyper-tuned) results from a previous SSN paper, then you are relying on SSN authors having done their job, but you are not doing yours. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for their feedback. Regarding the exposition, we will enhance the paper by providing clearer, more intuitive explanations of terms like “Lyapunov diagonally stable,” “Z-matrix,” “M-matrix,” and “indirect method of Lyapunov,” as suggested by reviewer ma5B. However, keeping the technical details is crucial. One of the core innovations of the paper is the technical approach that we developed to prove the stability of the recurrent circuit model (ORGaNICs), connecting the linear stability of a neural circuit model to the dynamic analysis of mechanical systems. As such, we are not merely going off on a tangent, but achieving a rigorous proof of stability for a neural circuit model by linking two disparate fields. This has important implications for the derivation of a Lyapunov function and thus a normative principle that makes ORGaNICs interpretable. We need to be technical enough to make this link precise and for the description of the approach to be correct and reproducible. We note that there are many mathematical papers published in NeurIPS that are richer in theorems and technical details/jargon than this work. We tried to strike a balance to reach a broader audience, so we welcome the reviewer’s comment and will clarify the writing further. Regarding the comparison to SSN, the results by Song 2022 were not hyper-tuned either. We used the same learning rate, optimizer configuration, and equal number of layers and number of units in each layer to make the comparison fair. We will state this clearly in the manuscript. Moreover, we will conduct hyperparameter tuning to facilitate future comparisons and ensure our results are robust. Nevertheless, we would like to underscore once again that the numerical comparison of ORGaNICs with SSN or LSTMs are not the main point of the paper. Specifically, the core contributions of the paper can be summarized at a lay level as follows: - We take an existing neurobiologically plausible RNN model designed to achieve divisive normalization (DN) exactly via recurrent excitation. This model has been shown to recapitulate a broad range of neurophysiological observations, not by design but as a result of imposing a specific circuit function (i.e., DN). Given the resemblance to LSTMs, we hypothesize that this model should also do well on ML tasks and ask what is the impact of imposing DN on the dynamics and trainability of an RNN. - We discover empirically that ORGaNICs is unconditionally stable under very mild constraints (this is highly nontrivial). We thus develop mathematical machinery to prove this unconditional stability in a couple of limiting cases (d=2 for all parameter values, and $\mathbf{W}_r = \mathbf{I}$ for generic d and all other parameter values) by expressing ORGaNICs as a mechanical (specifically, gyroscopically stabilized) system – to do this we even need to prove new results on mechanical systems. - But what are the implications of this unconditional stability, if any? Stability is crucial to ensure the trainability of RNNs that typically require gradient clipping/scaling. We discover that by virtue of its stability, ORGaNICs can be trained on a standard ML task (sMNIST and psMNIST) by naïve BPTT, without gradient clipping/scaling. In fact, naïve BPTT works so well that on a first attempt, it does about as well as LSTMs (trained with gradient clipping/scaling and hyper-tuned) – it is not luck, the result is robust with respect to re-initialization of the training. The point then is not simply that ORGaNICs is at least as good as LSTMs, but that imposing DN dynamically makes training effortless. Note that unlike standard RNNs, where normalization is imposed a-posteriori to the output layer, ORGaNICs implements normalization dynamically: this is the first time this idea is tested in ML and analyzed theoretically (as pointed out enthusiastically by reviewer ma5B). Moreover, ORGaNICs trained by naïve BPTT fares well when compared to SSN trained by a specialized technique. The reason we choose SSN is that this model is also neurobiologically inspired and it has also been shown to approximate normalization in certain parameter ranges. - Finally, the connection to mechanical systems allows us to derive a Lyapunov function that provides a normative principle for the dynamics of the models (i.e., a means of interpreting why and how the model tends to the stable fixed point). We will better summarize the paper’s key contributions in the paper’s introduction and discussion. We kindly ask that the reviewer assess our paper on its own merit, and not “reflecting a comparative rating taking into account the strength of the field across all my review assignments”. This request is in line with the NeurIPS FAQ that states “Q: Can I accept or reject all the papers in my stack?” “A: Please accept and reject papers based on their own merits. You do not have to match the conference acceptance rate.”	Summary: The paper discusses the development and analysis of "Oscillatory Recurrent Gated Neural Integrator Circuits" (ORGaNICs), a biologically plausible model of recurrent neural networks that implements divisive normalization. The authors prove the unconditional local stability of ORGaNICs with an identity recurrent weight matrix using the indirect method of Lyapunov function. They also demonstrate empirical stability for higher-dimensional circuits. The model's performance is evaluated on static and dynamic classification tasks, showing comparable results to LSTMs without specialized training strategies, thanks to its intrinsic stability properties. Strengths: 1. Biological Plausibility: ORGaNICs are designed with a structure and dynamics that are more aligned with biological neural circuits than traditional artificial neural networks. 2. Biological divisive normalization: Implemented divisive normalization using a more bioplausible network. 3. Trainability: ORGaNICs are a biophysically plausible extensions of LSTM and can be directly trained by BPTT. Weaknesses: 1. Generalization: The paper does not extensively discuss how well the results might generalize to more complex or different types of tasks beyond the tested benchmarks. 2. Performance on Benchmarks: The proposed model performance still has a gap compared to machine learning models. 3. Scalability: parameter size scales too fast with model size and there is no obvious way for impovement. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can this model provide more interpretability in relevant machine learning tasks? 2. Besides normalization and stability, can this E-I model offer more insights into neuroscience compared to machine learning models? 3. Many existing machine learning models have already implemented various forms of divisive like normalization such as layer normalization. What are advantages or innovations of the divisive normalization implemented in this model. 4. In comparison to standard machine learning models, biophysically plausible models typically offer superior interpretability. However, the paper offers limited discussion on this topic. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: 1. In the sequence modeling benchmarks, only 2 tasks are evaluated. It is recommended to include a broader range of sequence modeling tasks to allow for more comprehensive comparisons, such as those where SSM excels in long sequence modeling or where LSTM performs well in formal language tasks like the parity task. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: - “Generalization: The paper ... the tested benchmarks”. Answer: We expect ORGaNICs to perform well on other ML benchmarks, especially those concerning sequential data. This will be done in a future study, therefore we refrain from making any specific claims about performance on other benchmarks. In this paper, we sought to demonstrate the stability, trainability, and interpretability of ORGaNICs compared to other RNNs, and present proofs and evidence supporting these claims. Specifically, we were able to rigorously prove for the two limiting cases that ORGaNICs is absolutely stable and demonstrated empirically that this stability holds broadly. We achieved good performance on ML tasks using naive BPTT without resorting to gradient clipping/scaling or other techniques required for alternative neurobiological models like SSN, and without hyperparameter tuning. We derived a Lyapunov function for our model offering an interpretable optimization principle explaining the dynamics of ORGaNICs. - “Performance on Benchmarks: … machine learning models. Answer: There was no hyperparameter tuning nor an attempt to get SOTA performance in this paper. The top-performing ML models are designed with properties of low-dissipation and stability to solve the problem of vanishing and exploding gradients. In contrast, ORGaNICs is a biologically plausible circuit designed to implement divisive normalization (DN), a computational principle experimentally found in a wide range of cortical areas (i.e., different brain regions) across different species, to simulate cortical activity. Unlike ML models, stability in ORGaNICs is not engineered but it emerges from its neurobiological plausible design. Despite not being optimized for the tasks considered, ORGaNICs performs competitively with other RNN models. To our knowledge, no other neurodynamical biologically plausible model achieves this. - “Scalability: parameter size scales … way for impovement”. Answer: In future work, we will design the gains $\mathbf{b}$ and $\mathbf{b_0}$ to be dependent locally on $\mathbf{y}$ and $\mathbf{a}$, This will significantly reduce the number of parameters compared to the current model where the matrices $\mathbf{W_{by}}$, $ \mathbf{W_{ba}}$, $ \mathbf{W_{b_0 y}}$ and $ \mathbf{W_{b_0 a}}$ are $n \times n$. Questions: - “Can this model provide … machine learning models?” Answer: We draw a direct connection between ORGaNICs and systems of coupled damped harmonic oscillators, which have been studied in mechanics and control theory for decades. This connection allows us to derive an interpretable energy function for a high-dimensional ORGaNICs circuit (Eq. 128), providing a normative principle of what the circuit aims to accomplish (see Eq. 13 and the subsequent paragraph). For a relevant ML task, having analytical expressions for the energy function (which is minimized by the dynamics of ORGaNICs) allows us to quantify the relative contributions of the individual neurons in the trained model, offering more interpretability than other RNN architectures. For instance, Eq. 128 reveals that the ratio $\tau_y / \tau_a$ of a neuron pair ('y' and its corresponding 'a') indicates the "weight" a neuron assigns to normalization relative to aligning its responses to the input. This provides a clear functional role for neurons in the trained model. Furthermore, since ORGaNICs is biologically plausible, we understand how the different terms of the dynamical system may be computed biologically within a neural circuit (Heeger & Mackey, 2019). This bridges the gap between theoretical models and biological implementation, providing a framework to test hypotheses about neural computation in real biological systems. - “Many existing … implemented in this model”. Answer: Divisive normalization (DN) was introduced as a model of the steady-state response of neurons, functioning as a static nonlinearity similar to batch and layer normalization. The DN equation has also been shown to generalize batch and layer normalization, as detailed in the work by Ren, M., Liao, R., Urtasun, R., Sinz, F. H., & Zemel, R. S. (2016). However, in the brain, there are no static nonlinearities; it has thus been proposed that DN is achieved via a recurrent circuit. ORGaNICs is such a recurrent circuit designed so that the responses of neurons at steady state follow the DN equation. While batch and layer normalization are ad hoc implementations that do not affect the dynamics (they are applied to the output layer), ORGaNICs implements DN dynamically. Additionally, whereas batch and layer normalization do not inherently affect the stability of the model (because they do not influence the dynamics), our paper demonstrates that a model implementing DN naturally exhibits stability, which is greatly advantageous for trainability. This stability, derived from the dynamic implementation of DN, sets ORGaNICs apart by providing both output normalization and model robustness. - “In comparison… limited discussion on this topic”. Answer: We thank the reviewer for this suggestion. We will add a paragraph based on our answer to the first question. Limitations: - “In the sequence modeling … the parity task”. Answer: We thank the reviewer for the suggestions regarding additional tasks. To satisfy the reviewer, in a subsequent version of the paper, we will include ORGaNICs’ performance on the parity task. However, we stress that the primary focus of this paper is to demonstrate the stability, trainability, and interpretability of ORGaNICs compared to other RNNs, which we have achieved through a thorough analytical and numerical analysis, and not to provide an extensive benchmark of the model on ML tasks. A more comprehensive test of ORGaNICs on additional long sequence modeling tasks, along with hyperparameter optimization, will be pursued in future work.	Rebuttal 1: Rebuttal: We thank all the reviewers for dedicating their time and providing valuable feedback on our submission. Your insights and comments are greatly appreciated and will help improve the quality of our work. In response to the reviewers' comments, we have conducted additional experiments and analyses. Please find the following new results in the attached PDF: - $\textbf{Table 1:}$ Based on the suggestion from Reviewer 1, we demonstrate that ORGaNICs can be trained with fixed time constants on both sequential and permuted MNIST datasets without requiring gradient clipping or scaling. - $\textbf{Figure 1:}$ In response to Reviewer 1's inquiry about ORGaNICs' stability when the recurrent weight matrix defined as $\mathbf{W}_r=\alpha \mathbf{I}$ for $\alpha \in [0,1]$, we simulated 10,000 networks with random weights, inputs and values of $\alpha$, and with random initial values of $\mathbf{y}$ and $\mathbf{a}$. We found that all simulations converge to the fixed point identified by the iterative algorithm. These fixed points are stable, as the largest real part of the eigenvalues, for all the networks, is strictly negative. - $\textbf{Figure 2:}$ Based on the insight from Reviewer 5 that ORGaNICs, when trained on static input classification task (Section 6.1), acts as a deep equilibrium model (DEQ) (Bai, S., Kolter, J. Z., & Koltun, V. (2019)). DEQs are known to be prone to instabilities during training. We demonstrate that since ORGaNICs is intrinsically stable for all parameters and input in the conditions assumed in Section 6.1, it stays stable throughout the training. Specifically, the fixed points remain attractive as all of the eigenvalues, across all the test samples, have negative real parts throughout the training. Pdf: /pdf/90934c4cc4d9eb3af642b8b05659f33f4509d348.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: Recurrent neural networks are widely applied in both solving machine learning tasks and modeling neural recordings. However, they can be difficult to train due to exploding/vanishing gradients and other sources of instability, often requiring hyperparameter optimization. In this paper, the authors argue that the ORGaNICs model is capable of solving sequential tasks without the need for hyperprameter optimization due to the conjectured property of ``unconditional stability,'' which guarantees the existence of a stable fixed point in the dynamical system for any model parameters. The biological plausibility of the ORGaNICs dynamics also suggests that a similar dynamical system could guarantee stability in biological neural circuits for processing sensory information and generating motor sequences. The stability property is proven rigorously in the case where the recurrent weights are equal to the identity and in the case where only a single recurrent and inhibitory unit are included in the model. The authors provide new insights into the behavior of ORGaNICs by providing a Lyapunov function for the dynamics near the fixed point in the case where $\mathbf{W}_r = \mathbf{I}$. An algorithm for quickly finding a stable fixed point of the ORGaNICs dynamical system in general is provided. This algorithm is used to probe empirically for the fixed points of the ORGaNICs dynamics with general recurrent weights $\mathbf{W}_r$. The authors test their prediction empirically in numerical experiments where they train an ORGaNICs network on two tasks: Static and sequential MNIST classification. In both cases, they show that ORGaNICs obtains competitive performance without hyperparameter optimization. Strengths: This paper includes many interesting insights and clever mathematical arguments. The proof of unconditional stability through the indirect method of Lyapunov and by relating the problem to a second-order dynamical system is mathematically sound and provides interesting insights through the analogy to a damped harmonic oscillator. The finding that ORGaNICs achieves competitive performance without the need for hyperparameter optimization on the static and sequential MNIST tasks is also very promising. Weaknesses: The main theorem makes the very restrictive assumption that the recurrent weights $\mathbf{W}_r = \mathbf{I}$, such that they exactly cancel the intrinsic decay rate of the recurrent neurons. While there are neural circuits where all recurrent interactions take place through inhibitory interneurons (e.g. cerebellar granule cells interacting through inhibitory Golgi cells), the intrinsic decay must also be taken into account. If the assumption can be relaxed to say that $\mathbf{W}_r = \alpha \mathbf{I}$ for some $\alpha \in [0, 1]$ then I would be inclined to raise my score. Also, the argument for unconditional stability of the ORGaNICs model is not well-supported in the paper's current form. Beyond the case $\mathbf{W}_r = \mathbf{I}$ and the two-dimensional case, there this is tested only for the MNIST task with constant inputs. In these tests, it also appears that the fixed-point-finding algorithm is used to quickly discover the stable fixed point. However, it would be stronger to show that the dynamics naturally converge to the fixed point in a reasonable amount of time when started from randomly generated initial configuration. In fact, it is possible that the dynamics naturally lead the system to a stable limit cycle and never converge to the stable fixed point, even if it is guaranteed to exist. We ask the authors to clarify the methods used to locate the fixed point of the ORGaNICs network in their experiments and show that the system's dynamics tend toward this fixed point and not another limit cycle in a reasonable amount of time. The numerical training also enforces a maximum singular value of $\mathbf{W}_r$, which is a strong condition required for stability. ``Towards Unconditional Stability...'' would be a more appropriate title given the strong constraints required for stability guarantees in both the analytical proofs and numerical tests. For the sequential MNIST task, the authors claim that they are able to achieve LSTM-like performance without hyperparameter optimization or gradient clipping/scaling. During these experiments, the neural time constants are treated as parameters that can be learned. This result is different than showing that for any fixed values of the time constants, ORGaNICs can be trained without issue. The authors should verify that fixing the time constants and then training also leads stable training and no need for gradient clipping. This is important to maintain a claim of biological plausibility where time constants are intrinsic properties of the neurons and not trained. The static and sequential MNIST tasks also represent a rather limited scope in which to test for unconditional stability in practice. It's possible (though I agree, unlikely) that for other more unwieldy tasks unconditional stability will not hold. Either a proof of unconditional stability for general recurrent weight matrices or a wider range of numerical tests are necessary to make the case for unconditional stability of the ORGaNICs dynamics. Technical Quality: 3 Clarity: 3 Questions for Authors: What hyperparameters are present in the ORGaNICs model as trained on the static and sequential MNIST tasks? While the time constants are learned, how is the learning rate for naive BPTT selected? Would this hyperparameter plausibly need to be optimized to achieve performance competitive with a well-tuned LSTM? In the verification of the iterative fixed-point finding algorithm, how is the ``true'' fixed point determined? By running the dynamics from a random starting point to the fixed point, or by running the dynamics from an initial estimate obtained using the iterative fixed point finding algorithm? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Even for static inputs, there are scenarios in which a biological circuit may want to produce a non-static output. For example, in central pattern generators of motor sequences. If indeed ORGaNICs always converges to a stable fixed point, this may represent too constrained of a circuit to explain many interesting neural phenomena. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: - “The main … raise my score”. Answer: Our proof for the stability of the high-dimensional model (using the indirect method of Lyapunov) relies on the existence of analytical expressions for the fixed point. Such an expression exists when $\mathbf{W}_r=\mathbf{I}$, but does not exist when $\mathbf{W}_r \neq \mathbf{I}$. However, we were able to prove unconditional stability for the 2-dimensional model (see Fig.1, Theorem.5.1 and Theorem.5.2), for any positive value of $\alpha (w_r) \in (0,\infty)$ and any combination of choices for other parameters. Building on these two limiting cases, for which a rigorous proof of stability is tractable (i.e., for any $w_r$ in mean field, when d=2, and for $\mathbf{W}_r=\mathbf{I}$ when d>2, with no restrictions on other parameters), we conjecture at the end of Section 5 that the stability property holds in higher dimensions for systems where the largest singular value of $\mathbf{W}_r$ is less than 1 (therefore covering the case when $\alpha \in [0, 1]$). We provide empirical evidence in support of this conjecture in Fig. 4 and Fig. 5. To satisfy the reviewer’s question directly, we simulated 10,000 random networks with $\mathbf{W}_r=\alpha \mathbf{I}$ with $\alpha \in [0,1]$ and plot the distribution of the maximum real part of eigenvalues, all of which are found to be less than 0 (Fig.1 of pdf), indicating stability. We do not exclude that a statistical approach could be used to prove the stability of a general network “in expectation”, but 1) it would fall short of a rigorous proof of stability like ours; 2) it is sufficiently ambitious that it should be the subject of a subsequent paper. - “Also, the … amount of time”. Answer: We would like to point out to the reviewer an additional experiment at the end of Section 5 where we mention empirical evidence of stability for a more general case of the recurrent matrix. where the maximum singular value of $\mathbf{W}_r$ is constrained to 1. This conjecture is supported by empirical evidence showing consistent stability, as ORGaNICs initialized with random parameters and inputs under these constraints have exhibited stability in 100\% of trials Fig.4.” In Fig. 4 the “true” fixed point was found by simulating the network from a zero initial condition. In all of the trials where the conditions in the conjecture were met, the dynamics always converged to the stable fixed point found by the iterative scheme. To show that in all simulations there is no limit cycle and that we can start from a random initial condition, we re-performed this analysis (10,000 networks), starting from random initial conditions \in [0, 1] and fixed $\mathbf{W}_r=\alpha \mathbf{I}$ for some $\alpha \in [0,1]$. We find that the system is stable (Fig.1 of pdf) and the simulations always converge to the same fixed point as found by the iterative scheme. So, while we cannot prove the conjecture by the indirect method of Lyapunov, numerical evidence suggests that the conjecture is true almost surely. - “For the … and not trained”. Answer: This is an excellent point. We chose the time constants to be learnable to demonstrate that ORGaNICs is competitive with LSTMs on sequential MNIST tasks without hyperparameter optimization and without using ad hoc techniques for training, such as gradient clipping/scaling. As per the referee's suggestion, we have now conducted additional experiments where ORGaNICs is trained with fixed time constants (Table 1 of pdf). Our findings indicate that even with fixed time constants, ORGaNICs maintain stable training without the need for gradient clipping/scaling. - “The static and … ORGaNICs dynamics”. Answer: We have provided empirical evidence for unconditional stability for a general recurrent weight matrix, particularly when the maximum singular value of $\mathbf{W}_r$ is less than 1 (Section 5, Fig. 4, and Fig. 5). We prove unconditional stability, using a non-trivial approach, in two specific cases of $\mathbf{W}_r$ : first, when $\mathbf{W}_r = \mathbf{I}$, and second, for the 2D model. While a proof for unconditional stability for a general $\mathbf{W}_r$ would be most desirable, it is currently beyond reach as detailed in our response above. Therefore, we must rely on empirical evidence for the stability of a general $\mathbf{W}_r$. In a subsequent version of the paper, we will provide results for the formal language parity task to further corroborate our point. Questions: - “What hyperparameters … tasks?” Answer: Learning rate = 0.001 (static MNIST), 0.01 (sMNIST); Batch size = 256, Weight decay = $10^{-5}$ for both. This information has been added in the Appendix. - “While the … LSTM?” Answer: We do not have a specific scheme for selecting an optimal learning rate. Since the point of this paper is to showcase greater stability, trainability, and interpretability compared to other RNNs, and not to achieve SOTA on the benchmarks, we did not tune the hyperparameters. This will be explored in future work and should lead to better performance on ML benchmarks. - “In the verification … algorithm?” Answer: The “true” fixed point is found by simulating the network from a zero initial condition. However, in response to the reviewer, we have verified that for a random initialization as well, the simulation converges to the same fixed point. Limitations: - “Even for … phenomena”. Answer: This class of models has already been shown to produce non-static oscillatory activity as well as to reproduce a wide range of neural phenomena, including sustained activity, sequential activity, motor preparation and motor control (Heeger, D. J., & Mackey, W. E. (2019)); simulate the dynamics of V1 activity (Heeger, D. J., & Zemlianova, K. O. (2020), S. Rawat, D.J. Heeger, and S. Martiniani. Cosyne Abstracts 2024); predict the emergence of communication subspaces in interareal communication (S. Rawat, D.J. Heeger, and S. Martiniani. Cosyne Abstracts 2023). --- Rebuttal Comment 1.1: Title: Raising Score Comment: The authors have addressed most of my concerns with added experiments. I believe that this is a significant contribution even without a proof of unconditional stability in the most general case. I've raised my score to a 6. --- Reply to Comment 1.1.1: Title: Comment Comment: We thank the reviewer for their feedback and for increasing the score.	null	null	null	null	null	null
Reinforcement Learning Gradients as Vitamin for Online Finetuning Decision Transformers	Accept (spotlight)	Summary: The authors consider the problem of online fine-tuning decision transformers. Current approaches for this problem do not work well when the offline data is low-quality. The authors analyze the online DT algorithm theoretically and show why DT-induced policy update would not work if the RTG used for conditioning is much higher than RTG seen in the dataset. They then propose combining the ODT update with RL gradients, in particular, the ones obtained from the TD3 model. This approach was then tested through Adroit, Antmaze, and Mujoco experiments, where the method obtained strong results. Finally, the authors perform additional analyses and make improvements to their proposed approach. Strengths: - The work considers the important problem of finetuning decision transformers. I see this as a significant issue, as we need strong online fine-tuning methods to better leverage pre-trained models. - The proposed method seems to achieve quite good performance in the presented experiments, at least when not taking the statistical significance of the results into account. In my opinion, the authors also consider a sufficient set of benchmarks and baselines. - The study is backed by theoretical results. - Related work is well-described to the best of my knowledge. - The appendix includes many additional results and extended analysis that might be of interest to the reader. Weaknesses: - The empirical results seem to be noisy, and a more careful approach to statistical significance is needed. Looking at, e.g., Figure 5, I'm not sure if TD3+ODT is better than TD3 as the plots overlap quite a lot, but the corresponding Table 1 makes it look like TD3+ODT is a clear winner. - This includes bolding "the best solution" in some of the tables. If one of the solutions achieves a score of 101 and the other of 100, in the RL regime, it is usually not possible to tell which one is statistically better and the fair thing to do is to bold both. - I didn't find the information on how many seeds are used for each experiment, and this is very important in the RL context. - In summary, I would suggest the authors adopt the recommended practices of [1] to have a clearer picture of the results. - I think the baselines considered here should be made stronger - The TD3 baseline collapses right after fine-tuning starts in many cases. It looks like it might be caused by forgetting which was shown to be a problem in RL fine-tuning [2]. I wonder if combining TD3 with some forgetting mitigation technique would help. - The authors point out that their method improves on TD3+BC in two ways: architecture (MLP -> DT) and objective (BC -> RvS). I think these two axes of improvement should be investigated separately. That is, introducing two additional baselines would show us where the improvements come from: TD3+BC+Transformer and TD3+RvS (with MLP rather than DT). [1] Agarwal, Rishabh, et al. "Deep reinforcement learning at the edge of the statistical precipice." Advances in neural information processing systems 34 (2021): 29304-29320. \ [2] Wolczyk, Maciej, et al. "Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem." Forty-first International Conference on Machine Learning. Technical Quality: 2 Clarity: 3 Questions for Authors: Currently, I would be inclined to accept the paper if not for the problems with the evaluation. As such, I would like to ask the authors: - Please discuss and address the questions about the evaluation setting, as raised in the Weaknesses section above. If this is sufficiently resolved, I'd be happy to increase the score. - Please consider introducing stronger baselines. This is not as crucial, but it would make the paper stronger. Less important but interesting to discuss: - The authors say: "training a transformer-based value function estimator is quite hard." I agree with this opinion based on my own experiences, but I don't think there has been a thorough discussion of this in the literature. The papers referenced by the authors do not really address this issue well in my opinion. So, it's not a weakness of the paper, but I would like to ask the authors where they think the problems of value learning in transformers come from (or maybe even their online training in general). Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Although the authors do discuss the limitations of their work, I think there are at least two more points that should be discussed: * The method has not been tested on image-based datasets and environments. * There is an increased computational cost as compared to a regular ODT, since we have to train the value functions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for appreciating our work. Here are our responses: Q1. Results are noisy, and a more careful approach to statistical significance is needed. Thanks for pointing this out! We understand the importance of reliable evaluation and statistical significance. For this we reported standard deviations in our learning curves in all main results (Fig. 2,3,5) and most ablations (Fig. 6-7, 9, 16-17, 19). We provided Tab. 1 so that readers can focus on the final performance and its gain over pretraining. To make our reported results more reliable, according to the reviewer’s advice, we evaluated all main results (MuJoCo, adroit, maze, antmaze) using the rliable library, as suggested by the paper mentioned in the review. Results are provided in Fig. 3 of the global pdf, which shows: our method indeed outperforms all baselines on adroit, MuJoCo, and umaze/medium antmaze. Also, we will update Tab. 1-3, 10-11 by highlighting scores >=90% of the best reward with bold-face following MAHALO [1]. Q2. Number of random seeds. Thanks for pointing this out. We use 3 seeds for all experiments (a few ablations without standard deviation use only 1 seed). Note, even without ablations this amounts to a large number of experiments: 1) we use 46 different datasets for our main results [4 for maze, 24 for MuJoCo ({random, medium-replay, medium}x{hopper, halfcheetah, walker2d, ant}x{normal, delayed reward}), 12 for adroit ({pen, hammer, door, relocate}x{expert, cloned, human}), and 6 for antmaze]; and 2) we test 6 methods. This amounts to 800+ runs (46 datasets, 6 methods, 3 seeds), each of which requires 6-8h on average (IQL and TD3+BC need less time, but PDT and other methods on complex environments such as adroit require more). Q3. Try adding forgetting mitigation methods onto TD3 baseline. Great suggestion! The paper mentioned by the reviewer is pretty recent. It discusses 4 types of forgetting mitigation methods: Behavioral Cloning (BC), Kick Starting (KS), Episodic Memory (EM), and Elastic Weight Consolidation (EWC). Our TD3 baseline and the MLP TD3+BC baseline already use EM as a mitigation method (quote, “simply keeping the examples from the pre-trained task in the replay buffer when training on the new task”). Note, since TD3 is deterministic without log-likelihood, adapting EWC is out of scope because it requires the Fisher information matrix, which in turn needs the gradient of the log-likelihood. This is also true for BC and KS, which are auxiliary losses of KL-divergences between policies on states drawn from different samplers. However, since BC/KS are essentially soft constraints that prevent the current policy from being too different from the pretrained policy, we can implement BC/KS on continuous action space for deterministic policies by substituting KL divergence with Euclidean distance between actions predicted by the current and pretrained policies. Here, we add a $0.05\cdot\\|\\|a_{\text{new}}-a_{\text{old}}\\|\\|_2^2$ penalty term to the actor loss, which is a gradual shift from BC (sampled from offline dataset) to KS (sampled from online rollouts). To see whether other forgetting mitigation methods work, we also implemented jump-start RL [2]. Concretely, the pretrained policy is used for the first $n$ steps in an episode, before ODT takes over. We set $n=100$ (100% max episode length for pen, 50% max episode length for other adroit environments) initially, and apply an exponential decay rate of $0.99$ for every episode. We evaluate both methods on adroit environments with the expert dataset (where the forgetting issue is serious) and the cloned dataset (where the agent needs to improve from low-return policies). Results are summarized in Fig. 1. We find that 1) jump-start RL struggles (possibly because it does not directly prevent out-of-distribution updates), and 2) while BC/KS effectively improves both TD3 and TD3+ODT on expert datasets, it also hinders online policy improvement on most cloned datasets. Q4. Try other baselines such as TD3+BC+transformer and TD3+RvS. We would like to point out that our work is focused on analyzing and improving ODT. Thus the main baseline for this work is ODT instead of TD3+BC. That being said, we agree that TD3+BC+transformer and TD3+RvS are interesting baselines to test. The results are shown in Fig. 2 in the global pdf. We test the methods on adroit environments with the cloned dataset. The results show that TD3+RvS slightly improves upon TD3+BC, while TD3+BC+transformer improves more. We speculate that an MLP is not expressive enough to model the policy change over different RTGs. Note, both suggestions still fall short of our method. Q5. Where does the problem of value learning come from? Good question! Generally, we found that there is a tradeoff between more information and training stability: while longer context length allows the decision transformer to utilize more information, it also causes increased input complexity, i.e., noise from the environment leads to reduced stability and slower convergence in training. Note, this differs from LLM works, where context length improves generalization thanks to extremely high expressivity, huge amounts of data, and less noisy evaluation. This is especially the case with bootstrapping involved, as the fitting target depends on its own output. In fact, Parisotto et al. [4] suggest that if a recurrent agent can first learn a Markovian policy and value function, then the following training could be stabilized. Q6. Some limitations are left out, e.g., lack of image-based testbed and increased computational cost. Thanks for pointing these out! Those are important questions, and we will discuss those in a revised limitation section. Regarding computational cost, we discussed the impact of the RL gradient on training in Appendix H: we found that our method is only marginally (20%) slower than ODT. For the MuJoCo experiment, ODT needs ~6 hours to train on our machine. --- Rebuttal Comment 1.1: Comment: I appreciate the thorough response from the authors. After reading it as well as the other reviews, I decided to increase my score. Here's my detailed response: * I appreciate that the authors ran the rliable library to get more stable results. As for the results in Fig. 3 -- are these averaged over all the settings (e.g., cloned and expert in Adroit case)? It would be nice to also see this analysis on each setting. * I understand that having more seeds is expensive, but I think we need them to be sure about the results. Given some overlaps in Figure 3 presented by the authors I would recommend adding at least 2 more seeds (5 in total) * I really appreciate the forgetting mitigation experiments. It makes sense that BC works well on expert datasets but not on cloned ones. I wonder if one could apply some sort of selective forgetting mitigation on the cloned datasets to remember what is useful and forget all of the behaviors that are not useful. * It's quite interesting that the transformer architecture seems to be much more important than the conditioning (i.e., BC vs RvS). I wonder if the transformer is crucial or just having bigger MLPs or RNNs/SSMs would suffice. --- Reply to Comment 1.1.1: Comment: Thanks a lot for your timely reply and for appreciating our rebuttal. Below are our follow-up responses: Q1. About the results in Fig. 3. Yes, those results are averaged over all the settings due to the 1-page limit of the response pdf. For example, for adroit, we average over 12 different settings ({expert, cloned, human}x{pen, hammer, relocate, door}). We will update figures on each setting in the revised paper. To begin, here, we provide some numbers for IQM (which is more robust than median as suggested by the paper mentioned in the review) and optimality gap on the adroit pen environment, as the mean and standard deviation are already provided in the main paper. The numbers in parentheses are the 95% confidence interval. IQM Method \| pen-expert-v1 \| pen-cloned-v1 \| pen-human-v1 ---\|---\|---\|---\| TD3+BC \| 40.15 (24.00~48.17) \| -2.39 (-2.82~-1.55) \| 18.63 (8.49~34.17) IQL \| 149.44 (146.54~151.13) \| 76.06 (72.17~78.46) \| 98.13 (91.18~101.85) ODT \| 129.31 (125.10~133.52) \| 26.24 (19.74~32.73) \| 32.28 (25.23~39.33) PDT \| 27.91 (14.50~41.32) \| 14.05 (5.14~30.09) \| 2.67 (1.75~3.58) TD3 \| 69.48 (61.52~77.44) \| 70.52 (67.25~73.79) \| 32.28 (19.25~45.31) TD3+ODT (ours) \| 118.30 (102.80~135.14) \| 130.51 (124.43~135.65) \| 116.71 (110.05~120.61) Optimality Gap Method \| pen-expert-v1 \| pen-cloned-v1 \| pen-human-v1 ---\|---\|---\|---\| TD3+BC \| 59.85 (48.17~76.00) \| 102.39 (101.55~102.83) \| 81.36 (65.83~91.51) IQL \| 0.00 (0.00~0.00) \| 23.94 (21.54~27.83) \| 2.93 (0.00~8.81) ODT \| 0.00 (0.00~0.00) \| 73.75 (67.26~80.25) \| 67.61 (60.66~74.76) PDT \| 72.08 (58.68~85.50) \| 85.95 (69.90~94.85) \| 97.33 (96.42~98.24) TD3 \| 30.51 (22.56~38.48) \| 29.48 (26.21~32.75) \| 67.72 (54.69~80.74) TD3+ODT (ours) \| 0.00 (0.00~0.00) \| 0.00 (0.00~0.00) \| 0.00 (0.00~0.00) Q2. More seeds are needed for experiments. We completely agree, thorough evaluation is important. We started experiments with two additional seeds, and we will include those in the revised paper. Q3. Selective forgetting mitigation on cloned datasets. 1. “Forgetting all the behaviors that are not useful” might be suboptimal. While expert trajectories are certainly more useful, it is hard to say which behaviors are not useful at all in the offline dataset. For example, all trajectories may provide information on environment dynamics and possibly undesirable states. We speculate that it would be better if such data could be memorized in some way not reflected in the evaluation policy, e.g., dynamic models in model-based RL, areas of state-action space with low Q-values, or encoded in the policy but “hidden” by conditioning on high RTG. Our method is a combination of the latter two. 2. We think the ODT gradient in TD3+ODT during the online phase could be interpreted as a kind of selective forgetting mitigation. Since TD3+ODT (as well as our TD3 / original ODT) keeps the offline data in the replay buffer for ODT itself during the online phase, the supervised learning gradient on those offline data can be regarded as a regularization towards the RTG-conditioned empirical policy, which, ideally, should be equal to the pretrained policy if the algorithm converges during the offline phase. 3. Given the performance increase due to the KL regularizer used with TD3+ODT on adroit expert datasets, we agree that an RTG-weighted regularizer could be interesting to test and could potentially further improve results. We have started some experiments, and will add results on this in the revised paper. Q4. Is the transformer crucial, or bigger MLPs would suffice? Good question! While ODT is the main baseline of our paper as we mentioned before, we feel ablations on using bigger MLPs / RNNs for TD3+RvS are interesting. We started some experiments for our revised paper.	Summary: This paper presents a method for fine-tuning decision transformers (DT) online. The proposal is to integrate TD3 with DT such that the gradients from TD3 can help the online policy to explore highly rewarding trajectories, hence further improving the agent performance. It can be seen as a variant of TD3+BC with the MLP being replaced by a transformer. The experimental results show the proposed method outperforms several baselines, and is effective even when the offline dataset is of low quality, a setting which traditional online DT cannot do well. Strengths: - The topic of improving DT for online fine-tuning is worth exploring. The way this paper approaches the topic is reasonable. - The empirical results are good. Weaknesses: - The TD3 gradients for policy update are not clear. For example, in equation 2, it seems that $Q$ is a constant if you optimize $\mu^{RL}$? Or is $a_t$ predicted by $\mu^{RL}$? - Some notations are a bit confusing. For example, there are RTG, RTG_real, RTG_eval. Better to clearly define them first. Technical Quality: 2 Clarity: 3 Questions for Authors: - Could you please explain more about the point (2) in line 182: how does DT prioritize trajectories with higher return? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: - Compared to TD3+BC, the proposed method uses transformer instead of MLP. It costs much more computation to gain better performance. Such limitation is not mentioned in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for appreciating our work. Below are responses to questions: Q1. TD3 gradient for policy update is not clear. Thanks for pointing this out! There is a typo in Eq. (2) which should read as follows: $$\min\_{\mu^{\text{DT}}}\mathbb{E}\_{\tau\sim D}\Bigg[\frac{1}{T\_{\text{train}}}\sum\_{t=1}^{T\_{\text{train}}}\left[-\alpha Q\_{\phi\_1}(s\_t,\mu^{\text{DT}}(s\_{0:t},a\_{0:t-1},\text{RTG}\_{0:t},\text{RTG}=\text{RTG}\_{\text{real}})) + \\|\\|\mu^{\text{DT}}(s\_{0:t},a\_{0:t-1},\text{RTG}\_{0:t},\text{RTG}=\text{RTG}\_{\text{real}})-a\_t\\|\\|\_2^2\right]\Bigg],$$ i.e., the $a_t$ in Q should be the action generated by the decision transformer. Q2. Some notations are confusing. We introduced RTG in line 82-83 (“RTG, the target total return”), RTG_eval in line 88 (“a desired return RTG_eval is specified”) and line 96, and RTG_real in line 100 (the real return-to-go). We understand that they are not introduced concurrently, which makes it a bit confusing. We will modify this in the revised version. Q3. How does DT prioritize trajectories with higher return as mentioned in point (2) in line 182? DT prioritizes trajectories because the actions are learned to be generated conditioned on the Return-To-Go (RTG), and we use high RTG as the condition in evaluation. Consider as an example training with two sets of trajectories extending from the same state: one set with ground truth RTG being 0 and the other set with ground truth RTG being 1. If inference asks DT to generate trajectories conditioned on RTG=1, then the generated trajectory will be more similar to the latter set of training trajectories. Q4. The limitation that the use of transformers requires more computational cost is not mentioned. Thanks for pointing this out. We are happy to include a discussion in the limitation section in the revised version. Meanwhile, we would like to mention two points: 1. Our work aims to analyze and improve ODT. We thus compare computational cost primarily to ODT instead of MLP-based solutions; 2. In Appendix H we stated that our proposed solution is only marginally (20%) slower than ODT: while the use of RL gradients slows down training, the actor training overhead is negligible (since it only contains an MLP critic inference to get the Q-value), and the critic only takes about 20% time to train. We found that on our machine, for the MuJoCo experiment, ODT requires about 6 hours to train. --- Rebuttal Comment 1.1: Comment: Thanks for your rebuttal. Given the idea is interesting and the experimental results are thorough, I will adjust my score to 6. --- Reply to Comment 1.1.1: Comment: Thanks a lot for appreciating our work and rebuttal! We will revise our paper as discussed in the reviews for the next version.	Summary: The authors introduce a novel framework for improving the performance of Online Decision Transformer through adding TD3 gradients to the fine-tuning ODT objective. This is motivated by a theoretical analysis of ODT, that highlights an issue with low-reward, sub-optimal pretraining. The authors also provide an extensive empirical investigation. Strengths: * Paper is clearly written, and well motivated. The detailed explanation on the preliminaries was useful to the reader. * The contribution of TD3+ODT appears novel, original, and of high significance to the community. Weaknesses: * Overclaim of the contribution of “We propose a simple yet effective method to boost the performance of online finetuning of decision transformers.” The empirical results of relocate-expert-v1 show the opposite, where ODT outperforms TD3+ODT. Perhaps the authors can discuss this result and/or refine the claim to environments with particular properties where the expectation is that TD3+ODT outperforms ODT. * Minor: Section 3.1 could highlight the conceptual figure of the plot of Figure 1, c, to help the reader understand the graph at the beginning of section 3.1. Technical Quality: 3 Clarity: 3 Questions for Authors: * Can the authors comment or perform an ablation with ODT+DDPG for all the main experiments as well? * What is the impact of context length on your approach? * For the main experiments how many random seed runs are each experimental result over? E.g. in Figure 2. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are adequately discussed in section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for appreciating our work. Below are responses to questions: Q1. Overclaim of “boosting performance”, and the need to refine claims with properties where TD3+ODT should outperform ODT. Thanks for pointing this out. While TD3+ODT is generally better than ODT, we are happy to rephrase our claim to reflect that there are some cases where ODT performs better than TD3+ODT. To be specific, we expect ODT to struggle (either fail completely or fail to further improve) with medium-to-low quality offline data. This is supported by our result on adroit environments with cloned/human dataset, antmaze, and MuJoCo (especially medium-replay and random where there are many trajectories with low ground truth RTGs). In contrast, if offline data has good quality (e.g., adroit expert), or the reward signal is sparse (e.g., delayed reward in Appendix G.1) so that RL struggles, then we expect ODT to work as good or better. Q2. [minor] Figure in Section 3.1 could be improved for better understanding. Thanks a lot for the advice! We will modify the figure accordingly in the revised version. Q3. Ablations on DDPG+ODT. Great suggestion! Unfortunately, due to the rebuttal time limit, we can only provide ODT+DDPG results on some environments for now. We will provide results of ODT+DDPG on all additional environments in the revised version. The current results are shown in Fig. 2 of the global pdf. We find DDPG+ODT does not work. We speculate that this is due to the rougher landscape of the estimated Q-value without smoothing, delayed update, and double Q-learning by TD3. Q4. Impact of context length on the approach. We conducted ablations on both the context length at training time ($T_{\text{train}}$) and evaluation / rollout time ($T_{\text{eval}}$). The former is shown in Fig. 18 of Appendix G.3, and the latter is shown in Fig. 4 (b). Generally, we found that there is a tradeoff between more information and training stability: while longer context length allows the decision transformer to utilize more information, it also causes increased input complexity, i.e., noise from the environment leads to reduced stability and slower convergence in training. Note, this differs from LLM works, where context length improves generalization thanks to extremely high expressivity, huge amounts of data, and less noisy evaluation. This is especially the case when bootstrapping is involved, as the fitting target depends on its own output. In fact, in prior work, Parisotto et al. [4] suggest in their Sec. 3.1 that if a recurrent agent can first learn a Markovian policy and value function at the start of its training, then the training process could be stabilized. Q5. Number of random seeds. We use 3 seeds for all experiments (a few ablations without standard deviation use only 1 seed). Note, even without ablations this amounts to a large number of experiments: 1) we use 46 different datasets for our main results [4 for maze, 24 for mujoco ({random, medium-replay, medium}x{hopper, halfcheetah, walker2d, ant}x{normal, delayed reward}), 12 for adroit ({pen, hammer, door, relocate}x{expert, cloned, human}), and 6 for antmaze]; and 2) we test 6 methods. This amounts to 800+ runs (46 datasets, 6 methods, 3 seeds), each of which requires 6-8h on average (IQL and TD3+BC need less time, but PDT and other methods on complex environments such as adroit require more). --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed rebuttal response, and I particularly appreciate the new experimental ablations in the global response. Thank you for re-phrasing your claim. I look forward to seeing the results of ODT+DDPG on all additional environments in the revised version. I am still keeping my score the same. --- Reply to Comment 1.1.1: Comment: Thanks for your appreciation of our work and response! We have started the experiments and will surely have those results ready in our revised version.	Summary: This paper addresses the challenge of online finetuning of online decision transformers (ODT). Theoretical results are provided to show the target return-to-go can hamper finetune. The authors propose to have TD3 gradients added to ODT finetune and improve its performance especially when pretrained with low-reward data. Extensive empirical results are provided to show the proposed method can achieve stronger or competitive results across a large number of environments. Strengths: originality - The idea to combine TD3 gradient and ODT training is quite novel. The theoretical results can also be considered novel findings. quality - Overall good quality, the paper is well-written. Figures are easy to read and extensive experiments and ablations are provided. clarity - Overall clear and easy to follow. significance - The empirical results help improve our understanding of DT methods and how to better finetune them. The results are quite strong. The extensive experiments across different benchmarks make the results more convincing. The theoretical results add to significance. Weaknesses: - Maybe I missed something but I feel it might be good to have ODT baselines that try to tackle online finetuning with alternative (and more naive) methods. For example, authors argue that a main difficulty when finetuning DT is that when pretrained with low-reward data, the target RTG at finetuning stage is simply hard to obtain. But if we provide a realistic target RTG initially (based on pretrain data), and gradually increase it, will that help performance? Another thing is if we have a better exploration policy, will that help ODT finetune to the same extent? - Adding the TD3 component can make the training more complex and slower. Technical Quality: 3 Clarity: 4 Questions for Authors: - Will alternative strategies that improve exploration at finetune phase help ODT to the same extent? - The random datasets are not commonly tested in previous papers, when the baselines are tested on these datasets, are the hyperparameters tuned for these datasets? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Some limitations are discussed in section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for appreciating our work. Below are responses to questions: Q1. test ODT baselines that tackle online finetuning with alternative methods, e.g., better exploration policy and gradually increasing target RTG. Great suggestions! Currently, ODT explores due to the stochasticity of its policy. We enhance ODT in two ways and provide results in the global pdf: 1) ODT+gradually increasing target RTG (which is similar to curriculum learning in RL); and 2) a cheating baseline ODT with oracle exploration, i.e., jump-start RL [2], using as the guide policy an expert policy trained via IQL on expert data. Concretely, the expert policy is used for the first $n$ steps in an episode, before ODT takes over. We set $n=100$ (100% max episode length for adroit pen, 50% max episode length for other adroit environments) initially, and apply an exponential decay rate of $0.99$ for every episode. Results are summarized in Fig. 1 of the global pdf. Curriculum RTG does not work, probably because the task is too hard and cannot be improved by random exploration without gradient guidance. Also, even with oracle exploration, ODT is not guaranteed to succeed: it fails on the hammer environment where TD3+ODT succeeds, probably because of insufficient expert-level data and an inability to improve with random exploration. Q2. TD3 slows down training. Our method is only 20% slower than ODT without TD3 gradients. In Appendix H, we discussed the impact of the introduced RL gradient on training: the critic only takes about 20% of the time to train, while the actor overhead is negligible (since it only contains an MLP critic inference to get the Q-value). We found that on our machine, for the MuJoCo experiment, ODT requires about 6 hours to train. We will add part of the discussion in Appendix H to the limitation section of the revised paper. Q3. Are the hyperparameters tuned on the random datasets? No, as suggested in Tab. 9 in the appendix, we use the same hyperparameters for every variant (e.g., medium / medium-replay / random, expert / human / cloned) on the same environment. Note, our parameters on the random datasets are aligned with the ODT medium environment and not tuned further.	Rebuttal 1: Rebuttal: We thank the reviewers, ACs and SACs for valuable feedback. We are delighted that our idea was appreciated as novel (AdzZ, 1gFB), well-motivated (1gFB), valuable (NWdD, gVzs, 1gFB) and backed by theoretical foundations (AdzZ, NWdD), the literature review was appreciated as well-described (NWdD), and results were referred to as strong (AdzZ, gVzs) and sufficiently extensive (AdzZ, NWdD). Also, reviewers unanimously rate the presentation of our work positively (AdzZ gives “excellent”, all other reviewers give “good”). We answer common questions here: Q1. Number of random seeds (1gFB, NWdD). We use 3 seeds for all experiments (a few ablations without standard deviation use only 1 seed). Note, even without ablations this amounts to a large number of experiments: 1) we use 46 different datasets for our main results [4 for maze, 24 for mujoco ({random, medium-replay, medium}x{hopper, halfcheetah, walker2d, ant}x{normal, delayed reward}), 12 for adroit ({pen, hammer, door, relocate}x{expert, cloned, human}), and 6 for antmaze]; and 2) we test 6 methods. This amounts to 800+ runs (46 datasets, 6 methods, 3 seeds), each of which requires 6-8h on average (IQL and TD3+BC need less time, but PDT and other methods on complex environments such as adroit require more). Q2. Our design slows down training (AdzZ, gVzs, NWdD). We thank the reviewers for pointing this out, and we are happy to include the discussion in the limitation section. Meanwhile, we note two points: 1. Our work aims to analyze and improve ODT. We thus compare computational cost primarily to ODT instead of MLP-based solutions; 2. In Appendix H we stated that our proposed solution is only marginally (20%) slower than ODT: while the use of RL gradients slows down training, the actor training overhead is negligible (since it only contains an MLP critic inference to get the Q-value), and the critic only takes about 20% time to train. We found that on our machine, for the MuJoCo experiment, ODT requires about 6 hours to run. Q3. Additional ablations (AdzZ, 1gFB, NWdD). We also provide a variety of new experimental ablations, which are summarized in the global pdf (experiment 1 and 5 mentioned below are in Fig. 1; experiment 2, 3, 6 are in Fig. 2; experiment 4 is in Fig. 3). 1. Better exploration by “cheating”: letting an expert take over in early steps (reviewer AdzZ). Even with an oracle (“cheating”) exploration strategy, ODT can still fail in some cases where our method prevails, likely because of insufficient expert-level data and an inability to improve with random exploration. 2. ODT baselines which gradually increase target RTG (reviewer AdzZ). We find that such curriculum learning does not work, probably because the task is too hard and cannot be improved by random exploration without gradient guidance. 3. ODT+DDPG (reviewer 1gFB). We find that ODT+DDPG does not work. We speculate that this is due to a less stable Q-value landscape during training. 4. Better evaluation using rliable library (reviewer NWdD). We re-evaluate all our main results using the rliable library, and find that our method still generally outperforms other baselines, especially in adroit environments. 5. TD3 with forgetting mitigation issues implemented via KL regularizer and jump-start RL where the guide policy is the pretrained policy (reviewer NWdD). Jump-start RL with the pretrained policy does not work well, probably because it does not directly prevent out-of-distribution policy updates. KL regularizer effectively mitigates forgetting for both TD3 and TD3+ODT, but it also hinders policy improvement of the TD3+ODT policy pretrained on low-RTG offline data. 6. TD3+BC+transformer and TD3+RvS (reviewer NWdD). We find that TD3+BC+transformer works well in the adroit environment, albeit still worse than our proposed method. TD3+RvS does not work: it only slightly outperforms TD3+BC. We speculate that this is because an MLP is not expressive enough to model the policy change over different RTGs. Finally, we list the papers referred to in our responses: References [1] A. Li et al. MAHALO: Unifying Offline Reinforcement Learning and Imitation Learning from Observations. In ICML, 2023. [2] I. Uchendu et al. Jump-Start Reinforcement Learning. In ICML, 2023. [3] S. Emmons et al. RVS: What is Essential for Offline RL via Supervised Learning? In ICLR, 2022. [4] E. Parisotto et al. Stabilizing Transformers for Reinforcement Learning. In ICML, 2020. Pdf: /pdf/8e3d3921cbca99a04e856b765c48e0d79a6b4033.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On the Efficiency of ERM in Feature Learning	Accept (poster)	Summary: This paper considers a "feature learning" setting equivalent to structural risk minimization over a set of hypotheses of the form $\langle w, \phi_t(x)\rangle$. It demonstrate that the statistical error of this procedure converges to a quantity that depends on a natural empirical process defined only in terms of the classes which contain an optimal predictor. Thus, it is claimed that ERM is "efficient" at feature learning in a natural way. EDIT: I raise my score from a 5 to a 6. I encourage the authors to include the counterexample presented to me in the rebuttal in the main text as a form of motivation, and as a didactic example to express "what goes wrong". The authors can also remark about situations under which better dependence on $\delta$ is achievable. Thank you to authors for addressing my points. Strengths: This paper has a number of strengths that make it compelling. (1) The formalism studies is exceptionally clean and natural. It lends itself to study well in other settings as well. (2) The authors provide both asymptotic and nonasymptotic results, and the asymptotic ones are rather "sharp" in that they provide insight into both the limiting distributions, and do so in terms of universal, Gaussian empirical processes. (3) The authors proofs are effective and succinct, and demonstrate great command of the relevant technical machinery. (4) The authors use of asymptotic bounds allows them to build considerable intuition in the presentation, before introducing the non-asymptotic bounds which are considerably more involving to parse. Weaknesses: There are a couple weaknesses, however, that temper my excitement. (1) I apologize for the directness, but I do not find the qualitative finding particularly surprising. We know that ERM localizes, and as a consequence, if I have a predictor class of the form $\mathcal F= \bigcup_{t \in \mathcal{T}} \mathcal{F}_t$, and all optimal predictors lie in $\mathcal F_t, t \in \mathcal{T}^{\star}$, and moreover, the risk of an $f \in \mathcal{F}_t$ where, $t \notin \mathcal{T}^{\star}$ is lower bound away from zero, then localization should force the limiting behavior of the problem to only depend on the statistical complexity of $\bigcup_{t \in \mathcal{T}^\star} \mathcal{F}_t$ It is nice to quantify this rigorously, but again, the phenomenon does not seem to be to bring fundamentally new insight. (2) The non-asymptotic dependence on the probability of error $\delta$ is quite bad. Indeed, an $O(1/\delta)$ dependence is not even integrable, obviating in expectation bounds (perhaps in-expectation in too much to ask for, given the possibility that the covariance matrices become singular. Still, this seems like a major limitation. (3) While the problem setting is remarkably clean, Theorem 4 is not. This seems inevitable, and the authors both (a) explain its intuition and (b) instantiate it for a natural class of problems... (4) I think the authors should include some more intuition about the relevant increments and terms defined in the paper ($G_n(t)$, $\Lambda_n(t)$) and so forth. Citing that "such and such is just such and such in Bos[02]" does not help with intuition building. The authors might consider adding a section in the appendix that elaborates further, and also provides formal definitions of what it means for a class to be Glivenko Cantelli (this was relatively clear) and, more importantly, Donsker (as a reader with the relevant background, I know what is meant, but this could be less accessible to others). The authors might also consider remarking on what limitations the Donkser-ity of the problem entail (e.g. sufficient restriction on metric entropy). Technical Quality: 4 Clarity: 3 Questions for Authors: Can the authors please explain why the localization phenomenon presented in this work is surprising? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: (1) Weak dependence on error probability $\delta$ (2) Possible limitations due to the Donkser assumption (3) Restricted to the linear setting Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for engaging with our paper and for their feedback. We address some of their concerns below. --- **I apologize for the directness, but I do not find the qualitative finding particularly surprising. We know that ERM localizes, and as a consequence, if I have a predictor class of the form $\mathcal{F} = \bigcup\_{t \in \mathcal{T}} \mathcal{F}\_{t}$, and all optimal predictors lie in $\mathcal{F}\_{t}$, $t \in \mathcal{T}\_{\}$, and moreover, the risk of an $f \in \mathcal{F}\_{t}$ where $t \notin \mathcal{T}\_{\}$ is lower bound away from zero, then localization should force the limiting behavior of the problem to only depend on the statistical complexity of $\bigcup\_{t \in \mathcal{T}\_{\}}\mathcal{F}\_{t}$. It is nice to quantify this rigorously, but again, the phenomenon does not seem to be to bring fundamentally new insight.*** If one believes on some level that the performance of ERM is related to the complexity of the class on which it is performed, then our results are quite surprising. Our understanding is that this is a commonly held belief in the machine learning community. For a reader very familiar with localization, we can see how they would be skeptical that the above statement holds once the setting we propose is presented to them. For such a reader, perhaps the most surprising aspect of our result is that one can get away with extremely weak assumptions on certain empirical processes yet still obtain a strong localization phenomenon. Conversely, and perhaps just as surprising to such a reader, is that without such assumptions, the localization phenomenon can completely vanish. Here is a counter-example to the reviewer's statement (as we understood it, please correct us if we missed something) that we hope makes this last point clear. Let $\mathcal{F} := \\{x \mapsto \langle w, \phi(x) \rangle \mid w \in \mathbb{R}^{d}\\}$ be a linear class induced by a feature map $\phi$. We can write this $d$-dimensional subspace of functions as the union of its $1$-dimensional subspaces. More explicitly, let $\mathcal{T}$ be the half of the unit Euclidean sphere in $\mathbb{R}^{d}$ with non-negative first coordinate, let $\phi\_{t}(x) := \langle t, x\rangle$ for all $t \in \mathcal{T}$, and set $\mathcal{F}\_{t} := \\{x \mapsto a \cdot \phi\_{t}(x) \mid a \in \mathbb{R}\\}$, then $\mathcal{F} = \bigcup\_{t \in \mathcal{T}} \mathcal{F}\_{t}$. Now consider a well-specified linear regression problem under square loss over $\mathcal{F}$, i.e. $Y = \langle w\_{\} , \phi(X) \rangle + \varepsilon$ for some $w\_{\} \in \mathbb{R}^{d}$ and $\varepsilon \sim \mathcal{N}(0, \sigma^{2})$ independent of $X$. Then the optimal feature map is $t\_{} = \pm w\_{\}/ \\\|w\_{\}\\\|\_{2}$, and for each $f \in \mathcal{F}\_{t}$ for $t \neq t\_{\}$, the excess risk of $f$ is lower bounded away from zero, so if the statement of the reviewer held, the excess risk should be of order $1$ (proportional to the complexity of the class corresponding to the optimal feature map $\mathcal{F}\_{t\_{\}}$, which is one dimensional). Yet, it is a classical result that the excess risk in this problem is of order $d$, i.e. it depends on the complexity of the full class $\mathcal{F}$. This example shows that the intuitive reasoning that because we have a function class expressed as the union of other function classes we should expect localization might be misleading. More is needed, but perhaps the surprising thing is that not too much is needed: Glivenko-Cantelliness/Donskerity of certain empirical processes. To a certain extent, we agree that the most significant contribution of our work is to propose a tractable framework where feature learning can be studied rigorously, and we are excited to see how it can be applied to study other problems beyond the supervised learning setting we considered in this paper. Even if our results are not surprising to the reviewer, we hope that the simplicity and usefulness of the framework we introduce and the rigor of our work convinces them that the ideas in this paper are worth sharing. The non-asymptotic dependence on the probability of error is quite bad. Indeed, an $O(1/\delta)$ dependence is not even integrable, obviating in expectation bounds (perhaps in-expectation in too much to ask for, given the possibility that the covariance matrices become singular). Still, this seems like a major limitation.* We note that, as discussed in line 133, under our very weak assumptions, this is the best dependence on $\delta$ one can obtain for the performance of ERM. In fact, this bad dependence on $\delta$ has spurred the development of a susbtantial literature that aims at designing procedures with better dependence on $\delta$; please refer to the works cited in line 133. We chose to present our results under the weakest assumptions possible in order to maximize their range of applicability. We agree with the reviewer that under more stringent assumptions (e.g. boundedness of $X$ and $Y$), our analysis would yield a much improved dependence on $\delta$. --- Rebuttal 2: Title: Rebuttal (continued) Comment: *I think the authors should include some more intuition about the relevant increments and terms defined in the paper ($G\_{n}(t)$, $\Lambda\_{n}(t)$) and so forth. Citing that "such and such is just such and such in Bos[02]" does not help with intuition building. The authors might also consider remarking on what limitations the Donkser-ity of the problem entail (e.g. sufficient restriction on metric entropy).* We will aim at providing more intuition when defining our terms in the final version of our paper. In particular, we will replace the particular passage referred to by the reviewer by the following sentence: "The parameter $L$ characterizes the deviation of the supremum of the empirical process $\Lambda\_{n}(t)$ from its mean." As for Donskerity, we have refrained from referring to metric entropy in the paper except very briefly in lines (229-230). Our hope was to keep the paper accessible, and as such we preferred to describe Donskerity by analogy to the central limit theorem (lines 157-158). We will briefly mention that Donskerity can be established under appropriate metric entropy restrictions in the main paper, and defer a more in depth discussion to the Appendix. *The authors might consider adding a section in the appendix that elaborates further, and also provides formal definitions of what it means for a class to be Glivenko Cantelli (this was relatively clear) and, more importantly, Donsker (as a reader with the relevant background, I know what is meant, but this could be less accessible to others).* We agree with the reviewer's suggestion, and we will add such a section in the Appendix. --- Rebuttal 3: Title: discussion Comment: Dear Reviewer bjqG, Thank you very much for submitting your review report. The author(s) have posted responses to your review. Could you kindly provide comments on whether your concerns have been adequately addressed? Best regards, AC --- Rebuttal Comment 3.1: Title: Following up Comment: Thank you, authors, for your detailed feedback. As indicated in my updated review, I raised my score to reflect that my concerns were addressed.	Summary: This paper investigates the learning theory of empirical risk minimization (ERM) with feature learning. Under the setting where the optimal finite-sample feature is selected by minimizing the empirical risk over a class of features, the authors show that ERM with feature learning implies convergence of the excess risk under certain assumptions. Moreover, other statistical properties, such as asymptotic normality, are also derived. Strengths: The statistical theory of this paper regarding feature learning is solid, and the results of the main theorems (Theorem 3 and Theorem 4) seem to be correct, even though I have not had time to check the detailed proofs. In fact, the Glivenko-Cantelli assumptions on the empirical processes and the finite moment assumptions are widely used in learning theory, and the rate $1/n\delta \approx O(1/\sqrt{n})$ seems to be correct. Weaknesses: 1. The title "On the Efficiency of ERM in Feature Learning" seems misleading. Initially, it suggests that ERM should aid feature learning. However, after reading the paper, it appears that the authors are discussing the statistical performance of ERM in the context of feature learning, without clearly demonstrating how ERM aids in feature learning. 2. The authors have reviewed many statistical theory papers. However, in terms of feature learning or representation learning, they didn't conduct a thorough literature review. Specifically, for the final-layer feature, recent studies show that the last-layer feature will converge to an Equiangular Tight Frame (optimal feature for classification problems), which is known as neural collapse, as proposed by Papyan et al. in their paper "Prevalence of neural collapse during the terminal phase of deep learning training." There are some subsequent studies that show that the optimal feature is indeed the minimizer of a regularized ERM. These studies include but are not limited to "Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training" (Fang et al., 2021); "A Geometric Analysis of Neural Collapse with Unconstrained Features" (Zhu et al., 2021); and "Neural Collapse in Multi-label Learning with Pick-all-label Loss" (Li et al., 2024). For different loss functions, the authors may refer to "Neural Collapse Under MSE Loss: Proximity to and Dynamics on the Central Path" by Han et al. Regarding learning theory, the sample complexity under neural collapse is investigated by Wang et al. (2024) in their work "Neural Collapse Meets Differential Privacy: Curious Behaviors of NoisyGD with Near-perfect Representation Learning." 3. Some of the claims are confusing and may be over-claimed. Specifically, in the Conclusion, the authors claim that their theory might be used to explain the double descent phenomenon and generalization of deep learning with label noise in Zhang et al., 2021. However, the feature learning setting in this paper is far from being extended to the deep learning setting. Indeed, their assumptions here, such as the moment assumption, should hold for all feature maps in the hypothesis class, which is not verified for deep neural networks. However, existing feature or representation learning theory such as the Neural Collapse theory can partially explain the phenomenon in Zhang et al., such as the overfitted model can still generalize. 4. As the theory is not as enlightening as the authors claimed in their paper since their setting is not practical, this should be regarded as a purely theoretical paper. As a theoretical paper, there should be some room for improvement, such as verifying the assumptions for all feature maps (such as the moment assumptions in Theorem 4) for a certain class. Moreover, as a purely statistical theory paper, a lower bound showing that the obtained finite-sample rate is optimal is necessary for a high-quality publication. It would also be better for the authors to emphasize more technical difficulties compared to ERM without feature learning. In fact, under the Glivenko-Cantelli assumptions and the finite sample assumptions, both the asymptotic theory (Theorem 3) and the non-asymptotic theory (Theorem 4) look like a simple extension of Theorem 1 and Theorem 2 without feature learning, while Theorem 1 and Theorem 2 are not novel in learning theory. Technical Quality: 3 Clarity: 3 Questions for Authors: Please address my concerns in the weaknesses part. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to review our paper. We address some of their concerns below. --- *The authors have reviewed many statistical theory papers. However, in terms of feature learning or representation learning, they didn't conduct a thorough literature review. Specifically, for the final-layer feature, recent studies show that the last-layer feature will converge to an Equiangular Tight Frame (optimal feature for classification problems), which is known as neural collapse,...* We thank the reviewer for bringing this line of work to our attention; we will carefully discuss their relevance in the final version of the paper. That said, our work is squarely in the realm of statistical learning theory (STL), and as such we have focused our effort on reviewing the STL literature, outlining existing results and techniques, and explaining what is challenging in our newly proposed setting. --- *Some of the claims are confusing and may be over-claimed. Specifically, in the Conclusion, the authors claim that their theory might be used to explain the double descent phenomenon and generalization of deep learning with label noise in Zhang et al., 2021. However, the feature learning setting in this paper is far from being extended to the deep learning setting. Indeed, their assumptions here, such as the moment assumption, should hold for all feature maps in the hypothesis class, which is not verified for deep neural networks. However, existing feature or representation learning theory such as the Neural Collapse theory can partially explain the phenomenon in Zhang et al., such as the overfitted model can still generalize.* The only two sentences that relate our work to the experiments of [Zha+21] are: 1- "The most tantalizing aspect of our results is their potential* in explaining the experiments in [Zha+21]".* 2- "Formally connecting our statements to these experiments is beyond what we achieved here, yet, we believe that the new perspective we took might generate useful insights in this area." We believe the first sentence above leads to a misinterpretation of what we intended, and we will carefully revise this statement for clarity. --- *As the theory is not as enlightening as the authors claimed in their paper since their setting is not practical, this should be regarded as a purely theoretical paper.* We agree with the reviewer's second remark that our paper is indeed purely theoretical. However, our work provides a new perspective on the analysis of ERM and feature learning, and has the following main message: Learning feature maps is easy when only a small subset of them is good, as the bad ones are quickly discarded by the ERM procedure. We are not aware of prior works with the same conclusion. --- *It would also be better for the authors to emphasize more technical difficulties compared to ERM without feature learning.* We have a brief discussion about this in lines 72-78. We will elaborate more on these technical difficulties in the revised paper. --- *In fact, under the Glivenko-Cantelli assumptions and the finite sample assumptions, both the asymptotic theory (Theorem 3) and the non-asymptotic theory (Theorem 4) look like a simple extension of Theorem 1 and Theorem 2 without feature learning, while Theorem 1 and Theorem 2 are not novel in learning theory.* We will clarify the technical contributions and the novelty of the analysis in more detail in the revision. We briefly remark here that while Theorem 3 vastly extends Theorem 1, this extension is by no means straightforward: 1. It requires a new decomposition (eq. 19) of the excess risk that separates the error coming from the choice of feature map and the error coming from the choice of linear predictor. This is in contrast with Theorem 1 where the feature map is fixed, and there is only one source of error coming from the choice of linear predictor. 2. It requires an independent proof that asymptotically, the feature map chosen by ERM is an optimal one. In Theorem 1, we have a single fixed feature map, so there is no analogue to this statement in the proof of Theorem 1. 3. It requires a new limiting argument that relies on tools from empirical process theory (limiting gaussian process, continuous mapping theorem) to show that the excess risk is controlled by the complexity of the set of optimal feature maps $\mathcal{T}_{\}$. The analogous step in the proof of Theorem 1 only requires studying a single* empirical average, which is readily achieved by an application of the Central limit theorem. Similarly, while Theorem 4 is an extension of Theorem 2, this extension is again not straightforward: 1. It requires a new proof that bounds the suboptimality of the feature map picked by ERM non-asymptotically and relies on a localization argument of Koltchinskii [Kol06]. In Theorem 2, the feature map is fixed, so no analogous step exists in its proof. 2. It requires bootstrapping this localization result to show that the excess risk of ERM does not depend on the full complexity of the function class, but rather on the complexity of shrinking subclasses of it. This is done by studying the supremum of a certain empirical process. In Theorem 2, the analogous step requires controlling a single empirical average using Markov's inequality. 3. It requires showing that the shrinking subclasses of functions whose complexity controls the excess risk of ERM converges to the class of functions induced by the set of optimal feature maps (Lemma 1), recovering the result of Theorem 3 asymptotically. This is also new and unrelated to the proof of Theorem 2 where the feature map is fixed. --- We would be happy to answer any questions that may arise during the discussion period. --- References [Zha+21]: Zhang, Chiyuan, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. "Understanding deep learning (still) requires rethinking generalization." --- Rebuttal Comment 1.1: Comment: 1. The representation learning theory, such as neural collapse, is more related to [Zha+21], as I mentioned. Thus, it is connected to statistical learning theory (STL) since STL aims to derive statistical properties of certain phenomena in machine learning, not just the asymptotic properties or the error bounds. BTW, they can partially explain some phenomena occurring in deep learning, whereas classical (asymptotic or non-asymptotic) learning theory may fail to do so. 2. The most important part I mentioned, the sharpness of the derived bound, is currently not discussed in the rebuttal (maybe due to the limited space?). As I said, once the theoretical framework is formalized, a lower bound that matches the upper bound is essential for a purely theoretical paper, which I believe is common sense in STL (such as in those papers published in AOS). --- Reply to Comment 1.1.1: Comment: We thank the reviewer for their response. --- *The representation learning theory, such as neural collapse, is more related to [Zha+21], as I mentioned. Thus, it is connected to statistical learning theory (STL) since STL aims to derive statistical properties of certain phenomena in machine learning, not just the asymptotic properties or the error bounds. BTW, they can partially explain some phenomena occurring in deep learning, whereas classical (asymptotic or non-asymptotic) learning theory may fail to do so.* Our work is most closely related to the literature that aims at understanding machine learning procedures through upper and lower bounds on their performance, which we called statistical learning theory, but we understand that we might be using the term differently from the reviewer. This will be clarified in the paper once we include a discussion of the additional line of work pointed out by the reviewer. Our comments on the relationship between our work and the experiments [Zha+21] are a minor aspect of our paper. They should be seen merely as an invitation to the reader to look at the problem of explaining these experiments through a new perspective. --- *The most important part I mentioned, the sharpness of the derived bound, is currently not discussed in the rebuttal (maybe due to the limited space?). As I said, once the theoretical framework is formalized, a lower bound that matches the upper bound is essential for a purely theoretical paper, which I believe is common sense in STL (such as in those papers published in AOS).* The second statement of Corollary 1 in our paper provides matching upper and lower bounds on the asymptotic quantiles of the excess risk of ERM, and the gap between these bounds is a factor of two. The general statement of Theorem 3 only contains an upper bound, but it is straightforward to check that the argument behind the lower bound in Corollary 1 immediately extends to the general case. In this case however, it yields a lower bound on the quantiles of the excess risk of the same form as the upper bound in Theorem 3, but with the supremum replaced by an infimum. Roughly speaking, this gap is due to the fact that the sequence of ERMs can “oscillate” between optimal feature maps. This problem already appears if one considers only two features maps which are both optimal, and to the best of our knowledge no matching upper and lower bounds on the quantiles of the excess risk are known even in this simple case. On the non-asymptotic front, we note that even in the linear regression case, Theorem 2, there is no known matching lower bound to the upper bound we presented. One may sacrifice interpretability and instead use a tighter upper bound in terms of the quantiles of $\\|\|g(X, Y)\\|\|_{\Sigma^{-1}}^{2}$, which can be reversed up to an absolute constant and a different dependence on $\delta$, under the sample size restriction of the theorem. Moving from the linear regression setting to the case of multiple feature maps we study is more delicate however, and the iterative localization method of [Kol06] we used only yields upper bounds, and sheds little light on lower bounds. Despite this shortcoming, as we have emphasized in the paper, the upper bound we derived in Theorem 4 is asymptotically tight in that it is consistent with the asymptotic behavior of the excess risk we derived in Theorem 3 and Corollary 1. If the reviewer thinks the above discussion on lower-bounds is interesting, we would be happy to include it in the final version.	Summary: This paper consider the problem of regression over the linear classes induced by a collection of feature maps. They study both the asymptotic and the non-asymptotic behavior of the empirical risk minimizer. Surprisingly, although the linear classes has a complexity much higher than that of only one linear map, the authors find that when there is a unique optimal feature map, ERM actually behaves similar with the oracle procedure (knows a priori the optimal feature map). General results for non-unique or even infinite feature map is also provided. The authors also apply their non-asymptotic result on finite feature map cases. Strengths: The results in this paper is both novel and significant. They show that even though non-optimal feature map exists in training process, the actual upper bound of the excess risk depends on the size of optimal feature maps. From theoretical perspective, the proofs are solid. The writing is also good, clearly states the results and the intuition, and how their results improve beyond previous classical results where the feature map set is a singleton. Case study is also provided, giving a comprehensive review of how their general framework can be applied. Weaknesses: It will be good if more case studies are provided. Technical Quality: 3 Clarity: 3 Questions for Authors: Is it possible to consider some infinite feature map set with some structure, so that you can also compute the constants in your results? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No limitations are stated. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for reading our paper and for their positive comments. We address their question and comments below. *It will be good if more case studies are provided.* As mentioned in another comment, one additional example we wanted to include of some practical relevance is the following. A popular statistical learning problem, which motivates the LASSO procedure, asks the following question: in a regression problem under the square loss, learn the best linear predictor from a $d$-dimensional feature map, knowing that the optimal linear predictor is $k$-sparse, i.e. the optimal weights have only $k$ non-zero entries for $k \ll d$. Our results allow us to tackle the following more general problem: under no assumptions on the optimal weight vector, learn the best linear predictor from a subset of size $k$ of the $d$ features. This problem fits in our framework neatly by taking $\mathcal{T}$ to be given by the set of all subsets of size $k$ of $[d]$ with $\|\mathcal{T}\| = {d \choose k}$, and our theorems imply, among other things, that the asymptotic quantiles of the excess risk of ERM on this problem are, up to a factor of two, the same as the oracle procedure that knows a priori the best subset of size $k$ of the $d$ features (assuming it is unique). --- *Is it possible to consider some infinite feature map set with some structure, so that you can also compute the constants in your results?* As we described in another comment, upper bounds on the various expected suprema appearing in our results can be obtained under an assumption on the metric entropy of the set $\mathcal{T}$ under an appropriately chosen metric. One may then use an $\varepsilon$-net argument along with our finite-case results and an argument that controls the approximation error to obtain an upper bound on these expected suprema. However, obtaining such metric entropy estimates is a highly non-trivial task for any given problem, and requires specialized techniques that leverage the particular structure of the problem. We are currently trying to derive such entropy estimates for two-layer neural networks but we have yet to succeed. We believe that this is a good direction for future work. --- *Limitations* Please also note that we discussed the limitations of our work in the last paragraph of the paper. If the reviewer thinks something is missing, we would be happy to discuss it. --- Rebuttal Comment 1.1: Comment: Thanks for your reply.	Summary: This paper studies the novel setting where we are give a collection of predefined feature maps (indexed by a set T) and we choose one of these feature maps and then, learn a linear predictor on top of the chosen feature map. The authors derive upper bounds on the excess risk that depend on the number (size) of "optimal" feature maps and not the size of the set T. Strengths: - The author propose a setting for feature learning that is novel and I find it interesting. - The result in this setting very satisfying. To the best of my knowledge, this is the first result that shows that the excess risk (or at least the upper bound on the excess risk) depends on the size of "optimal" features. - The proof outline and strategy seems correct. I have not fully checked all the details of the proofs. But, the ones I have checked are all correct and solid. The paper is also very well written. Weaknesses: - The analysis ignores the role of the learning algorithm and looks at the problem from a purely statistical perspective. The role of implicit bias of the algorithm is not seen here. On a high level, from the set of optimal features, some features are easier to learn/ achieve than others. This might also shrink the effective size of the features. - The suprema term in the expression of Theorem 4 is not very interpretable. The case of finite features in the next section makes it more clear. However, I find the finite case not that interesting. Is there an interesting non-finite case that one can analyze to get an interpretable result? - In general, the features are also learned from data and this gives extra dependencies and goes beyond the setting of this paper (unless some sample splitting is done). - Can the authors comment on the tightness of Theorem 3 and 4? An the potential challenges of coming up with lower bounds? - The paper will greatly benefit from a simulation result to support the main finding of the paper. For example, for a simple finite feature case. A theoretical example can also help. Technical Quality: 4 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to read our paper and for their positive evaluation. We address their comments below. --- *The analysis ignores the role of the learning algorithm and looks at the problem from a purely statistical perspective. The role of implicit bias of the algorithm is not seen here. On a high level, from the set of optimal features, some features are easier to learn/ achieve than others. This might also shrink the effective size of the features.* At a high-level, we agree. Our guarantees are for the worst-case ERM and it is conceivable that a learning algorithm can avoid such worst-case scenario. Please note however that, as we have briefly hinted at in the conclusion, it is difficult to discuss learning algorithms in our setting given its extreme generality: the index set $\mathcal{T}$ is not even equipped with a topology. One way to study a notion of "(implicit) bias" in our setting is to introduce some ordering on the set $\mathcal{T}$ and consider the procedure that picks the smallest ERM in this ordering. However, we strongly believe that it is more appropriate to study this question for specific model classes and learning algorithms of interest, rather than in the general setting we consider here, as the insights obtained for the above-described procedure might not be transferrable to cases of practical interest. --- *The suprema term in the expression of Theorem 4 is not very interpretable. The case of finite features in the next section makes it more clear. However, I find the finite case not that interesting. Is there an interesting non-finite case that one can analyze to get an interpretable result?* We agree with the reviewer on this limitation; the current Theorem 4 yields interpretable results only in the finite case. Obtaining interpretable results in the non-finite regime is an interesting and important direction that we leave for future work. One way forward would be to consider for example any set $\mathcal{T}$ satisfying a metric entropy estimate (in an appropriately selected metric). One may then use an $\varepsilon$-net argument along with our finite-case results and an argument that controls the approximation error to obtain an upper bound on the expected suprema of Theorem 4. However this approach only provides an upper bound, and requires some work for it to be practically interesting. --- *Can the authors comment on the tightness of Theorem 3 and 4? An the potential challenges of coming up with lower bounds?* On the asymptotic side, when the optimal feature map is unique, the second statement of Corollary 1 (of Theorem 3) already provides matching upper and lower bounds on the asymptotic excess risk (and the gap between the bounds is a factor of two). The general statement of Theorem 3 only contains an upper bound, but it is straightforward to check that the argument behind the lower bound in Corollary 1 immediately extends to the general case. In this case however, it yields a lower bound on the quantiles of the excess risk of the same form as the upper bound in Theorem 3, but with the supremum replaced by an infimum. Roughly speaking, this gap is due to the fact that the sequence of ERMs can "oscillate" between optimal feature maps. This problem already appears if one considers only two features maps which are both optimal, and to the best of our knowledge no matching upper and lower bounds on the quantiles of the excess risk are known even in this simple case. If the reviewer thinks the above-discussed lower-bound is interesting, we would be happy to include it in the final version. On the non-asymptotic front, we note that even in the linear regression case, Theorem 2, there is no known matching lower bound to the upper bound we presented. One may sacrifice interpretability and instead use a tighter upper bound in terms of the quantiles of $\\|\|g(X, Y)\\|\|_{\Sigma^{-1}}^{2}$, which can be reversed up to an absolute constant and a different dependence on $\delta$, under the sample size restriction of the theorem. Moving from the linear regression setting to the case of multiple feature maps we study is more delicate however, and the iterative localization method of [Kol06] we used only yields upper bounds, and sheds little light on lower bounds. Despite this shortcoming, as we have emphasized in the paper, the upper bound we derived in Theorem 4 is asymptotically tight in that it is consistent with the asymptotic behavior of the excess risk we derived in Theorem 3 and Corollary 1. --- *The paper will greatly benefit from a simulation result to support the main finding of the paper. For example, for a simple finite feature case. A theoretical example can also help.* One example we wanted to include of some practical relevance is the following. A popular statistical learning problem, which motivates the LASSO procedure, asks the following question: in a regression problem under the square loss, learn the best linear predictor from a $d$-dimensional feature map, knowing that the optimal linear predictor is $k$-sparse, i.e. the optimal weights have only $k$ non-zero entries for $k \ll d$. Our results allow us to tackle the following more general problem: under no assumptions on the optimal weight vector, learn the best linear predictor from a subset of size $k$ of the $d$ features. This problem fits in our framework neatly by taking $\mathcal{T}$ to be given by the set of all subsets of size $k$ of $[d]$ with $\|\mathcal{T}\| = {d \choose k}$, and our theorems imply, among other things, that the asymptotic quantiles of the excess risk of ERM on this problem are, up to a factor of two, the same as the oracle procedure that knows a priori the best subset of size $k$ of the $d$ features (assuming it is unique). --- References [Kol06]: Koltchinskii, Vladimir. "Local Rademacher complexities and oracle inequalities in risk minimization." --- Rebuttal 2: Comment: I thank the authors for their very thorough response. The only reason I'm not giving a higher score is that as I discussed in my review, the setting of the paper is very general (as the authors also point out in their paper) and I'm not 100% sure how much the results of this very general setting can shed light on what is happening in practical scenarios. However, this is a very solid paper, giving theoretically neat results for a very general setting, and should be accepted at NeurIPS.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper considers a linear regression problem where an ERM learner tries to learn a linear predictor and a feature map (over a countable set of feature maps) under some specific assumptions on the set of feature maps and the property of minimizers. The paper analyzes both asymptotic and non-asymptotic behaviors of the excess risk in the problem above and suggests that it matches the rate of the case when we only learn linear predictors on top of a fixed feature map, which is surprising. Strengths: The paper is well-written and easy to follow. The motivation of this paper is. The key results and directly related results in prior works, along with their intuition behinds, are explained very clearly and concisely. There are a few potential typos (see Weaknesses) but it should not be a problem. The proofs look good to me, though I did not check very carefully. Overall, I enjoy reading the paper. The only reason why I do not give a higher score is that the setting is too niche (see Weaknesses), which makes the results not so ground-breaking. However, it is still a good paper, and I advocate to accept it. Weaknesses: The main concern about this paper. 1. The setting is narrow: when I think about feature learning, I imagine we first learn a feature map that gives a good representation of the data. We then want to use the learned feature map that adapts it to a downstream task, which is a linear regression with squared loss in this case. The setting this paper proposed is somewhat the way around: (1) Given a feature map, learn the best linear predictor, (2) Select the best feature map. It seems to me that this paper is considering a non-linear regression problem, not feature learning. More precisely, the paper is trying to solve the problem learn the best feature map specified for the linear regression task. What can we tell about the learned feature map on other downstream tasks? Why the title of the paper is "On the Efficiency of ERM in Feature Learning", instead of "On the Efficiency of ERM in Non-linear Regression"? 2. If we look at the picture from that perspective, prior results (Theorem 1, Theorem 2) seem good enough for me. Given a learned feature map, under some conditions, I have a fast rate of learning a linear predictor. Of course, it is not feature learning at all, but it tells something about the learned feature map on a specific downstream task (linear regression), which is fair enough. There is no need (at least for me) for a result explaining if I learn a linear predictor with squared loss along with a feature map specifically designed for linear regression, what the sample complexity is. Comments on the Conclusion. 1. The claims on potential explanations for generalization in DNNs: To the best of my knowledge, the reason behind that should be explained by the geometry of the loss landscape of over-parameterized models and the implicit bias of (stochastic) optimization algorithms used. I would not go into detail since it goes beyond the scope of this paper. However, this paper: (1) considers an optimization oracle for ERM, and (2) does not assume any geometry of the set of feature maps indexed by $\mathcal{T}$. Therefore, linking the results in this paper to generalization in DNNs seems unnecessary and inappropriate for me. Minor comments 1. In the Appendix, it might be helpful if the authors first give a proof sketch for each result for readability. Minor typos: 1. Line 164, a comma missing after the inequality. 2. Line 501, should the RHS be $\frac{1}{2}\|\|\nabla{R(w^*)}\|\|_{\Sigma^{-1}}$? 3. Multiple commas, dots after (in)equalities missing in the Proof of Theorem 1, 2. After all, it might be too harsh to undervalue this paper based on the points above. I still think it is a good paper, and the results do not have to be connected to Feature Learning and Generalization to be meaningful. Technical Quality: 3 Clarity: 4 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive evaluation and detailed feedback. We will make sure to fix the typos for the final version of the paper. We address the reviewer's main concern below. --- *The setting is narrow: when I think about feature learning, I imagine we first learn a feature map that gives a good representation of the data. We then want to use the learned feature map that adapts it to a downstream task, which is a linear regression with squared loss in this case. The setting this paper proposed is somewhat the way around: (1) Given a feature map, learn the best linear predictor, (2) Select the best feature map. It seems to me that this paper is considering a non-linear regression problem, not feature learning. More precisely, the paper is trying to solve the problem learn the best feature map specified for the linear regression task. What can we tell about the learned feature map on other downstream tasks? Why the title of the paper is "On the Efficiency of ERM in Feature Learning", instead of "On the Efficiency of ERM in Non-linear Regression"?* Please note that our usage of the term "feature learning" in our setting is aligned with recent theoretical work, e.g. [Ba+22, Dam+22, Fre+23]. In plain terms, every predictor in the model classes we consider is jointly determined by a choice of feature map and a linear predictor on top of it, and as such, every learning method that operates on such model classes learns a feature map, i.e. it is performing feature learning. We agree that other methods exist to learn feature maps outside of the supervised setting we consider, and under different data access models such as transfer learning or self-supervised learning. Nevertheless, this does not exclude our setting, which captures arguably the simplest instantiation of the feature learning idea. Finally, we agree that it would be nice to obtain performance guarantees within our newly proposed framework in settings other than the supervised learning setup we consider, and we are excited to see what can be done with the new abstractions we introduce in our work. However, given that the framework we propose is new, and studying the performance of ERM on the function classes we introduce already requires the development of new ideas, we leave this to future work. --- *If we look at the picture from that perspective, prior results (Theorem 1, Theorem 2) seem good enough for me. Given a learned feature map, under some conditions, I have a fast rate of learning a linear predictor. Of course, it is not feature learning at all, but it tells something about the learned feature map on a specific downstream task (linear regression), which is fair enough. There is no need (at least for me) for a result explaining if I learn a linear predictor with squared loss along with a feature map specifically designed for linear regression, what the sample complexity is.* We argue that in the scenario described by the reviewer, one does not know that the learned feature map is good. This feature map is itself learned from data through some procedure, which incurs an estimation error. In the simplest case (our case), through ERM on data from the given task, but more generally, through ERM on data from another task in transfer learning as an example, or through contrastive learning in self-supervised learning as another example. Quantifying the estimation error of this learned feature map is an important problem, and part of our work is dedicated to answering this question in our setting (e.g. first statements in Theorems 3 and 4). --- *The claims on potential explanations for generalization in DNNs: To the best of my knowledge, the reason behind that should be explained by the geometry of the loss landscape of over-parameterized models and the implicit bias of (stochastic) optimization algorithms used. I would not go into detail since it goes beyond the scope of this paper. However, this paper: (1) considers an optimization oracle for ERM, and (2) does not assume any geometry of the set of feature maps indexed by $\mathcal{T}$. Therefore, linking the results in this paper to generalization in DNNs seems unnecessary and inappropriate for me.* We agree with the reviewer that we did not establish a formal link between our results and DNNs; see e.g. lines 336-337 "Formally connecting our statements to these experiments is beyond what we achieved here, yet, we believe that the new perspective we took might generate useful insights in this area." We pointed out in a paragraph in the conclusion that our results show a disconnect between the complexity of the function classes we consider (which share the ability to select a feature map in a data-dependent way with DNNs), and the excess risk of ERM on them, which is one of the main surprising empirical findings in the experiments of [Zha+21]. To the best of our knowledge, there is currently no consensus or overwhelming evidence for any of the existing explanations of the results of these experiments, and the paragraph we included is meant as an invitation to the reader to look at the problem from the lense of the new framework we introduced. If the reviewer believes our wording is misleading, we are open to suggestions on how to make our message clearer. --- References [Ba+22]: Ba, Jimmy, Murat A. Erdogdu, Taiji Suzuki, Zhichao Wang, Denny Wu, and Greg Yang. "High-dimensional asymptotics of feature learning: How one gradient step improves the representation." [Dam+22]: Damian, Alexandru, Jason Lee, and Mahdi Soltanolkotabi. "Neural networks can learn representations with gradient descent." [Fre+23]: Frei, Spencer, Niladri S. Chatterji, and Peter L. Bartlett. "Random feature amplification: Feature learning and generalization in neural networks." [Zha+21]: Zhang, Chiyuan, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. "Understanding deep learning (still) requires rethinking generalization." --- Rebuttal 2: Title: Reply to the rebuttal Comment: I thank the authors for the detailed feedback. I am still not convinced with the "feature learning" setting and for me, it is more of a non-linear regression setting. I am also aware of the works that the authors refer to, but to be honest, I do not really like those works and their settings. However, I know that I am being biased, and it also does not affect the contributions of this paper. As for the conclusion, it would be nice if the authors spent more time discussing the linking between the results and generalization in DNNs in multiple views: (1) how it (potentially) explains generalization (2) the drawback of current settings (and assumptions) and how it conflicts with other potential explanations of generalization in DNNs. Overall, I still find it a good paper, and I will keep my origin evaluation. --- Rebuttal Comment 2.1: Comment: We thank the reviewer for their feedback and appreciate their comments. We will make use of the additional space in the final version of the paper to address the two points raised by the reviewer.	null	null	null	null	null	null
Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts	Accept (spotlight)	Summary: This paper investigates an interesting task (Text2CAD) in CAD automated applications, which achieves generating parametric CAD models from text prompts. Specifically, the authors first introduce a data annotation pipeline to generate suitable text prompts for CAD models in public DeepCAD dataset (including about 170K CAD models) via conducting LLM models such as Mistral and LLaVA-NeXT. The generated text prompts mainly consist of abstract CAD descriptions and detailed specifications. Second, the authors further design a transformer-based auto-regressive network to learn the gap between generated text prompts and CAD models. Finally, the effectiveness of proposed method is validated under a mixture of metrics to show its potential in AI-aided CAD design. Strengths: S1: The paper is well organized and easy to understand. S2: The visualization of data annotation pipeline is clear to view. S3: The structure of the paper is clear and the logic is rigorous. Weaknesses: W1: Some typos need to be corrected (“Computer-Aided Design (CAD) play”). W2: Insufficient quantitative experiments are conducted, which currently appears to be less impactful. W3: The figures in the manuscript could be enhanced. For example, it would be better to use more colors to show the generated CAD model. W4: Lack the training computational costs and the inference time. Technical Quality: 2 Clarity: 3 Questions for Authors: Q1: Given the design of network architecture, it seems that both text prompts and CAD sequences need to be used as inputs for forwarding. But in the inference stage, text prompts as only inputs for generating CAD models according to Figure 5. The reviewer is curious about is this just not showing the CAD tokens in Figure 5 or CAD tokens is completely unused during the inference stage? This is crucial for text2CAD generation problem, as it only makes sense when only text prompt is used as input. Q2: For those numerical numbers within CAD commands, it is very difficult to encode accurately, which is still an open problem in the NLP. Would author kindly further give more details about how to encode them and maintain their validity after decoding? This is important to demonstrate the proposed framework is able to generate CAD models precisely from textual prompts. Q3: The reviewer puts other questions in Weaknesses and Limitations. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: L1: This paper does not discuss the diversity and bias of the generated textual prompts, is it possible to generate the same/different CAD models using similar/different style of textual prompts? Discussing this issue would be better to show how the framework performs. L2: Given the experiments are only conducted on DeepCAD dataset, it would be better to show the extend results on other datasets to demonstrate its generalized capability. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's careful consideration of our work and the thoughtful responses. Below are our responses to the specific points you raised. # Response to Weaknesses (W1) Reviewer: Some typos need to be corrected ... We have corrected the typo. (W2) Reviewer: Insufficient quantitative experiments are conducted, which currently appears to be less impactful. We acknowledge that this is a limitation in the text-to-3D domain and remains an area of active research. Given the lack of standardized benchmarks in text-to-3D domain that can be directly applied to text-to-CAD tasks, we assess the quality of generated CAD sequences against ground truth, using a multi-faceted evaluation approach. This includes Sequence-Level, GPT-4 [1], and Human evaluations to provide a comprehensive assessment of our model's performance. [1] Wu et al. GPT-4V(ision) is a Human-Aligned Evaluator for Text-to-3D Generation, CVPR 2024 (W3) Reviewer: The figures in the manuscript could be enhanced... In the final version, we will update the colors for better visibility. (W4) Reviewer: Lack the training computational costs and the inference time. 1. Training Computation Cost: - GPU: 1 A100-80GB Nvidia GPU - Training Time: 2 days. - Trainable Parameters: ~23M 2. Average Inference Time per Sample: ~ 0.3s (RTX 4090) # Response to Questions (Q1) Reviewer: Given the design of network architecture, it seems that both text prompts and CAD sequences need to be used as inputs for forwarding... Training: Yes, we train the model using both the CAD sequence and text prompt. We use a teacher forcing strategy which is a common practice to train an autoregressive network as it converges faster. Since our model architecture is autoregressive, it learns to predict the next token in the CAD sequence given the textual contexts. Inference: However, during inference, our model only requires text prompts as input. Since the architecture is autoregressive, the model doesn't generate the complete CAD sequence in one forward pass. For example, the model takes as input: $(C_{1:1}, T_{adapt})$, where $C_{1:1}$ is the start token and $T_{adapt}$ is the adapted BERT encoding of the text prompts. The model then outputs: $(C_{2:2})$, where $C_{2:2}$ is the second token of the output sequence. In the next iteration, it takes as input: $(C_{1:2}, T_{adapt})$, and outputs: $(C_{3:3})$, the third token of the output sequence. This process continues until the end of sequence token is generated. Hence, during inference only the start token and the user's text prompt are required to generate the complete CAD sequence. This complete CAD sequence can then be rendered into a 3D CAD model (which is shown in Figure 5). We concur with the reviewer that it only makes sense when only text prompts used as input. Accordingly, our model generates the final CAD sequence exclusively from text prompts, using a text-conditioned autoregressive approach. (Q2) Reviewer: For those numerical numbers within CAD commands, it is very difficult to encode accurately, which is still an open problem in the NLP... We acknowledge the reviewer's point about the challenges in encoding numerical values in NLP. To address this, our approach avoids direct arithmetic operations on numbers extracted from text. Instead, we discretize the numerical values in our ground truth CAD sequences. Our model is then trained to predict these discrete values directly from the text prompts (e.g., "$0.0208$" → $142$). As all text prompts are encoded using the BERT encoder, it is possible that BERT may encode some of the numerical values (e.g., "0.0208", "-0.1512") into the UNK token. To mitigate this issue, we use the Adaptive layer and downsampling layer. These components fine-tune the text features of BERT to better align with continuous values, specific vocabulary, and structural requirements of CAD instructions. The Layerwise Cross-Attention mechanism in the decoder learns the mapping between quantized and continuous values. The effectiveness of our approach is conclusively demonstrated in Table 1. Our model achieves low Chamfer Distance (CD) for expert-level prompts (L3), which contain the highest number of continuous values and complex geometric descriptions. The lower CD indicate that the reconstructed CAD model is very similar to the ground truth CAD model. This is further supported by our ablation study, where directly using BERT encoding without our specialized layers results in higher CD. These results indicate that our model successfully encodes numerical values and maintains their validity after decoding. # Response to Limitations (L1) Reviewer: This paper does not discuss the diversity and bias of the generated textual prompts, is it possible to generate the same/different CAD models using similar/different ... prompts? The diversity of our generated textual prompts is influenced by dataset variety and the performance of Mistral and LLaVA-Next. We note in Section 6 of the main paper that the DeepCAD dataset, the only large-scale dataset with a full design history, predominantly features simpler rectangular, and cylindrical shapes. To enhance prompt diversity, we've shifted focus from generating mere object names to detailed shape descriptions using VLM. For instance, a ring might be described as a circular object with a cylindrical hole in the center. This approach allows for the generation of identical CAD models using varied textual styles. In the attached pdf, Figure 3 shows examples of the same CAD models being produced from diverse text prompts. (L2) Reviewer: Given the experiments are only conducted on DeepCAD dataset... To showcase our cross-dataset performance, we will conduct experiments on the Fusion360 [1] dataset where both DeepCAD and Text2CAD are trained on the DeepCAD dataset. We will add it to the main matter in the final version. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' feedback, which addresses most of my concerns. Given the text2CAD is indeed an interesting problem rarely covered by previous efforts, the reviewer would raise the score up to Weak Accept. --- Rebuttal 2: Comment: Thank you for reconsidering our work and raising the score to Weak Accept. We're happy to address any further questions you may have.	Summary: The paper proposes Text2CAD, a framework designed to expedite the prototyping of complex computer-aided design (CAD) models. The proposes method uses designer-friendly text instructions to generate parametric CAD models, making it accessible for all skill levels. To facilitate this, a data annotation pipeline is introduced, which generates text prompts based on natural language instructions for the DeepCAD dataset using Mistral and LLaVA-NeXT. The Text2CAD employs an end-to-end transformer-based autoregressive network to generate parametric CAD models from input texts. Authors have evaluated the performance of their method through a combination of metrics. s. Strengths: - Authors have done nice work to simplify the annotations in natural language - The use of VLM and LLM in generating prompts for CAD images is an interesting direction. - The results with the proposed method are decent. Weaknesses: - The authors assert that the proposed method Text2CAD is the first AI framework for CAD image generation. However Text2CAD is not the first AI method for generating CAD images. A previous work SketchGen [1] exists which was the first to generate CAD images. Therefore, authors should include this work and enlist how different is their work compared to SketchGen. Surprisingly SketchGen also generates CAD sketches auto-regressively, which makes the novelty of proposed approach weaker. - While there are 4 levels of language prompts used in the method. However, there is no separate analysis on how each type of prompt is helping the final generation. There is an ablation on alignment module, however, there is no analysis on the use of 4 prompts. This is important to validate the role of each prompt in aiding the generating process. I am just wondering why we need 4 levels of prompt abstractions. In general, the baselines used to compare the results are limited. - While the layerwise cross attention is not a novel component in the proposed method, did the authors try to explore alternative methods for information fusion? Authors are encouraged to provide solid explanations to the above queries. References: [1] Wamiq Para et al. "Sketchgen: Generating constrained cad sketches". NeurIPS 2021 Technical Quality: 3 Clarity: 3 Questions for Authors: Please check the weaknesses section for the questions. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Authors have provided the limitations statement. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's careful consideration of our work and the thoughtful responses. Below are our responses to the specific points you raised. # Response to Weaknesses and Questions (W1) Reviewer: The authors assert that the proposed method Text2CAD is the first AI framework for CAD image generation. However...SketchGen also generates CAD sketches auto-regressively... We appreciate the reviewer's comments and welcome the opportunity to clarify the fundamental contributions of our method. Our approach, Text2CAD produces parameterized 3D CAD models directly from textual descriptions where SketchGen[1] only generates 2D sketches. This capability supports our claim as pioneers in the field of parametric CAD model generation from text. Additionally, the 3D CAD models created by our method can be further edited using commercial CAD software. Below, we outline the key distinctions and novel contributions of Text2CAD compared to SketchGen[1]. 1. *Scope and Focus - Generative vs Modality-Translation*** - SketchGen is a generative network focused on generating 2D-constrained CAD sketches. It tackles a generative-modeling problem for 2D sketches. - Text2CAD is the first AI framework to generate parametric 3D CAD models in a sketch-extrude format conditioned upon textual descriptions. It tackles a modality-translation problem. 2. *Input Modality - Unconditional vs text* - SketchGen generates 2D sketches from scratch (unconditional generation) due to its generative nature, but its real-world applicability is limited due to the lack of user-guided sketch generation. - Text2CAD leverages text prompts to guide the generation of 3D CAD models, enabling more practical applications in CAD design workflows. 3. Output Modality - 2D vs 3D** - SketchGen produces only 2D sketches with constraints. Since extrusion parameters are not predicted, the final models are not actual 3D CAD models. These can be only visualized as an image. - Text2CAD generates actual 3D CAD models in sketch-extrude format which can be visualized as boundary representations (Brep) or meshes and be further edited in CAD software. 4. Novelty of Text2CAD - Text2CAD's novelty lies in its ability to interpret text descriptions and generate parametric 3D CAD models, which involves a different set of challenges compared to generating 2D-constrained sketches from scratch. While both works utilize autoregressive generation, their scope, input, output, and applications are fundamentally different. We appreciate the reviewer's suggestion to include this work. We already included SketchGen[1] in Line 120 (Section 2 Related work) of the main paper and we will update the explanation on how Text2CAD differs in the final version. We are happy to discuss more and clarify any further questions on this matter. (W2.1) Reviewer: While there are 4 levels of language prompts used in the method. However...why we need 4 levels of prompt abstractions. Below, we explain the role and importance of each of the four levels of prompts within our framework. 1. Single Input Prompt for Training and Inference: Figure 3 in the main paper may imply that all four prompts are used simultaneously to generate the final CAD model. However, our model uses only a single text prompt of any level as input during both training and inference. In Figure 3, the four prompts are only for visualization purposes. In reality, these are four different training instances. Each training instance uses one prompt level at a time, resulting in ~600k training samples, as mentioned in Line 250-252 (Section 5) of the main paper. We will update the Figure 3 caption in the final version to clarify this. 2. Purpose of Four Prompt Levels: The four levels of language prompts—abstract, beginner, intermediate, and expert—were developed during our data annotation process to accommodate users of all skill levels. Unlike the text prompts in existing text-to-3D methods, which are typically more object-centric, text-to-CAD methods must handle a range of prompts, from simple object-centric descriptions to detailed parametric instructions (Line 48-53 of the main paper). 3. Analysis of prompts: Since each text prompt, regardless of its level, can independently generate the final CAD model, there is no context of helping the final generation. However, in Table 2 of the main paper, we have provided the performance of our model on each level separately using both GPT-4V and Human evaluation. We hope this clarifies the reviewer's question. In summary, our model can generate a 3D CAD model from an abstract shape description (e.g, generate a star) or a detailed parametric one. It doesn't require all four level prompts. We are happy to discuss more on this matter. (W2.2) Reviewer: In general, the baselines used to compare the results are limited. We acknowledge the limited baselines in our paper. We agree with Reviewer ovUL's understanding of the challenges in establishing baselines - Since this is the first work for large-scale text2cad generation, there wasn’t too many baseline to compare to, this is understandable and should not be taken as a major weakness (ovUL). (W3) Reviewer: While the layerwise cross attention is not a novel component in the proposed method, did the authors try to explore alternative methods for information fusion? We acknowledge that layerwise cross-attention mechanism is a well-established technique. In our current work, we have primarily focused on demonstrating the feasibility and effectiveness of the Text2CAD transformer using this approach. However, we recognize the potential for improvement in this area. In future research, we plan to explore alternative fusion mechanisms more specifically suitable for text-to-CAD domain for integrating textual and geometric information. [1] Wamiq Para et al. "Sketchgen: Generating constrained cad sketches". NeurIPS 2021 --- Rebuttal 2: Comment: Thanks to authors for a detailed explanation of my queries. I am satisfied with the answers. However it would be interesting to see how authors would take into consideration the last point of fusion in the upcoming version. After a comprehensive analysis of the responses from authors, I decide to increase my rating to weak Accept. --- Rebuttal 3: Comment: Thank you for reconsidering your rating of our work to Weak Accept based on our responses. We greatly appreciate your thorough review and feedback, especially regarding alternative fusion mechanisms. This is a core part of the architecture we want to improve in our future work. We're happy to address any further questions you may have. --- Rebuttal 4: Comment: Dear Reviewer HTpR, Thank you for your thoughtful feedback on our paper. In your comment, you mentioned increasing your rating to "weak Accept". However, we noticed that the rating remains unchanged (5) in the system. We're wondering if this might be a technical issue. Could you kindly check if the score in the system accurately reflects your intended rating? We appreciate your time and attention to this matter. Regards, Authors	Summary: This paper proposes Text2CAD, the first approach for generating parametric CAD models from different levels of designer-friendly language instructions. The critical challenge of generating CAD models from text is the lack of high-quality paired data. The paper's main contribution lies in a well-designed LLM-assisted data annotation pipeline, which augments the existing DeepCAD dataset with rich text descriptions at four skill levels. Based on the paired text-CAD data, the paper proposes an autoregressive transformer to generate CAD construction sequences from text inputs. Text2CAD outperforms an adapted DeepCAD baseline and produces promising qualitative results. Strengths: 1. Generating CAD models from different levels of textual description is an interesting direction. Existing CAD generation methods have limited flexibility on the condition format, and this paper makes an exciting attempt that can bring practical impact. 2. The data annotation pipeline is creative and reasonable. The authors leverage LLaVa and Mistral with well-designed prompts to generate different levels of text descriptions for DeepCAD data. The data quality looks good in the visualized examples. 3. The Text2CAD transformer is properly designed to generate CAD construction sequence in an autoregressive manner conditioned on the textual descriptions. 4. A baseline is built up by adapting DeepCAD and training with the curated Text2CAD data. The proposed Text2CAD transformer model consistently outperforms DeepCAD across quantitative and qualitative comparisons. Weaknesses: 1. The main paper and the appendix provide only a small number of qualitative examples. More examples should be provided to better demonstrate Text2CAD's generation capacity. 2. Some typical failure examples should be presented and discussed, which could provide more insights for future works. Technical Quality: 3 Clarity: 3 Questions for Authors: In general, this paper presents an interesting and promising attempt to generate CAD models from language descriptions at various skill levels. I don't have major concerns about the data curation process and model design. For experiments, more qualitative results (including failure cases) should be included. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The potential social impact is discussed in Section 1 of the paper, and the limitations are properly discussed in Section 6 of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's careful consideration of our work and the thoughtful responses. Below are our responses to the specific points you raised. # Response to Weaknesses and Questions (W1) Reviewer: The main paper and the appendix provide only a small number of qualitative examples. More examples should be provided to better demonstrate Text2CAD's generation capacity. We will provide more qualitative samples in the appendix of the final version. We have added some qualitative samples in Figure 1 in the attached pdf. (W2) Reviewer: Some typical failure examples should be presented and discussed, which could provide more insights for future works. We have identified two types of failure cases for our model: 1. Invalidity: In this scenario, the model fails to generate any CAD model from the text prompts. As reported in Table 1 of the main paper, this occurs in approximately 1% of the test samples. In these cases, the model predicts invalid sketch or extrusion parameters, such as the same start and end points for lines or arcs, or zero values for extrusion depth on both sides. 2. Discrepancy: Discrepancy refers to situations where the generated model does not precisely match the shape described in the text prompts. This is relatively more prevalent in our model than invalidity and is harder to quantify. We notice that this occurs when prompts are more focused on rare object names (e.g., spatula, paddle) in the dataset rather than parametric descriptions. To showcase this, we have provided some samples in Figure 2 of the attached pdf. We will add these samples in the appendix of our final version. For future work, we plan to address these challenges through: - Developing more robust parameter prediction methods by imposing syntax-based criterion during training phase to reduce invalidity cases. - Generate more text prompts using interpolated prompt generation (The method is shown in Section 11 and Figure 9 and 10 - Appendix) for rare objects. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed discussion of the failure case. It would be good to include a brief version of the failure case figure and the analyses in the main paper. After reading other reviews and the corresponding author responses, I don't have further questions and will keep the accept rating. --- Rebuttal 2: Comment: Thank you for your Accept rating and valuable suggestions. We'll include a brief analysis of the failure cases in the main paper as the space allows. We are glad that our responses have addressed all your concerns. --- Rebuttal Comment 2.1: Comment: Thank you for the detailed rebuttal. Most of my concerns are addressed. I will maintain my score of accept.	Summary: This paper proposes the first text2cad framework. This includes a CAD model annotation pipeline using VLM + LLM. And an autoregressive Transformer for generating sketch and extrude. Authors extended DeepCAD dataset with different level of text description using their method and demonstrate promising results for text-based CAD generation. Strengths: Most CAD dataset like DeepCAD and ABC do not contain text description. The challenge lies in annotation CAD mechanical parts by non-expert user. This grealy hinders the advancement of text2cad. The method proposed in this paper use the advantage of VLM and LLM to automatically annotate different level of expert textural description for DeepCAD. This is a big improvement and solves a challenging task. Authors further demonstrate promising text2cad results by training an autoregressive Transformer to generate the sketch and extrude cad parameters via cross-attention. Overall I think the motivation is strong, the annotation method is novel and beneficial for the whole community, and the initial results are very promising. Weaknesses: Since this is the first work for large-scale text2cad generation, there wasn’t too many baseline to compare to, this is understandable and should not be taken as a major weakness. On the other land, I think the paper could possibly improve by providing an analysis of the annotation quality. A simple test could be given one model and several text annotations, can gpt-4v selects the correct annotation corresponding to that model. This would test if the annotated text contains enough critical information for identifying the correct model. Technical Quality: 3 Clarity: 3 Questions for Authors: What is the quality of the annotated text? Can authors provide some way to evaluate it? Can the cross-attention model used in the paper be modified for classifier-free training? Could the annotation be very different for the same model? How would the author address this discrepency? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's careful consideration of our work and the thoughtful responses. Below are our responses to the specific points you raised. # Response to Weaknesses and Questions (Q1) Reviewer: What is the quality of the annotated text? Can authors provide some way to evaluate it? We thank the reviewer for suggesting the use of GPT-4V to assess the quality of our annotations. Below, we describe the experimental setup and discuss the results. Experimental Setup: We provided GPT-4V with four text prompts—three incorrect and one correct—along with the corresponding minimal JSON and multi-view stacked images of the CAD model. We then asked GPT-4V to select the text prompt that best matches the minimal JSON, utilizing both its understanding of the minimal JSON and the multi-view stacked images. Experimental Results: We conducted this experiment using Expert level (L3) text prompts, which contain the highest level of parametric details and numerical values, ensuring a one-to-one correspondence between the CAD design history and the text prompts. The matching accuracy was 99.72%, indicating that the annotated texts at the Expert level (L3) provide sufficient critical information for precise 3D CAD model reconstruction. However, we found that the GPT-4V results for Abstract (L0), Beginner (L1), and Intermediate (L2) level prompts are unreliable. This unreliability arises because these prompts are designed to progressively lose parametric information to better suit different user levels. Consequently, the same text prompt can lead to different CAD models (one-to-many), or different text prompts can lead to the same CAD model (many-to-one). This variability makes matching the correct text prompts to the minimal JSON or the multi-view images a one-to-many or many-to-one problem, which is inherently ill-posed even for humans. It is worth noting that the merit of our proposed data pipeline remains significant as it is model-agnostic. Our pipeline leverages open-sourced LLM and VLM, and during the course of this work, we utilized the best available models, namely LLaVA-NeXT and Mistral-7x8B. However, the landscape of available open-source models has evolved rapidly, with some achieving GPT-4 level performance. Therefore, replacing the current models with the latest LLM and VLM can provide annotation quality approaching the GPT-4 level. We consider this an important extension for future work. (Q2) Reviewer: Can the cross-attention model used in the paper be modified for classifier-free training? Currently, we do not have a definitive answer to this question as we have not yet explored modifications of the cross-attention model for classifier-free training. However, we recognize this as a potential direction for future research. We would be glad to discuss this further. (Q3) Reviewer: Could the annotation be very different for the same model? How would the author address this discrepancy? Our data annotation pipeline includes different annotations (because of top-p sampling from VLM and LLM) for identical CAD models to enhance data diversity and capture the multiple ways users might describe these identical models. This diversity in the training samples helps our transformer model to learn to generate similar CAD models from different types of prompts. For example, in our dataset a CAD model of a pipe can be described as both "a pipe" and "a thin cylindrical shape with a hole." These variations reflect the real-world scenario where different users describe the same object differently based on context or personal preference. In the attached pdf, Figure 3 shows two examples of the same CAD models being produced from three different text prompts.	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for their thorough evaluation and insightful comments. We are delighted that the reviewers recognize the significance of our work. 1. Text2CAD is acknowledged as the first framework (ovUL) for generating parametric 3D CAD models from different levels of designer-friendly language instructions (hkcu). 2. Our work is seen as having high potential for practical impact in the CAD industry and design processes (hkcu, ovUL). 3. Our novel data annotation pipeline using VLMs and LLMs is praised for its creativity and effectiveness in generating high-quality paired text-CAD data (ovUL, hkcu, HTpR). This is a big improvement and solves a challenging task (ovUL). Also this is an interesting direction (HTpR). 4. The autoregressive Text2CAD transformer model is properly designed (hkcu) for CAD construction sequence generation from text prompts. The initial results are very promising (ovUL) and the model consistently outperforms DeepCAD across quantitative and qualitative comparisons. (hkcu) 5. The paper is well organized and easy to understand and the structure of the paper is clear and the logic is rigorous (mcF8). Our work makes a significant contribution to the field of AI-assisted CAD modeling and opens up a new research domain in text-to-CAD modeling. We appreciate the reviewers' constructive feedback and have addressed each point in the individual responses below. The comments and subsequent revisions have strengthened our paper considerably. We hope we have answered all the questions properly within the given time-frame and are open to further discussion if additional questions arise. Please find attached the pdf for the figures. Pdf: /pdf/312d83829f2f1ef0a6be7cbfb6d639b5d962c7fc.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis	Accept (poster)	Summary: This paper presents a hybrid program synthesis approach that strengthens the classic bottom-up search with an LLM-guided prior distribution, given as a probabilistic context-free grammar (PCFG). The focus of this paper is to solve complex PBE tasks where programs in unfamiliar DSL are difficult for LLMs to generate directly, and traditional combinatorial search is infeasible. The key insight is that LLM-predicted programs (although incorrect by themselves) can provide valuable intuition regarding which operator or component to use in the correct program. As such, HYSYNTH samples a set of full programs from LLM to train an approximation PCFG model. To perform the bottom-up synthesis, HYSYNTH uses an off-the-shelf or custom synthesizer for the specific domain. The experimental results show its improvement against both vanilla LLM generators and non-LLM-based synthesizers. Strengths: - It proposes a promising hybrid approach that combines the efficiency of formal program synthesis and the power of LLMs. - Compared to training new models for every new DSL/domain, the proposed approach only needs to extract a small PCFG from a few LLM samples, which is lightweight and robust. - The approach is general; in particular it can be applied to various domains (reasoning about grid puzzles, tensor manipulations, and string manipulations) and with different LLMs. It can outperform baselines substantially according to the experiments. - The problem is well-motivated, and the limitations and related work are discussed in an insightful and thorough manner. Weaknesses: - The datasets are small and domain specific, and requires implementing a synthesizer. It is unclear whether such an approach can scale to more complex programs, for example to program with loops, or even general-purpose programs. - It mentions in the limitation section that sampling from LLM is costly, and it uses different models like GPT3.5, GPT-4, and DeepSeek, but did not provide a comparison of the costs. Technical Quality: 4 Clarity: 4 Questions for Authors: - Using PCFG as the approximation model seems too simple at first glance, but it works well in the evaluated benchmark. It is also surprising to see that only 10 samples can often achieve comparable performance as the full 100-sample approach (Appendix C). This leads me to wonder whether the model only captures very superficial patterns from the LLM samples, and whether an even simpler surrogate model could be used. For example, It is mentioned in line 151-157 that GPT-4o predicts the relevant components with high accuracy, and never uses irrelevant components in the example. If we directly remove such irrelevant components from the CFG and directly use the search algorithm, how would such a baseline perform? - I am not entirely convinced by the superior performance of HYSYNTH over using LLMs purely without search. For example, for the ARC task, the HYSYNTH approach uses a divide-and-conquer strategy; it would be interesting to see how the LLM baseline performs when it also applies such a strategy. Also, in the paper ExeDec, Section 5.2, the experiments show that presenting DSL programs as Python functions can enable the LLM to better leverage its general knowledge from pretraining and improve the result. I wonder if such a prompting technique would improve the performance of pure LLMs. [1] Shi, Kensen, et al. "ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis." The Twelfth International Conference on Learning Representations. (ICLR 2024) Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The limitations are well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and thoughtful comments. Please note our top-level comment with additional experimental results. Below we address specific comments and questions. Q: The datasets are small and domain specific, and requires implementing a synthesizer. Can HySynth scale to more complex programs, for example to program with loops, or even general-purpose programs? It is true that we have not evaluated the approach on more complex programs with loops, but there is relevant work in the program synthesis community on extending bottom-up search to programs with local variables and loops [1,2]. [1] LooPy: interactive program synthesis with control structures https://dl.acm.org/doi/pdf/10.1145/3485530 [2] Efficient Bottom-Up Synthesis for Programs with Local Variables https://dl.acm.org/doi/pdf/10.1145/3632894 Q: It mentions in the limitation section that sampling from LLM is costly, and it uses different models like GPT3.5, GPT-4, and DeepSeek, but did not provide a comparison of the costs. The inference cost of the LLMs is dependent on the API pricing; with GPT3.5 and DeepSeek being cheaper than GPT-4o and the number of samples drawn per LLM are roughly the same. We are happy to include a cost analysis across LLMs in an updated version of the paper. Q: If we remove irrelevant components not present in LLM solutions from the CFG and use the search algorithm, how would such a baseline perform? This is an interesting experiment and we have now performed it on the String and Tensor domains. More specifically, we evaluated an ablation, where the surrogate model is a “binary PCFG”, i.e. a CFG that only includes components mentioned by the LLM. You can see the graph of the results in the pdf attached to the Global Response. It performs worse than HySynth because the search might exclude essential components from the grammar. This experiment further highlights the balance achieved by HySynth in terms of prioritization based on LLM guidance. Q: for the ARC task, HySynth uses a divide-and-conquer strategy; how would an LLM baseline that also applies such a strategy perform? If we understand correctly, the reviewer is suggesting to query the LLM for filters and transforms separately and then combine the results in some way. We argue, however, that such a technique would not count as an LLM baseline, but rather as a different hybrid algorithm (only slightly less advanced than HySynth). Like HySynth, it would combine LLMs with a semantics-based pruning technique originally proposed in the program synthesis community (in this case, divide-and-conquer synthesis [3]). [3] Scaling Enumerative Program Synthesis via Divide and Conquer https://www.cis.upenn.edu/~alur/Tacas17.pdf Q: Will presenting DSL programs as Python functions enable the LLM to better leverage its general knowledge from pretraining and improve the results? This is an interesting suggestion that could perhaps improve the performance of LLMs. However, we consider non trivial prompting and fine tuning techniques as being orthogonal to our contribution, which focuses on using existing LLM solutions more effectively. --- Rebuttal 2: Comment: Thanks for the responses and adding the new experiments – I'll keep my score and raise the confidence as the new "binary PCFG" baseline addresses my concern and highlights the effectiveness of design components.	Summary: The paper introduces a new approach for solving structured prediction and reasoning tasks by leveraging the programming and planning capabilities of LLMs to enhance bottom-up search in program synthesis. Specifically, HYSYNTH initially generates preliminary solutions directly from the LLM. Since direct LLM sampling performs poorly in unfamiliar DSLs, HYSYNTH does not use these samples directly but calculates the probabilities for a PCFG grammar from them, incorporating LLM guidance into the bottom-up search (CFG -> PCFG). Experiments show that HYSYNTH consistently outperforms both the baseline synthesizers and ablations, and is effective even with a smaller number of samples. Strengths: The HYSYNTH method can be implemented directly on existing LLMs without additional model fine-tuning, allowing its performance to improve as the models advance. HYSYNTH performs well with fewer samples, consistently surpassing current methods, making it robust and efficient. Experiments demonstrate that HYSYNTH consistently outperforms both baseline synthesizers and ablations across various configurations. Weaknesses: The use of LLMs in HYSYNTH appears simplistic and seems not to fully leverage the capabilities of LLMs. HYSYNTH's performance depends on the LLM's capabilities, including additional inference time and prediction accuracy. Technical Quality: 3 Clarity: 3 Questions for Authors: The experiments show that HYSYNTH solves more problems within the same time limit. Does the time reported include the time taken for sampling from the LLM and training the surrogate model? What percentage of the total time does LLM inference take? For GPT-4o's 0% performance on the percentage of syntactically valid completions in the STRING domain, can a more detailed analysis be provided? Could this be improved with a simple CoT strategy or other methods? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and thoughtful comments. Please note our top-level comment with additional experimental results. Q: The experiments show that HySynth solves more problems within the same time limit. Does the time reported include the time taken for sampling from the LLM and training the surrogate model? What percentage of the total time does LLM inference take? See Global Response Q1. Q: For GPT-4o's 0% performance on the percentage of syntactically valid completions in the STRING domain, can a more detailed analysis be provided? Could this be improved with a simple CoT strategy or other methods? See Global Response Q2.	Summary: This paper introduces HySynth, an approach to program synthesis that combines (1) sampling programs from a large language model (LLM) to train a probabilistic context free grammar (PCFG) with (2) applying bottom-up enumerative search guided by the PCFG to solve programming by example tasks. It applies this approach to three domains: the abstract reasoning corpus (ARC), tensorflow expression synthesis, and string expressions (SyGuS). By learning the weights of the PCFG from the samples drawn from the LLM, HySynth produces state of the art results on all three domains, solving more problems in less wall-clock time. The main contribution is the use of an LLM to choose weights to guide a bottom-up enumerative program synthesis search system. Strengths: The paper is the first (concurrent with Li et al [29]) to use pretrained LLMs to guide search-based program synthesis. The approach introduced by HySynth -- to sample programs from a pretrained LLM to learn the weights for a PCFG, and then to guide bottom-up enumerative search according to these weights -- is a natural and simple idea, and the paper demonstrates that it works well across a variety of domains. The simplicity of the approach is a strength, making it possible to apply to a range of domains -- wherever a CFG can describe the DSL used in the domain. The simplicity and generality of the approach make the contribution significant: it is quite likely that additional methods that learn to perform a fast search guided by an LLM will follow in future research, as program synthesis with LLMs is an important and growing area. The experimental results are robust, showcasing the HySynth method across a good diversity of domains and yielding clear improvements in each. (See also Weakness 1 however.) The paper is clearly written, including a robust and clear background section, a clear statement of the method and experimental results, and well contextualized in the context of the literature. Weaknesses: For the direct sampling baseline, there is either a clarity or methodological issue. It is unclear how many samples are drawn from the LLM in the direct sampling approach, and if the number of samples is more than 1, it is not clear what is done to produce diversity (i.e. setting the temperature during sampling, or sampling without replacement e.g. with unique randomizer [1]). From the horizontal lines in Figure 4a, 4b, and 4c, it seems overwhelmingly likely that only a single sample is drawn from GPT4o for the direct sampling baseline, which places this baseline at a disadvantage since >>1 samples can be drawn within 10 minutes. The disadvantage is compounded because direct sampling is trivially parallelizable, so the number of samples that could be drawn in this time is large. And as one further point in favor of drawing multiple samples for the direct sampling baseline, the HySynth approach itself uses as many as N = 100 samples from the LLM (line 273). I recognize this may be a costly baseline to run (I have not computed the cost), and suggest a reduced time limit if the cost is prohibitive for the 10m time limit. [1] Incremental Sampling Without Replacement for Sequence Models https://arxiv.org/pdf/2002.09067 The other weaknesses I observed are touched upon in the limitations section of the paper. The main limitation of the significance of the work is (as stated in the paper) that is requires a DSL, and particularly one that has a CFG, and so a custom synthesizer for each DSL, for the method to be applied. This requires meaningful work to apply HySynth to a new domain. Another concern (also addressed in the limitations section) is the possibility of data leakage in the evaluations. I share the authors' view that this is not a major issue _provided_ they address the first weakness I state above for the direct sampling baseline. Technical Quality: 3 Clarity: 3 Questions for Authors: Is the time to draw N samples from the LLM and construct the PCFG included in the plots in Figure 4a, 4b, 4c? If not, I would encourage doing so. Why do you include few-shot examples when prompting for the tensor domain, but not for the string domain (or have I misread this?)? It seems more important for the string domain given that 0% of the LLM-generated programs in the string domain are valid completions. nit: Figure 2 suggests the I/O examples are provided as an input to the synthesizer, rather than just the PCFG. I don't believe this is correct or intended. Is it correct to state that the PCFG is reasonably approximating the LLM's condition distribution over output programs for a given prompt (line 65)? Though it is trained to approximate this, I expect it is quite a poor approximation. And then the approach does not sample from the PCFG, but rather enumerates from it, making it unclear to me that a good unbiased approximation of the LLM's conditional distribution is what is desired (certainly if the PCFG could perfectly mimic this distribution that would be great!, but perhaps there is better; beyond wanting to somewhat approximate the LLM's condition distribution, increased functional diversity over the 10m search period is also important.) Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations section (4.3) captures the main limitations of the system well and articulates them clearly. The Broader Research Impacts appendix (K) touches on the research, albeit not societal, impact of the work. This seems sufficient to me. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and thoughtful comments. Please note our top-level comment with additional experimental results. Below we address specific comments and questions. Q: For the direct sampling baseline, clarification about how many samples are drawn from the LLM in the direct sampling approach, and if the number of samples is more than 1, what is done to produce diversity? There has been a misunderstanding: the results of the direct sampling approach are based on 100 samples (see line 251), the same number that is used to learn the PCFG. We also report the temperature and other prompting settings in appendix J – we sample with a high temperature to get diverse solutions. Please see Global Response Q1 for why we decided to separate LLM sampling time from symbolic synthesis time. Q: The main limitation of the work is that is requires a DSL, and particularly one that has a CFG, and so a custom synthesizer for each DSL, for the method to be applied. Can this be overcome? Our approach does require a CFG for the DSL but a custom algorithm is optional. In this paper, for the String domain we use a generic synthesizer that takes a grammar (and an interpreter) for the DSL as an input. Q: Is the time to draw N samples from the LLM and construct the PCFG included in the plots in Figure 4a, 4b, 4c? See Global Response Q1. Q: Why do you include few-shot examples when prompting for the tensor domain, but not for the string domain? It seems more important for the string domain given that 0% of the LLM-generated programs in the string domain are valid completions. See Global Response Q2. Q: Why are I/O examples provided as an input to the synthesizer, rather than just the PCFG? The I/O examples need to be provided as input to the synthesizer in order to test the correctness of generated programs and prune the search space based on observational equivalence (see lines 5, 6, and 8 in Algorithm 1). Q: Is it correct to state that the PCFG is reasonably approximating the LLM's condition distribution over output programs for a given prompt? You are right that a PCFG is still a poor approximation (even for the conditional distribution). What we meant to say here is that, for a specific task, a PCFG is able to capture just enough signal from the LLM that it can guide enumerative search. We’re happy to update the phrasing! --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: Thank you for your rebuttal, and in particular Global Responses 1 and 2. Regarding Global Response Q1 (Q1, Q3 in "Rebuttal by Authors"): > The time to draw those samples is excluded from the graphs in Fig 4. Thank you for clarifying this. This is a very important detail to have omitted, and it does meaningfully change the results of the paper / what is conveyed in Figure 4. The baselines ARGA, Probe, TF-Coder, and Unguided are being given a 268-285 second disadvantage in Figure 4 compared with HySynth and No-Search. Eyeballing it, it looks like HySynth still outperforms the baselines even when this adjustment is accounted for, but not as immediately or decisively as Figure 4 currently conveys. My recommendation for making this clear would be to include the sampling time in the Figure 4 plots, giving all methods the same total time (not just the same search time) for a fair comparison. Making clear the total time spent by each method (not just the time spent searching) is necessary for a fair comparison of the methods if you wish to claim that HySynth outperforms the other methods (a key claim of the paper, line 86). > That said, we did compute the time it took to sample 100 solutions, and it amounts to 285 secs for tensor domain and 268 secs for the string domain. Thanks for measuring this. This suggests that running the No-Search / "GPT-4o" baseline for 10 minutes would involve drawing no more than 225 samples, and so the cost might be quite affordable. A back-of-the-envelope calculation suggests the total cost would be between $3 - $4 USD per domain (so up to $12 USD total). If you do this, the Fig. 4 plot for the "GPT-4o" baseline would then show progress over time like the other methods, rather than being a fixed horizontal line. > 1) LLM sampling incurs not only time cost, but non-trivial monetary cost, while symbolic search is virtually free Yes, it's true that sampling is more expensive than search per unit time, but I don't think this justifies ignoring the sampling time. (e.g. the caption for Fig 4 currently reads as "Number of benchmarks solved by HYSYNTH as a function of time", not even specifying search time; similarly the language used in 4.2 is "We compare the time to solution for the main HYSYNTH configuration, baseline synthesizers, and the two ablations; the results for the three domains are shown in Fig. 4", again giving the impression that this measures total time.) Regarding Global Response Q2 (Q4 in "Rebuttal by Authors"): > Providing in-context examples is unlikely to help because these examples would have a different grammar, and only confuse the model. This is possible, but I would not be surprised if in-context examples did help, even if they came from different grammars. (The way I expect you would prompt the model, the grammars would be included in the in-context examples.) -------- Q2: Thank you for clarifying. Q5: Thank you for clarifying; that was my mistake. Q6: Thank you for your comment and agreeing to update the text. --- Reply to Comment 1.1.1: Comment: Non-LLM baselines (ARGA, Probe, TF-Coder, and Unguided) are at a disadvantage: > Sorry, our explanation was confusing: the time reported in the response is for the entire dataset; it only takes ~4 seconds to sample 100 solutions for a single problem from gpt4o. This makes the disadvantage quite small, and our claim that HySynth outperforms all the non-LLM baselines would definitely still hold if we added this sampling time to HySynth’s results. We will make a plot with the sampling time incorporated in future revisions in order to convey the full picture. LLM baseline is at a disadvantage because we did not sample until timeout: > In light of the last paragraph, sampling for 10 min per problem would actually be quite expensive (but we can perform this experiment with a shorter timeout, if the reviewer thinks it’s necessary). In general, we still believe that cost is a major limiting factor for the LLM baseline, rather than just time (especially since, as you mentioned, samples can be generated in parallel). So, to put all techniques on equal footing, it makes more sense to limit both time and cost for all the techniques. In our paper we use sample size (100) as an approximation of the cost limit, but we could also conduct experiments with an actual cost limit, if the reviewer finds this necessary (in which case, the sample size will be different for different problems and LLMs).	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their time, valuable comments, and encouraging feedback! In the global part of the response, we answer shared questions and provide a list of changes we plan to make in a revised version of the paper. Q1 (Reviewers FZiE and XXbQ): Does the time reported in Fig 4 include the time taken for sampling from the LLM and training the surrogate model? What percentage of the total time does LLM inference take? The LLM baseline is at a disadvantage because many more samples could be drawn within the timeout of 10 minutes. In our experiments, the results for both the LLM baseline and HySynth are based on 100 samples from the LLM. The time to draw those samples is excluded from the graphs in Fig 4. Reviewer FZiE is right that many more samples could be drawn within 10 minutes. Indeed, one way to conduct our experiments would be to give all techniques the same amount of total time, which can be used towards either LLM sampling or symbolic synthesis. We thought such setup would not be fair for the following two reasons: 1) LLM sampling incurs not only time cost, but non-trivial monetary cost, while symbolic search is virtually free; 2) the time cost of LLM sampling depends on many incidental factors out of our control (e.g. network latency, server load). We chose instead to completely separate the sampling stage from the synthesis stage and to only report the synthesis times in the figures (which is why the LLM-only line is always horizontal, as the reviewer notes). That said, we did compute the time it took to sample 100 solutions, and it amounts to 285 secs for tensor domain and 268 secs for the string domain. Learning the PCFG takes less than 1 sec for all domains. We will report these times in the next version of the paper. Q2 (Reviewers FZiE and XXbQ): Why does GPT-4o have 0% correctness and syntactic validity on the String domain? Could this be improved with more advanced prompting techniques? Why do we not provide in-context examples for this domain, like for Tensor? This is due to the specifics of the SyGuS benchmark, from which our String problems are drawn. In this benchmark, each problem comes with a custom restricted grammar; all grammars are subsets of the full SyGus grammar but purposefully impose additional restrictions on the solution (for example, the synthesizer is usually allowed only a handful of string constants, which excludes trivial solutions that simply reproduce the given output examples). In the experiments reported in the paper, we judge syntactic validity against these custom grammars (which is how the SyGus competition is also judged), and the LLMs turn out to be quite poor at following these syntactic restrictions, even though the grammar is given in the prompt. Providing in-context examples is unlikely to help because these examples would have a different grammar, and only confuse the model. If we were to relax the definition of syntactic validity, ignoring the custom grammars and allowing the full SyGuS grammar, LLM solutions would achieve a syntactic validity of 60.5% and solve 20/70 problems correctly (which is better than Unguided search but worse than Probe and HySynth.). We will include these additional results in the paper along with a detailed analysis. Change list 1. We will report the time and monetary cost of sampling from the different LLMs in our experiments (as mentioned in Q1 above). 2. We will clarify why LLMs get 0% syntactic validity on the String domain and add the results of our new experiments with a relaxed notion of validity (as explained in Q2 above). 3. We will include the results of an ablation study using a simpler surrogate model, which simply excludes the components not used by the LLM (as suggested by Rev MWWa; preliminary results reported in the attached PDF). Pdf: /pdf/4570a2467bfdd4a324dad4e527b4a60cdd79b8a8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On the Adversarial Robustness of Benjamini Hochberg	Accept (poster)	Summary: Benjamini-Hochberg provides a means for controlling the false detection rate (FDR) in multiple hypothesis testing. This paper explores how the FDR region can be perturbed by an adversary. Specifically, the author(s) propose two adversarial attacks Increase-C and MOVE-1 to efficiently perturb c scores and 1 score, respectively. The author(s) prove analytical bounds on their attacks' guarantees. The author(s) also provide simulated experiments demonstrating their attacks effectiveness on synthetic z-scores. Strengths: I was not familiar with Benjamini-Hochberg before reading this paper (I did not bid on the paper). However, with significant effort and outside reading, the paper was still attainable. That is a strength of the work in general. The Balls in Bins warm-up was definitely helpful to a reader. I provide comments on how this warm-up could be improved below. ### Novel Area and Good Execution To the extent of my knowledge, no existing work has studied the adversarial robustness of methods to control FDR. I discuss why I believe that is below in the limitations section. Nonetheless, the authors provide reasonable first-pass attack methods that are intuitive and performant. They even provide an optimal method in the case of singleton perturbations. The authors also provide thorough theoretical analysis. Weaknesses: ### Very Little Analysis of Section 5's Empirical Results Sections 5.1 and 5.2 are essentially just an overview of the two experimental setups. There is no real analysis or discussion of what the experimental results show or why it's important. While Section 5.3 has some basic analysis, it is again quite minimal. I understand that the page limit forces difficult decisions over content but so little in the way of commentary on the results is a disservice to the reader and overall a poor choice. I would have preferred empirical results on real datasets as opposed to the toy datasets in the paper, but I understand the author(s)'s choices here, especially given related work. ### Move-1 Relegated to the Appendix Beyond a very definition in Section 1.5, all discussion of MOVE-1 including how the method works and why its optimal are relegated to the appendix. There should at least a basic overview of the key ideas of MOVE-1 including some basic intuitions how it differs from INCREASE-c in the main paper. ### Making the Work More Intuitive Section 2 in my view is a warmup for readers to help build intuitions and understanding of BH, even if they are not familiar with the related background. Section 2 could be significantly improved if figures were used to better illustrate the key ideas. These figures need not be large and any additional camera ready page could be used to visualize the ideas. Even if the authors chose not to include such figures in the main paper, they could be in the appendix. ### Limited Potential Impact There is undeniable merit and utility in studying the robustness of a method that provide tools to control FDR, such as Benjamini-Hochberg. Nonetheless, this work falls essentially is a niche (BH) within a niche (adversarial analysis). Hence, I view it as unlikely that this work will have a significant impact. Technical Quality: 3 Clarity: 2 Questions for Authors: None. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: This paper describes two adversarial attacks and identifies the brittleness of Benjamini-Hochberg. While such work inherently creates a non-zero risk, there are no serious negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Addressing the ``Weaknesses" $\ldots$ 1. Very Little Analysis of Section 5's Empirical Results Sections 5.1 and 5.2 are essentially just an overview of the two experimental setups. There is no real analysis or discussion of what the experimental results show or why it's important. While Section 5.3 has some basic analysis, it is again quite minimal. I understand that the page limit forces difficult decisions over content but so little in the way of commentary on the results is a disservice to the reader and overall a poor choice. Rebuttal: In terms of Sections 5.1 and 5.2, which concern simulating INCREASE-c on iid and PRDS p-values respectively, we had thought the figures and tables illustrated the effectiveness of INCREASE-c far better than words could. Indeed, the page limit did severely constrain some discussions, but if given the opportunity to, we would absolutely correct this with more elaboration, as you suggest. For what it's worth, what we had intended for the reader to understand from Figure 1, was that with both colors and lines, we were indicating how frequently INCREASE-c not only increased the FDP (above the $45^\circ$ line) but also raised it above the theoretical control level of $\pi_0 \cdot q$ (above the horizontal line), or in other words, broke BH's FDR control. Likewise in Table 1, the high average FDP numbers and perturbed rejection counts illustrate the effectiveness of INCREASE-c. Figure 4 and Table 2 communicated the similar conclusion but for PRDS conformal p-values. I would have preferred empirical results on real datasets as opposed to the toy datasets in the paper, but I understand the author(s)'s choices here, especially given related work. Rebuttal: Thanks for your input. For us, given the choice between a real dataset and a simulation-procured, synthetic dataset of Bates, Candes et al 2023 -- with publicly available code which they used to conduct their experiments -- we felt our message on BH's fragility would be stronger by repurposing their code for comparison's sake. Consequently, in light of the page limit, we had to compromise on the real datasets. 2. Move-1 Relegated to the Appendix Beyond a very definition in Section 1.5, all discussion of MOVE-1 including how the method works and why its optimal are relegated to the appendix. There should at least a basic overview of the key ideas of MOVE-1 including some basic intuitions how it differs from INCREASE-c in the main paper. Rebuttal: We thank you for your interest and appreciation of MOVE-1. Of course we appreciate it too, but it not only pertains to the very special case of $c = 1$ but also does not strongly outperform INCREASE-1 empirically per se, despite its theoretical optimality in this case. Indeed, the INCREASE-c algorithm is arguably easier to describe and implement, and of course handles all possible $c$ value cases. Our theoretical performance analysis guarantee was also designed specifically for INCREASE-c. Hence, for all these reasons, and in light of the 9 page limit, we did not feel as though discussing MOVE-1 in detail had good bang for buck. 3. Making the Work More Intuitive Section 2 in my view is a warmup for readers to help build intuitions and understanding of BH, even if they are not familiar with the related background. Section 2 could be significantly improved if figures were used to better illustrate the key ideas. These figures need not be large and any additional camera-ready page could be used to visualize the ideas. Even if the authors chose not to include such figures in the main paper, they could be in the appendix. Rebuttal: Your suggestions to use illustrations to enhance the readability of the mathematical description of the Balls into Bins perspective on BH are duly noted. Thank you! 4. Limited Potential Impact There is undeniable merit and utility in studying the robustness of a method that provides tools to control FDR, such as Benjamini-Hochberg. Nonetheless, this work falls essentially is a niche (BH) within a niche (adversarial analysis). Hence, I view it as unlikely that this work will have a significant impact. Rebuttal: Thank you for your assessment. The BH is the tool of choice for large scale hypothesis testing, with over 100K Google scholar citations of the original Benjamini and Hochberg "Controlling the false discovery rate: a practical and powerful approach to multiple testing" paper. As well, in the intro to our paper, we also cite a handful of more modern works in AI/ML that are leveraging BH in some way for out-of-distribution-detection (OOD) - among them, Lieu et al 2024 (accepted at the recent ICML 2024) and Bates, Candes et al 2023. So, we believe a paper examining BH's fragility is relevant and important. ---	Summary: This paper studies the adversarial robustness of the Benjamini-Hochberg (BH) procedure, introducing simple adversarial test-perturbation algorithms. The experiments show that BH's control can be significantly compromised with minimal perturbations. The analysis uses a combinatorial perspective and generalized ballot problems to derive non-asymptotic lower bounds. Strengths: This paper addresses a novel and significant aspect of the BH procedure—its adversarial robustness. This research probelm is relatively unexplored, providing new insights into the vulnerabilities of widely used statistical methods. Weaknesses: This paper would benefit from a discussion of potential mitigation strategies that could enhance the robustness of the Benjamini Hochberg procedure against attacks such as INCREASE-c. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to the weaknesses. Confidence: 1 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors' justification is "Yes, we discuss to what extent our analysis could be extended and further generalized, but were not done due to space constraints.". However, even in Appendix, there is not any analysis of limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Addressing the Weakness comment $\ldots$ This paper would benefit from a discussion of potential mitigation strategies that could enhance the robustness of the Benjamini Hochberg procedure against attacks such as INCREASE-c. Rebuttal: Thanks for asking about mitigation strategies, which we view as an important follow-up question to this paper's thesis that BH's FDR control can be susceptible to minimal adversarial perturbation. In fact, we are in the midst of preparing this follow-up work. We kindly refer you to our answer to Reviewer pEvv's question #2, in which we briefly comment on interesting dynamics to the attacker-defender interaction, including some insights into mitigation strategies studied in this said follow-up work. Given that our submission as it stands now rides right up to the 9-page limit, with a moderate amount of appendix material already, we believe that the study of mitigation strategy extensions would be better positioned as a separate paper.	Summary: In this paper, the authors have tested the Benjamini-Hochberg (BH) 's adversarial robustness, as this procedure is deployed in critical applications such as drug discovery, forensics, and anomaly detection. Specifically, the authors develop a class of simple and easily implementable adversarial test-perturbation algorithms to analyze under what conditions BH does and does not exhibit adversarial robustness. Next, to support their findings, the authors provide non-asymptotic guarantees on the expected adjustments to the FDR due to these adversarial attacks. Their technical analysis involves a combinatorial reframing of the BH procedure as a "balls into bins" process. It connects this to generalized ballot problems to utilize information-theoretic approaches for deriving non-asymptotic lower bounds. Finally, the authors also conducted experiments to support their findings. Strengths: - In this paper, the authors provided a detailed theoretical analysis of the BH procedure's robustness against adversarial attacks. They also introduced the INCREASE-c algorithm, which gives methodical mathematical bounds and probabilities for different cases (large and small alternative means). - I liked how advanced probabilistic and statistical tools, such as KL divergence and Pinsker's inequality, are used to provide a solid mathematical foundation for the analysis. The proofs and the lemmas were thoroughly justified. - Finally, the paper presentation is well-structured. The paper's organization is clear and logical. In each section, a comprehensive analysis is done and well-presented. Weaknesses: - In practical scenarios, an adversary manipulates the input directly, not the z-scores; thus, the insight into how perturbations applied to z-scores translate back to the original data samples or vice-versa is missing in the paper. - My other concern is that since z-scores are computed only utilizing the means and the standard deviation of the sample, the manipulation at the data level can get lost at the z-scores level. Technical Quality: 3 Clarity: 4 Questions for Authors: - In practical scenarios, an adversary manipulates the input directly, not the z-scores; thus, could the authors provide more insight into how perturbations applied to z-scores translate back to the original data samples or vice-versa? Or, when the inputs are perturbed, what will impact the proposed analysis of the BH algorithm (given the z-scores computed from the input samples)? - A statement on Page 4: "A benefit of this approach is that perturbation at the level of z-scores places the corruption more directly at the point of data collection than does perturbation of the p-values." Could the authors answer: Won't the perturbation at the data collection directly place the corruption at the data level? Also, why was it not assumed to be corruption at the data level, then the computation of either the z-scores or the p-values? How this mapping between the data and the z-scores is computed? - The evaluation results are a bit difficult to navigate in the paper. For example, Figure 1 and Figure 4 are not understandable. Could the authors make the evaluation section comprehensible? - Could the authors add more evaluations to test how the proposed method performs with real-world data? Post-rebuttal: I have changed my rating from borderline reject to borderline accept, confidence from 3 to 4, and soundness of the paper from 2 to 3. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Same as Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Q1 Re: Indeed, $z$ scores are classically $z_i := \frac{\sum_{j=1}^n x_{ij}/n - 0}{s/\sqrt{n}}$, so that perturbations to the "samples" $x_{ij}$ translate to perturbation of the z-score $z_i$, and ultimately the p-value $p_i:= 1 - \Phi(z_i)$. Since p-values are ultimately the input processed by BH, our model is not concerned with how perturbations of the input samples $x_{ij}$ are made, so long as the $p_i$ value is affected. For a broader perspective on this matter of samples versus z-scores versus p-values, we call to mind the application of ``conformal p-values" by Bates, Candes, et al 2023 that strongly motivated this work, and for which we reference several times, and implement their experiments within ours in Section 5.2. In particular, they consider a statistical wrapper that transforms each $X_i$ in a collection $(X_j)_{j=1}^n$ of signals (/measurements) into a p-value $p_i \in [0,1]$, the total collection $\{p_j\}_{j=1}^n$ of which corresponds to $n$ hypothesis tests to be conducted simultaneously, each of which determines whether a signal $X_i$ is an anomaly, or outlier. Therefore, in this context, perturbations to a score $X_i$ is direct. # Q2 Re: As far as theoretical analysis, our work is largely free of distributional assumptions, so there is little "impact" from "data-generating distribution changes." For instance, it is w.l.og. that the null p-values are Uniformly distributed. As for the alternative distributions, we made few assumptions throughout. Indeed, Theorems 3.1, 3.2 after all do not make any distributional assumptions on the alternatives. The bound provided in Theorem 4.4 is a function of the $\mu_1^i$, which as we discussed in Section 4.2, can also be replaced with max $\mu_1:= \max_{i \in \mathcal{H}_1} \mu_1^i$ for a more conservative bound. As far as numerical analysis, it was for the sake of experiments that we made assumptions on the alternative distribution, as do all papers dealing with the BH procedure - see Bates, et al 2023 for example. # Q3 Re: To the best of our understanding, we address your concern in our answer to Question 1, so we kindly refer you to the answer we provided above. # Q4 Re: With regards to the Type II error, since all the z-scores will be in the rejection region, there would be no failure in rejecting any $i \in \mathcal{H}_1$ (i.e. any test for which the ``alternate hypothesis is true"); hence, there would be no Type II error, by definition. For what it's worth, although Type II error is of note, false discovery rate control is concerned with Type I error. # Q5 Re: If we're not mistaken, in the "same case above", the alternative distributions' being "close" to the null is a non-factor. Indeed, you provide an event in which all $N$ z-scores are rejected, so the Type II error is 0, as above. As for the FDP in this event, it would be forced to be $\pi_0$ because this is the fraction of scores out of $N$ that are null. If you are asking about FDR control in the case of alternative distributions being "close" to the null distribution, this is precisely the subject of Section 4.2, wherein we provide a theoretical lower bound (Theorem 4) on how much the adversary can increase the FDR with INCREASE-c. This bound is plotted in Figures 2 and 3 of Section 5, in which we examine, respectively, the extreme case when the alternative distributions are identical to the null distribution and a case where they are not identical but roughly speaking, about a quarter standard deviation away from each other. The point to these plots is to illustrate the fragility of the BH FDR control when the alternative and null distributions are "close". # Q6 Re: The frequency color of a plotted dot at $(FDP[BH_q; p]$, $FDP[BH_q; p_{+c}]) \in \mathbb{R}^2$ indicates how many of the $10^4$ simulations produced that combination of before-adversary and after-adversary FDP. Figures 1 and 4 illustrate how very frequently the adversary's INCREASE-c not only increased the FDP (above the $45 \deg$ line) but also raised it above the theoretical control level of $\pi_0 \cdot q$ (above the horizontal line). This high frequency illustrates a breaking of BH's FDR control. # Q7 Re: We believe you are referring to Section 5.1, line 311 which states $z_{i \in \mathcal{N}_0} \sim N(\mu_1, 1)$. There was a typo in the subscript, and should actually read: $z_{i \in \mathcal{N}_1} \sim N(\mu_1, 1)$. # Q8 Re: We followed the approach taken in most BH-procedure papers, where the experiments are over a range of $\mu_1$. Unfortunately, not all parameter choices could be explored/presented within a 9-page paper. For what it's worth, follow-up experiments do not seem to indicate anything that we haven't showcased already. # Q9 Re: We executed the algorithm just as described on page 5. We "move the largest $c$ (ties broken arbitrarily) in the (N+1)-th bin to bin $\tilde{k}_{+c}$." More precisely, as our z-scores were stored in memory with vectors, ties were broken by taking whichever z-score occurred earlier in the vector's indexed-ordering. # Q10 Re: Section 5.3's experimental setup involved repeated simulations of $N = 10^3$ p-values drawn just as described in Section 5.1. Regarding the experiment that produced Figure 2, as the caption indicates, we had parameter settings of $\mu_1 = 0$, $\pi_0 = .90,$ and $c = 1.$ As for $q$, as the figure indicates, this was now varied across a grid of $q \in (0,1)$ (mesh spacing of .01). For each $q,$ we performed the repeated simulations, in each of which we computed $FDP[BH_q; z_{+c}] - FDP[BH_q; z]$, so that averaged across all simulations we obtained $\Delta_1(q)$. Repeated for all $q \in (0,1)$, we obtained the blue curve that is labeled $\Delta_1$. The red curve is $L_c$ from Theorem 4.4 as a function of $q$. Regarding the experiment that produced Figure 3, as the caption indicates, we had parameter settings of $\mu_1 = 0.25$, $\pi_0 = .95,$ and $c = 1.$ The rest of the details are just as described above. --- Rebuttal Comment 1.1: Comment: Dear Authors, Thank you for your response. The rebuttal clarified the majority of my questions. - However, there remains an open concern (from all the reviewers) about how well the proposed method would perform with real-world data. Given the page limit of 9 pages, could the authors experiment with an MNIST/FashionMNIST OOD-classifier to demonstrate the efficacy of the INCREASE-c algorithm, as suggested by reviewer pEvv, and add it in the appendix? It will help strengthen the paper. - This will also show the practical relevance of the proposed approach, as real-world adversarial attacks are more likely to target the raw input data. --- Reply to Comment 1.1.1: Comment: Yes, we would be happy to add experiments with real-world data to the appendix. We have provided an experiment in our "global" official comment above, which serves as an example of things we could provide in the appendix, if given the opportunity to do so. Thank you for your suggestion!	Summary: The paper explores the adversarial robustness of the Benjamini-Hochberg procedure. In particular, the authors theoretically show that it is possible to perturb test scores to cause the BH procedure to not be robust to adversarial attacks. BH is reframed as a "balls into bins" problem, and the authors propose an algorithm (INCREASE-c) to increase the rejection count. The authors also provide experiments to support their theoretical findings. Strengths: - The technical contribution seems solid and sound. - There is a clear novelty aspect, as the adversarial robustness of BH has not been considered before. - The authors show an interesting gap between distributional robustness (which was shown to hold (ref [29] in the paper)) and adversarial robustness for BH procedure. - The (synthetic) statistical experiment supports the theoretical claims. Weaknesses: - In the introduction, the paper (correctly) highlights the importance of hypothesis testing to various safety applications such as OOD detection. However, the experimental section does not, in any capacity, consider an end-to-end application such as OOD detection. In fact, it reads purely as a statistical simulation (which is fine as a synthetic experiment). But a minimal experiment with a MNIST/FashionMNIST OOD-classifier, to show that indeed the Z-scores can be affected by INCREASE-c could make the paper much stronger. - The attack is dependent on the algorithm (BH) used for FDR -- stronger results in adversarial robustness are often model/algorithm-independent. - There seems to be a certain lack of justification for choices in the problem setup (addressed in questions below). Technical Quality: 3 Clarity: 4 Questions for Authors: - Problem setup questions: - What is the rationale for the attacker having access to z-scores at that stage of the pipeline? Usual studies consider attackers having access to data, and it may take quite a lot of changes in data points to change a z-score. - Why is the attacker's knowledge of $q$ realistic? I assume that in some models, the deployment will come with a (public) guarantee for $q$ -- but this should be at least mentioned. - It would be good to have a comment justifying the choice of perturbation ( $\|\|\cdot\|\|_0$ ), which I suppose is because BH is an algorithm for order statistics. What happens when considering a different type of budget (e.g., arbitrary $\|\|\cdot\|\|_p$ norms)? - l.128-132: When are these assumptions met in practice? - What is the impact of attacking BH procedure on the whole OOD system? Can you break OOD detection in a meaningful way? **** Post-rebuttal: I have increased my confidence score from 2 to 3 following the authors' response and other reviews. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: I think some of the problem setup questions above would address some limitations of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Addressing Weaknesses 1. In the introduction, the paper (correctly) highlights the importance of hypothesis testing $\ldots$ However, the experimental section does not $\ldots$ consider an application such as OOD detection \ldots Rebuttal: We did in fact experiment on an ``OOD-classifier." We refer you to our Section 5.2's experiments on the non-parametric outlier detection method proposed in the paper by Bates, Candes, et al 2023 based on conformal p-values. Specifically, we repeated the experiment in their paper that produced conformal p-values with a trained, SVM one-class classifier, only now we added small corruption to see how badly their OOD system might mistake inliers for outliers. 2. The attack is dependent on the algorithm (BH) used for FDR -- stronger results in adversarial robustness are often model/algorithm-independent. Rebuttal: BH is the tool of choice for large scale hypothesis testing, with over 100K Google scholar citations of the original paper. BH is also emerging in AI/ML for out-of-distribution-detection (OOD) - for example, see Lieu et al 2024 (accepted at the recent ICML 2024) and Bates, Candes et al 2023. In other words, BH would seem to be a natural and relevant choice for an adversarial study. Further, BH belongs to the family of step-up procedures, which means the rejection region is decided via a stopping time (see our analysis), which since our paper shows can be manipulated, it means other ``step-up" procedures" are similarly prone. Addressing Questions Question 1. What is the rationale for the attacker having access to z-scores at that stage of the pipeline? Usual studies consider attackers having access to data, and it may take quite a lot of changes in data points to change a z-score. Rebuttal: Thank you for your comment. For what it's worth, changes in the data translate into changes at the z-score level. As for the size of changes, in Page 16, Appendix Table 3, the algorithm MOVE-1 on average moved the z-score .551, .492, and .139 in the settings of $\mu_1 = 0, 1,$ and $2$ respectively. Granted, the effect of data-point perturbations on z-scores would ultimately also be affected by matters like sample size and standard deviation. That being said, we are more focused on applications in outlier detection or out-of-distribution-detection (OOD), as in Bates et al 2023, which we referenced several times and use as an experimental baseline in Section 5. More precisely, in this context, they consider a statistical wrapper that, broadly speaking, transforms each $X_i$ in a collection $X_1, X_2, \ldots X_n$ of signals (measurements) into a p-value $p_i \in [0,1]$, the total collection $(p_j)_{j=1}^n$ of which then corresponds to $n$ hypothesis tests to be conducted simultaneously, each of which determines whether a signal $X_i$ is an anomaly/outlier. Therefore, perturbations in this context affect the $X_i$ scores directly. For further discussion on this, we kindly refer you to our answer to reviewer L7r5's question \#1. Question 2. Why is the attacker's knowledge of $q$ realistic? $\ldots$ Rebuttal: Indeed, your point about a "public guarantee" that would grant the adversary knowledge is realistic, with values of $q = 0.05, 0.10$ a standard practice. On a related note, while it is possible for the adversary to not precisely know the true control level $q$ to be implemented by the decision maker, this opens the door to some really interesting attacker-defender dynamics between the adversary and decision maker that we are in fact studying in a follow-up work. In any case, we confirm that there are measures that can be taken by the adversary towards harming FDR control even without perfect knowledge of $q.$ Question 3. It would be good to have a comment justifying the choice of perturbation $(\|\|\cdot\|\|_0)$, which I suppose is because BH is an algorithm for order statistics. What happens when considering a different type of budget (e.g., arbitrary $\|\|\cdot \|\|_{p}$ norms)? Rebuttal: Viewed from the perspective of outlier detection, our study is considering an adversary attempting to make the decision maker confuse inlier signals for outlier signals or vice versa. For some particular motivating applications, consider candidate screening, spotting frauds/intrusions, and forensic analysis - applications discussed in Bates, Candes, et. al 2023 and Jin, Candes 2023. In such contexts, we think it is natural that the number of hypothesis tests (equiv. signals) that an adversary can influence is bounded - hence, the modeling choice of a budgeted $\\|\cdot \\|_0$. This choice was sufficient to demonstrate a fragility to BH's FDR control. More precisely, the attacker can break the BH guarantee by moving few z-scores. But, yes, other measures ($\\|\cdot \\|_p$ norms) of corruption effort might be considered as well. Question 4. l.128-132: When are these assumptions met in practice? Rebuttal: The lines l.128-132 do not detail assumptions; rather, all we mean to say in these lines is that the corruption model we propose is most interesting in scenarios where not too many tests get rejected, for otherwise any attack would have to be massive in order to have an effect on the FDR. Indeed, the number of rejections $R_p$ comprises the denominator in the false detection proportion $FDP[\mathcal{A};p]:= \frac{a_p}{R_p \vee 1}$. Regardless, the theoretical results hold true under any values of $N$, $q$, and the collection of $\mu_1^i$'s. Question 5. What is the impact of attacking BH procedure on the whole OOD system?... break OOD detection? Rebuttal: Figure 4 and Table 2 detail the experiments we conducted on the OOD "system" that Bates, Candes, et al 2023 studied. In particular, in Table 2, the average proportion of the reported outliers that are in fact inliers is reported. As one can see, there can be significant impact on OOD systems based on BH (of which there are several recent publications on). --- Rebuttal Comment 1.1: Title: Rebuttal Response Comment: Dear authors, Thank you for the detailed rebuttal and the clarifications. Regarding the OOD application, my comment was more in line with reviewer L7r5, in that an experiment on real-world data (rather than a synthetic one) would significantly strengthen the paper. Regarding the algorithmic-dependent results, a remark in the spirit of the last sentence of your response to this weakness (Further, BH belongs to the family of step-up procedures, which means the rejection region is decided via a stopping time (see our analysis), which since our paper shows can be manipulated, it means other ``step-up" procedures" are similarly prone.) would be worth including in a future version, perhaps along with more general theorem statements/corollaries. My last remaining question for the author is whether they would be willing/able to add justifications/explanations in the paper in line with my and other reviewer's comments and questions on z-scores perturbation, choice of norm, adversary's power, etc. (which would make the paper clearer, a bit more solid and easier to approach for people who are not as familiar with the literature), and whether they have or will perform an experiment on real-world data as suggested by L7r5 and myself. --- Reply to Comment 1.1.1: Title: Adding Justifications/Explanations and Experiments Comment: Regarding the OOD application, my comment was more in line with reviewer L7r5, in that an experiment on real-world data (rather than a synthetic one) would significantly strengthen the paper. Response: Understood, thanks for the suggestion. We have now posted a "global" official comment above that discusses a real-world data experiment concerning credit card fraud detection. We are happy to include more real-world data experiments along these lines in the paper, if given the opportunity. Regarding the algorithmic-dependent results, a remark in the spirit of the last sentence of your response to this weakness "Further, BH belongs to the family of step-up procedures, which means the rejection region is decided via a stopping time (see our analysis), which since our paper shows can be manipulated, it means other ``step-up" procedures" are similarly prone.) " would be worth including in a future version, perhaps along with more general theorem statements/corollaries. Response: Yes, we appreciate this suggestion and agree that this additional discussion surrounding the generality of attacking "step-up" procedures would be worthwile. My last remaining question for the author is whether they would be willing/able to add justifications/explanations in the paper in line with my and other reviewer's comments and questions on z-scores perturbation, choice of norm, adversary's power, etc. (which would make the paper clearer, a bit more solid and easier to approach for people who are not as familiar with the literature), and whether they have or will perform an experiment on real-world data as suggested by L7r5 and myself. Response: Yes, we are absolutely willing and able to add the justifications/explanations we offered to the review team on matters like z-score perturbation, norm choice, adversary's power, etc. And yes, we have performed one experiment on real-world data and offered it in the "global" official comment above, and we are willing to perform additional experiments on real-world data as suggested by L7r5 and yourself.	Rebuttal 1: Rebuttal: We thank the reviewers for their questions and feedback. We have provided individual responses to each and welcome the opportunity to engage further in the discussion period.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models	Accept (spotlight)	Summary: The paper proposes token-level generalization bounds for large language models (LLMs), such as LLaMA2-70B, using less restrictive compression techniques like Monarch matrices, Kronecker factorizations, and post-training quantization. The authors argue that traditional document-level bounds are vacuous at this scale and introduce a method leveraging martingales for deriving tighter bounds, which not only hold theoretically but are also demonstrated through empirical validation. Strengths: 1. Originality: The paper introduces a novel approach to computing generalization bounds at the token level, which is a significant departure from the document-level bounds prevalent in prior works. 2. Technical Soundness: The use of martingales and non-restrictive compression methods to derive generalization bounds is both innovative and robust, providing a solid theoretical framework backed by empirical results. 3. Significance: The ability to provide non-vacuous generalization bounds for LLMs as large as 70 billion parameters is highly significant, as it pushes the boundary of what is understood about LLM generalization in practical settings. 4. Clarity: The paper is well-written, with clear explanations of the methods and their implications, making it accessible to readers who may not be experts in the specific sub-field of machine learning. Weaknesses: No major weaknesses Technical Quality: 3 Clarity: 3 Questions for Authors: Can the generalization bounds proposed be integrated into the training regimen to enhance model generalization directly? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper lacks intuitive explanations for the proposed bounds, which might hinder understanding for readers not familiar with advanced statistical concepts in machine learning. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Many thanks for your encouraging and thoughtful feedback! We respond to your questions below. An intuitive application of our bounds to downstream tasks: Inspired by your comments, we provide another intuitive application of our bounds. In this case, to a downstream scientific task. Our token-level generalization bounds are particularly descriptive of antibody design in biology. An antibody sequence is usually composed of 20 different amino acid tokens to bind to a target of interest. In therapeutic antibody design, biologists propose mutations to existing antibody sequences by changing the amino acid tokens at specific positions in the sequence. Recent works have shown that LLMs pretrained on large antibody datasets can be used to propose mutations conditioned on starting antibody sequences. Our token-level generalization bounds match the antibody design setting by bounding the expected next amino acid token negative log likelihood averaged over training contexts that serve as starting sequences for iterative mutations. In the table below, we show that language models based on the Mistral 7B architecture pretrained on a processed subset of the Observed Antibody Sequences (OAS) from scratch achieves non-vacuous token-level generalization bounds: \| Compression Approach \| BPD Bound \| Top-1 Error Bound \| Validation Loss \| \| --- \| --- \| --- \| --- \| \| Mistral 377M \| 2.41 \| 31.60 \| 0.28 \| \| Mistral 212M \| 2.06 \| 26.25 \| 0.30 \| \| Mistral 94M \| 1.62 \| 19.40 \| 0.30 \| \| Random Guess \| 4.86 \| 96.56 \| 1.46 \| Using generalization bounds for model optimization: It is indeed possible to integrate the elements of our bounds directly into the training regimen to enhance model generalization directly. One example is quantization-aware training, which we in fact already use for post-training quantization of all pretrained models except the LLaMA models. In this scenario, we wish to map the pretrained weights of the neural networks into a significantly smaller number of quantization clusters. The quantized vector $\hat{w} = [\hat{w}_1,\dots,\hat{w}_d]$ can be constructed from the original weights vector $w = [w_1,\dots,w_d]$ by assigning these weights to different clusters $c = [c_1,\dots c_L]$, where $\hat{w}_i =c_q$ such that ${q= \operatorname{argmin}_k \|w_i-c_k\|}$. The quantization clusters $c$ are learned alongside $w$, such that we optimize the empirical risk and the compressed size of the model as well. We will add more details about our quantization-aware training procedure in the appendix. The focus of our work however is to compute non-vacuous generalization bounds that we can use to understand when and why LLMs generalize and provide a prescription for how to design better models in practice. We provide additional experiments in the general response showing that not only is the quantity we bound predictive of downstream performance, but the bounds themselves positively correlate with downstream performance. We also demonstrate that the trade-off between the empirical risk and the compressed size of the model highly depends on the compression scheme and the size of the training data. The additional figures can be found in the Rebuttal PDF. Inspired by your feedback, we also made sure to include additional intuitive explanations and details throughout the revised manuscript to make sure that the paper's content is accessible to readers. Thank you again for your supportive and positive review. We made a significant effort to revise our paper and run additional experiments in light of your feedback. We would appreciate it if you would consider raising your score in light of our response. Please let us know if you have any additional questions or comments. --- Rebuttal Comment 1.1: Comment: I appreciate the authors response, which addressed all my questions. I increase my score to 7.	Summary: This paper develops nonvacuous generalization bounds for modern language models. Specifically, this paper proves a token-level generalization bound, and applies different techniques (LoRA, 2 Kronecker Product, Monarch Matrices, and post-training quantization) to control the capacity of model class. Strengths: This paper is well-written: Theorem 3.1 is clean, and many compression techniques (parameter-efficient tuning, post-training quantization,etc.) and modern language models (GPT, LLAMA, etc.) are analyzed. I believe these are nice contributions. Weaknesses: The main weakness in my opinion is the left-hand side of eq. (2), since it uses contexts from the training data. I agree it is still a meaningful result, but it is also a little hard to interpret. Line 190 claims that "This figure confirms our intuition that the next token distribution is particularly diffuse at the beginning of a sentence, while it decreases for later tokens but remains relatively high. Given how diffuse the distribution is and the large number of possible sentences, it is broadly infeasible to make predictions on new resampled tokens from the empirical distribution alone." I am not fully convinced by this, since we can add some fixed prompt and only measure generalization error for the generated part. Technical Quality: 3 Clarity: 4 Questions for Authors: Figure 2 (right) claims that the left-hand side of eq. (2) is correlated with downstream performance. Specifically, the left y-axis plots accuracies of GPT-2 models, while the y-axis uses samples from the LLAMA model; if I understand correctly, the LLAMA model is treated as an oracle model here, since we do not know the true data-generation process. However, I feel this is a little indirect; can you instead compare downstream performance with your generalization bound (i.e., right-hand side of eq. (2)) for the same model? If this also works, then we can claim that not only the quantity we try to bound is meaningful, but the bound itself is also meaningful. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We really value your thoughtful and supportive feedback! We provide several additional results inspired by your comments. An intuitive application of our bounds to downstream tasks: Inspired by your comments, we provide another intuitive application of our bounds. In this case, to a downstream scientific task. Our token-level generalization bounds are particularly descriptive of antibody design in biology. An antibody sequence is usually composed of 20 different amino acid tokens to bind to a target of interest. In therapeutic antibody design, biologists propose mutations to existing antibody sequences by changing the amino acid tokens at specific positions in the sequence. Recent works have shown that LLMs pretrained on large antibody datasets can be used to propose mutations conditioned on starting antibody sequences. Our token-level generalization bounds match the antibody design setting by bounding the expected next amino acid token negative log likelihood averaged over training contexts that serve as starting sequences for iterative mutations. In the table below, we show that language models based on the Mistral 7B architecture pretrained on a processed subset of the Observed Antibody Sequences (OAS) from scratch achieves non-vacuous token-level generalization bounds: \| Compression Approach \| BPD Bound \| Top-1 Error Bound \| Validation Loss \| \| --- \| --- \| --- \| --- \| \| Mistral 377M \| 2.41 \| 31.60 \| 0.28 \| \| Mistral 212M \| 2.06 \| 26.25 \| 0.30 \| \| Mistral 94M \| 1.62 \| 19.40 \| 0.30 \| \| Random Guess \| 4.86 \| 96.56 \| 1.46 \| Correlation between our bounds and the performance on downstream tasks: Following your suggestion, we compute the direct correlation between our bounds and downstream performance on the tasks reported in Table 6 of the paper. In Figure 1(left y-axis) reported in the attached Rebuttal PDF, we plot the average zero-shot error (Error), defined as 1 - the accuracy, and the perplexity (PPL) achieved by GPT2 small, medium and large on downstream tasks, as reported in Radford et al. [1]. On the right y-axis, we plot the token-level bounds achieved by the GPT2 models with different sizes on the OpenWebText dataset that they were partially trained on. Our token-level BPD bounds achieve 98.9% and 99.4% correlation with the downstream perplexity and error, respectively, and are indeed predictive of generalization on downstream tasks. Given the significance of these results, we included them in our revised manuscript. The distribution over next tokens being diffuse: As you mentioned and as shown in Figure 2 (middle), the average entropy of the next token distribution conditioned on fixed contexts decreases for later token positions compared to the beginning of text token. Therefore, if we have a fixed prompt, this distribution might be less diffuse for the first token that we predict given the prompt. However, the main takeaway from this experiment actually pertains: even for later token positions, the distribution does not collapse entirely and the entropy is non-zero. For instance, if the prompt is composed of 127 tokens, then the average entropy for the next token is equal to 3 bits after transformation to the logarithm base 2, which corresponds to 8 choices for the next tokens. The number of choices is far greater than a single predetermined token, which implies that it is infeasible to make predictions on new resampled tokens from the empirical distribution alone. Therefore, non-vacuous token-level bounds are indicative of generalization beyond the training data. Thank you again for your detailed review. We put a significant effort into our response and would appreciate it if you could consider raising your score. ______ References: [1] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al. Language models are 454 unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.	Summary: This paper presents a novel approach that computes non-vacuous compression-based generalization bounds for LLMs at the billion-parameter scale. Prior works could only achieve vacuous bounds for these large-scale models and rely on the assumption of IID documents. By leveraging the vast number of tokens in LLM training sets and properties of martingales, the authors derive non-vacuous bounds for LLMs that generate high-quality texts. Further, they showcase the tightness of the bounds by examining compression schemes including Monarch matrices and Kronecker factorizations, and post-training quantization techniques. Strengths: 1. The paper tackles an important problem that aims to give guarantees on the generalization abilities of LLMs, which are getting more powerful but the good performance is extremely hard to interpret and assess. 2. Prior works compute non-vacuous bounds on LLMs but rely on assumptions of IID documents and therefore can only be applied to those that generate poor text quality. This work presents a novel approach based on properties of martingales and gives much tighter bounds on LLMs of much more practical capabilities. Further, it does not require altering the pretraining pipeline of the LLMs being analyzed. 3. It investigates the generalizations by examining compression schemes including Monarch matrices and Kronecker factorizations, and post-training quantization techniques. The results also give interesting insights for practitioners. Weaknesses: 1. Since the utilization of martingales is one main theoretical contribution of the work, I feel some background and proof sketch on how they are being used would be better included in the main text. 2. The main models being examined are of the LLaMA and GPT-2 family of models. In particular, the experiment on the chat version of LLaMA is interesting as the generalization gets worse from the supervised finetuning. It would be interesting to see if this is generally true and why this is the case. More experiments on other finetuned LLMs would provide more evidence. 3. From table 2 and 3, it seems that given larger model sizes, the derived bounds get closer to random guess performance. Why is this the case and does it mean the bound would potentially be no-longer meaningful if the model gets large enough? Maybe I misunderstood something and would appreciate some clarification on this point. Technical Quality: 3 Clarity: 3 Questions for Authors: Stated in the prior part. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors note the limitations of the current work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your thoughtful and supportive feedback! We respond to your questions below. Effect of finetuning on downstream task performance: As per your suggestion, we run additional experiments where we finetune the GPT2 large model (774M parameters) – pretrained on the WebText dataset – on multiple downstream tasks: grade school math, stack exchange and planning. These datasets are publicly available on HuggingFace. We report the results in the table below. While finetuning visibly improves the model’s performance on the downstream task, it negatively affects the performance on the pretraining dataset and hence leads to a worse upstream bits-per-dimension (BPD) bound since the compressed size of the model remains approximately constant. \| Model \| Compressed Size (MB) \| Downstream Empirical Risk \| Upstream Empirical Risk \| Upstream BPD Bound \| \| --- \| --- \| --- \| --- \| --- \| \| Pretrained GPT2 Large \| 424.07 \| 3.44 (on planner); 4.56 (on stackexchange) ; 0.45 (on grad school math) \| 4.65 \| 10.47 \| \| GPT2 large finetuned on grade school math \| 420.65 \| 0.016 (on grad school math) \| 6.92 \| 12.72 \| \| GPT2 large finetuned on Stack Exchange \| 424.06 \| 0.14 (on stackexchange) \| 4.80 \| 10.62 \| \| GPT2 large finetuned on Planning \| 424.07 \| 0.001 (on planner) \| 4.78 \| 10.56 \| Trade-off between the empirical risk and the compressed size of the model: Our generalization bounds, as described by Equation 2, can be conceptually written as: $$\text{Expected Risk} \leq \text{Empirical Risk} + \sqrt{\text{Compressed Model Size} / \text{Train Data Size}}.$$ Therefore, the trade-off between the empirical risk and the compressed model size heavily depends on the compression approach and the size of the dataset. Having a dataset that contains a higher number of tokens puts a bigger emphasis on the empirical risk compared to the compressed model size, and vice versa. Likewise, if the compression approach is very aggressive, it will significantly reduce the compressed model size while potentially causing a deterioration of the empirical performance. The rate at which each element of the bound changes dictates whether the bound will increase or decrease for larger models. We demonstrate this effect in Figure 2 in the attached Rebuttal PDF, where we see that an aggressive compression scheme for the GPT2 models consisting of pretraining them in restricted SubLoRA spaces with 25k parameters only then quantizing them leads to an improvement in the token-level bounds as we increase the size of the original model. In contrast, quantizing the LLaMA2 models using QuIP# maintains a good empirical performance but does not reduce the compressed size significantly compared to an approach like SubLoRA, therefore the bounds deteriorate as the models become larger. If the Amber dataset contained more tokens, one would expect that the bounds would improve for larger models given the improvement in the empirical risk. An intuitive application of our bounds: In addition to the above experiments, we ran additional experiments for the antibody design downstream task to provide more intuition on the quantity that we bound .In fact, biologists propose mutations to existing antibody sequences by _changing the amino acid tokens at specific positions in the sequence_. Recent works have shown that LLMs pretrained on large antibody datasets can be used to propose mutations _conditioned on starting antibody sequences_. Our token-level generalization bounds match the antibody design setting by bounding the expected next amino acid token negative log likelihood averaged over training contexts that serve as starting sequences for iterative mutations. We provide results in the general response showing that these bounds are non-vacuous. Background and proof sketch of the main theorem: We provide a sketch of the proof in Appendix B.1, and it consists of three components: set up the empirical loss as the average of a martingale difference sequence, apply Azumas inequality for each hypothesis assigning failure probability proportional to the prior p(h), and then apply a union bound to relate the concentration around the many for individual hypotheses to any given hypothesis that can depend on the training data. We take your point that it would be beneficial to have this summary and some additional background in the main text, to provide additional context for the bounds we use. We will update the paper accordingly. We provide additional experiments in the general response showing that not only is the quantity we bound predictive of downstream performance, but the bounds themselves positively correlate with downstream performance. Thank you again for your supportive feedback. We made a significant effort to address your comments and run several new experiments inspired by your feedback, which we believe have improved our paper. We would appreciate it if you would consider raising your score in light of our response. We would be happy to engage in additional discussion if there are further questions. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I will maintain my score and recommend for acceptance.	null	null	Rebuttal 1: Rebuttal: We thank all the reviewers for their very supportive and helpful feedback. Inspired by the reviewers’ comments, we now report additional results that highlight the following contributions: (i) we complement our other understanding-oriented experiments through an antibody design setting where the task itself naturally matches the definition of the expected risk in our bounds since one would use LLMs pretrained on large antibody datasets to propose mutations conditioned on starting antibody sequences from the training dataset; (ii) we run additional experiments to further investigate the effect of finetuning models on their upstream performance and upstream generalization guarantees; (iii) we show that our bounds are predictive of downstream performance, achieving 98.9% and 99.4% correlation with downstream perplexity and error, respectively; (iv) we empirically demonstrate that the trade-off between the empirical risk and the compressed size of the model depends on the size of the dataset and the compression approach. The new figures and table can be found in the attached Rebuttal PDF. We have also incorporated reviewer feedback to provide additional background on Martingale bounds as well as a proof sketch in the main text in the revised manuscript. We begin with a general response and then address reviewers individually as separate posts. Intuitive interpretation of our bounds for antibody design: We provided several experiments in our submission that help interpret the bounds. We thought it would also be especially exciting to consider a downstream task beyond next-word prediction. We believe this is the first time generalization bounds for language models have been used in such a way — to guarantee downstream performance on a scientific problem. Our token-level generalization bounds are particularly descriptive of antibody design in biology. An antibody sequence is usually composed of 20 different amino acid tokens to bind to a target of interest. An example of an antibody sequence from the Observed Antibody Space (OAS) database is the following: “SETLSLTCTVSGGSMSSY…” [1]. In therapeutic antibody design, biologists propose mutations to existing antibody sequences by changing the amino acid tokens at specific positions in the sequence. In our example sequence, a mutation is introduced if we change one or many amino acid tokens. The next-token prediction task in language modeling thus has a natural interpretation of predicting mutations at position i conditioned on positions <i. Recent works have shown that LLMs pretrained on large antibody datasets can be used to propose mutations conditioned on starting antibody sequences. Our token-level generalization bounds match the antibody design setting by bounding the expected next amino acid token negative log likelihood averaged over training contexts that serve as starting sequences for iterative mutations. In the table below, we show that language models based on the Mistral 7B architecture pretrained on a processed subset of the Observed Antibody Sequences (OAS) from scratch achieves non-vacuous token-level generalization bounds: \| Compression Approach\|BPD Bound\|Top-1 Error Bound\|Validation Loss\| \| ---\| ---\| --- \| --- \| \|Mistral 377M \|2.41\|31.60\|0.28\| \|Mistral 212M\|2.06\|26.25\|0.30\| \|Mistral 94M\|1.62\|19.40\|0.30\| \|Random Guess\| 4.86\| 96.56\|1.46\| Further investigating the effect of finetuning LLMs on upstream performance: In our work, we show that chat versions of the LLaMA models obtain worse bounds on the Amber dataset. We extend these experiments to GPT2 models finetuned for different purposes, namely to answer grade school math questions, coding questions, and to do planning. These downstream datasets are publicly available on HuggingFace. We report the results for pretrained GPT2 Large (774M) in Table 1 of the attached Rebuttal PDF. While finetuning visibly improves the model’s performance on the downstream task, it negatively affects the performance on the pretraining dataset and hence leads to a worse upstream bits-per-dimension (BPD) bound since the compressed size of the model remains approximately constant. Correlation between our bounds and the performance on downstream tasks: Following the suggestion of reviewer mAFv, we compute the correlation between our bounds and downstream performance. In particular, we compute token-level bounds for GPT2 small, medium, and large pretrained with SubLoRA with an intrinsic dimensionality of 25,000 and a LoRA rank of 4 on the OpenWebText dataset. In Figure 1(left y-axis) reported in the attached Rebuttal PDF, we plot the average zero-shot error (Error) and the perplexity (PPL) achieved by GPT2 models on downstream tasks, as reported in Table 6 of the original submission. On the right y-axis, we plot token-level bounds achieved by GPT2 on OpenWebText. Our token-level BPD bounds achieve 98.9% and 99.4% correlation with the downstream perplexity and error, respectively, and are indeed predictive of generalization on downstream tasks. Trade-off between the empirical risk and the compressed size of the model: We demonstrate in Figure 2 in the attached Rebuttal PDF that the trade-off between the empirical risk and the compressed model size heavily depends on the compression approach. Summary: We are thankful for the supportive feedback from the reviewers and believe that their input has made a positive impact on our paper. We make a significant contribution in our work by computing non-vacuous bounds at the LLaMA-70B scale and use our bounds to derive insights about generalization in LLMs, highlighting the remarkable ability of transformer models in capturing longer range correlations, and distinguishing between memorization and reasoning. ___ Reference: [1] Olsen TH, Boyles F, Deane CM. Observed Antibody Space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences. Protein Science. 2022; 31: 141–146. Pdf: /pdf/9e892c01f4ee475da48e3897b7db03e268ecdc75.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Peri-midFormer: Periodic Pyramid Transformer for Time Series Analysis	Accept (spotlight)	Summary: This paper introduces Peri-midFormer to capture multi-periodicity in time series data. Specifically, it designs a pyramid structure and attention mechanisms to effectively model complex temporal variations. The proposed methods demonstrates great performance in several time series analysis tasks in the author's experiments. Strengths: 1. This paper compares a variety of cutting-edge methods. 2. Overall, this paper is solid and the authors conducted thorough experiments. Weaknesses: None. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Were the results of DLinear and PatchTST obtained with a lookback of 96? From the results in Table 12, it seems they were achieved with the default input of 336. Correspondingly, your assertion on line 236 is incorrect. As far as I know, DLinear, PatchTST, and FITS do not use a default input length of 96. FITS has a default lookback of 720. 2. Regarding the Anomaly Detection experiment, did you use point adjustment techniques and manual thresholds? Combining point adjustment techniques with very low manual thresholds (e.g., 0.5%) can easily achieve an F1 score greater than 90 on these datasets [1], even with a randomly generated anomaly score list. Please do not simply state that you followed previous work, as this only exacerbates the legacy issue. Clearly state your stance and clarify this issue appropriately in the paper. 3. Was the ablation study in Table 5 conducted only on the ETTh2 dataset? As far as I know, the ETTh2 dataset is small in scale, has relatively weak periodicity (compared to ETTh1, or multi-period datasets like Electricity and Traffic), and suffers from significant distribution drift. Under such circumstances, any hyperparameter adjustment or other random factors can greatly affect model performance. 4. The Exchange dataset belongs to the financial domain and typically exhibits weak periodicity (due to the unpredictability of financial data). This is evidenced by findings in the DLinear paper, where simply copying the last point can achieve SOTA performance. This contradicts two claims in your paper: (1) Peri-midFormer improves predictive accuracy by extracting periodicity, so how did it achieve SOTA performance on this weak-period dataset? (2) Your Limitation section mentions that Peri-midFormer is not good in scenarios with weak periodicity, yet your experimental results show that your method performs well in this scenario. [1] Wagner, Dennis, et al. "Timesead: Benchmarking deep multivariate time-series anomaly detection." Transactions on Machine Learning Research (2023). Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: As mentioned above, the authors have already discussed the limitations, though I have some concerns for it. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer ECSx Thank you for your detailed review and questions. Please find our answers below. ### Q1: Were the results of DLinear and PatchTST obtained with a lookback of 96? We have reviewed the results for DLinear and PatchTST in Table 12 and confirmed that you are correct—their look-back windows are not 96. We apologize for this mistake. Since most of Table 12's content was referenced from Table 14 of the GPT4TS paper, we assumed that the same look-back window was used across comparison methods. We also apologize for the incorrect assertion in line 236. Upon review, DLinear, PatchTST, and FITS indeed do not use the default look-back window length of 96. We have adjusted the experiments and used longer look-back windows (512 and 720) for these methods to ensure fairer comparisons. We have included the updated experiments in the second part of the global reply. ### Q2: Regarding the Anomaly Detection experiment, did you use point adjustment techniques and manual thresholds? You are correct that our brief statement about following the TimesNet approach was problematic. To address your question first, we did use point adjustment techniques with manual thresholding, and this framework was consistently applied across all comparison methods. Here is a detailed explanation of our anomaly detection strategy: Training Phase: During training, we applied a simple reconstruction loss to help the model learn the distribution of normal data. Testing Phase: For testing, we used the following code (showing only the main parts): ```python def test(self, setting): attens_energy = [] self.anomaly_criterion = nn.MSELoss(reduce=False) # (1) stastic on the train set with torch.no_grad(): for i, (batch_x, batch_y) in enumerate(train_loader): outputs = self.model(batch_x, None, None, None) score = torch.mean(self.anomaly_criterion(batch_x, outputs), dim=-1) attens_energy.append(score) attens_energy = np.concatenate(attens_energy, axis=0).reshape(-1) train_energy = np.array(attens_energy) # (2) find the threshold attens_energy = [] test_labels = [] for i, (batch_x, batch_y) in enumerate(test_loader): batch_x = batch_x.float().to(self.device) outputs = self.model(batch_x, None, None, None) score = torch.mean(self.anomaly_criterion(batch_x, outputs), dim=-1) attens_energy.append(score) test_labels.append(batch_y) attens_energy = np.concatenate(attens_energy, axis=0).reshape(-1) test_energy = np.array(attens_energy) combined_energy = np.concatenate([train_energy, test_energy], axis=0) threshold = np.percentile(combined_energy, 100 - self.args.anomaly_ratio) # (3) evaluation on the test set pred = (test_energy > threshold).astype(int) test_labels = np.concatenate(test_labels, axis=0).reshape(-1) gt = test_labels.astype(int) # (4) detection adjustment gt, pred = adjustment(gt, pred) return ``` The `adjustment` function is defined as: ```python def adjustment(gt, pred): anomaly_state = False for i in range(len(gt)): if gt[i] == 1 and pred[i] == 1 and not anomaly_state: anomaly_state = True for j in range(i, 0, -1): if gt[j] == 0: break else: if pred[j] == 0: pred[j] = 1 for j in range(i, len(gt)): if gt[j] == 0: break else: if pred[j] == 0: pred[j] = 1 elif gt[i] == 0: anomaly_state = False if anomaly_state: pred[i] = 1 return gt, pred ``` We applied point adjustments using `gt, pred = adjustment(gt, pred)` to correct some false positives and false negatives. Additionally, we set a manual threshold with `threshold = np.percentile(combined_energy, 100 - self.args.anomaly_ratio)`, where `combined_energy` is the combined energy score of the training and testing sets, and it is a hyperparameter. The `anomaly_ratio` used for the different datasets are shown in the table below: \| Datasets \| SMD \| MSL \| SMAP \| SWaT \| PSM \| \|:---------------:\|:----:\|:---:\|:----:\|:----:\|:---:\| \| anomaly_ratio \| 0.5 \| 1 \| 1 \| 1 \| 1 \| As shown in the table, we did not apply excessive manual intervention for specific datasets. Of course, this information should have been included in the original paper, which is a shortcoming of our work. We will add this detailed explanation in the revised version. ### Q3: Was the ablation study in Table 5 conducted only on the ETTh2 dataset? Following your suggestions, we have added ablation experiments on several additional datasets, and we have included the results in the third part of the global response. ### Q4: The Exchange dataset belongs to the financial domain and typically exhibits weak periodicity. Our method does indeed maintain strong performance on some datasets with weaker periodicity, which may seem inconsistent with our focus. However, our approach involves more than just the periodic pyramid; it also incorporates time series decomposition. The success of our method on trend-dominant datasets like Exchange is largely due to this approach. Also as you mentioned the DLinear method, which also employs the operation of temporal decomposition, so in our experiments DLinear achieved very good performance on the Exchange dataset. We have provided a detailed response in the global reply, specifically in the fourth section. We had overlooked this phenomenon in the original paper, and your insight has helped us recognize it. Thank you once again for your valuable feedback. If you have any further questions, please feel free to ask. --- Rebuttal Comment 1.1: Comment: Thanks for the response. It is necessary to include the new results and analysis in the revised paper. I have updated my rating. --- Reply to Comment 1.1.1: Comment: Thank you very much for recognizing our work, your suggestions make it more complete! We will add new experimental results, theoretical proofs, and further analysis to the revised version. Thank you for your time in reviewing the manuscript!	Summary: In this paper, the authors proposed a new method Peri-midFormer, which uses the multi-periodicity of time series and modeling the periodic part of a time series in a pyramid way. They further proposed an attention mechanism to use the neighborhood relation in the pyramid. Extensive experiments on different tasks show the effectiveness of the proposed method. After reading the rebuttal, I raise my rating from 5 to 6. Strengths: - The idea of using attention in the pyramid structure seems to be novel. - Extensive experiments are conducted for different tasks on benchmark datasets. - The proposed model is light-weight. Weaknesses: - Some important related works are missing. - There are other related works utilizing the idea of modeling time series in multi-scale, e.g. [1] [2] [3]. - Experiments could be improved. - The authors should clearly state whether they reproduce the results for comparison methods or they copy the numbers from the paper. - In the original paper of PatchTST, the authors use a longer context window than 96. The authors are suggested to tune this parameter for all comparison methods (as many papers did) and provide the best results, rather than fix the context window. - The authors should provide the script for hyper-parameter tuning to improve the reproducibility of the paper. - The ablation study (Table 5) is only shown on one dataset (ETTh2). The authors are suggested to include results in more datasets. - The training and inference complexity and actual time could be analyzed, as well as memory efficiency. [1] TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. ICLR 2024. [2] PERIODICITY DECOUPLING FRAMEWORK FOR LONGTERM SERIES FORECASTING. ICLR 2024. [3] Disentangling Structured Components: Towards Adaptive, Interpretable and Scalable Time Series Forecasting. TKDE 2024. Technical Quality: 3 Clarity: 4 Questions for Authors: - Apparently, the proposed method focused on periodic part forecasting. However, it works pretty well on the non-periodic forecasting task (short-term forecasting). For instance, there is no periodic in the 'Yearly' dataset, the proposed method still performs quite well. Can authors explain the possible reasons? - It is not clear why the proposed methods have a smaller number of parameters than PatchTST. Can the authors explain more? - It seems that the trend part is not used for classification tasks (from Figure 7). Can the author explain more? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Focused only on periodic signal Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer jAZw Thanks for your valuable comments. We will explain your concerns point by point. ### Q1: Some important related works are missing Thank you for your reminder. We have carefully read the paper you provided and found them very helpful for improving the background of our paper. We will include these works in the related work section of the revised version. ### Q2: Experiments could be improved Thank you very much for your valuable feedback. Here are our additions or modifications: 1. The authors should clearly state whether they reproduce the results for comparison methods or they copy the numbers from the paper. We apologize for the oversight in the original paper. We have reproduced some of the experimental results of the comparison methods, but due to time constraints, we also included results from the papers of these methods. Specifically, for the long-term forecasting task, we reproduced results for TSLANet and FITS, while other results were taken from previous papers: (1) Time-LLM and iTransformer are from Table 13 of the TSLANet paper. (2) GPT4TS, Dlinear, PatchTST, TimesNet, FEDformer, Autoformer, Stationary, ETSformer, LightTS, Informer, and Reformer are from Table 14 of the GPT4TS paper. We will include these details in the revised version. 2. The authors are suggested to tune this parameter for all comparison methods and provide the best results. We apologize for this oversight. To ensure fair comparison, we have adjusted the look-back window for other comparison methods as per your suggestion and re-conducted the experiments. We have included this information in the second part of the global reply. 3. The authors should provide the script for hyper-parameter tuning to improve the reproducibility of the paper. Thank you for your valuable feedback. We will explain our hyperparameter tuning script with the following example and will release the script along with all the code to improve reproducibility: For the long-term forecasting task on the Electricity dataset with a prediction length of 96: ``` for d_model in 64 128 256 512 768 do for layers in {1..5} do for top_k in {2..5} do for batch_size in 4 8 16 32 64 do for learning_rate in 0.0001 0.0002 0.0005 0.001 0.002 do python -u run.py \ --layers $layers \ --d_model $d_model \ --top_k $top_k \ --learning_rate $learning_rate \ --batch_size $batch_size \ --...... done done done done done ``` Based on this script, we recorded results from multiple experiments and selected the best-performing hyperparameters. 4. The ablation study is only shown on one dataset (ETTh2) Based on your suggestion, we have added more ablation experiments, and we have included this information in the third part of the global reply. 5. The training and inference complexity and actual time could be analyzed, as well as memory efficiency. Thank you for your suggestion. We have added the experiments as you recommended,and show it in the first part of the global reply section. ### Q3: There is no periodic in the 'Yearly' dataset, the proposed method still performs quite well Thank you for your insightful comments. Our method does indeed maintain good performance on datasets with weaker periodicity, which might seem mismatched with the focus of our approach. However, due to our use of time series decomposition, our method can effectively capture future trends by predicting the trend part even in datasets like Yearly that do not have obvious periodicity but exhibit significant trends. This is illustrated by the "trend part" in Figure 7 of the original paper. We have provided a detailed response in the global reply, specifically in the fourth section. ### Q4: It is not clear why the proposed methods have a smaller number of parameters than PatchTST Thank you for your insightful comments. The main reasons for this phenomenon are: 1. Number of patches: If we consider each periodic component in our method as a patch, our total number of patches is smaller than that of PatchTST. Specifically, PatchTST uses a fixed number of 64 patches (there are two versions, 64 and 42, and the 64 version was used in the experiments), while the number of patches in our method is variable. In the experiments shown in Figure 16 of the original paper, we set $k$ to 3 for both tasks, resulting in 40 and 32 patches after periodic decomposition, which is less than the 64 patches in PatchTST. 2. Model Size Variance: Not all comparison methods use the same model size. We refer to what was done in the TimeaNet experiments and retain the hyperparameters of the models of the original method, thus maximizing the best possible performance of the individual models. In the experiments shown in Figure 16, the PatchTST model has a dimension of 512 and 3 layers, while our model has a dimension of 16 and 2 layers. This difference in model size is another reason for the smaller parameter count in our method compared to PatchTST. ### Q5: It seems that the trend part is not used for classification tasks. Thank you for your insightful comments. As you noted, the trend part is not used in classification tasks (without time series decomposition) but follows the path indicated by the red arrow. In classification tasks, reconstructing the original data is unnecessary, so the trend part is not extracted and added back. Additionally, the trend part is a crucial discriminative feature for classification data, so it should not be separated from the original data before feature extraction. We acknowledge that this aspect was not explained in detail in the original paper, and we will provide a more thorough explanation in the revised version. Thank you again for your valuable time and constructive feedback. If you have any further questions, please feel free to ask. --- Rebuttal Comment 1.1: Comment: Thanks for the authors for the response in the rebuttal. My main concerns about experiments are addressed. Therefore, I would like to raise my rating from 5 to 6. --- Reply to Comment 1.1.1: Comment: Thank you very much for recognizing our work, your suggestions are very helpful in improving the quality of our work, and thank you for your valuable time in reviewing the manuscript!	Summary: The paper introduces Peri-midFormer, a novel transformer-based architecture designed for time series analysis. By leveraging the multi-periodicity inherent in time series data, the model constructs a Periodic Pyramid structure that decouples complex periodic variations into inclusion and overlap relationships among different periodic components. The proposed method incorporates self-attention mechanisms to capture dependencies between these periodic components, achieving state-of-the-art performance across five mainstream time series tasks: short- and long-term forecasting, imputation, classification, and anomaly detection. Strengths: S1. The concept of decoupling time series data into a Periodic Pyramid is a valid point. This new representation seems to capture the multi-periodicity of time series effectively. S2. The effectiveness of the proposed method is extensively verified on five tasks. Weaknesses: W1. While the Periodic Pyramid and self-attention mechanisms are well-explained, the model's complexity might pose challenges for practical implementation and scalability, especially for users with limited computational resources. It would be useful if the authors could report in the main body of the paper, at least the main conclusion of the training and inference time for the proposed method, compared with existing baselines, across the five tasks. W2. The improvements provided by Peri-midFormer seem to be insignificant when compared with Time-LLM and GPT4TS on forecasting, imputation, and anomaly detection tasks. It would be better if the author could further discuss this in the main body of the paper (e.g., discuss the time-accuracy tradeoff as shown in Appendix E.4 Complexity Analysis). Technical Quality: 3 Clarity: 4 Questions for Authors: Q1 (cr. W1). Add in the main body of the paper a discussion about the training and inference time, as well as some critical results and main conclusions. Q2 (cr. W2). Add in the main body of the paper a discussion about the significance of the improvements or about the time-accuracy tradeoff for Peri-midFormer against Time-LLM and GPT4TS on forecasting, imputation, and anomaly detection tasks. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: L1. It is not clear how the proposed periodic pyramid can be effectively and efficiently integrated into multi-dimensional time series. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer GBMh Thanks for your valuable comments. We will explain your concerns point by point. ### Q1: (cr. W1). Add in the main body of the paper a discussion about the training and inference time, as well as some critical results and main conclusions. Thank you very much for your valuable feedback. Concerns about computational complexity and scalability were also raised by other reviewers. Therefore, we have supplemented our experiments in this area, including the complexity, time, and memory efficiency of training and inference. Please refer to the first part of the global reply for details. We will include this content in the revised version to make the main text more complete. ### Q2: (cr. W2). Add in the main body of the paper a discussion about the significance of the improvements or about the time-accuracy tradeoff for Peri-midFormer against Time-LLM and GPT4TS on forecasting, imputation, and anomaly detection tasks. Thank you very much for your valuable feedback. Our method does indeed perform worse than Time-LLM and GPT4TS on some tasks. However, due to page limitations, we placed the complexity validation in the appendix, which may affect readers' understanding of the paper. Additionally, we did lack validation and discussion on the time-accuracy trade-off in the original paper. We have addressed this in the first part of the global reply section and will include it in the revised version. In addition, due to the limitation of the amount of content that can be displayed, we currently only supplemented our experiments on the long-term forecasting task, and if necessary, we will also supplement more experiments on the imputation and anomaly detection tasks. ### Q3: L1. It is not clear how the proposed periodic pyramid can be effectively and efficiently integrated into multi-dimensional time series. We apologize for not clearly explaining the operational mechanism of our method in the original paper. The figures only showed the operation for a single channel, and although we mentioned retaining the original channels in line 132, this did not effectively convey our approach. We should have emphasized our use of a channel-independent strategy and clarified that the figures only illustrate operations for one channel. This will be addressed in the revised version. Thank you again for your valuable feedback, which has helped make our work more complete. If you have any further questions, please feel free to ask. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for addressing my concerns; I remain positive about this work as earlier. Look forward to future versions with more detailed discussions and new results as proposed. --- Reply to Comment 1.1.1: Comment: Thank you for recognizing our work! In future versions, we will add all new experimental results, theoretical proofs, and further discussions to make our work more complete. Thank you again for your valuable time in reviewing our paper!	Summary: The abstract succinctly introduces the Peri-midFormer, a novel approach designed for time series analysis, acknowledging the challenges posed by the discrete nature of time series data and the complexity of capturing periodic variations directly. It proposes a method to address these challenges by decomposing complex periodic variations into hierarchical periodic components, termed the periodic pyramid. This approach leverages inclusion and overlap relationships among these components, mimicking the natural pyramid structure observed in time series data. Strengths: Innovative Approach: The concept of a periodic pyramid to model time series data is innovative and promises to address the limitations of traditional methods that struggle with capturing complex periodic patterns. Hierarchical Representation: By representing time series as a pyramid with progressively shorter periodic components, the model potentially enhances the understanding of multi-scale temporal relationships. Self-Attention Mechanism: Incorporating self-attention into the periodic pyramid allows capturing intricate relationships among periodic components, which is crucial for tasks like anomaly detection and forecasting. Weaknesses: Complexity and Scalability: The introduction of a hierarchical pyramid structure combined with self-attention could potentially introduce computational complexities and scalability issues, especially with larger datasets or real-time applications. Addressing these concerns in the paper would strengthen its practical utility. State-of-the-Art Comparison: It is not explicitly stated whether Peri-midFormer achieves state-of-the-art (SOTA) performance when compared to strong baseline models. Without clear comparative results, it is difficult to ascertain if the proposed method truly represents an advancement over current leading approaches in time series analysis. Interpretability: The abstract lacks interpretability regarding why the periodic pyramid structure exists in the applications and why it plays a critical role in improving forecasting. Providing a rationale or theoretical justification for the efficacy of the pyramid structure in capturing temporal patterns would enhance the understanding and acceptance of the proposed method. Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: see weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer W65V Thank you for your detailed review and questions. Please find our answers below. ### Q1: Complexity and Scalability We have adopted the suggestions from you and two other reviewers to include additional experiments, covering the complexity, time, and memory efficiency of both training and inference, to further validate the method's complexity and scalability. Please refer to the first part of the global reply section for details. ### Q2: State-of-the-Art Comparison In the original paper, we stated that our method achieved SOTA results (see line 77). Additionally, we compared the performance of various methods in the explanation of each experiment's results, highlighting the outstanding performance of our method. However, this may still be insufficient and could require more thorough discussion of the comparative results. If possible, we will provide a more comprehensive analysis of the comparative results in the revised version. ### Q3: Interpretability The abstract indeed lacks an explanation of the periodic pyramid structure in time series, which affects the understanding of the method. We will include this content in the revised version. Additionally, the lack of a clear explanation of the fundamental principles of the periodic pyramid makes it seem unsupported. Therefore, we have re-examined our method and analyzed its fundamental principles as follows: To demonstrate the essence of attention computation among multi-level periodic components, we need to analyze how the interactions between periodic components at different levels affect the final feature extraction. In time series analysis, different periodic components correspond to different time scales. This means that through decomposition, we can capture components of various frequencies within the time series. The essence of the periodic pyramid is to capture these different frequency components through its hierarchical structure. Using single-channel data as an example, and given that we adopt an independent channel strategy, this can be easily extended to all channels. Assume the time series $x(t)$ can be decomposed into multiple periodic components ${x_n}(t)$ : $$x(t) = \sum\limits_{n = 1}^N {{x_n}} (t) \tag{1}$$ Taking two different periodic components as examples: $${x_i}(t) = {A_i}\sin \left( {\frac{{2\pi t}}{{{T_i}}} + {\phi _i}} \right),{x_j}(t) = {A_j}\cos \left( {\frac{{2\pi t}}{{{T_j}}} + {\phi _j}} \right) \tag{2}$$ where $A$ is amplitude, $T$ is period, and $\phi $ is phase. Due to the overlap and inclusion relationships between different periodic components, we employ an attention mechanism in the periodic pyramid to capture the similarities between different periodic components, focusing on important periodic features. When applying the attention mechanism, we have: $${Q_i} = {W_Q}{x_i}(t),\quad {K_j} = {W_K}{x_j}(t),\quad {V_j} = {W_V}{x_j}(t) \tag{3}$$ where ${W_Q}$ 、${W_K}$ and ${W_V}$ are learnable weight matrices. From equations (2) and (3): $${Q_i} = {W_Q}{A_i}\sin \left( {\frac{{2\pi t}}{{{T_i}}} + {\phi _i}} \right),\quad {K_j} = {W_K}{A_j}\cos \left( {\frac{{2\pi t}}{{{T_j}}} + {\phi _j}} \right) \tag{4}$$ Further, the dot-product attention can be expressed as: $${Q_i}K_j^T = {A_i}{A_j}\left( {{W_Q}\sin \left( {\frac{{2\pi t}}{{{T_i}}} + {\phi _i}} \right)} \right){\left( {{W_K}\cos \left( {\frac{{2\pi t}}{{{T_j}}} + {\phi _j}} \right)} \right)^T} \tag{5}$$ Using the trigonometric identity $\sin (a)\cos (b) = \frac{1}{2}[\sin (a + b) + \sin (a - b)]$, the dot-product ${Q_i}K_j^T$ can be further expressed as: $${Q_i}K_j^T = \frac{1}{2}{A _i}{A _j}\left\\{ {{W _Q}\left[ {\sin \left( {\frac{{2\pi t}}{{{T _i}}} + {\phi _i} + \frac{{2\pi t}}{{{T _j}}} + {\phi _j}} \right) + \sin \left( {\frac{{2\pi t}}{{{T _i}}} + {\phi _i} - \frac{{2\pi t}}{{{T_j}}} - {\phi _j}} \right)} \right]} \right\\}{\left( {{W_K}} \right)^T} \tag{6}$$ Based on this, considering the periodicity and symmetry of $\sin (a + b)$ and $\sin (a - b)$, when the periods of two time series components are close / same (intra-level attention in the pyramid, see the right side of Figure 3 in the original paper) or have overlapping / inclusive parts (inter-level attention in the pyramid, see the right side of Figure 3 in the original paper), the values of these two sine functions will be highly correlated, resulting in a large ${Q_i}K_j^T$ value. This indicates that the periodic pyramid model can effectively capture similar periodic patterns across different time scales. Next, incorporating this into the calculation of the attention score: $${s _{ij}} = \frac{{\exp \left( {\frac{{{Q _i}K _j^T}}{{\sqrt {{d _k}} }}} \right)}}{{\sum\limits _{m} {\exp } \left( {\frac{{{Q _i}K _{m}^T}}{{\sqrt {{d _k}} }}} \right)}} \tag{9}$$ where $m$ denotes the index of all key values, including $j$. It can be seen that the attention scores between highly correlated periodic components will be higher, which we have already validated in Figures 13 and 14 of the original paper. From the above derivation, it can be seen that the attention mechanism can measure the similarity between different periodic components. This similarity reflects the alignment between different periodic components in the time series, allowing the model to capture important periodic patterns. By capturing these periodic patterns, the periodic pyramid can extract key features of the time series, resulting in a comprehensive and accurate time series representation. This representation not only includes information across different time scales but also enhances the representation of important periodic patterns. Due to character limitations, there are parts of the derivation will be explained inside the official comment. The above proof partially explains the effectiveness of the periodic pyramid feature extraction method. We hope this addresses your concerns. If you have any further questions, please feel free to ask. --- Rebuttal Comment 1.1: Comment: Thank the authors for the detailed reponse. I have raised my rating accordingly. --- Reply to Comment 1.1.1: Comment: Thank you very much for recognizing our work! Your review comments have made it more complete, especially with the additions to the theoretical proof. We will incorporate these improvements into the revised version. Thank you for taking the time to review our manuscript! --- Rebuttal 2: Title: Additions to replies Comment: # Additions to the answer to the third question (Q3: Interpretability): ### 1. Qualitative analysis of the essence of the periodic pyramid: Qualitatively analyzing the essence of feature extraction through the periodic pyramid lies in the combination of its multi-level structure and attention mechanism: (1) Multi-level Structure: By decomposing the time series into multiple periodic components, the periodic pyramid can capture features at different time scales. This decomposition allows the model to handle both short-term and long-term dependencies at various levels. (2) Attention Mechanism: The attention mechanism can adaptively focus on the most relevant parts of the periodic pyramid, enhancing the model's focus on important features. (3) Feature Aggregation: By aggregating features from different levels, the periodic pyramid can generate a comprehensive representation of the time series that includes information from all periodic components. This aggregation ensures that the model can fully capture the complex dynamic patterns in the time series. ### 2. Additional proof of continuation of Equation (9): Further, the attention vector ${{\bf{a}}_i}$ of ${x_i}(t)$ can be obtained as: $${{\bf{a}}_i} = \sum\limits_m {{s _{im}}} {V_m} \tag{10}$$ where ${V_m} = {W_V}{x_m}(t) = {W_V}{A_m}\cos \left( {\frac{{2\pi t}}{{{T_m}}} + {\phi _m}} \right)$, therefore ${{\bf{a}}_i}$ can be expressed as: $${{\bf{a}} _i} = \sum\limits_m {\frac{{\exp \left( {\left\\{ {{W _Q}\left[ {\sin \left( {\frac{{2\pi t}}{{{T _i}}} + \frac{{2\pi t}}{{{T _m}}} + {\phi _i} + {\phi _m}} \right) + \sin \left( {\frac{{2\pi t}}{{{T _i}}} - \frac{{2\pi t}}{{{T _m}}} + {\phi _i} - {\phi _m}} \right)} \right]} \right\\}{{\left( {{W _K}} \right)}^T}/\sqrt {{d _k}} } \right)}}{{\sum\limits _{m} {\exp } \left( {\left\\{ {{W_Q}\left[ {\sin \left( {\frac{{2\pi t}}{{{T _i}}} + \frac{{2\pi t}}{{{T _{m}}}} + {\phi _i} + {\phi _{m}}} \right) + \sin \left( {\frac{{2\pi t}}{{{T _i}}} - \frac{{2\pi t}}{{{T _{m}}}} + {\phi _i} - {\phi _{m}}} \right)} \right]} \right\\}{{\left( {{W _K}} \right)}^T}/\sqrt {{d _k}} } \right)}}} {W _V}{A _m}\cos \left( {\frac{{2\pi t}}{{{T _m}}} + {\phi _m}} \right) \tag{11}$$ where $m$ is the same as in Equation (7) in the original paper and is used for selecting components that have interconnected relationships with ${x_i}(t)$. Equation (11) is the expanded form of Equation (7) in the original paper, which leads to an explanation for the good performance of the Periodic Pyramid Attention Mechanism in capturing the periodic properties of the different levels in the time series.	Rebuttal 1: Rebuttal: # General Responses We thank the Reviewers for the insightful comments and detailed feedback. Here's the global reply. ### 1. Validation of Computational Complexity and Scalability Following the suggestions of several reviewers, we have supplemented our experiments with tests on computational complexity and scalability, specifically including training and inference complexity, actual time, and memory usage. We conducted these experiments on larger datasets (Electricity and ETTh1) compared to the ETTh2 dataset used in the original paper. The results are presented in Tables 1 and 2 (Due to the long training time of Time-LLM, the relevant metrics for its inference on the Electricity dataset have not been collected yet). As shown, our proposed Peri-midFormer demonstrates a significant advantage in computational complexity on the Electricity dataset, without the excessive inference time concerns raised by several reviewers, and achieves the lowest MSE. Similarly, on the ETTh1 dataset, the computational overhead and inference time of Peri-midFormer do not pose a disadvantage. In fact, while achieving an MSE second only to Time-LLM, Peri-midFormer’s computational overhead and inference time are substantially lower than those of Time-LLM. This analysis demonstrates that our method has notable advantages in terms of computational complexity and scalability. ### 2. Adjustment of Lookback Window for Comparison Methods Due to our oversight, the original paper inaccurately described the lookback window for some comparison methods and used inappropriate lookback windows for others, leading to less rigorous experiments. Based on suggestions from several reviewers, we have adjusted the lookback windows for some comparison methods. Specifically, we have set the lookback windows for FITS, DLinear, PatchTST, TimesNet, and Pyraformer to 512, consistent with our Peri-midFormer. Additionally, since FITS originally had a lookback window of 720, we have included experiments with a 720 look-back window for comparison with Peri-midFormer. The results are presented in Table 3. As seen from the table, several comparison methods benefit from the extended look-back window, showing some improvement in prediction performance compared to the results in Table 12 of the original paper. However, there is still a noticeable gap compared to our Peri-midFormer. Furthermore, in the comparison with a 720 look-back window, FITS does not perform as well as our Peri-midFormer. This adjustment ensures a fairer experiment and highlights the advantages of Peri-midFormer. ### 3. Ablation Study Adjustment In the original paper, our ablation study was only validated on the ETTh2 dataset. Based on suggestions from several reviewers, we have expanded the experiments to include more datasets. We have supplemented the ablation study with experiments on the ETTh1, Electricity, Weather, and Traffic datasets, as shown in Table 4. It can be seen that each module we proposed performs effectively across multiple datasets, further demonstrating the superior performance of Peri-midFormer. ### 4. Explanation for Peri-midFormer's Strong Performance on Non-periodic Datasets In the original paper, we mentioned that Peri-midFormer excels on datasets with strong periodicity, and performs poorly on those with weak periodicity. However, as pointed out by several reviewers, Peri-midFormer has shown outstanding performance on the Exchange dataset in long-term forecasting tasks, and on the Yearly dataset in short-term forecasting tasks, both of which lack clear periodicity. This contradicts our initial description. Upon careful examination of the Exchange and Yearly datasets, we found that while they indeed lack obvious periodicity, they exhibit strong trends, as shown in Figure 1-6. This explains why Peri-midFormer performs well on these datasets. Peri-midFormer employs a temporal decomposition strategy, which involves separating the trend part from the original data before partitioning the periodic components. The trend part is then added back after the output of Peri-midFormer, as illustrated in Figure 7 of the original paper. To be candid, we adopted temporal decomposition to mitigate the influence of the original data's periodic characteristics, thereby enhancing Peri-midFormer's effectiveness. This approach was repeatedly explained in the original paper. Thanks to temporal decomposition, Peri-midFormer can leverage trend prediction to achieve excellent performance on datasets like Exchange and Yearly, which exhibit strong trends. We sincerely thank the reviewers for their valuable suggestions, which have made our work more comprehensive. If possible, we will incorporate all supplementary content in the revised version. Once again, we appreciate the reviewers' time and insightful feedback! Pdf: /pdf/b3afb5cd4122e58fbb0a0bdbbf49105417243969.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Testably Learning Polynomial Threshold Functions	Accept (poster)	Summary: This paper studies testably learning an n-dimensional Polynomial Threshold function using a reduction proved in a previous work called fooling. The authors give an analysis of a construction of fooling multilinear PTF and then further for fooling arbitrary PTF. The paper completes itself with proof that push-forward cannot learn PTF and that fooling is the best they can have. Strengths: The topic is very interesting, as there is a line of work on general testable learning and testable learning in specific classes like halfspace. Testable learning PTF is the natural next step. Though the authors use the technique of reducing testable learning to fooling from previous work, it seems the analysis of the construction of a fooling for multilinear PTF is also an important technique. As the authors mentioned, the previous fooling construction [GKK23] only works for degree 2 PTF, and the construction from [Kane19] needs some careful analysis to work for PTF with more than constant degree. The paper is complete as they also show fooling is necessary as another more direct approach cannot testably learn PTF. Weaknesses: The last part of the paper shows that push-forward cannot learn PTF seems very interesting, but it seems hard for me to determine how significant the contribution is from the analysis of fooling multilinear PTF. Technical Quality: 3 Clarity: 3 Questions for Authors: I am wondering what are the original techniques the authors want to emphasize. Like using Taylor expansion to bound additional error terms or reducing PTF to multilinear PTF or anything else? I am wondering if the authors can emphasize a bit more. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We want to thank the reviewer for the kind feedback. We appreciate that the reviewer thinks of testable learning of PTFs as a natural problem to consider. The primary motivation for including the impossibility result for proving testable learning guarantees via the push-forward is to show that a straightforward approach (which yields reasonable results for halfspaces [RV23]) does not yield anything for PTFs. This in turn motivates our use of the (more complicated) techniques of Kane [Kan11] (i.e., fooling). Thereby, it partially explains the difference in dependence on $d$ between our result and existing results for agnostically learning PTFs. Regarding the question raised by the reviewer, we think the following are the main technical contributions of our paper, and we plan to emphasize these more in the revision: - First, as already alluded to by the reviewer, we bound the size of additional error terms arising from a Taylor expansion (see lines 290-307 of our paper), which do not appear in [Kan11]. We could imagine that our analysis of these terms could also be applied to different testable learning problems in the future. - Second, we generalize the arguments given by Kane to move from multilinear to arbitrary PTFs (lines 335-346 of our paper). There, we showed that even under the weaker assumption of approximate moment-matching, we can show that his construction also works in our setting. In particular, we were able to circumvent the use of moments of high degree (i.e., depending on $n$), which would have given us only quasi-polynomial runtime, i.e., $n^{\log(n)}$. On the contrary, we showed that fooling arbitrary PTFs needs the same degree of moment-matching as fooling multilinear ones, which allowed us to conclude our main result. [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity [RV23]: Testing Distributional Assumptions of Learning Algorithms, Ronitt Rubinfeld, Arsen Vasilyan, Proceedings of the 55th Annual ACM Symposium on Theory of Computing --- Rebuttal Comment 1.1: Comment: Thank you for your response. I think this paper will be worth reading for people interested in testable learning and more enjoyable if the authors improving their writing in the final version in the way they claimed in this rebuttal. I raised my score to 7.	Summary: Background: Agnostic learning is a well-studied framework that models learning when no function is some hypothesis class F describes the data perfectly. Specifically, the agnostic learning framework requires the learning algorithm to give a hypothesis whose classification error is at most opt+$\epsilon$, where opt is the best prediction error among all hypotheses in the class F. Almost all existing agnostic learning algorithms are distribution-specific, i.e. they assume the examples are drawn from some distribution, for example a Gaussian distribution. A distribution-specific agnostic learning algorithm lacks in reliability, because it is allowed to output an extremely poor classifier if the examples do not come from e.g. Gaussian (or some other assumed distribution). Yet, fully eliminating such assumptions is shown to be impossible for many basic function classes (based on well-established cryptographic assumptions). Testable learning is a framework that aims to mitigate the above mentioned limitation, by allowing the algorithm to abstain on a specific dataset if the examples do not come from the assumed distribution. Overall, this allows a user to be confident that the classifier indeed has error of at most opt+$\epsilon$, as required by the agnostic learning framework. Testable learning has been a focus of many works in recent years (see the paper for references). The paper studies testable learning of polynomial threshold functions (PTFs) under the Gaussian distribution. I.e. the function class F considered in this work consists of functions of the form sign$(p(x))$, where p is a degree-$d$ polynomial. The work in n dimensions, with accuracy parameter $\epsilon$, the paper gives an algorithm for testable learning of constant-degree PTFs with a run-time of $n^{poly(1/\epsilon)}$}. The paper is based on the moment-matching framework of [Gollakota, Klivans, Kothari ‘23], and shows that the direct approach used in [Vasilyan, Rubinfeld ‘23] to handle linear threshold functions cannot be extended to PTFs. In order to apply the moment-matching framework of [Gollakota, Klivans, Kothari ‘23], the paper shows that PTFs are “fooled” by distributions whose low-degree moments are close to Gaussian moments. To do this, the paper expands on the approach of [Kane ‘11] that proves a less general statement that PTFs are “fooled” by distributions for which the marginal of every k coordinates equals to the k-dimensional Gaussian. The proof first first considers multilinear PTFs, and then reduces the case of general PTFs to that of multilinear PTFs. Strengths: - Polynomial threshold functions are an extremely well-studied class of hypotheses that has been the focus of many works in learning theory (including in works that appeared in NeurIPS). - Previously no testable learning algorithms were known even for degree-2 polynomial threshold functions. - The run-time of $n^{poly(1/\epsilon)}$ qualitatively matches the best run-time for the agnostic learning of polynomial threshold functions. For example, existing hardness results preclude run-times such as $poly(n/\epsilon)$ or $n^{polylog(1/\epsilon)}$. - Studying polynomial threshold functions naturally extends previous works that study linear threshold functions. Weaknesses: -The run-time dependence of the algorithm on the degree d of the PTF can conceivably be sub-optimal. As explained on page 12, the run-time is $(n \epsilon)^{O_d(\epsilon^{-4d7^d})}$, whereas it is conceivable that this run-time could potentially be improved in the future to $(n \epsilon)^{poly(d/\epsilon)}$ Technical Quality: 4 Clarity: 4 Questions for Authors: The paper mentions that the analysis does not require to extend the Carbery-Wright inequality to distributions whose low-degree moments match those of the Gaussian. Could you give some more high-level intuitive explanation for why this is the case? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: I think that the limitations are discussed adequately Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We wish to thank the anonymous reviewer for their kind feedback. We are encouraged that the reviewer believes the problem we study is well-motivated and naturally extends previous work and that the reviewer appreciates that our work qualitatively matches existing lower bounds even for the agnostic setting. We agree that it is conceivable (but likely difficult) that the runtime dependence on $d$ could be improved (or potentially lower bounds could be established). We address this in more detail in the general rebuttal and would refer the reviewer there. We think of this as an interesting direction for future research. Regarding the question raised by the reviewer, we were also somewhat surprised that we do not need an analogue of Carbery-Wright for moment-matching distributions. On a high-level, the reason is as follows. The point in the analysis when we need such an anti-concentration result is once we have shown that $\mathbb{E}[\mathrm{sign}(p(X) \pm O_d(\varepsilon^d))] \approx \mathbb{E}[\mathrm{sign}(p(Y))] \pm O(\varepsilon)$ (where as in our paper $Y$ is Gaussian and $X$ is approximately moment-matching). Now, we would like to say that the left-hand side is roughly the same as $\mathbb{E}[\mathrm{sign}(p(X))]$, which would exactly be an analogue of Carbery-Wright for moment-matching distribution (i.e. showing that the probability that $p(X)$ is small is low). However, the trick here (which was already used in [Kan11]) is to apply the above to the polynomial $p \mp O_d(\varepsilon^d)$ and thus shift the additional factor to the side with the Gaussian $Y$, where we then can apply Carbery-Wright (for Gaussians). Thus, once we have a result relating $\mathrm{sign}(p(Y))$ and $\mathrm{sign}(p(X) \pm O_d(\varepsilon^d))$, by changing the polynomial slightly, we can shift the additional factors to the $Y$ side which allows us to use Carbery-Wright for the Gaussian instead of an extension to approximately moment-matching distributions. [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity --- Rebuttal Comment 1.1: Comment: Thank you for your response.	Summary: This paper studied the problem of testably learning polynomial threshold functions (PTFs). The authors aimed to answer the question of whether PTFs are qualitatively harder to learn in the testably learning model, compared to the agnostic learning model. The authors answered the question in the negative, showing that the degree-d PTFs can be testably learned up to $\epsilon$ with respect to the standard gaussian in time and sample complexity $n^{poly(1/\epsilon)}$, which qualitatively matches the $n^{O(d^2/\epsilon^4)}$ sample complexity of agnostic learning degree-d PTFs. To prove the above result, the authors linked testable learning with distribution fooling, building upon previous results on polynomial approximations. The authors also showed that it is impossible to testably learn PTFs with the techniques from [RV23]. Strengths: 1. The paper provided the first sample and computational complexity on testably learning PTFs, showing that testably learning PTFs is qualitatively similar in hardness to agnostic learning PTFs. To reach this result, the authors overcome a handful of technical obstacles that arose in adapting the fooling techniques to testable learning. Critically, the authors constructed a new low-degree polynomial to approximate the PTFs based on [Kane11] with refined approximation error bounds. The techniques the authors applied here could be of independent interest. Weaknesses: 1. Though the final sample complexity is $n^{poly(1/\epsilon)}$, the $poly(1/\epsilon)$ is of order $\epsilon^{-7^d}$, in other words, it is substantially worse than agnostic learning in terms of the order of $1/\epsilon$. It might be too harsh to say this is a serious weakness of this paper, as this is the first paper that provided these kinds of complex results; I think it would be an interesting future work to reduce the order of $1/\epsilon$ that is truly comparable to agnostic learning. Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness above. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: the authors have addressed the limitations properly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First, we would like to thank the reviewer for their kind feedback. We are encouraged by the fact they appreciate that we give the first result on testably learning PTFs. We agree with the reviewer that it would be interesting future work to try to improve the runtime, potentially to $n^{\mathrm{poly}(d/\varepsilon)}$. We address the dependence of the runtime on $d$ in the general rebuttal. Briefly, our worse runtime dependence on $d$ (w.r.t. the agnostic model) is inherited from the result of [Kan11], which we build on. An improvement of the runtime using our techniques would directly improve the result of [Kan11], which has stood for over 10 years. Furthermore, under widely believed hypotheses, the best runtime we could hope for is $n^{\mathrm{poly}(d/\varepsilon)}$. [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity	Summary: The authors study the problem of testing polynomial threshold functions in the agnostic setting in the testable learning paradigm. They present a testable learning algorithm that matches with the asymptotic bound (in terms of n) known for agnostic learning Strengths: Testing of polynomial threshold functions is a very natural problem and the testable learning paradigm is also very natural. The authors present the first such testable learning algorithm in the agnostic learning setting for PTFs. Weaknesses: The dependence on epsilon and the degree is very bad. This makes the results completely useless is practice. In fact the degree os the polynomial on the exponent is exponentially dependent on d. Technical Quality: 3 Clarity: 3 Questions for Authors: Can you please argue why the dependence on d is so bad and is there possibility of improving it. Can a better bounds be obtained for smaller d, like d=2,3. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We wish to thank the anonymous reviewer for their feedback and questions. We are happy that the reviewer finds testable learning, particularly of PTFs, to be a very natural topic. The primary concern of the reviewer appears to be the dependence of our results on the error $\varepsilon$ and the degree $d$ of the PTF. In particular, that this dependence impacts the practical relevance of our results. While this is certainly a fair criticism, we would like to point out that: - By considering the testable model (rather than the standard agnostic model), we are actually taking a step towards practicality. Indeed, in the standard agnostic model, one makes assumptions on the distribution of the data which cannot be verified algorithmically (neither in theory nor in practice). On the other hand, in the testable model, one relies only on properties of the data which can be verified directly (in our case, via moment matching). The testable model is therefore harder (leading to worse dependences on the problem parameters), but also a better reflection of practical reality. - Even in the (easier) agnostic model, the runtime dependence of known algorithms on $\varepsilon$ and $d$ is quite bad, and moreover, this dependence cannot be improved much under typical hardness assumptions. In particular, under these assumptions, it is not possible to find an algorithm which is polynomial both in $n$ and in $1/\varepsilon$. For instance, [Kan11a] shows a runtime of $n^{O(d^2/\varepsilon^4)}$, which is beyond practical computation already when, say, $d=2$ and $\varepsilon = 0.25$. - The primary motivation of our paper is to gain further theoretical understanding of which learning problems might be 'hard' or 'easy' in some asymptotic sense. This follows a long line of papers in learning theory, some of which appeared in earlier editions of NeurIPS. Our main goal was to show that, for any fixed $d$ and $\varepsilon$, PTFs can be testably learned in polynomial time in $n$, thus qualitatively matching the agnostic setting. We believe that our result, while not immediately applicable in practice, is nonetheless of interest to the audience of NeurIPS. We give a more detailed explanation of why our dependence on $d$ is worse than in the agnostic model in our general rebuttal. In short, we inherit our dependence from [Kan11], and it seems nontrivial to improve them. We think determining the best-possible dependence is an interesting direction for future research. The reviewer asks specifically whether improvements are possible for small values of $d$. This is as an interesting suggestion. We rely on the 'fooling' result [Kan11] because it applies to PTFs of arbitrary degree. For PTFs of degree $d=2$ an earlier paper [DKN10] achieves a similar result to [Kan11], but with better (and more explicit) dependence on $\varepsilon$. It would be interesting to see if the result of [DKN10] can be translated to testable learning to achieve better dependence for $d=2$. However, we note that such a translation would likely require substantial additional technical effort. As our focus in this paper was to achieve a result for all choices of $d$ simultaneously, we did not pursue this direction. [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity [Kan11a]: The Gaussian Surface Area and Noise Sensitivity of Degree-d Polynomial Threshold Functions, Daniel M. Kane, computational complexity vol. 20 [DKN10] Bounded Independence Fools Degree-2 Threshold Functions, Ilias Diakonikolas, Daniel M. Kane, Jelani Nelson, 2010 IEEE 51st Annual Symposium on Foundations of Computer Science --- Rebuttal Comment 1.1: Comment: Thanks for your detailed response.	Rebuttal 1: Rebuttal: First and foremost, we would like to thank the reviewers for their time and valuable feedback. We appreciate that many of the reviewers consider testably learning PTFs as a natural and well-motivated problem. Several reviewers correctly pointed out that the dependence of our runtime on $d$ (the degree of the PTF) is worse than in the (standard) agnostic setting. We inherit this dependence from the result in [Kan11] since we modify their construction. We note that the dependence in our result is no worse than in [Kan11], even though it holds in a strictly more general setting. Any improvement to our dependence would immediately imply an improvement over the results in [Kan11]. While in [Kan11] it was speculated that such a better dependence is conceivable, this was not achieved since the first dissemination of the result over ten years ago. In light of this, achieving a better dependence on $d$ is a great, but likely difficult open question for future work. It is also related to the intriguing question whether a strictly larger runtime is necessary for testable learning over agnostic learning. We would also like to highlight that our impossibility result (Section 2.4) rules out a natural, "simpler" approach to prove guarantees for testably learning PTFs. In some sense, this shows that achieving better runtime dependences for testably learning PTFs is likely difficult without also improving the result of [Kan11]. Furthermore, we would like to stress that even for $d = 2$, our result is the first for testably learning PTFs. Moreover, as pointed out by reviewer zio4, for any fixed $d$, our dependence on $\varepsilon$ is qualitatively optimal in the sense that it matches known lower bounds (which hold even in the simpler agnostic setting). In particular, for a fixed $d$, our runtime scales as $n^{\mathrm{poly}(1/\varepsilon)}$ and known lower bounds (either in the SQ model [DKPZ21] or under standard cryptographic assumptions [Tie23]) imply that this is necessary. For example, these lower bounds rule out runtimes such as $\mathrm{poly}(n, 1/\varepsilon)$. Kind regards, the Authors [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity [DKPZ21]: The Optimality of Polynomial Regression for Agnostic Learning under Gaussian Marginals in the SQ Model, Ilias Diakonikolas, Daniel M. Kane, Thanasis Pittas, Nikos Zarifis, Proceedings of Thirty Fourth Conference on Learning Theory [Tie23]: Hardness of Agnostically Learning Halfspaces from Worst-Case Lattice Problems, Stefan Tiegel, Proceedings of Thirty Sixth Conference on Learning Theory	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper investigates testable learning of polynomial threshold functions (PTFs) with respect to the standard Gaussian distribution. The authors extend previous work on testable learning of halfspaces to show that PTFs of arbitrary constant degree can be testably learned up to excess error \epsilon in time n^poly(1/\epsilon), matching the best known guarantees in the agnostic model. The key technical contribution is showing that distributions which approximately match the moments of a Gaussian up to degree poly(1/\epsilon) fool constant-degree PTFs. Strengths: - The paper is well written and easy to follow. - The paper makes significant progress on an open problem in learning theory by extending testable learning to PTFs. Weaknesses: - It is not clear how important the problem of testable learning for PTFs is and if the results and/or techniques have applicability to other learning theory problems. Technical Quality: 3 Clarity: 3 Questions for Authors: - It seems like the testable learning setting is similar to the samples coming from a distribution close to the distribution D in some sense? Is there any model of learning theory that explicitly studies this? - Do you see a path to improving the sample complexity to polynomial in both n and 1/\epsilon, rather than n^poly(1/\epsilon)? - The authors have looked at Gaussian distribution in this work. What other distributions could this result be extended to? - Can you provide intuition for why the runtime dependence on d is so much worse than in the agnostic model? Do you believe this gap is inherent or an artifact of the analysis? - Did the authors look into getting any sort of lower bounds for this problem? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their kind feedback and insightful questions. We appreciate that the reviewer found our paper easy to follow, and that they believe we make significant progress on an open problem in learning theory. The main concerns of the reviewer appear to be the relative importance of the problem we study (testable learning of PTFs), and the potential applicability of our results and techniques to other problems in learning theory. To motivate the importance of our main result, we wish to briefly discuss the importance of PTFs and the testable learning model: - As reviewer zio4 mentions, PTFs are very well-studied in (theoretical) computer science, and in particular in learning theory. They are a natural extension of linear threshold functions, introducing non-linearity while maintaining some amount of structure. For this reason, they are often used as a test-case to determine the boundaries of efficient learning algorithms, which is also their role in our paper. - The testable learning model was introduced relatively recently as an extension of the (standard) agnostic model. It has already attracted significant attention, evidenced by several publications in leading conferences (including NeurIPS) [DKK+23, GKK23, GKSV23, GKSV24, RV23]. A recurring theme in these works is an attempt to determine whether testable learning comes at an additional computational cost with respect to agnostic learning. Our work continues on this theme by proving that, qualitatively speaking, the class of PTFs can be testably learned at no additional cost (for any fixed $d$). No such results were previously available, even for $d=2$. Beyond the inherent importance of our results, we believe there is potential for future applications of our proof techniques (as mentioned by reviewer 2r1G). We refine an earlier error analysis of [Kan11], thereby translating a fooling result for $k$-independent Gaussians to a testable learning guarantee. Our methods could prove useful in future translations from existing results in approximation theory to testable learning, as well. Finally, we wish to reply to the questions by reviewer pSUk: 1. This is the right intuition. In the context of our paper, where we only consider approximate moment-matching to test the data, testable learning corresponds to agnostic learning w.r.t. the class of distributions whose low-degree moments are close to those of a Gaussian. However, the testable model is more general than this, as it makes no assumptions on the type of testing algorithm used. This means we do not have such a correspondence in general. Agnostic learning w.r.t. classes of distributions that contain (but are broader than) the Gaussian has been considered in the literature before. A key distinction is that previous approaches typically focus on a class with nice mathematical properties (e.g., log-concave distributions), whereas in the testable model one considers classes for which membership can be verified efficiently from a small sample. 2. For this question, we would refer the reviewer to our general rebuttal. In short, there is evidence that this is not possible, even for the easier agnostic learning model with respect to the Gaussian. This is not made sufficiently clear in the paper, and we will improve this in the revision. 3. This is an interesting direction for potential future work: The most natural generalization that we see, based also on other works on testable learning, e.g., [GKK23], [GKSV24], [GKSV23], is to strongly log-concave distributions. However, the result in our paper does not directly generalize to any other distribution than Gaussian. 4. For this question, we would again refer to our general rebuttal. In short, it is not clear whether the gap is inherent to the testable model or arises from our proof techniques. We inherit our dependences from [Kan11], and it seems nontrivial to improve them. Intuitively, one expects worse dependences in the testable model as one needs a stronger notion of polynomial approximation than in the (standard) agnostic model for the polynomial regression algorithm to (provably) work (compare Thm. 6 to Thm. 7). In the paper, we work with the notion of 'fooling', which also corresponds to a strong notion of polynomial approximation (namely 'sandwiching', cf. Lines 228-232). 5. We did not study lower bounds specifically for testable learning of PTFs beyond the impossibility result for the approach used by [RV23] (cf. Section 2.4). (As mentioned in our answer to question 2, there is a lower bound for agnostic learning that also applies to the testable setting). However, we think it is an interesting future research direction to either prove a computational gap between agnostic and testable learning (w.r.t. $d$), or prove that no gap exists. [DKK+23]: Efficient Testable Learning of Halfspaces with Adversarial Label Noise, Ilias Diakonikolas, Daniel M. Kane, Vasilis Kontonis, Sihan Liu, Nikos Zarifis, Advances in Neural Information Processing Systems 36 (NeurIPS 2023) [GKK23]: A Moment-Matching Approach to Testable Learning and a New Characterization of Rademacher Complexity, Aravind Gollakota, Adam R. Klivans, Pravesh K. Kothari, Proceedings of the 55th Annual ACM Symposium on Theory of Computing [GKSV23]: Tester-Learners for Halfspaces: Universal Algorithms, Aravind Gollakota, Adam R. Klivans, Konstantinos Stavropoulos, Arsen Vasilyan, Advances in Neural Information Processing Systems 36 (NeurIPS 2023) [GKSV24]: An Efficient Tester-Learner for Halfspaces, Aravind Gollakota, Adam R. Klivans, Konstantinos Stavropoulos, Arsen Vasilyan, The Twelfth International Conference on Learning Representations [Kan11]: k-Independent Gaussians Fool Polynomial Threshold Functions, Daniel M. Kane, 2011 IEEE 26th Annual Conference on Computational Complexity [RV23]: Testing Distributional Assumptions of Learning Algorithms, Ronitt Rubinfeld, Arsen Vasilyan, Proceedings of the 55th Annual ACM Symposium on Theory of Computing	null	null	null	null	null	null
Categorical Flow Matching on Statistical Manifolds	Accept (poster)	Summary: The paper extends Flow Matching to discrete ("categorical") variables, similarly to a whole flourish of other papers that appeared in the last months, all of which work on the "probability simplex", i.e. they pass from a discrete alphabet A to the finite dimensional probability space P(A), which can be identified with a simplex with vertex set A. Here the emphasis is on the Fisher-Rao metric over this simplex, which the authors claim gives a natural riemannian metric that is well-suited to the flow-matching formalism. Strengths: The method is well founded and the geometric insights are inspiring and seem "right". It seems that the method performs well, as indicated by experiments. The paper is well written and clear, easy to read execept for a few hiccups (mentioned in "questions" section). Weaknesses: The Riemannian metric on the subset of the sphere $\{x\in \mathbb R^n: \|x\|^2=2, x_i\geq 0 \text{ for }1\le i\le n\}$ is not that different (in fact a $C^1$-close deformation of) the one on the actual "flat" simplex. So why would the results be so much different between the two cases? The grounding for a big gap from this setup to the standard one is not clear. It is also not clear why this method would have to perform better than Stark et al. "Dirichlet flow" paper. Technical Quality: 3 Clarity: 3 Questions for Authors: By the way, what's the difference between this paper and https://arxiv.org/abs/2405.14664 ? I understand and follow well the core paper, so I have only a few minor questions: 1) line 94: what does it mean to "condition on the delta measure"? A measure is not a random variable right? I don't fully follow what this means. 2) line 151: "we assume a single data dimension" -- what does that mean? also line 152 "extended to any data dimension" is kind of confusing, but I'm sure this is minor.. please clarify? 3) for section 3.3, line 171 and following, says "the Fisher information metric defines a second-order optimization [...]".. I don't fully understand what this means exactly. A metric is a metric, it doesn't have agency and so it doesn't define anything. I think this could be clarified by adding a few more formulas and by expanding on what the authors actually mean. 4) also for section 3.4 I don't fully follow: what Optimal Transport are the authors considering, how is that set up and how is that justified? At the moment I don't find this section very useful to a reader. 5) about section 3.5: why would one want to calculate NLL? The reason for that is missing, so I think it would make sense to shorten the derivation of the NLL formula to a minimum (move them to the appendix?) and actually explain where this section is coming from, as it's not crucial to actual flow matching objectives, and at the moment seems a bit artificial tbh Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: see "weaknesses" section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your high recognition of our work's novel geometric insight and clear delivery. We will address your questions and concerns as follows. ## Q1 Advantage of Riemannian structure Unlike LinearFM which assumes a flat simplex and straight flows, SFM considers the Riemannian structure induced by the Fisher information metric for categorical distributions. The naive Euclidean assumption may fail to capture the true geometry of the statistical manifold. Furthermore, we demonstrated in Sec.3.3 that following the vector field induced by the Fisher information metric coincides with the steepest direction of the natural gradient that decreases the KL divergence, which may also contribute to a better performance from the optimization point of view. Another possible advantage comes from the curved geodesic in Fig.1. Unlike straight geodesic under the Euclidean setting, the Riemannian vector fields are curved towards the boundary (more parallel to the boundary), making it hard to overshoot across the boundary. ## Q2 Comparison to DirichletFM We have included comparisons to DirichletFM in Appendix A.4 to provide some insights into SFM's better performance. DirichletFM considered the specific probability path in which the target distributions are always assumed to be Dirichlet. Such an assumption is not applicable to more complex target distributions and fails the Swiss roll toy example in Fig.3. Additionally, its probability path does not reflect the shortest geodesic distance on the statistical manifold, which may lead to suboptimal flows and vector fields. In contrast, our SFM framework constructs flows based on the geodesic on the statistical manifold, where vector fields follow the steepest direction of decreasing KL divergence. Our method can applied to any source and target distribution, making it more flexible and efficient. ## Q3 FisherFlow We thank you for mentioning FisherFlow, which is a concurrent work closely related to ours. We note that the FisherFlow paper was uploaded on May 23 after the NeurIPS submission ddl. Therefore, SFM and FisherFlow should be considered concurrent work. Both SFM and FisherFlow explored the Riemannian structure of the statistical manifold to establish flows and vector fields by constructing isomorphic mappings between the simplex and the sphere. We both demonstrated the theoretical connection to natural gradient and explored the minibatch-OT formulation during training. We took a step further to derive the exact likelihood over the probability space and a tight ELBO for the likelihood of discrete data. We carefully choose the prior, reference measure, and divergence to ensure comparability with other discrete diffusion/flow models. While FisherFlow predominantly focused on bioinformatic tasks of DNA design, we conducted more comprehensive experiments across various domains with additional baselines to explore the superior performance and versatility of our method. We believe both works are important explorations on the use of Riemannian geometry structure on discrete generative tasks. ## Q4 Conditional on measure We apologize for the typo. We meaned "condition on $x_1$" where the conditional probability path has $p_1(x\|x_1)\approx \delta_{x_1}$ at $t=1$, as described in the flow matching paper. In our case, each target point $x_1$ is a one-hot distribution $\mu_1$. ## Q5 data dimension The data dimension refers to the length of data. As an example, a DNA sequence with $D$ bases has data dimension $D$. We can represent this data as a matrix $X \in [0,1]^{D \times n}$, where $n=4$ reflects the four different categories (A, T, C, G). For simplicity, we assumed $D=1$ in the derivations, which is a standard practice. It is straightforward to extend this to multi-dimension by jointly modeling $D$ categorical probability measures. ## Q6 Connection to natural gradient Please see the common rebuttal. ## Q7 Optimal transport We consider minibatch-OT similar to [20,43,62] but on the statistical manifold. We match between a minibatch of the noises from the prior distribution and the samples from the target distribution with the smallest transport cost (defined line 189) based on the statistical distance defined in Eq.3. A thorough investigation of the theoretical benefit of minibatch OT can be found in the original paper. For Markov-chain based models like D3PM and MultiFlow, it is not possible to derive such a distance measure due to the discrete jump between Markov states. ## Q8 Significance of NLL calculation Generative models are trained to capture data distribution $p_\theta\approx p_\text{data}$. Therefore, it is natural and crucial to calculate the likelihood for a given data sample $p_\theta(x)$ as both the evaluation metric for generative models and confidence quantification for data. When evaluated on ground truth data, the NLL serves as a natural evaluation metric of how closely the generative model fits the data. Many common evaluation metrics including cross-entropy, perplexity, and bits-per-dimension are derived from likelihood estimation. With the ability to calculate likelihood, we can provide an intrinsic evaluation of flow models. Additionally, NLL can be used to measure the confidence of given inputs and facilitate RLHF of flow models. Many policy-based RL requires explicit log-likelihood, and ELBOs that are loose (e.g. with impromptu noise schedule) could impact the effectiveness of RLHF. Our exact NLL formulation is more accurate and our ELBO definition over discrete data is arguably much tighter, potentially benefiting applications that explicitly rely on likelihood. --- Rebuttal 2: Comment: I thank the authors for the rebuttal, but I am not understanding one of the answers. About the sphere metric, I don't see how taking a simplex and curving the interior (mapping from a flat simplex to a curved one) can prevent overshooting over the boundary. Could you please elaborate? To clarify my doubt, think of the case of 3 categories. We then compare a flat equilateral triangle with straight lines, to a "quadrant" of a sphere, where geodesics on a sphere are so to say, "pieces of equators", or maximal circunferences on the sphere. Then the segment going from a vertex of the straight triangle to the opposite side, makes an angle $\alpha\in[\pi/3, \pi/2]$ with the boundary whereas, in the curved setting, the geodesic from a vertex of the quadrant to the opposite side makes an angle $\alpha=\pi/2$ with the boundary. So why does one trajectory overshoot less than the other? I can see the performance of your experiments, but to me the geometry of the sphere quadrant and of a simplex still seem quite similar. In fact, adding to what I said in the review: the two metric spaces are not just $C^1$-diffeomorphic, but actually have finite distortion one with respect to the other, and I think the distortion constant is smaller than $2$. So, what would warrant such a big difference of performance then? Given the above comparison of agles of geodesics, I don't think the answer is related to overshooting, it must be something else. Emphasizing again, to be clear: this is just an important curiosity for me, because I can see the better performance of the experiments. But still I'm not convinced about the actual principle/reason beyond the better performance at the moment. --- Rebuttal Comment 2.1: Comment: We thank your feedback for our rebuttal and we are more than happy to further discuss the potential benefits of considering the curved geometry induced by the Fisher information metric. We will further elaborate on our hypothesis of overshooting for SFM versus LinearFM. In the middle column of Fig.1, we plotted the geodesics between the same source and target pairs under the Euclidean setting (flat simplex) and the Riemannian settings (using the Fisher information metric). Besides the curved geodesics, we also noted that the spacing between adjacent points is also different (though they are linearly spaced with respect to the geodesic distance). It can be seen from the figure that points are more clustered near the boundary than those near the middle region. Also note that the ground truth conditional vector fields have constant length for all timestep $t$ for both the flat simplex (always $\mu_1-\mu_0$) and the sphere (the arc length between $x_0,x_1$). In this way, a predicted vector field on the sphere with the same norm near the boundary will instead move the point for a smaller distance on the simplex to avoid overshooting. The Euclidean vector field, on the other hand, remains constant near the boundary. Furthermore, for points near the boundary but not at the vertex (which may occur in our toy example), it can be seen from Fig.1 that the vector fields and the geodesic are curved to be more parallel to the boundaries. It can be also demonstrated mathematically (in our response to Reviewer xYE5) by looking at the direction term $\sqrt{\mu\odot\nu}-\langle\sqrt{\mu},\sqrt{\nu}\rangle \mu$. For $\mu$ close to the boundary with component $\mu_k\approx 0$, its corresponding vector field will also have a close to $u_k\approx 0$ component, which is different from linear flow matching's fixed $\nu-\mu$. In this way. the Riemannian vector field avoids further pushing the points outside the boundaries. We further noted that the geodesic distance in Eq.3 cannot be bounded by the Euclidean distance. More rigorously, there does not exist a finite constant $C$ such that $d\_\text{cat}(\mu,\nu)\le C\\\|\mu-\nu\\\|\_2,\forall \mu, \nu$. In other words, there isn't a finite distortion constant due to the singularity of the transform on the boundary. This can be demonstrated with Taylor expansion at $\mu$ as: $$ d_\text{cat}(\mu,\mu+\Delta\mu)=2\arccos\left(\sum_{i=1}^n\sqrt{\mu_i(\mu_i+\Delta\mu_i)}\right) \approx 2\arccos\left(\sum_{i=1}^n\mu_i+\frac{\Delta\mu_i}{2}-\frac{\Delta\mu_i^2}{8\mu_i}\right) =2\arccos\left(1-\sum_{i=1}^n\frac{\Delta\mu_i^2}{8\mu_i}\right) \approx2\sqrt{2}\sqrt{\sum_{i=1}^n\frac{\Delta\mu_i^2}{8\mu_i}}=\sqrt{\sum_{i=1}^n\frac{\Delta\mu_i^2}{\mu_i}} $$ Compared to the Euclidean distance $\\\|\Delta\mu\\\|=\sqrt{\sum_{i=1}^n\Delta\mu_i^2}$, it is clear that $d_\text{cat}(\mu,\mu+\Delta\mu)$ cannot be bounded when some $\mu_i$ is close to zero. An alternative theoretical benefit of SFM is its connection to natural gradient, as we have mentioned in the common rebuttal. We also noted that similar non-flat assumptions can be found in previous work including DirichletFM which follows the specific path of Dirichlet distributions and achieves better performance than LinearFM. Again, we sincerely thank you for bringing in such an inspirational discussion and we hope our explanation addresses your questions.	Summary: The paper proposes an approach to generative modelling of discrete data based on the Riemannian Flow Matching [1] algorithm where the authors use the Fisher metric to define the geometry of this space. In detail, the authors define the generative modelling from the discrete data as sampling from the empirical distribution on the statistical manifold (the space of categorical distributions). There every point is a categorical distribution on a fixed state space and the generative model maps the noise distribution to the data distribution on the manifold, i.e. it operates with distributions of categorical distributions. In order to apply Riemannian Flow Matching, the authors define the metric tensor, tangent space, geodesics, logarithmic and exponential maps. Furthermore, for computational stability, the authors map the statistical manifold to the sphere and redefine all the differential geometric tools there. Finally, the authors perform an extensive empirical study applying their method for language, discrete image generation, and sequence modelling in bioinformatics. The method demonstrates competitive results with prior works. [1] Chen, Ricky TQ, and Yaron Lipman. "Riemannian flow matching on general geometries." arXiv preprint arXiv:2302.03660 (2023). Strengths: The paper presents a complete study, i.e. it approaches an important problem in the field (generative modelling for discrete data) proposes a reasonable approach, presents the proposed approach in a comprehensible way, and provides minimal necessary empirical study of the idea. Weaknesses: Although the paper does not raise major concerns and overall satisfies the criteria of a NeurIPS publication, the practical motivation of the method is insufficient. Indeed, the problem of generating categorical distributions is interesting and the paper answers this question definitively (to a reasonable extent). However, then the authors quickly jump to the conclusion that a natural way to model discrete data is to sample a categorical distribution rather than to sample from a categorical distribution. Indeed, in Algorithm 2, the authors don’t even mention how they sample these categorical distributions (as I understand they consider every sample to define a point mass on the statistical manifold). First, the transition between sampling categorical distribution (in theory) and sampling discrete data (in practice) should be stated and explained clearly. Second, this transition requires a practical motivation (perhaps an empirical study). Indeed, why would one consider the Fisher-Rao distance to be a natural distance between two sentences in a natural language? Why one would need to consider a sentence to be a categorical distribution on the statistical manifold rather than a sample from such distribution? The paper does a poor job of answering these questions and describing how the proposed method works with discrete data. There are some minor concerns regarding the presentation: - The name Statistical Flow Matching does not match the title of the paper. - Line 29. It is not clear what `assumptions that fail to capitulate` mean. - Line 52. It is not clear what `existing models often fail due to impromptu prior assumptions` means. - Line 105 requires citations. - In lines 112-117, please specify that equipping the manifold with the Fisher-Rao metric. - The sampling notation in Eq. (8) is confusing due to the diffeomorphism applied to a probability density. - There is a typo in line 118. The tangent space is called `the target space`. - Section 3.3 reads as a preliminary section on the natural gradient rather than a connection of the proposed method to the optimization literature. Technical Quality: 3 Clarity: 3 Questions for Authors: I suggest adding practical examples where the data is a categorical distribution that is not concentrated in a single point. This would be a significant motivation for the proposed method. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper lacks a discussion of the applicability of the proposed abstraction to the generation of discrete data (see Weaknesses). Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your high recognition of our work's significance in the realm of discrete generative modeling and for raising interesting questions and suggestions. We will fix the typo and citation in the revision. Our responses to your questions are as follows: ## Q1 Motivations ### Sampling over categorical distributions instead of from a single categorical distribution We agree that both formulations can be effectively used on discrete sampling, but their compatibility with different generative models may vary. It is important to note that SFM (and the majority of diffusion/flow models, VAEs, and GANs) are not autoregressive (AR). Unlike LM which predicts logits for the next single token deterministically given context, diffusion and flow models operate directly on the joint distributions of all tokens. Therefore, it is inefficient to learn a single set of deterministic categorical distributions to capture diverse joint distributions. Alternatively, the majority of non-AR methods seek to transform from prior noise distribution (e.g. multivariant Gaussians) to target joint distribution, effectively resulting in a two-step formulation where stochastic samples over joint distribution space are drawn first (i.e., logits) and then decoded with a second sampling step (argmax or softmax). Secondly, diffusion/flow models are naturally defined over continuous space. To leverage such a framework, existing discrete diffusion/flow models utilize relaxation to operate on logits which are then projected to probability simplex, similar to the statistical manifold we defined but without a properly defined geometry. In other words, the two-step sampling in SFM is not a niche design choice but rather universal for discrete diffusion and flow models (e.g., D3PM, SEDD, BFN, LinearFlow). Our distinction to prior works is the proposal of a better objective for training the flow model over the categorical distributions (first step), where we equip the manifold with a Riemannian metric. For the second step, we always used argmax when decoding discrete samples. ### Fisher-Rao metric We would first like to clarify a potential misunderstanding. During training, we did not use the Fisher-Rao metric for calculating the distance between two sequences (i.e., discrete samples). Instead, a metric is necessary for defining probability paths along the geodesic from the prior distribution (usually uniform over the simplex) towards the target distribution (concentrated around the vertex). The role of the Fisher metric is to equip the manifold of categorical distributions with a Riemannian structure on which vector fields can be calculated as the learning target. The Fisher information metric is the natural canonical Riemannian metric on the statistical manifold. It also enjoys the benefit of following the "steepest" direction of the natural gradient of decreasing KL divergence, as we have mentioned in Sec.3.3. Nevertheless, we respectfully disagree that treating discrete data as onehot distributions and using Fisher-Rao distance over them is unnatural. In fact, discrete data have been commonly treated as distributions in classic discrete generative model training, when using cross-entropy loss. Fisher-Rao distance is yet another metric between distributions, which indicates the steepest direction for KL divergence minimization. We will make this clear in the revision to better motivate this setup as suggested by the reviewer. ## Q2 Name of SFM & title As mentioned in the previous text and Sec.3, our proposed SFM is a general generative framework for modeling measures (distributions) over the statistical manifold of a family of parameterized distributions. In this work, we presented the realization of our framework on categorical distributions which have wide applications in various discrete generation domains. Therefore, we wish to keep the general naming of statistical FM and the categorical in the title to inform the audience about the application. We will think of a better title to avoid confusion. ## Q3 Impromptu assumptions in prior work For Line 29, we want to convey that the naive Euclidean assumption of the probability simplex or the logit space does not capture the true geometry of the statistical manifold. Similarly, in Line 52, we want to emphasize that the Euclidean assumption in previous existing models does not have solid mathematical grounds, which may lead to worse performance on statistical manifolds. ## Q4 Fisher-Rao metric The Fisher information matrix for categorical distributions is $g_{jk}(\mu)=\frac{\delta_{jk}}{\mu_j}+\frac{1}{\mu_n}$ for $1\le i,j,\le n-1$, where $\delta_{jk}$ is the Kronecker delta. Substituting this into $\langle u,v\rangle_\mu$ leads to the Riemannian inner product in Eq.4. ## Q5 Sampling in Eq.8 We will rewrite the sampling as $x_0\sim\pi_(p_0(\mu))$ where $\pi_(p_0)$ denotes the standard pushforward measure induced by the diffeomorphism $\pi$. This is equivalent to first sampling $\mu_0\sim p_0(\mu)$ and then taking $x_0=\pi(\mu_0)$. ## Q6 Connection to natural gradient Please see the common rebuttal for details. ## Q7 Non one-hot data We really appreciate your recognition of the potential usage of SFM on general categorical distributions besides one-hot. In the paper, we have provided the Swiss Roll example as a demonstration, where other baselines with strong priors may fail. We believe that SFM could be of interest to Bayesian inference, distributional RL, and other tasks that involve sampling over distributions. We defer these use cases to our future work to extend SFM due to limited time and resources. --- Rebuttal Comment 1.1: Title: the motivation is still unconvincing but I consider the method to be important Comment: Thank you for the response. The presentation of the proposed method as the right way to generate text remains unconvincing to me. I'm not claiming that this is the wrong approach, I feel that the paper does a poor job of convincing the reader in this. However, I recognize the significance of generative modeling on the manifold of distributions equipped with the Fisher-Rao metric (which I'm well aware of). I agree, that the Fisher-Rao metric is a fundamental concept, and that's why I think it doesn't require further motivation through the discussion of the natural gradient. --- Reply to Comment 1.1.1: Comment: We first thank your recognition of the significance of our SFM for generative modeling on the manifold of distributions and for approving the validity of our approach. We would like to further clarify why our choice of modeling discrete data as one-hot probabilities is natural and widely adopted in prior work on discrete flow/diffusion. We will make sure to improve the arrangement of sections to provide more background to the audience earlier in the text. As described in our rebuttal, previous diffusion and flow models also predominantly viewed discrete data as a distribution, and the generative task for these models effectively learns a distribution over distributions. For example, DDSM and Dirichlet FM both assumed the discrete data as probabilities following some Dirichlet distribution (points on the simplex). D3PM, MultiFlow, and SEDD viewed each token as continuous logits (points on the probability simplex after softmax). We provided extended discussions about this setting in Sec.5 in our original paper. Similarly, our SFM views tokens as points on the simplex, differing only by providing a more mathematically meaningful geometry for the probability simplex. Regarding the "right way" to generate text, to the best of our knowledge, all current discrete diffusion and flow models for natural language modeling were based on the probability path on the simplex or related to the path on the simplex using logits. In this sense, we believe that working with simplex is more of the "natural" and "standard" way to the diffusion community following existing work. We thank your insightful comment, which reminds us that such an assumption may diverge from that in other text modeling communities. We provided an explanation to the possible gaps and misunderstandings in our rebuttal, and we will add them to our revised manuscript to make it more accessible to general audiences with different backgrounds. We would also appreciate it if you could provide pointers to works that do not view text tokens as one-hot categorical distributions (e.g., do not use cross-entropy loss), and we will include discussions on them. Again, we sincerely thank your review and response and we hope we have fully addressed your questions and comments. We will better organize our paper in the revised manuscript.	Summary: This work introduces statistical flow matching, an improved method for discrete generation on the simplex by utilizing the geometry induced by the Fisher information metric. By mapping points on the simplex to points on the sphere, flows wrt the Fisher metric can be efficiently computed. SFM is tested on toy examples in the simplex as well as on binarized MNIST, Text8, and a Promotor design application. Strengths: - Great figures with above average effort and conceptual clarity that make it very easy to understand the method and theory geometrically. - Simple, but seemingly effective addition to flow matching on data supported on a simplex. - Could be quite significant if it is truly competitive to diffusion-based approaches. Weaknesses: - I’m quite concerned about the “Pseudo log likelihood” metric being compared to log likelihood. While the authors know and say it is incomparable, it is for some reason still compared in Tables 1 and 2 directly to the discrete NLL. I don’t understand why this is done. I might recommend an evaluation similar to MultiFlow for text8, where the evaluation is done on generated samples, and can be used to compare fully discrete, vs. the continuous state approach taken here. Its extremely concerning how much the pseudo-NLL changes based on the choice of a seemingly arbitrary threshold $t_{max}$ and the number of steps (Table 4). To me that means this score should either not be used, or should at most only be used to compare like methods. - Qualitatively, the text8 generations look extremely poor relative to other recent related work. I’m very uncertain that “SFM achieved the second-best performance among all diffusion or flow baselines”. It would be great to include qualitative samples from other established methods for comparison. I think it is essential to establish the performance of SFM on the text8 dataset since this is arguably the only non-toy dataset and most established benchmark. If this text8 baseline can be fixed and these results shown to hold on valid and comparable metrics, I would substantially increase my score. Even if the text8 result is not as exciting, as long as it improves over existing flow-based approaches (Including MultiFlow), I believe this paper could still be of interest to the community, and I would still consider raising my score. MultiFlow: https://arxiv.org/pdf/2402.04997 Technical Quality: 3 Clarity: 4 Questions for Authors: Questions: - I don’t understand why the change of measure term is undefined at the boundary. Isn’t the transformation the square or sqrt transformation? Why is this undefined at the boundary? Could the authors elaborate on this point? - I don’t understand the “forward noising process” $q(\tilde{\mu}_1 \| \delta)$ as far as I can tell this is not defined in this work. - Table 4,5 “ODE solver” is what solver? Euler is an “ODE solver”. Could the authors specify which solver was used? - Why is LinearFM quite different between 300 steps and “ODE Solver” in Table 5? - Table 5 does not look like an FID score to me, which must be strictly positive. Perhaps this is an NLL? - Why is Multiflow cited but not compared to in the text8 task? Minor comments: - “we define the ODE log-likelihood as” — This is more accurately described as a change in likelihood perhaps? - “Swill roll” —> “Swiss roll” perhaps is meant in multiple places? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your recognition of our work's clear presentation and novel geometric perspective for discrete generation. We'd like to clarify that Text8 results of MultiFlow were added on Jun 5 after the NeurIPS ddl, thus we were unable to compare them although it was cited. We wish to emphasize that our major contributions are beyond the Text8 experiment. SFM is a general generative framework for measures of categorical distributions, and we have demonstrated the effectiveness of SFM across diverse domains including computer vision (binarized MNIST), bioinformatics (DNA design), and NLP(Text8). We also demonstrated performance improvements using evaluation metrics independent of likelihood calculation (FID & Sei MSE). We further established the underlying connection of SFM to natural gradient and proposed likelihood formulation with comparability in mind. We respectfully disagree that Text8 is the only non-toy dataset. DNA design has a tangible societal impact, and DDSM & DirichletFM all predominantly focused on this task. The data dimension of MNIST (784) and DNA (1024) is also larger than Text8 (256). Nonetheless, we understand the importance of text8 evaluation and now include additional results in the common rebuttal following your suggestions. Below we provide more details and justifications in response to your comments. ## Q1 Choice of evaluation metrics We're afraid that you might have misinterpreted our evaluation setup. For Swiss roll and MNIST, all baselines are diffusion or flow models with continuous parameterization. We never compared NLLs with discrete-space likelihood from discrete models (e.g. autoregressive) in these tasks. Secondly, all NLLs used in discrete tasks (Tab.1,2,4,5) are NOT pseudo-likelihoods. They are based on the variational bond described in Eq.14 and are comparable to ELBOs from the diffusion models we compared. Only BPC for Text8 (Tab.2) was calculated based on pseudo-likelihoods as described in Appendix B.2. According to the original PLL paper, it provides a reasonable evaluation metric between non-autoregressive models. We used the identically variational formulation in the DDSM paper to be comparable. We also provided FID scores that do not rely on NLL calculation, for which our models still outperformed other baselines by a margin. More discussion on the validity of such a variational likelihood can be found in Sec.3.5 of DDSM. We justified our choice of $t_\text{max}=0.995$ in the calculation of the variational bound for NLL in Appendix D.1. We carefully chose this value to make sure it was comparable to DDSM settings. Our ablation study on the choice of $t_\text{max}$ agreed with that in DDSM, where $t$ closer to data gives a tighter estimated bound (it can't get arbitrarily low). We believe it is reasonable to compare DDSM and LinearFM using similar $t_\text{max}$. ## Q2 Additional results on Text8 We provide GPT-NLL vs token entropy, and representative samples for SFM and MultiFlow with $T=0.8,0.9,1$ for visual quality check. SFM has closer to data token entropy and comparable GPT-NLL with other methods in MultiFlow's result, although slightly worse on the latter. We respectfully disagree that our samples "look extremely poor", especially when compared with MultiFlow results with matching entropy ($T=1$). The perceptual quality of SFM and MultiFlow are very close, with noticeable misspellings in both. We also found GPT-NLL very unreliable and can be easily fooled by random strings. As an example, the following sample from MultiFlow at $T=1$ looks even worse despite a GPT NLL of 6.648. ``` she hill obhalarnnach eochrans eileann munthar cearlomha mhaonna tardare mho mord tore lian linn mu phaile gael cangallauig laothuis guilleith leois glion guildh lara gall innonte tilbonne guilecht shuachtansert guillaste guatnaoic asthache cuichant conai ``` We will include these results on Text8 for a more comprehensive evaluation as you suggested, and tone down on our original claims. We sincerely hope that our contributions besides text8 can be valued properly. ## Q3 Change of measure term The change of measure term describes the change of the log-likelihood between different manifolds as $\log p_S(x)=\log p_\Delta(\mu)+\log \|\det d\pi(\mu)\|$ where $x=\sqrt{\mu}$. Although the density is always well-defined, the log-likelihood is undefined at the boundary, as it involves taking the logarithm of a zero probability density (see Eq.37 & 38). ## Q4 Forward probability $q(\mu\|\delta)$ The detailed definition of the forward probability $q(\mu\|\delta)$ was provided in Eq.29 & 30 in Appendix B.1. The forward diffusion process $q_t$ defines a small neighborhood for variational estimation. In DDSM, the authors followed the fixed probability path of the Jacobi diffusion process with known forward probability. In our flow-based setting, we can use simpler indicator measures as the linear interpolation between the delta measure and the uniform distribution on the simplex. ## Q5 ODE solver As described in Sec.4.2, all generated results and NLL calculations are based on the Dopri5 solver unless otherwise specified. The Dopri5 solver is an adaptive solver with a good balance between accuracy and efficiency. ## Q6 LinearFM Sensitivity to solvers We noted that LinearFM tends to have very negative divergence around the one-hot distribution. Therefore, Euler step may overestimate the contribution of divergence in this case while adaptive solvers can provide a more accurate result. In contrast, as the SFM vector field is defined on the sphere, it is less sensible to solvers and results in a more stable divergence. ## Q7 Typos & minor issues The results in Tab.5 are NLLs. The ODE likelihood defined in Eq.12 is indeed the change of log-likelihood: $\log p^\text{ODE}:=\log p(\mu_1)-\log p(\mu_0)$. We thank you for pointing out these issues, and we will fix them in the revised manuscript. --- Rebuttal Comment 1.1: Title: Alternative BPC Calculation Comment: We hope our previous rebuttal has adequately addressed your questions. To further support our discussion, we would like to introduce a potentially more comparable BPC formulation inspired by a new concurrent work MDLM [1]. MDLM provided a general form of ELBO for the continuous flow setting with a more formal proof. This ELBO applies to both continuous flow models and flow models that rely on discrete jumps with logits. Specifically, the ELBO has the form: $$ \mathcal{L}=-\mathbb{E}_{\mu_1\sim q(\mu)}\left[\int_0^1\frac{1}{1-t}\log\langle \mu_t,\mu_1\rangle dt\right] $$ where $\mu_t$ follows the predicted inverse trajectory starting from $\mu_1$. We believe BPC calculated with this ELBO is comparable to SEDD/D3PM/Multiflow (as they were compared in [1] as well). The BPC based on this variational bound is provided in the following table. Our BPCs were averaged over the first 1000 sequences of length 256 in the test set. \| Model \| BPC↓ \| \| -------------- \| ------------- \| \| SFM w/ OT \| 1.412 ± 0.006 \| \| SFM w/o OT \| 1.410 ± 0.004 \| \| LinearFM \| 2.197 ± 0.008 \| \| D3PM-absorb \| 1.45 \| \| D3PM-uniform \| 1.61 \| \| BFN \| 1.41 \| \| SEDD-absorb \| 1.32 \| \| SEDD-uniform \| 1.41 \| \| Discrete Flow \| 1.23 \| \| Argmax Flow \| 1.80 \| \| MultiFlow η=0 \| 1.41 \| \| Transformer XL \| 1.08 \| Similar to our evaluation results using GPT-J-6B, our SFM reached a comparable BPC with MultiFlow, BFN and SEDD-uniform, outperformed many existing discrete diffusion/flow models including D3PM, LinearFM, and Argmax flow. We will update the results in the revised manuscript as such BPCs are more comparable than pseudo-BPCs. We sincerely hope that our discussion addressed all your concerns and we look forward to further discussion. [1] Sahoo, Subham Sekhar, et al. "Simple and Effective Masked Diffusion Language Models." arXiv preprint arXiv:2406.07524 (2024).	Summary: This paper proposes flow matching on the manifold of discrete distributions, where each point represents a probability mass function (pmf). In essence, this is done by parameterizing a vector of size n for each n-class categorical distribution. Instead of using Euclidean geodesic,s the authors discuss the use of Riemannian geodesics given the Fisher information matrix. To address the constraint that this should represent a probability mass function, the authors suggest using constraint \|\|x\|\|^2 = 1 (i.e. the normalization is over L2), which allows geodesics to be well-defined on when points lie on the boundary. Furthermore, since this is within a continuous normalizing flow framework, authors claim that it is possible to compute exact log-likelihoods, as opposed to related diffusion model counterparts. Experiments are carried out over a range of datasets where n=2, 26, 4. Strengths: - Well written. I found the writing clear and to the point. - A natural extension of existing works to flow matching on the manifold of discrete pmfs. Weaknesses: Overall, I feel the idea is a straightforward adoption of prior work on riemannian FM in terms of novelty, but the paper is nicely presented and packaged together. The topic of generative flows with discrete data is also timely. Technical Quality: 4 Clarity: 4 Questions for Authors: ### Main questions / concerns: _Regarding likelihood definition._ In Sec 2.1, you use Radon-Nikodym derivative $p=\frac{d\mu}{d\nu}$ to define the density p, where $\nu$ is used to denote the reference measure. In Sec 3.5, you state the change of measure as if it is a normal density function rather than treat it as a Radon-Nikodym derivative. - How do you chose $\nu$? - Can you say that the Radon-Nikodym derivative is defined for all $\mu$ that you use in practice? - Does this change of measure (Eq 11) correspond to some $\frac{d\mu_t}{d\nu}$, i.e., fixed reference measure? I feel that even though $p$ is defined through a Radon-Nikodym derivative, there isn't actually a formal treatment of $p$. In the text, $p$ seems to just be treated as a density function. This should be okay specifically because $x$ lies in a finite dimensional space, but this special case of the Radon-Nikodym derivative should be stated. _Regarding choice of metric in likelihood computation._ In Tables 1 & 2 involving NLL values, are you sure these values are comparable between methods? For instance, the NLL depends on the choice of metric g (as in Eq 11), so if the $div_g$ is replaced with Euclidean divergence, this will have very different NLL values. I see that in both tables, LinearFM is directly compared to SFM. Even if different g's are used during training, they can still be compared if using the same g for NLL computation. - Did you use the same g for NLL computation? - Which g did you use (Euclidean or Riemannian)? _Regarding choice of metric in desiging conditional u_t._ In Figure 1, it might be a good idea to also visualize the velocity field u_t. I think there's also an interesting property of the Riemannian u_t that isn't discussed, and it's that this u_t is parallel to the boundary. Whereas the Euclidean u_t will have a nonzero inner product between the u_t and the normal direction when evaluated on / close to the boundary. Is this true? _Regarding the connection to natural gradient._ It's unclear what the point of Section 3.3 is. Right now, this section feels pointless as it doesn't say anything about the proposed method. What your point is should be made more explicit. - Why should we care if the Fisher information matrix shows up as the Hessian of KL? What does this imply? - Can you show that the geodesic under this Riemannian metric is following the "natural gradient"? Is this what you wanted to claim here? ### Minor suggestions: - State in the main text that you are using minibatch OT couplings. Sec 3.4 does not mention how OT is used in the actual training algorithm. (I had assumed it is done over the full training set until finding Alg 2.) - The "target" space --> "tangent" in Sec 3.1. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Authors adequately addressed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your appreciation of our work's clear presentation as a natural and timely extension of Riemannian FM on discrete data. We thank your thoughtful suggestions and will fix typos and clarify minibatch-OT in the revision. We'd like to address your comments as follows. ## Q1 Regarding likelihood definition We mentioned the Radon-Nikodym derivative in 2.1 as the definition of probability density functions on general manifolds. In our likelihood derivation (Sec.3.5), the density $\frac{dP}{dV}$ can be well-defined over the statistical manifold which is finite-dimensional ($n-1$ dimensional for $n$ categories) using the Riemannian volume as the reference measure. This is in contrast to works like functional FM where data points are continuous functions lying on the infinite-dimensional Hilbert space. We will clarify this as you suggested. In terms of the choice of measure, we make sure to use a consistent reference measure for all change-of-measures (COM) in Sec 3.5. Note that the Riemannian volume on the sphere coincides with the Euclidean one (see response to Q2), and only the COM for the two transformations $\pi$,$\pi^{-1}$ will be impacted by the choice of measure. To make the NLL comparable to Euclidean flows, we have kept the reference measure consistent as the Euclidean volume form in the equations in Appendix B.1. We will make these mathematical details clearer and explicitly state the choice of volume in the revised manuscript. For categorical distributions $\mu$, we used the canonical counting measure over the discrete sample space $\mathcal{X}=\{0,1,\dots,n\}$. ## Q2 Regarding likelihood computation We thank your insightful question regarding the divergence in NLL calculation. You are completely right that even though vector field is learned with Riemannian geometry, we can still compute likelihood with Euclidean divergence. As a matter of fact, we have taken comparability into consideration in our derivation of likelihood. Note that for manifolds that can be embedded in the ambient Euclidean space (e.g., simplex and sphere), the Riemannian divergence is the same as Euclidean divergence (for embeddable manifolds, we have $\log\|\det g\|\equiv0$). By using the Euclidean volume form as the reference measure in all COMs and computing prior likelihood on Euclidean simplex, we effectively used Euclidean divergence on the simplex for SFM and LinearFM to guarantee that the results are comparable. It is possible, though, to calculate the Riemannian likelihood for SFM. To do this, we need to make several adaptations. The prior likelihood should be adjusted to $p_0=\Gamma(n)/\sqrt{\|\det g\|}$ such that the integral $\int_\Delta p_0dV=\int_\Omega p_0\sqrt{\|\det g\|}d\mu_1\cdots d\mu_{n-1}=1$. The change of measure term in Eq.37 & 38 also needs to be adjusted as $\log\|\det d\pi\|=(n-1)\log 2$ and $\log\|\det d\pi^{-1}\|=-(n-1)\log 2$. ## Q3 Regarding choice of metric in designing conditional $u_t$ Thank you for mentioning the nice property of Riemannian $u_t$ being parallel to the boundary. We'd like to mention that our visualization of the logarithm map can be viewed as a demonstration of $u_t$'s direction (up to a constant scaling factor that depends on $t$), as the vector field can be calculated in terms of the logarithm map as $u_t(\mu_t\|\mu_0,\mu_1)=\log_{\mu_t}(\mu_1)/(1-t)$ (assuming a linearly decreasing geodesic distance). We indeed noticed that the Riemannian structure induced by the Fisher information metric leads to vector fields more parallel to the boundaries. This can also be demonstrated mathematically by looking at the logarithm map in Eq.23. Consider the direction term $\sqrt{\mu\odot\nu}-\langle\sqrt{\mu},\sqrt{\nu}\rangle \mu$. For $\mu$ close to the boundary with $\mu_k\approx 0$, its corresponding vector field will also have a close to $u_k\approx 0$ component, which is different from linear flow matching's fixed $\nu-\mu$. We hypothesize that one potential benefit of such a curved geometry over the naive Euclidean geometry is that the former helps prevent overshooting across the boundaries. Specifically, consider a target point on the boundary. The Euclidean vector field will continue to push the points outside the manifold, whereas the Riemannian vector field tends to travel parallel to that boundary to prevent going across the manifold. The sphere exponential map also naturally ensures the point (after transformation back) lies on the simplex. Once overshooting happens during sampling, the model may exhibit undefined behavior as it was never trained on points outside the manifold. We will emphasize this benefit more clearly in our future revision. ## Q4 Regarding the connection to natural gradient Please see the common rebuttal for details.	Rebuttal 1: Rebuttal: Dear all Reviewers, We sincerely appreciate your reviews that help make our work more concrete and comprehensive. We thank you for recognizing our novel geometric perspective of discrete generation and our paper's clear presentation. Here we address some of the common questions. ## More details regarding connection to natural gradient As per request by Reviewer xYE5, WPWv, and hmQs, we elaborate more on the connection to natural gradient in addition to Sec.3.3. We want to demonstrate the direction of the Riemannian vector field induced by the Fisher information metric coincides with the steepest direction of minimizing local KL divergence (the natural gradient). From the optimization view point, one objective for generative modeling of categorical distributions would be the KL divergence $D_\text{KL}(\tilde{\mu}\\\|\mu)$ between the generated distribution $\tilde{\mu}$ and the target distribution $\mu$. The fact that the Fisher information metric is the Hessian of KL divergence allows us to locally expand the change of KL divergence in Eq.10 in terms of Fisher information metric as $D_\text{KL}(\mu(\theta)\\\|\mu(\theta_1))\approx \frac{1}{2}\sum_{jk}\Delta\theta_j\Delta\theta_j g_{jk}=\frac{1}{2}\\\|\Delta\theta\\\|^2_g$ where $\Delta\theta=\theta-\theta_1$. The steepest direction $\Delta\theta$ decreasing the KL divergence is known as the natural gradient in the existing literature. From the geometric view point, the geodesic, by definition, is a (locally) length-minimizing curve with respect to the corresponding Riemannian metric. Therefore, by following the direction of the vector field that decreases the geodesic element $ds^2=\\\|d\theta\\\|^2_g$, we are indeed following the steepest direction of natural gradient that minimizes the local KL divergence. This theoretical connection may contribute to the better performance of SFM. ## Additional results for Text8 We first noted that, though we cited MultiFlow, their Text8 experiment was not added until Jun 5 (>2 weeks after NeurIPS ddl). Nevertheless, as requested by Rewiewer SAto, we now add results with the GPT-J-6B model in addition to BPC, following the setting in MultiFlow. We emphasize that pseudo-LL is only used in Text8 BPC and all discrete NLL comparisons follow the similar ELBO derivation in DDSM. We only compared NLLs to diffusion/flow models with comparable ELBO formulations (see the response to SAto) and used consistent divergence (see the response to xYE5). We also provided non-NLL metrics (FID, Sei MSE) for MNIST/DNA data which are non-toy high-dimensional real-world datasets, where SFM outperformed all baselines with a margin. We sincerely hope that these contributions can still be valued. As GPT-J-6B was not trained on Text8, its NLL does not necessarily reflect Text8 distribution fitness. From the MultiFlow results, such an NLL can be made artificially low by duplicating high-frequency words e.g. numbers, thus joint comparison of GPT NLL and token entropy was used, where an entropy closer to data is preferred. We show in the table below that SFM tends to produce samples with data entropy closer to data, and the GPT NLLs are still comparable to (slightly worse) MultiFlow/D3PM when entropy is close. As SFM directly samples categorical probabilities instead of making multiple discrete jumps with logits, it is incompatible with temperature-tuning. Thus in the table below, we compare with $T=1$ which matches the entropy of SFM. We also noted that GPT NLL can be easily fooled by randomly generated strings based on letter frequency. Additionally, low-T MultiFlow variants achieved lower NLLs than the ground truth data, making this metric less credible. We also noticed that the low NLL samples for MultiFlow often consist of repeated numbers with little semantic meaning. To enable fair comparison, we provide multiple generated samples (above/below/just-on-average NLL for each model) and their GPT NLLs in the PDF. Different temperature settings for MultiFlow are also included using checkpoint provided on Github. We noted the big impact of temperature on MultiFlow as it generated more numbers repeatedly with lower temperature. In contrast, SFM achieved perceptually similar or even better results with more diverse vocabulary. We thank the authors of MultiFlow for generously providing their raw data in Fig.2. The complete figure is provided in the pdf. \| Model \| GPT-J-6B score ↓ \| Token entropy ↑ \| \| - \| - \| - \| \| Data \| 4.099 \| 7.479 \| \| Random \| 5.827 \| 5.519 \| \| SFM w/ OT \| 7.071 \| 7.347 \| \| SFM w/o OT \| 7.154 \| 7.388 \| \| LinearFM \| 7.490 \| 7.118 \| \| SEDD mask \| 6.49 \| 7.17 \| \| SEDD absorb \| 6.28 \| 6.97 \| \| MultiFlow \| 6.728 \| 7.387 \| \| MultiFlow η = 0 \| 6.729 \| 7.325 \| \| D3PM \| 6.929 \| 7.379 \| \| Autoregressive \| 6.296 \| 7.165 \| We thank Reviewer SAto for proposing an alternative evaluation. We will include these results on Text8 for more comprehensive evaluation, and tone down on our original claims. Nevertheless, we'd like to note that the general formulation of SFM over continuous statistical manifold has potential beyond NLP discrete sampling, and could be of interest to Bayesian inference, distributional RL, and other tasks that involve sampling over distributions. NLL used to be one of the common metrics for generative models as it reflects closeness to the data distribution. It is until recently that this metric became underlooked due to arguably loose and diverse ELBO derivation in diffusion, and researchers resort to third-party models (InceptionV3, GPT, AF2, etc.) fitted on different data that may be subject to artifact and bias. We'd like to call for better community efforts on this issue. We provided exact NLL and a tight ELBO formulation with a goodwill that this could be a step towards better evaluation of flow models, and we have tried our best to make it as comparable as possible. Pdf: /pdf/067649711cae83db2465e34d6f9a0b8b61961445.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Imitating Language via Scalable Inverse Reinforcement Learning	Accept (poster)	Summary: This paper investigates the possibility of applying inverse reinforcement learning for language imitation learning problem. Specifically, the paper reformulates IQLearn as a temporal difference regularized extension of MLE. This basically bridges inverse reinforcement learning with MLE with a coefficient $\lambda$ that can be tuned to change the weight of inverse RL loss versus MLE loss. In experiments on XSUM and GMS8K, the proposed method outperforms vanilla MLE consistently across various base models from 250M to 3B parameters. Time profiling is also reported and it is shown that the proposed offline IQLearn induces minimal computational overhead compared to vanilla MLE. Strengths: The paper is well-written and easy to follow. The problem is well-motivated as inverse RL is already shown to work better than vanilla MLE is many robotics control problems. The idea of applying inverse RL to the language setting seems novel to me. However, I am not an expert in inverse RL so not sure about the novelty of reformulating IQLearn as a temporal difference regularized extension of MLE. The proposed method is extensively validated with various base models on two common language benchmarks: XSUM and GMS8K. The section of time profiling is particularly helpful and it is nice to see offline IQLearn enjoys a similar training efficiency as MLE. Weaknesses: As admitted in the paper in Line 193 as well, the performance gain is not big (although consistent across different base models) and around 1% accuracy on GMS8K and 1 ROUGE-1 score. It would also be nice to see how this method works on other metrics for text generation such as BertScore[1]. Would be nice to have more intuitive understanding of why adversarial methods such as GAIL do not work very well. There are some places where the writing can be improved. Please see questions. [1]Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, & Yoav Artzi (2019). BERTScore: Evaluating Text Generation with BERT. CoRR, abs/1904.09675. Technical Quality: 3 Clarity: 3 Questions for Authors: Line 31, "Including" -> "This includes" The drawing of Figure 1 can be improved, currently it is hard to tell from the figure what the differences between MLE, offline IRL, and online IRL are. It seems that MLE and offline IRL are exactly the same while online IRL has some blue boxes instead of grey boxes. It is unclear to me what that means even with the help of the caption. It seems that adversarial imitation learning method GAIL is always outperformed by non-adversarial one IQLearn? What might be a reason for that? In Figure 2 and 3, is the result of IQLearn offline or online? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are discussed in the Discussion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the constructive feedback. By addressing these helpful comments and questions, the paper has improved and gained in clarity. Please mention further questions or clarifications and we will work to address them ASAP. As the review mentions, our small performance gains are consistent and considerably larger in some domains, in particular for GSM8k. They further go hand-in-hand with crucial diversity improvements across tasks. We took this feedback seriously and additionally looked into further analysis as well as additional tasks. In addition to XSUM, GSM8k, and TLDR, we added WMT22 results (with results in the general rebuttal). We further analyzed the extracted rewards as described in the general rebuttal and are able to demonstrate correlation between learned rewards and task performance metrics. The particular suggestion of BertScore as an additional metric is highly appreciated. We are looking into experiments with BertScore but will unlikely be able to generate results before the rebuttal session ends due to the computational cost of recreating experiments. Adversarial imitation is generally known to be less stable (e.g. [1,2]) and often additional heuristics are required for stabilization [3] while hyperparameter tuning becomes increasingly expensive. Due to these aspects, we have only been able to obtain results for the T5 models. This is mentioned in parts of the paper already (see e.g. Sections 3.1 and A.3.3) and we have further expanded the corresponding sections in the submission. Minor points: - We have further expanded our discussion of related work to include the suggested papers. - We have improved figure 1 and the corresponding caption. - We have fixed the remaining spelling and other mistakes. Thanks for pointing these out. - The main IQLearn experiments are offline as we mention in sections 2 and 3. We have emphasized and clarified this throughout all sections. [1] On the Algorithmic Stability of Adversarial Training, NeurIPS 2021, Yue Xing, Qifan Song, Guang Cheng [2] Language GANs falling short, ICLR 2020, Massimo Caccia, Lucas Caccia, William Fedus, Hugo Larochelle, Joelle Pineau, Laurent Charlin [3] Creating Multimodal Interactive Agents with Imitation and Self-Supervised Learning, 2022, DeepMind Interactive Agents Team: Josh Abramson, Arun Ahuja, Arthur Brussee, Federico Carnevale, Mary Cassin, Felix Fischer, Petko Georgiev, Alex Goldin, Mansi Gupta, Tim Harley, Felix Hill, Peter C Humphreys, Alden Hung, Jessica Landon, Timothy Lillicrap, Hamza Merzic, Alistair Muldal, Adam Santoro, Guy Scully, Tamara von Glehn, Greg Wayne, Nathaniel Wong, Chen Yan, Rui Zhu --- Rebuttal Comment 1.1: Comment: Thank the author for the rebuttal. Some of the concerns are clarified. However, the improvements in the main result are still not compelling enough with around 1% across many benchmarks at the cost of a more complicated method and more hyperparameters. I will therefore stay with my original score. --- Reply to Comment 1.1.1: Title: Additional Information for Official Comment by Reviewer pjSm Comment: Thank you very much for the quick response. We appreciate the ability to continue this discussion. Trading off complexity and performance improvements is highly important and we would like to contextualize our contributions with this background. 1) The IQLearn offline variant used across most of our experiments does not require online sampling and has a single extra hyperparameter which can be set to 0.1 for most tasks with reasonable gains. 2) A key advantage of IRL in comparison to MLE based training is the extraction of reward information which can provide various opportunities for future research and in particular increasing the robustness of RLHF via better reward models. 3) The single digit absolut performance gains should be seen in relation to SFT performance. Relative gains are more crucial, e.g. between 10-20% for the PaLM2 models and up to 30% for the smallest T5 models for GSM8k. Furthermore, please be aware that the initial pretrained models already have much greater than 0 performance and are shared across methods such that we only change the algorithm for a comparably small amount of overall training data (due to the computational cost of pretraining experiments). 4) It is correct that some domains have smaller improvement, in particular XSUM. Including them serves 2 purposes. They demonstrate impact on increased diversity but also enable us to discuss IRL benefits in relation to task/data properties; it is more suited for some tasks than others. 5) In general, we add strong diversity improvements in addition to performance which can be highly relevant for creative applications as well as for online data generation such as in RLHF. Given the simplicity in the offline setting (Pt 1) with a single hyperparameter, these aspects represent considerable advantages.	Summary: The paper introduces a RL-centric perspective to imitation for LLMs, with an novel imitation learning algorithm for fine tuning LLMs, that is derived from IQLearn, forming a dynamics dependent temporal difference regularized variant of MLE. The authors provide an extensive analysis of other potential inverse RL algorithms for imitation learning with LLMs, and demonstrate IQLearns ability to perform well at tasks whilst still having good diversity. Their approach is a promising offline IRL approach. Strengths: * Clarity: The introduction is clear and well motivates the problem. Overall the paper has good clarity, bar a few typo's. * Originality: The RL-centric perspective to imitation for LLMs and the IQLearn objective and formulation appears novel and significant to the community. Weaknesses: * The paper could benefit from more experimental tasks such as CommonGEN, ROCStories, EMNLP2017, and COCO, as done in TextGAIL. * Missing error bars in Figure 4 and Figures 2 & 3 could benefit from error bars as well, perhaps with a different version in the appendix. * Minor: The first paragraph could benefit from references to ground the reader. * Minor: L90. Reading the references cited in this paragraph, they do not mention the term $(1-\gamma)$; perhaps find a reference for it, remove this term, or explain its inclusion in a footnote or an appendix. * Minor: Equation 8 could put an enclosing bracket for the min operator to help guide the reader. * Minor: The presentation of the figures could be improved by making the font size larger and using vector graphics. Typos: * L40: “behavior cloning [7], its” -> “behavior cloning [7], and its” * L123: “further simplify” -> “further simplify it” * L168: “temporal different regularization” -> “temporal difference regularization” * L174: “we a short” -> “we use a short” * L174: “pure MLE beneficial” -> “pure MLE being beneficial” * L:202: “full trajectories” -> “full of trajectories” * L:218: “Right” -> “Left” Technical Quality: 4 Clarity: 3 Questions for Authors: * Are the T5 model’s pre-trained? I presume they are, if so it could be helpful to state in the main paper in section 3.2.1 that the T5 models are pre-trained, and details of this. * As stated, IQLearn with Lambda = 0 “retrieves standard MLE.” Can we then interpret Figure 2 as IQLearn (Lambda=0.0) as MLE? If so, it is interesting how IQLearn always outperforms MLE, even for small Lambda. Could you perform ablations for small non-zero values of Lambda to show if it converges with MLE on the task results? * How do the conclusions of Figure 4 hold if Lambda is varied? What is the expectation? * Figure 6: can you quantify the uncertainty estimate used? * How would you see future work extending your approach to use human pairwise preference datasets? That is combine approaches perhaps such as Direct Preference Optimization. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: These were adequately discussed in Section 5. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the constructive feedback. By addressing these helpful comments and questions, the paper has improved and gained in clarity. Please mention further questions or clarifications and we will work to address them ASAP. Uncertainty estimates and error bars: We always plot the standard error (standard-deviation / sqrt(number-of-seeds) in all plots. We added a comment to the paper explaining our uncertainty estimation. Figures 2 & 3 already include error bars; these are very small due to only minimal variability from the random seeds. We updated the color of these error bars to make them more legible. Due to computational requirements, we cannot add further experiments to add error bars for Figure 4 before the rebuttal deadline, but have added it to our prioritization list to follow-up when other experiments are completed. We strongly acknowledge the benefits of extending analysis and a broader task set. While experiments, in particular with hyperparameter sweeps and multiple seeds are expensive, we have added results on WMT22 training (with results in the general rebuttal). We continue looking into further experiments that can be performed with limited costs and have, in parallel, extended our analysis to obtain further insights without additional training runs. For this, we have investigated the rewards extracted by IRL as described in the general rebuttal comment. Regarding the extension to preference learning, this is a very good point and we quickly discuss opportunities in the discussions section. A key benefit is the ability to extract reward information from both demonstration and preference datasets, which could bring benefits regarding the robustness of RLHF. While these are untested for language modeling, there exist early examples for classic robotics tasks [1]. Ablations on \lambda to interpolate between MLE and IQLearn are of crucial interest and included (with limited settings), for example in Figure 2. We know from early tests that much smaller values critically reduce performance and \lambda = 0 algorithmically reduces to pure MLE. However, since these experiments are computationally expensive we are unable at this point to add a more detailed study. Similarly, we made the decision to sweep over the quality-diversity tradeoff for PaLM2 models via inference time settings, rather than separate training experiments, to minimize computational costs and provide better coverage of different mechanisms for the tradeoff. We agree this would be interesting to investigate, but have to rely on the related sweeps for Figures 2 and 3 for the time being. Minor points: - The T5 models are in fact pretrained. Thank you for pointing this out; we have expanded the model discussion in the paper and clarified this aspect. - We have fixed the remaining spelling and other mistakes, as well as added further clarifications as suggested. Thank you for pointing these out. [1] Learning Reward Functions from Diverse Sources of Human Feedback: Optimally Integrating Demonstrations and Preferences, 2020, Erdem Bıyık, Dylan P. Losey, Malayandi Palan, Nicholas C. Landolfi, Gleb Shevchuk, Dorsa Sadigh --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed rebuttal response, clarifications and improvements to the paper. The new results on WMT22 are a welcomed addition. As my concerns have been addressed I have raised my score.	Summary: This paper investigates using inverse reinforcement learning (IRL) to directly optimize sequences for fine-tuning large language models. Moreover, this work reformulates inverse soft-Q-learning as a temporal difference regularized extension of maximum likelihood estimation (MLE), which bridges IRL and MLE in supervised fine-tuning tasks. The experiments demonstrate the clear benefits of IRL-based imitation for retaining diverse responses while maximizing task performance. Strengths: 1. This work investigates the potential of inverse reinforcement learning for tuning LLMs. The strengths of IRL are evaluated in terms of performance, diversity, and computational requirements. The experiments demonstrate the computationally cheaper offline IRL can obtain crucial performance gain over MLE-based LLM tunning. 2. IQLearn is a method that can work online and offline. The re-formulation of IQLearn to LLM tunning enables large performance gains compared to MLE methods. Moreover, IQLearn can be regarded as a regularized extension of MLE, which bridges IRL and MLE for LLM tunning. Weaknesses: 1. IRL methods, even for offline IRL methods, cost more time than MLE for LLM tunning. 2. More experimental details are required to reproduce the results. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The original GAIL is an online IRL method, does GAIL execute offline or online in the experiments? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The discussion section discussed the potential limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the constructive feedback. By addressing these helpful comments and questions, the paper has improved and gained in clarity. Please mention further questions or clarifications and we will work to address them ASAP. Thank you for pointing out computational requirements. We crucially reduce computational costs by training offline and thus removing the additional sampling cost for IQLearn in all sections but the online to offline comparison in section 3.3.1. The remaining difference in computational costs are partially due to a non-optimized implementation in our codebase, which shares many components between offline and online IRL algorithms. We have further emphasized the implementation dependence in the paper. We have started to recreate a pure and clean offline implementation but will be unable to verify results with the new implementation before the rebuttal ends. We have added more general details on the experiments. We would be glad to provide further clarifications and experiment details if the reviewer could specify what additional information or results would be needed from their perspective. We have further expanded the results, adding WMT22 where offline IQ-Learn shows increased performance in comparison to MLE-based training (with results in the general rebuttal). Finally, a key benefit for IRL-based models lies in additionally extracting reward information. To explore this, we have added analysis for the correlation between extracted rewards and performance metrics, with results summarized in the general rebuttal. Minor points: - GAIL is an online method that requires training the discriminator between the online agent and offline demonstration dataset. We have emphasized this in the paper. --- Rebuttal Comment 1.1: Comment: Thank the authors for the rebuttal. I maintain my score as is. --- Reply to Comment 1.1.1: Title: Additional Information for Official Comment by Reviewer 6J5H Comment: Thank you for the quick response. Our team has worked hard to answer your questions and address your comments. Could you please provide further information or open questions underlying the kept rather than updated score to enable us to address them.	Summary: In this paper the authors aim to cast fine-tuning LM as inverse RL from the perspective of distribution matching. In order to avoid online generation in existing IRL algorithms such as GAIL, the authors leverage an offline IRL algorithm, IQL, and reformulate it as a temporal difference regularized extension of MLE. This derivation enables the use of expert samples solely in the form of distribution matching. Experiments in XSUM and GSM8K show that their method outperforms both MLE and online IRL algorithms in terms of task performance and diversity. Strengths: The reformulation of IQL into distribution matching is novel and it leads to a simple and efficient version of IQL. The simplicity of the new derivation of IQL may potentially make it an impactful learning objective like MLE. Weaknesses: - The improvement is somewhat marginal especially in the task performance on the summarization benchmark. - It is not clear the proposed method is better than MLE in terms of quality-diversity trade-off. - The diversity evaluation for lower open-endedness tasks like GSM8K does not make much sense. Technical Quality: 2 Clarity: 2 Questions for Authors: - Could the authors provide an analysis for quality-diversity trade-off [1] of the algorithms? - Why does the temporal regularization term encourages higher diversity? - It would be helpful if the paper contained pseudo code of the proposed algorithm. [1] LANGUAGE GANS FALLING SHORT Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: see weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the constructive feedback. By addressing these helpful comments and questions, the paper has improved and gained in clarity. Please mention further questions or clarifications and we will work to address them ASAP. One of the review’s key points regards limited performance improvements on summarization. Here, we would like to emphasize that the goal of our paper is to establish in which tasks IRL-like learning is most impactful. Overall, the performance gains are stronger across the other domains, and in summarization we additionally see considerably improved diversity of generations. We have further emphasized the limitations of MLE/BC with respect to long target/trajectory lengths known from the imitation learning literature (e.g. DAgger [A]). To further improve the empirical evaluation, we have added results on WMT22 where offline IQ-Learn shows increased performance in comparison to MLE-based training (with results in the general rebuttal). Finally, a key benefit for IRL-based models lies in additionally extracting reward information. To explore this, we have added analysis for the correlation between extracted rewards and performance metrics, with results summarized in the general rebuttal. Regarding the connection between IRL and higher diversity, equation 3 in the paper describes this best. The original state-distribution-matching formulation builds on entropy-regularized RL and in particular includes a term for causal entropy. We have further emphasized this connection in the paper. Thank you in particular for enabling us to strengthen the connection to [1]. Our temperature sweeps as well as the sweep over different entropy regularisations enable the generation of the kind of quality diversity evaluations that [1] is arguing for. While we move away from purely adversarial inverse RL and expand to saddle-point-based methods like IQLearn, their insights are highly relevant and we have clarified additional heuristics required to enable better performance for our implementation of GAIL (which are already included in the appendix but were previously skipped in the main paper). Minor: - We add pseudo code to the appendix to minimize required space in the main paper. - We further expanded our discussion of related work to include the suggested paper. [A] A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning, AISTATS 2011, Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell --- Rebuttal Comment 1.1: Title: Follow up for reviewer aRfw Comment: Thank you for your original review. Our team has worked hard to answer your questions and address your comments. Since the original rebuttal, we further expanded the reward analysis to GSM8k and TLDR. With the discussion ending soon please ask any remaining questions or comments soon and we will answer ASAP; or share if previous answers and additional experiments have addressed your concerns.	Rebuttal 1: Rebuttal: We would like to thank our reviewers for their valuable feedback. The paper has already considerably improved and we will work hard to address further comments and questions to conclude the rebuttal process to everyone’s satisfaction. The reviews generally appreciate the discussion of inverse RL in the context of language modeling as well as the derivations and reformulation of IQLearn as temporal difference regularized MLE. Multiple reviews ask for the addition of further tasks and, while the computational cost is non-negligible, we have added results on the large WMT22 English-German task (285M examples). IQLearn gains are between 2.2-2.4 over MLE with beam search decoding. Early results with computationally cheaper decoding via temperature sampling indicate larger gains (evaluation is ongoing and we will add the complete results via comments ASAP). We are investigating further options for tasks as well. We aim to address questions around the key benefits of IRL. Here, in addition to training a policy (i.e. the generative LLM), IRL enables the extraction of reward functions from demonstration data. We discuss multiple benefits for this additional source of reward information in the discussion section and this includes the potential to mitigate reward model overoptimization during RLHF [1], and we have added more analysis on this (see below). In IQLearn, in particular, we can recover rewards from the Q-function, as we describe in line 136 (via $r_t=Q(s_t, a_t) - \gamma V(s_{t+1})$). We have added correlation analysis between the extracted rewards and task performance metrics. As an example, current results show both Pearson Correlation Coefficient and Spearman's Rank Correlation Coefficient between 0.21 and 0.44 for IRL-extracted rewards and BLEU and between 0.27 and 0.48 for ChrF on WMT. For comparison, the corresponding correlations for MLE are respectively around 0.04-0.05 and 0.004-0.005. The complete results have been added to the paper. PDF updates are unfortunately not possible during the rebuttal process, so please reach out if there are further questions about these results. Further specific details for each review are included in our individual answers. Please do not hesitate to point out further questions or comments. [1] Scaling Laws for Reward Model Overoptimization, 2022, Leo Gao, John Schulman, Jacob Hilton	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper looks at the existing IQLearn algorithm and makes some connections with standard maximum likelihood learning in the context of sequence based language models. On the empirical front, results are given showing different model performance along diversity (measured via self-Bleu) and some accuracy measure (such as rouge). These are given on a few public datasets. The empirical part of the paper lacks focus on what is trying to be optimised, and the actual metrics that are reported leave open some potential that the performance is not necessarily better. For example self-bleu may increase, but it isn't clear this is a better model. The diversity may be erroneous, but still more diverse and still scoring well on the token overlap measures of rouge etc. The other graphs and figures fail to leave the reader with a clear story of why using imitation learning at the sequence level, rather than MLE or RLHF (imitation learning at the token level) is necessarily better. Strengths: * A solid theoretical analysis is given which draws links between MLE and forms of IRL. Weaknesses: * The story of the paper is not that clear, and the empirical results do not tell a clear story. The paper shows that the IRL methods can be used to obtain better models on the Pareto type plots of quality and diversity, however it's not clear that these are actually better models. * The data used in the empirical section only appears a long way into section 3. It would be much clearer if the empirical section had a clearer introduction of what types of problems are being looked at, and what data is being used for such. Technical Quality: 2 Clarity: 2 Questions for Authors: What was the conclusion for which tasks and dataset properties is IRL most relevant for? (This is a question at line 160, however it's not clear that it is answered). Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: No concerns. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the constructive feedback. By addressing these helpful comments and questions, the paper has improved and gained in clarity. Please mention further questions or clarifications and we will work to address them ASAP. The included experiments show that IRL methods (GAIL/IQLearn) improve both performance and, crucially, diversity metrics, compared to common next token likelihood. The performance improvements are small for XSUM but consistent across domains. If the reviewer could specify what additional information or results would be needed to fulfill specific additional criteria for 'actually better models', we would be glad to provide further clarification. In addition, in the paper we have clarified the exact metrics used, and how to interpret the results. In particular for Self-BLEU, we will expand the captions to emphasize that smaller numbers indicate better diversity (lower self-similarity) and that the y-axis is flipped. Furthermore, if the point above is meant in relation to the comment “The empirical part of the paper lacks focus on what is trying to be optimised”, we can clarify. Section 2 describes the optimization (aspects such as the relation of reverse KL to MLE, regularized chi2 for Iq-learn). However, evaluating these divergences once trained is hard, because we do not have access to the distribution underlying the dataset. In classic imitation learning papers, the underlying ground truth reward function can be used for evaluating the quality of the imitating agent (typically, the expert is an RL agent trained on some known reward). Yet, in our case, we do not have such a reward function, and the expert is a human. We therefore have to rely on these proxy metrics, and we consider classic ones in the context of language models. In parallel to clarifying this point, we have expanded our experiments to include results on WMT22 (with results in the general rebuttal). Furthermore, IRL-based models have the added benefit of extracting reward information from demonstration data. We have added further analysis on the correlation between extracted rewards and performance metrics to visualize this advantage, with exact numbers added in the general rebuttal. Regarding the question of most relevant tasks for IRL, we have emphasized the theoretic connection to target/trajectory length known through the imitation learning literature (e.g. DAgger[1]). Practically, the performance gains are stronger with longer target lengths (GSM8k) and partially with smaller datasets (ablations on smaller subsets of XSUM), while improved diversity of generations is persistent across tasks. We have expanded our discussion of these aspects in the paper. Minor points: - We will introduce the used datasets and tasks earlier in Section 3. - RLHF: We do not provide an alternative to RLHF but rather emphasize the RL perspective during SFT, which could in the future enable better connections between SFT and RLHF data sources. As both method types use different data sources, a direct, fair comparison between IRL-based approaches and RLHF is not possible. Instead, these are complementary. [1] A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning, AISTATS 2011, Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell --- Rebuttal Comment 1.1: Title: Follow up to reviewer tAhh Comment: Thank you for your original review. Our team has worked hard to answer your questions and address your comments. Since the original rebuttal, we further expanded the reward analysis to GSM8k and TLDR. With the discussion ending soon please ask any remaining questions or comments soon and we will answer ASAP; or share if previous answers and additional experiments have addressed your concerns.	null	null	null	null	null	null
Position Coupling: Improving Length Generalization of Arithmetic Transformers Using Task Structure	Accept (poster)	Summary: This paper proposes a method called "position coupling" to enhance the length generalization of Transformer models, specifically targeting arithmetic tasks such as integer addition. The authors claim both empirical success and theoretical guarantees for their approach. The method involves assigning the same position IDs to semantically related tokens to better embed the task structure into the Transformer’s positional encoding. Strengths: 1. Novelty: The idea of coupling positional embeddings to reflect semantic relationships is interesting and novel. 2. Empirical Results: The paper provides detailed empirical results showing improved performance on length generalization for integer addition tasks. 3. Theoretical Analysis: The authors offer theoretical insights into why position coupling should work, which is a positive aspect of the paper. Weaknesses: 1. Scope of Application: While the paper demonstrates the effectiveness of position coupling on integer addition tasks, the generalizability to broader and more complex tasks is not convincingly shown. The examples provided are limited and do not cover a wide range of real-world applications. 2. Complexity and Practicality: The proposed method introduces additional complexity in the positional encoding process. This complexity may limit the practical applicability of the method, especially for larger and more diverse datasets. 3. Proper Citation of Related Work: In section 3, lines 113-115, the randpos method for length extrapolation, introduced by "Randomized Positional Encodings Boost Length Generalization of Transformers" (ACL 2023), is not properly cited. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can you provide insights on whether "position coupling" can be generalized to other positional encoding schemes beyond Absolute Positional Encoding? It would be beneficial to understand if and how this method can be adapted or integrated with other popular positional encodings, such as Rotary Positional Embedding (RoPE). Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: As mentioned above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for the reviewer’s effort and constructive feedback. Below we summarize your feedback/questions and address these one by one. > W1. Scope of Application: the generalizability of position coupling to broader and more complex tasks is not convincingly shown. - We agree that the applicability of position coupling to broader tasks has not yet been explored in this paper. However, we want to note that enhancing the length generalization ability of models, even for the addition tasks, is considered an important problem as addressing and improving the arithmetic abilities of models can lead to a better understanding of the model’s capabilities. - Definitely, the next step of our work is to extend the application of position coupling to complex, real-world tasks. To do so, we plan to develop a method called “automatic coupling”. Unlike position coupling which requires manual assignment of position ID for each token, it will automatically learn the task structure and assign appropriate position IDs without human intervention. This advancement will enable broader applications beyond simple arithmetic/algorithmic tasks. We believe our findings can serve as a stepping stone towards this future development. > W2. Complexity and Practicality: Position coupling introduces additional complexities in PE, which may limit its practical applicability. - The introduction of position coupling does add complexity, particularly in designing task-specific couplings. However, once the design is determined, the additional overhead due to its implementation is negligible, as explained in lines 659-662. - Furthermore, in contrast to approaches such as [1] and [2] that employ index hinting which requires doubling the input sequence length, position coupling does not increase the sequence length. As a result, we believe that there are no significant additional complexities (e.g., time, memory) introduced during the training phase. - Additionally, we believe that the development of “automatic coupling”, as mentioned earlier, has the potential to fully address the complexity issue. - If there are remaining concerns regarding additional complexity introduced by our method, please let us know without hesitation. > W3. Proper Citation of Related Work: “Randomized Position Encodings Boost Length Generalization of Transformers” (ACL 2023) is not properly cited. - We agree that there are some similarities in the underlying concepts, and therefore we will add the citation and provide a more detailed comparison. - We note that our method (described in lines 113-115) differs from that of Ruoss et al., 2023. Our method randomly selects the start position ID and assigns a sequence of consecutive numbers. In contrast, Ruoss et al., 2023 assigns a sequence of increasing integers, which are generally not consecutive. > Q1. Can position coupling be generalized to other PE schemes than APE? - Thank you for an insightful question. During the rebuttal period, we realized that our proposed coupling method could be extended to Relative PEs such as RoPE and T5’s relative bias. The relative position between the query and key is determined by the difference in position IDs that were assigned by our position coupling method. Specifically for RoPE, we conducted experiments and observed that position coupling enhances the length generalization capability of RoPE (See Fig. 4 of the PDF file in our Global Response). We believe that this approach has significant potential for the research of length generalization and we will add experiments in our next revision. - Instead of adapting our method to other PE methods than APE, it is also possible to integrate position coupling with existing RPE methods such as standard RPEs, RoPE, or FIRE [3]. In this approach, RPE methods are used independently alongside position coupling, which may provide hope for application to general LLM models. Thus, we think this is another promising direction to further improve the applicability of our proposed methodology, so we will conduct some experiments and consider adding the results to our final manuscript. We hope our response has adequately addressed the reviewer's concerns, and we would appreciate it if you could reconsider your assessment. --- References [1] Zhou et al., What algorithms can transformers learn? a study in length generalization. ICLR 2024. [2] Zhou et al., Transformers can achieve length generalization but not robustly. arXiv preprint, 2024. [3] Li et al., Functional interpolation for relative positions improves long context transformers. arXiv preprint, 2023. --- Rebuttal Comment 1.1: Comment: Thank you for your feedback. I will raise my score from 5 to 6. --- Reply to Comment 1.1.1: Comment: Thank you for your feedback and for reconsidering the score. We will be sure to incorporate your suggestions in our next revision. If you have any further questions or comments, please feel free to share them.	Summary: This paper proposes a new way to bake in the positional structure of problems for transformers. The authors also analyze the potential for models with and without their proposed positional coupling to solve problems of arbitrary size. They also show empirically that their method helps a small transformer learn addition. Strengths: 1. Originality: To the best of my knowledge the theoretical analyses are novel and the methods are novel (up to concurrent works). 1. Quality: The work is detailed and thorough. 1. Clarity: The paper is well written and clear. Weaknesses: 1. The significance is limited. i. I think often algorithm learning papers can come across as limited in impact and I don't mean to bring this up. Rather, this paper specifically makes a portion of its contributions around 1-layer transformers where some of the claims seem particularly limited. For example, the proof of impossibility of a 1-layer decoder only transformer is interesting, but I don't think it applies to even a 2-layer model weakening the motivation for the use of this positional coupling in practical settings. ii. The authors state the limitation that this method requires a priori understanding of the problem structure. I appreciate that this is acknowledged as it feels important to me. With enough a priori knowledge of problem structure, one can often solve the problem without ML. This paper is a cool demonstration of learning addition from data, but it is limited in this sense. iii. The fact that operands are zero padded to match in length is also a limitation here. This is another instance of requiring problem-aware data processing. The test set is only made up of operands of the same size (clarifying questions below). Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Is my understanding of the test set construction accurate? All addition problems are of the form A + B where A and B have the same number of digits? What happens when we test these models on things like "20345 + 34 = " for example? 1. Why is only the answer reversed? In related work, reversal is often discussed, but this particular scheme requires even more structure-aware processing. I think blindly reversing all numbers that a model encounters would be more compelling than knowing to only reverse the answer. Do the methods in this paper perform better/worse with more consistent data processing? 1. In Section 4, deeper models are shown to perform worse than the shallower models, which the authors attribute to the difficulty of training deep networks. Do the deeper models fit the training data as well as the shallow ones? It seems this way from the plot and if it is the case, then I think another explanation is needed here as the optimization does not seem to be the problem. Perhaps there is an argument to be made about over-fitting, but the rate of degradation seems to vary somewhat nicely with added depth raising more questions about what could be happening. Can the authors offer any other insights here? Maybe some understanding of what is happening at operand lengths where accuracy is ~80%, i.e. which examples are correct/wrong would lend some clarity here. I'm excited to discuss my review with the authors during the rebuttal period and remain open to improving my score with answers to my questions. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed all limitations in the limitations section. The significance is limited in ways I addressed above, but these are relatively minor. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer's valuable questions and rich comments, and we hope our response relevantly addresses all points raised in the review. > W1. Some of the contributions are limited to 1-layer Transformers. - We first highlight that our theoretical construction (Thm 5.1) is not limited to 1-layer models because it naturally extends to multi-layer cases. - As you pointed out, our impossibility result on NoPE may not hold for a multi-layer model. However, the main purpose of this result is to show a provable gap in the expressive power between position coupling and NoPE. Therefore, we believe that the inapplicability of the impossibility result in multi-layer models does not weaken the motivation for using position coupling. > W2. With enough knowledge of the problem, we don’t need ML. - We highlight that length generalization is a crucial issue in LLM. Plenty of existing research [1–4] (including us) regards arithmetic/algorithmic tasks as manageable but interesting test beds for studying length generalization because LLMs fail to length-generalize even on such simple tasks. - The main message from our work is that incorporating task structure when designing positional encodings can effectively improve the length generalization capabilities. We believe our findings can serve as a stepping stone for future research in this area. > W3. Zero-padding is a problem-aware data processing. - Note that zero-padding to match the operand lengths is prevalent in this research area ([3–5]). - One can implement our method without relying on zero-padding. We present experimental results using a no-padding scheme in the attached PDF file (Fig. 1) in our General Response. While the no-padding position coupling approach fails in in-distribution generalization when trained on a 1-layer model (due to the increased complexity of the algorithm that the model should learn), combined with proper reversing of the number(s), it functionally extrapolates the performance of deeper models. > Q1. Should the operands in every testing example have the same length? - You’re correct: we sampled the operands to be the same length while testing. To test the model (trained on zero-padding format) with the sample “20345+34=”, we could input it as “20345+00034=”. - To address the concern, we tested on operands sampled with different lengths. See Fig. 3 of our PDF attached to the General Response. Each axis corresponds to the lengths of each operand. The results show that the model is capable of solving tasks even when the two operands have different lengths, although zero-padding is applied to ensure consistency in the input format. We will add this result to our revision. > Q2. Reversing the answer solely is a problem-aware data processing. - Note that solely reversing the answer is also a common practice in this research area [3–5]. From now on, let us compare two different formattings: (a) solely reversing the answer and (b) reversing all the numbers. - We first empirically compare (a) and (b) while applying position coupling (see Fig. 1 of the PDF in our General Response). For 1-layer models, both formats exhibit near-perfect length extrapolation and show little difference. Conversely, for 6-layer models, a noticeable performance gap emerges. With zero-paddings, (b) performs better than (a); but if there’s no padding, (a) performs much better than (b). There seems to be no clear winner between the two. - The similarity between (a) and (b) for 1-layer models is expected. If we assign the position IDs based on the significance of the digits as usual, there is NO effective difference between (a) and (b) for a 1-layer model in terms of its prediction, which can be deduced from Prop 5.2. Accordingly, our Theorem 5.1 based on (a) can also be applied to (b) without any modification. - The difference between (a) and (b) for deeper models is also expected. In multi-layer models, the causal attention mask causes each token embedding to depend on the embeddings of preceding tokens after the first layer. Thus, unlike in the previous case, the predictions of the model may differ depending on the input format. > Q3. The optimization might not be the reason why deeper models perform worse. - There are two key aspects of optimization: convergence and implicit bias. Our explanation primarily concerns implicit bias. To answer your first question, deeper models do fit the training samples just as well as the shallower models. Thus, convergence is not the issue. We also think that overfitting is not the case, as the trained models only struggle with longer sequences while achieving perfect accuracy on in-distribution samples. - Therefore, we believe that the issue lies in the implicit bias of the models. Among the infinite number of solutions that achieve zero training loss, shallower models seem to possess a better implicit bias, which allows them to find solutions that generalize better for longer sequences. - Specifically, we conjecture that the outstanding performance of the 1-layer model is due to its relatively restricted expressivity and that the model has no way to fit the training samples other than learning the true algorithm. However, for deeper models, the model can still fit the training samples without necessarily learning the true algorithm due to greater expressivity. - For a detailed discussion on this, please refer to our General Response. --- References [1] Jelassi et al. Length generalization in arithmetic transformers. arXiv preprint, 2023. [2] Kazemnejad et al.. The impact of positional encoding on length generalization in transformers. NeurIPS 2023. [3] Zhou et al., What algorithms can transformers learn? a study in length generalization. ICLR 2024. [4] Zhou et al. Transformers can achieve length generalization but not robustly. arXiv preprint, 2024. [5] Lee et al. Teaching arithmetic to small transformers. ICLR 2024. --- Rebuttal Comment 1.1: Title: Reviewer Response Comment: I have read the detailed reply from the authors. The additional clarity and the details in the general response and the PDF have addressed my concerns. I have changed my score from a 6 to a 7. At this point, I think the paper is clearly above the acceptance threshold. --- Reply to Comment 1.1.1: Comment: Thank you for your feedback and for reconsidering the score. We are glad to hear that our response addressed your concerns and that you view the paper as above the acceptance threshold. We would also be happy to hear if you have any additional thoughts or suggestions.	Summary: This paper proposes "position coupling", a novel technique to improve the length generalization ability of decoder-only Transformers. Unlike standard positional embeddings, position coupling assigns the same position ID to semantically related tokens across the input sequence, directly embedding task structure within the model. This approach achieves near-perfect length generalization on integer addition, extrapolating from training on up to 30-digit sums to successfully solving 200-digit additions. The authors theoretically prove the capability of a 1-layer Transformer with coupled positions to perform additions with exponentially long operands. They further demonstrate the effectiveness of position coupling on tasks like addition with multiple summands, N×2 multiplication, and copy/reverse operations, providing a theoretical construction for the multiplication task as well. The paper explores the application of position coupling in 2D tasks, showcasing its potential beyond 1D sequences. Strengths: - While inspired by index hinting, position coupling offers a novel and more elegant solution for incorporating task structure into Transformers. It directly embeds this information within the positional encoding, eliminating the need for augmenting the input sequence and simplifying model training. - The paper demonstrates a high level of technical rigor. The proposed method is well-motivated and thoroughly evaluated on a variety of tasks, including both empirical analyses and theoretical constructions. The experimental design is comprehensive, with thorough comparisons against relevant baselines and ablations on various architectural choices. - The paper is well-written and easy to follow. The authors clearly articulate the problem, their proposed solution, and the key contributions. The use of figures and examples effectively illustrates the concepts and makes the theoretical constructions more accessible. - This work addresses a crucial challenge in Transformer-based learning: length generalization. The impressive results on arithmetic tasks, particularly the significant extrapolation achieved in addition, highlight the potential of position coupling for enabling Transformers to learn algorithms and generalize far beyond their training data. The theoretical analyses provide valuable insights into the mechanism of position coupling and its role in achieving length generalization. The extension to 2D tasks further broadens the applicability and impact of this work. Weaknesses: - While Theorem 5.1 provides a strong theoretical foundation for the capabilities of 1-layer Transformers with position coupling, the paper lacks a theoretical understanding of why deeper models might perform worse despite their greater expressivity. Further theoretical analysis on the interaction between position coupling and depth, especially on tasks like N×2 multiplication where deeper models are necessary, would significantly strengthen the work. - The success of position coupling heavily relies on the specific input format (reversed response, zero-padding, etc.). It remains unclear how robust the method is to variations in input format and whether it can be applied to tasks where such specific formatting is not possible or desirable. Exploring alternative position coupling schemes that are less sensitive to the input format or evaluating the method on tasks with diverse input structures would strengthen the claims of generalizability. - The paper primarily compares position coupling against basic positional embedding techniques. Including a broader range of recent length generalization techniques in the experimental comparison, such as those based on relative positional encodings would provide a more comprehensive understanding of the method's effectiveness and potential advantages. - The minesweeper generator task serves as a preliminary investigation into the potential of position coupling for multi-dimensional tasks. However, exploring the applicability and effectiveness of position coupling on a wider range of 2D or even higher-dimensional tasks, potentially with more complex structures, would further highlight the significance and generalizability of the proposed method. Technical Quality: 2 Clarity: 3 Questions for Authors: - The paper acknowledges the reliance on a specific input format for optimal performance. Could you elaborate on the sensitivity of position coupling to variations in input format? Have you experimented with alternative formats and if so, what were the outcomes? - While the paper explores several arithmetic and algorithmic tasks, it would be beneficial to understand the limitations of position coupling. Are there specific task characteristics or structures that might render position coupling less effective or even inapplicable? - The design of the position coupling scheme relies on an intuitive understanding of the task structure. Could you formalize the notion of task structure and provide guidelines for designing appropriate position coupling schemes for different tasks? - For the 2D task, why does using the same embedding layer for both position coupling modules perform better than separate layers? Is this specific to the task or a general observation? - Would it be possible to extend the evaluation of position coupling to more complex 2D tasks, such as image-related tasks or tasks involving graphs or other non-sequential structures? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful comments. We provide our response to the reviewer's concerns. > W1. Theoretically, why do deeper models perform worse despite their expressivity? - For a broader answer to the question, see our General Response. - We hypothesize that the performance degradation is due to the bad implicit bias of deep models (learning shortcuts to only achieve in-distribution generalization) when learning a simple algorithm to solve the task. We believe exploring a theoretical explanation for the bad implicit bias of large models on low-complexity tasks is a promising research direction. > W2+Q1. How robust is position coupling to input formats? Can it be applied to tasks when specific formatting is undesirable? - See our General Response for details on the robustness to input formats. - As we emphasized there, proper use of the input format is crucial for solving the task even with position coupling. Also, the model’s performance depends on the choice of input format. Thus, if we cannot apply any formatting, we should not expect significant success in solving the given task, not just in terms of length generalization. > W3. Compare with recent techniques for length generalization. - Thank you for your suggestion. One notable length generalization result for the addition task before us appears in [1], combining FIRE (a relative PE method) and index hinting. They achieve near-perfect length generalization up to operand length 100 by training on up to 40-digit additions. (Recall that we achieve 30-to-200 generalization with a 1-layer model.) Despite their great performance, they require doubling the input length because of index hinting. In contrast, our method doesn't require doubling the input sequence, so we believe our method is more efficient to run. - We also conducted experiments combining RoPE and our method and achieved an improved length generalization compared to vanilla RoPE: see Fig. 4 of PDF in our General Response. - We will add these comparisons in our final manuscript. - [1] Zhou et al., Transformers can achieve length generalization but not robustly. arXiv, 2024. > W4. Can position coupling be applied to a wider range of 2D or higher-dim tasks? - As demonstrated in the paper, our method significantly helps Transformers to length-generalize on the tasks with a clear structure between token positions. Extending this to various multi-dim tasks is an interesting future work. - A challenge in multi-dim tasks (not only for length generalization and for applying our method) is the exponential growth (in task dimension) of the number of tokens, which makes it difficult for the model to analyze queries and generate responses. Overcoming this dimensionality problem is an interesting future direction. > Q2. Are there specific task structures that render position coupling less effective or inapplicable? - Let us give you some examples. Our preliminary experiments weren’t very successful for some tasks including (1) addition with a varying number of operands and (2) multiplication with varying lengths of both operands. We tried couplings similar to the ones applied to simple addition and Nx2 multiplication, respectively, but they weren’t effective (although we did not invest much effort in making it work). - The algorithm for solving (1) needs to attend to a varying (over problem instances) number of positions for generating tokens. - The algorithm for solving (2) needs to attend to positions in varying relative distances in terms of the “coupled” position IDs. On the contrary, our theoretical construction for simple addition and Nx2 tasks does not suffer from these difficulties, which is the key reason for successful length generalization; thus, we do not expect length generalization on tasks (1) and (2) without any advance in input formats. - Some tasks don’t have any structure between specific positions (e.g., sorting and mode), thereby we cannot directly apply position coupling. (See our response to Reviewer ow6J.) > Q3. Can you formalize the notion of task structure and provide guidelines for designing proper coupling for different tasks? - Defining task structure involves the relationship between the query and response, focusing on which tokens in the query influence the determination of each token in the response. Designing position coupling relies on human intuition, making it challenging to provide concrete guidelines. However, for tasks with unclear coupling structures, assigning position IDs (piece-wise) consecutively may be worth trying. It is shown to be empirically effective in the Nx2 multiplication task (which is later theoretically backed by Thm 6.1), although it's not intuitive how the coupled positions capture the task structure. > Q4. Why does using the same embedding layer for both position coupling modules perform better than separate layers? - As mentioned in lines 655–657, we do not have a clear explanation for why sharing the embedding layer performs better. We conjecture it is due to the row-column symmetry of the Minesweeper generator task, meaning that transposing the board still results in a valid problem. We believe that tasks where rows and columns have different semantic structures might not benefit from using the same embedding layer. > Q5. Can we extend the evaluation of position coupling to more complex 2D tasks? - Although our study focuses on simple arithmetic/algorithmic tasks, we are eager to explore more complex multi-dimensional tasks. - For image-related tasks (e.g., involving ViT), we could apply position coupling similar to what we did for the Minesweeper generator task. For graphs and other non-sequential tasks, the position coupling scheme may need refinement, but extending it to these complex tasks is promising future work. We hope our response has adequately addressed the reviewer's concerns, and we would appreciate it if you could reconsider your assessment. --- Rebuttal Comment 1.1: Comment: Dear Reviewer CVfs, Thank you for taking the time to review our work and for providing such insightful and constructive feedback. We understand that you may have a busy schedule, but we wanted to follow up to ensure that our responses have sufficiently addressed your concerns. If you have any further questions or comments, we would be glad to hear them.	Summary: This work considers the problem of length generalization of Transformers and proposes injecting the task structure through positional embeddings for improving length generalization. Task structures are known and therefore, the authors come up with a (relatively) general heuristic to leverage this structure. The paper relies on the observation that for tasks like addition, there are “groups” of tokens that should be treated similarly as they carry the same semantics. I.e., the digits of the summands from least significant to most significant should be embedded with the same positional embedding (called position coupling) so that the model can take advantage of this structure and carry out the sum correctly. Position coupling is used in conjunction with a number of other tricks such as reversing the sum, zero padding, and using BOS/EOS. The proposed method is then evaluated comprehensively on the addition task (for which it was proposed), and strong length generalization is observed. Ablations on the number of layers, and different positional embeddings are carried out to further emphasize the importance of position coupling. The results are also backed by theory, and interestingly the attention patterns predicted by theory are observed in the experiments. Other than the addition task, Multiplication ($N\times2$) and a 2D task are considered in the experiments as well. Strengths: - The empirical results are quite strong and proper ablations and baselines are considered. - The definitions, presentation, and method are quite clear. - The method is backed by theory. - The method is relatively general, however, it is only applicable to a class of tasks where there is a clear structure to be exploited. - The predictions of the theory are further supported in the experiments (the attention patterns for carry detection) - The experiments go beyond addition to multiplication and a 2D game (where more explanation is required) Weaknesses: The proposed method, though general for some tasks, seems to be very much geared towards a specific class of tasks of interest in length generalization. In particular, tasks like sorting, mode, and parity seem to be automatically out of reach for positional coupling and limit the generality of its applicability. Technical Quality: 4 Clarity: 4 Questions for Authors: None. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: See weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive review and valuable comments. Below, we address the reviewer’s concern. > W1. Position coupling is geared towards specific tasks in length generalization. First, as mentioned in the conclusion of our paper, we focus on certain tasks with a handy structure between specific positions of tokens. Since vanilla Transformers often fail to learn the true structure of the tasks—even for simple algorithmic tasks—as observed by plenty of previous works [1–10], we aim to guide the model to properly learn the structure by coupling the positions, thereby improving length generalization. The structure between positions is important for employing our method, so we admit that it is not easy to apply our method to some tasks. However, we address the three tasks raised by the reviewer. Parity. - Given a binary query sequence, the objective is to output 1 if there is an odd number of 1s in the query and to output 0 otherwise. - Without any data processing, it is a difficult task for Transformers to achieve length generalization [1,4,5,7,9]. However, with the help of a scratchpad generated by leveraging the idea of unrolling the query, Transformers can achieve length generalization [9]. To utilize the power of position coupling, we opt for a different scratchpad from that used in [9]. - By unrolling the query, we can generate a sequence of partial results. Let’s put the partial result from the 1st to n-th token of the query into the n-th token of the response sequence. For example, if the query is given as 010011, the response with scratchpad would be 011101. Then, the last token of the response immediately becomes the answer for the original task. We train the model to generate the whole response (including scratchpad) for each query to solve the task step-by-step with next-token prediction. Now, we can naturally couple the n-th positions of the query and the response when we apply our method. For example, given an input sequence “010011=011101”, we can assign (3,4,5,6,7,8,2,3,4,5,6,7,8) as position IDs (when the starting ID is randomly chosen as 3). - To showcase the efficacy of position coupling on the parity task with a proper input format, we compare 4 different settings: position coupling with/without scratchpad and NoPE with/without scratchpad. For the “position coupling without scratchpad” setting, we naively couple all the positions (except for the ‘=’ token) with the same position ID (e.g., “010011=1” can get (3,3,3,3,3,3,4,3)). We train the models on the queries of lengths 1–20 and test the lengths up to 100. We measure exact-match accuracy (including scratchpad if applicable) as well as the accuracy only for the single token at the position of the last token of the response except for the EOS token (called “parity accuracy”). The result of the experiments is shown in the PDF file attached in our General Response. Without scratchpads, both NoPE and position coupling cannot even achieve good in-distribution performances. Even with our scratchpads, without any position embeddings (NoPE), a 6-layer 8-head model showcases a very restricted length generalization capability up to the length of ~30. The model performs worse than random from the length of 40 because the model sometimes outputs tokens other than 0 or 1. Most importantly, our 1-layer 4-head model with position coupling and scratchpad achieves perfect length generalization up to length 100! We strongly believe that thanks to the combination of coupled position IDs and scratchpad enabling Transformers to learn a simple algorithm for solving the task with next-token prediction, we could achieve another remarkable length generalization result. - We will add this result to our final manuscript with more ablations. Sorting and Mode. - The sorting task aims to generate a response equal to the sorted sequence of the given query; the mode task aims to find the most frequent token appearing in the given query. - In both tasks, there is no exploitable structure between the positions of tokens. Thus, it is not straightforward to apply position coupling to solve these tasks, thereby we did not test our method on these tasks. - Not only is it unnatural to couple the positions, but it is also unnecessary to do so. Vanilla Transformers already length-generalize well on these tasks [9]. In short, position coupling is an effective method if we can create a clear structure between positions with proper usage of input format; however, it is inapplicable if there is no such structure. Nonetheless, there are a lot of real-world tasks whose underlying structure between positions is vague or unavailable. We leave the research direction of extending our idea to such tasks by automatically discovering appropriate couplings of the positions as interesting future work. Please let us know without hesitation if you have further questions or comments. --- References [1] Bhattamishra et al., On the ability and limitations of transformers to recognize formal languages. arXiv preprint, 2020. [2] Kim et al., Have you seen that number? investigating extrapolation in question answering models. NeurIPS 2021. [3] Nye et al., Show your work: Scratchpads for intermediate computation with language models. arXiv preprint, 2021. [4] Chiang and Cholak., Overcoming a theoretical limitation of self-attention, arXiv preprint, 2022. [5] Delétang et al., Neural networks and the Chomsky hierarchy, ICLR 2023. [6] Kazemnejad et al.. The impact of positional encoding on length generalization in transformers. NeurIPS 2023. [7] Ruoss et al., Randomized positional encodings boost length generalization of transformers, ACL 2023. [8] Lee et al., Teaching arithmetic to small transformers. ICLR 2024. [9] Zhou et al., What algorithms can transformers learn? a study in length generalization. ICLR 2024. [10] Zhou et al., Transformers can achieve length generalization but not robustly. arXiv preprint, 2024. --- Rebuttal Comment 1.1: Comment: Thank you for your efforts and your detailed rebuttal, especially for coming up with how to apply position coupling to the parity task and showing successful results on it (and thanks for acknowledging the limitation of the method w.r.t. tasks like sorting and mode). I'd add that the 2D experiment could benefit from a bit more explanation and how position coupling is applied to it. I am happy with the rebuttal and like the paper, thus I'll maintain my score, and would just encourage the authors to see if they can extend their image experiment to other tasks with more structure, but regardless I find the work solid. As a side note, the discussion on length generalization getting worse with deeper models seems to be somewhat connected to a concurrent work [1], it might be worth seeing if there's indeed any relation. [1] https://arxiv.org/pdf/2402.04875 --- Reply to Comment 1.1.1: Comment: Thank you for taking the time to provide valuable feedback. We appreciate your suggestions on clarifying the 2D experiment, and will certainly consider these points in our future work. We also appreciate the reference you provided and will look into the potential connection.	Rebuttal 1: Rebuttal: We deeply appreciate all reviewers for their insightful and detailed reviews, questions, and comments on our work. We assure the reviewers that all the answers and discussions will be incorporated into our final manuscript. We are encouraged to see that the reviewers recognized that our method is novel and significant (CVfs, ksWg, U8WH), our experiments are thorough and detailed (ow6J, CVfs, ksWg, U8WH), and our theoretical analysis is insightful and well-motivates our approach (ow6J, CVfs, ksWg, U8WH). Please check out our PDF file containing: * Fig. 1: ablations on the input formats and the model size. * Fig. 2: Parity task solved with position coupling + scratchpad. * Fig. 3: Testing on separate operand lengths. * Fig. 4: RoPE + Position Coupling. Now, we will provide our response to two commonly raised questions. ## 1. Ablations on Input Formats * Reviewers raised a question on the robustness of our method to input formats (CVfs) and a concern that input formatting is problem-aware processing (ksWg). * We would like to clarify that our input format is primarily selected to simplify the algorithm of solving the addition task, not through extensive ablation studies. Thus, we are not arguing that our choice of input format is empirically optimal for training Transformers. * However, we note that applying proper input formatting is crucial and natural in general ML. Even in a simple image classification task, appropriate standardization and augmentation are often useful. When applying these techniques, we often utilize the fact that typical image pixels range in [0, 255] (to apply standardization) and that appropriate augmentation methods may differ by the image type. Hence, enough understanding of the task leads to proper input processing that helps the model to effectively solve a given task. * Our additional experiments on position coupling with various input formats show that the model’s performance varies with different input formats (refer to the attachment). It is expected, as the complexity of the algorithm that the model should learn changes according to the input format. * Small models (1-layer 4-head) achieve near-perfect generalization when the numbers are zero-padded and the answer or all numbers are reversed. We believe this is because the combination of zero-padding and reversing enabled a small Transformer to learn a simple length-generalizing algorithm. If we flip the answer or all the numbers without zero-padding, small models exhibit a bit worse in-distribution performance and a restricted length generalization capability. Without reversing, the models perform poorly. * Larger models (6-layer 16-head) perform better than the small model when the numbers are no longer zero-padded (especially when the answer is reversed). We believe this is because the task-solving algorithm with reversing and without zero-padding that the model should learn is more sophisticated, which larger models can learn more easily. Contrarily, we observe a degradation in performance when we add zero-padding in the larger model, which suggests that the model may have learned a "shortcut" due to its (overly) strong expressive power relative to the problem's complexity. (Refer to the next question for more details on this matter.) ## 2. Deeper models seem to perform worse. Why? * Some reviewers expressed concerns that position coupling performs worse when applied to deeper networks. Here, we provide our thoughts on this phenomenon. * The first thing to note is that position coupling enables the model to solve the addition problem using a much simpler algorithm. As proven in Theorem 5.1, even a 1-layer model is sufficiently expressive. We hypothesize that the reason for the outstanding performance of a 1-layer model is that the architecture is simple so the model has no way to fit the training samples other than by learning the true (or length-generalizable) function. * In contrast, we believe that deeper models perform poorly because their larger expressive capacity allows them to learn shortcuts to fit the training distribution that may not generalize well across different lengths. This phenomenon differs from classical overfitting in that these models still generalize well to in-distribution samples. * A similar observation was made in [1]: there exist non-extrapolating models that generalize well to in-distribution samples but struggle with longer ones. The authors interpreted this phenomenon as indicating that there might be unexpected ways to solve the problem and the model may rely on shortcut-like solutions that work for in-distribution samples but fail on longer samples. * The superior performance of the position coupling scheme without zero-padding in deeper models also aligns with our interpretation. In this scenario, the query sequence becomes less consistent: the lengths of two operands may differ and some position IDs appear only once. This makes the functions the model needs to learn more complex. This complexity is evident from 1-layer experiments that the no-padding scheme fails for even in-distribution generalization. We believe that such a complex structure of the target function compels the model to learn the true function rather than a shortcut-like solution, resulting in strong length generalization performance. * Combining the answers for the first two questions, we do not believe that position coupling itself is bad for deep models. The performance degradation on deep models can also be attributed to the input formats, which may make the task-solving algorithm way simpler, and other training details such as small dataset size, which can make the model easily overfit to training distribution. Again, we deeply thank all reviewers for their time and effort in reviewing our work. We are excited to hear more feedback. Warm regards, Authors --- References [1] Zhang et al., Unveiling transformers with lego: a synthetic reasoning task. arXiv preprint, 2022. Pdf: /pdf/b6aad85e8cb095e39168e807ba2a651718da3d68.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
SIRIUS : Contexual Sparisty with Correction for Efficient LLMs	Accept (poster)	Summary: The paper introduces a sparse LLM correction mechanism (SIRIUS) designed to improve the inference efficiency of large language models through contextual sparsity. Contextual sparsity reduces computational cost by pruning parameters that are less relevant based on the input context. However, this approach degrades performance in tasks requiring high-level reasoning and deduction. SIRIUS addresses this by selectively correcting key tokens, using the full model to rewrite the sparse model's Key-Value cache. Experiments show improvements in strict-match EM scores, particularly on mathematical reasoning tasks. Strengths: 1. The correction mechanism that selectively rewrites the KV cache is novel. 2. The paper presents an interesting finding where sparse models fail on tasks that require reasoning and deduction. 3. The paper demonstrates improvements in performance metrics on benchmarks like GSM8K. Weaknesses: - The KV cache rewriting mechanism requires an available full LLM to be of the same architecture as the sparse model. - The criteria for when to trigger KV cache rewriting are not clearly detailed, making it difficult to assess the practicality and reliability of the approach. - It is unclear whether using the full model’s KV cache to rewrite the sparse model guarantees that the rewritten KV cache can be accurately interpreted and utilised by the sparse model, due to the differences in the sparse and full model. - The hypothesis that the gap between the full and sparse models can be bridged by correcting a few key tokens (L44-45) is not sufficiently validated beyond the provided examples. It remains unclear whether this hypothesis holds across a broader range of tasks and datasets. - The writing needs to be improved for clarity, for example, - there are a few missing citations (e.g., L114. L284) - the tables can be clarified with the best performing variant per task in boldface, and a caption which describes the metrics in the table. - Algorithm 1's functions can be better described, for example, the descriptions in "normal forward function FORWARD, forward function with direct cache rewrite KVFORWARD, likelihood judge function of LIKELIHOOD" are unclear. Technical Quality: 2 Clarity: 2 Questions for Authors: - How are the likelihood thresholds for detecting and correcting errors empirically tuned? Can the authors provide more details on this process and its impact on performance? - How does SIRIUS handle long-term dependencies and corrections over extended sequences of text? Are there scenarios where the KV cache rewriting might not be sufficient to correct extended errors? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: not observed negative societal impact to my understanding. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your time and attention on the paper. We are very grateful for you to be interested in the KV Cache correction. We will try our best to answer your questions. 1. Full and Sparse of the same architecture Contextual Sparsity is the focus of Sirius. Contextually Sparse (CS) models are dynamically subsampled from the normal LLM. The subsampling process won't change the compatibility of the original model's cache with the newer subsampled sparse model. Therefore, the original cache is naturally usable by CS models. However, we are interested to see whether Sirius can be applied to model pairs outside the CS realm. 2. When to rewrite KV Cache We appreciate your suggestions, and we rewrote the method part of the paper to take into consideration KV rewriting confusion. The rewriting happens when the full model is correcting the sparse generated tokens. The full model is called once every kernel size, usually 16. Throughout the inference, the KV cache is shared between the sparse and the full model. The KV cache is mostly populated by the sparse model, which is called for every token. During correction, the full model takes in the last kernel size of tokens and generates its KVs for the past 16 tokens in parallel, these KVs are directly written to their corresponding positions in the shared KV Cache. 3. Will the full model's KVs make semantic sense to the sparse model We appreciate the brilliant question. Empirically, we found that the KV cache of the large model seems to always help the small model generation quality. Below shows one of the example on GSM8K (20% subsampled) \| GSM8K 20% \| \| GSM8K 20% \| \| \|-------------------------------------------\|----------\|-----------------------------------------------------\|----------\| \| Llama3-8B-Instruct \| 0.7538/0.7538 \| Llama3-8B-Instruct \| 0.7538/0.7538 \| \| Llama3-8B-Instruct + Griffin \| 0.3674/0.3674 \| Llama3-8B-Instruct + CATS \| 0.5644/0.5644 \| \| Llama3-8B-Instruct + Griffin + KV Cache Correct \| 0.4735/0.4735 \| Llama3-8B-Instruct + CATS + KV Cache Correct \| 0.6629/0.6629 \| You can see that doing KV Cache correction alone brings substantial improvement to both settings. Empirically, the sparse model is clearly able to extract more insightful semantics from the full model's KVs. 4. Whether correcting minor portion leads to performance recovery is applicable to more datasets We appreciate the comments. Although we only show for GSM8K in the text. However, in the newer version of the paper. We evaluated Sirius across diverse datasets, ranging from arithmetic reasoning, common sense reasoning, and coding. Sirius managed to achieve competitive efficiency metrics on most of them. These results show that for general text generation that requires challenging step-by-step correlation, the minor portion is the only part that needs to be corrected to recover the original performance. 5. Writing is unclear We deeply appreciate the suggestions on improving the writing quality of the paper. All of the suggested parts are revised in the newer version of the paper. 6. How to determine thresholds. Threshold determination is a crucial process of balancing the performance and efficiency of Sirius. Here we provide a much more in-depth ablation on how increasing the threshold exhibits the trend that improves the performance but hurts the efficiency. The efficiency of Sirius is measured by the average advance length out of a kernel size (usually 16 tokens), the higher the better. \| Threshold \| Performance Score \| Efficiency Metrics \| \|-------------------------------\|-------------------\|--------------------\| \| Original Full Model \| 0.7803/0.7828 \| \| \| No correction (threshold 0) \| 0.5884/0.5960 \| \| \| 0.05 \| 0.7247/0.7247 \| 15.2098 \| \| 0.1 \| 0.7399/0.7424 \| 14.6451 \| \| 0.2 \| 0.7247/0.7273 \| 13.2329 \| \| 0.3 \| 0.7399/0.7449 \| 11.6134 \| \| 0.4 \| 0.7551/0.7601 \| 10.0037 \| \| 0.5 \| 0.7677/0.7702 \| 8.56022 \| \| 0.6 \| 0.7758/0.7778 \| 7.44126 \| \| 0.7 \| 0.7702/0.7753 \| 6.26547 \| \| 0.8 \| 0.7753/0.7803 \| 5.25639 \| \| 0.9 \| 0.7652/0.7677 \| 4.20542 \| \| 0.95 \| 0.7626/0.7652 \| 3.56315 \| \| 1.0 \| 0.7601/0.7626 \| 1.2685 \| In practice, threshold of 0.1 generally works well across various models and datasets. 7. Long range dependencies Sirius allows the large model to corrects the small model on a basis of kernel size. Besides KV rewriting, roll back mechanism is especially important for generating longer text. The rollback works as follows, when full model parallel verifies, If there is any token within the range of the kernel size that the full model regards as unlikely (rejects by likelihood threshold), then the token will be removed. Therefore, regardless of the generation length, all the verified tokens will be regarded by the full model as "enough likely" given the context. --- Rebuttal Comment 1.1: Comment: The rebuttal addresses some of my concerns. I have increase my score to 6. --- Reply to Comment 1.1.1: Title: Thank you for your comment Comment: We would like to express our deep appreciation for your time reading our work. Your insightful comments on KV cache rewrite mechanism really help us reflect on our previous description of Sirius, and it motivates us to refine it for more readers to understand more easily. Thank you for your comments.	Summary: The paper introduces SIRIUS, a novel method designed to enhance the efficiency and accuracy of sparse Large Language Models in reasoning and deduction tasks. SIRIUS employs contextual sparsity to reduce computational costs and incorporates an efficient correction mechanism that recovers sparse model performance by correcting only a minimal number of key tokens. Experimental results demonstrate that SIRIUS significantly improves the performance of sparse models on complex reasoning tasks while maintaining efficiency. Strengths: 1. The paper introduces SIRIUS, a novel and efficient correction mechanism specifically tailored for sparse models 2. It provides a echnically rigorous approach to improving the inference efficiency of LLMs, supported by experimental validation. Weaknesses: 1. There are too few experiments, so no experiments were conducted on other LLMs， and no experiments were conducted on more contextual sparsity methods 2. The Mc annotation in pseudocode is unclear Technical Quality: 2 Clarity: 2 Questions for Authors: If the correction requires the use of a full model, won't the speed of correction slow down compared to directly using the sparse model? Is the memory usage the same as using a full model directly? If the memory usage is the same as using a full model directly, there is no advantage in memory savings Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The author improved the performance of the sparse model in reasoning and understanding tasks by adding a full model correction, which requires the use of the entire full model. Although the average parameters used per token increase is minimal, the memory usage is the same as using the full model directly Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the time taken to read the paper and your insightful comments. We are thankful that you suggested we add more experiment data to evaluate Sirius. We will try our best to address your concern. If you feel that your concern has been taken into account, please consider raising your score. 1. Too few empirical experiments We provide additional empirical results and analysis of Sirius on various models and different downstream datasets. As discussed in the paper, the weakness of the contextual sparse models seems to occur upon difficult reasoning tasks. Besides, we collect experiment results on different reasoning tasks. In Arithmetic reasoning, besides GSM-8K, we also show a difficult AQuA-RAT COT from Google. In the Common Sense reasoning, we follow the COT paper to evaluate CSQA, StrategyQA, Sports, and Dates. Also, we found that contextual sparse models cannot do well in code generation. We then run Sirius on HumanEval. We collect additional experiment results on six models (Llama-3-8B, Llama-2-7B, Llama-2-13B with their instruction fine-tuning counterparts). Due to the word limit, we only show AQuA, CSQA, and HumanEval on some models. Full results and Average advance length will be in the new version of the paper. L means Llama. O means Original S means Sparse SS means Sparse + Sirius \| AQuA RAT \| L-3-8B-I-FSparse \| L-3-8B-I-CSparse \| L-2-7B-Chat-FSparse \| L-2-7B-Chat-CSparse \| \|-----------------------------------\|-----------------------------\|-----------------------------\|-------------------------\|-------------------------\| \| O \| 0.51 \| 0.51 \| 0.25 \| 0.25 \| \| S \| 0.42 \| 0.27 \| 0.28 \| 0.22 \| \| SS \| 0.42 \| 0.46 \| 0.24 \| 0.25 \| \| \|L-2-13B-Chat-FSparse \| L-2-13B-Chat-CSparse \| \| \| \| O \| 0.23 \| 0.23 \| \| \| \| S \| 0.25 \| 0.20 \| \| \| \| SS \| 0.27 \| 0.26 \| \| \| \| CSQA \| L-3-8B-I-FSparse \| L-3-8B-I-CSparse \| L-2-7B-Chat-FSparse \| L-2-7B-Chat-CSparse \| \|-------------------------------------\|-----------------------------\|-----------------------------\|-------------------------\|-------------------------\| \| O \| 0.70 \| 0.70 \| 0.62 \| 0.62 \| \| S \| 0.61 \| 0.64 \| 0.61 \| 0.52 \| \| SS \| 0.69 \| 0.72 \| 0.63 \| 0.60 \| \| \| L-2-13B-Chat-FSparse \| L-2-13B-Chat-CSparse \| \| \| \| O \| 0.68 \| 0.68 \| \| \| \| S \| 0.53 \| 0.55 \| \| \| \| SS \| 0.65 \| 0.67 \| \| \| \| HumanEval \| L-3-8B-I-FSparse \| L-3-8B-I-CSparse \| L-2-7B-Chat-FSparse \| L-2-7B-Chat-CSparse \| \|------------------------------------\|-----------------------------\|-----------------------------\|-------------------------\|-------------------------\| \| O \| 0.56 \| 0.56 \| 0.14 \| 0.14 \| \| S \| 0.45 \| 0.20 \| 0.13 \| 0.07 \| \| SS \| 0.58 \| 0.55 \| 0.13 \| 0.15 \| \| \| L-2-13B-Chat-FSparse \| L-2-13B-Chat-CSparse \| \| \| \| O \| 0.18 \| 0.18 \| \| \| \| S \| 0.14 \| 0.12 \| \| \| \| SS \| 0.17 \| 0.17 \| \| \| Sirius can also work effectively on the Llama-3-70B-Instruct, model with large parameters. \| L-3-70B-I \| GSM8KCOT \| \|---------------------------------------------------\|----------------\| \| L-3-70B-I \| 0.90 \| \| L-3-70B-I CSparse \| 0.74 \| \| L-3-70B-I CSparse \| 0.87 (15.41) The bracket is the average advance length of 16 kernel size. 2. Other questions. a. We appreciate the suggestions on Pseudo code, we have rewritten it in the newer version. b. Correction step by full model is slower than sparse, but the cost is amortized by the kernel size of 16, which is less significant. c. Contextual sparsity subsample sparse model from the full model dynamically according to input, so the full model weight will always be in GPU memory. Adding correction won't increase the memory usage of the sparse model. d. Please note that the latency is mainly affected by the memory loading not the shear memory size on VRAM. Full model is loaded only once per kernel size. Please refer to the hardware speedup on the new version. --- Rebuttal 2: Title: Kindly Asking to Reconsider Your Score Comment: Thank you again for your feedback. We have added the necessary material to the author's rebuttal to address the points you raised. Given that we only have four reviewers, each reviewer's score significantly impacts the overall assessment of our work. We believe that the current overall assessment does not adequately reflect the contribution of our work. Therefore, we kindly request you reconsider your score. Thank you again for your time and effort in reviewing our paper. --- Rebuttal Comment 2.1: Title: Response by Reviewer ArEw Comment: Thank you for your detailed explanation and rich experimental supplements. It alleviated my concerns to a certain extent. I raise my score to 5. --- Reply to Comment 2.1.1: Title: Thank you for your comments Comment: We would like to express our deep appreciation for your time reading our work. The comment on lacks of empirical results help us make our study on Contextual Sparsity correction more concrete.	Summary: This paper aims to improve contextual sparsity (CS) approaches: LLM sparsification / parameter pruning methods where the sparsification strategy is conditioned on the input sequence / prompt itself. The paper begins by reproducing contextual sparsity baselines, but applied to modern LLMs: LLama 2 and 3 (7B & 8B respectively), and verifying that their reproduction successfully replicates previous CS techniques on summarization and question answering tasks. The authors then demonstrate that the technique is much less effective on more challenging reasoning tasks such as GSM8K and MMLU-CoT. Through investigation the authors observe that often, by correcting a small amount of tokens during the generation (~10%), the sparse models can significantly improve their performance to be on-par with the full dense model. This motivates their main proposed technique: SIRIUS, which employs a speculative decoding like technique to use the full dense model to score every k tokens (k=16-24 in practice) during the decoding process in order to rollback and use the dense model to correct a token when necessary. By doing so, the method is able to mitigate the losses on reasoning tasks while maintaining most of the efficiency of the sparsification method. Strengths: 1. The authors explore the viability of CS techniques on much more difficult reasoning tasks than previously in the literature, providing an important datapoint on whether CS approaches (at least ones studied in this paper) work in general. 1. The authors run thorough quality experiments on modern LLMs, and demonstrate that their technique effectively rectifies the issues they first identify (significantly improving reasoning performance.) 1. Using the full model to do something like self-speculative decoding with the drafting model being a sparse version of the same model is novel and promising. It aligns with contemporaneous works that explore this idea [1]. 1. The method is efficient w.r.t the number of parameters used on average during decoding. 1. The paper is overall well structured and the investigation is easy to follow. 1. The authors plan to release the code for their experiments. [1] LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding Weaknesses: 1. The main spirit of this paper is whether we can use some small amount of additional compute budget over the sparse model alone (e.g. via limited calls to the full model) to improve the decoding from the sparse model. However, the authors do not compare against any baseline approaches to do this same thing, such as speculative decoding. Even though the authors claim in Related Work that speculative decoding does not suit this problem since it’s not efficient enough, the authors do not provide any empirical evidence of such. 1. This paper is very much an efficiency paper, which tries to introduce some small amount of corrections from the full model to the dense model, while not doing too much as to nullify the gains from the sparse method. However, the authors only analyze Average Parameters Used Per Token (APU), a sparsity measure, instead of any other efficiency metric (importantly wall-time). - The authors make an argument for why this is in Section 4.1, using findings from previous papers. While this may be a reasonable intuitive argument, the method in this paper has significant differences like KV-cache correction that still makes this valuable to empirically verify. - Moreover, even if I intuitively believe from the author’s argument that the wall time latency of this approach would be somewhere between the sparse and full model’s wall time, it is unclear where on this spectrum it lies, how it compares to other techniques like speculative decoding, or how the tradeoff plays with varying sparsity levels without such measurement and comparison. - To place this into context, the CATS paper, which is one of the methods this work experiments with (FSparse) directly states “Activation sparsity of a model is not sufficient to directly enable wall-clock time inference Speedups”, and provides wall-clock time analysis. 1. The paper is poorly written, with various spelling errors (in the title), grammatical errors (for example: line 1, Section 4.2 title, and more), placeholders for citations (line 114), and difficult to read citation formatting. In many ways the presentation of this work is significantly not publication ready. 1. Without the code release, the description of the method is not very detailed, even though Algorithm 1 is provided, and would make it very hard for the reader to understand how this method is implemented from reading the paper alone, which hurts reproducibility. 1. For example: the description of the correction algorithm is only described in one paragraph in Section 4.3 with insufficient detail on how the KV cache is updated. 1. Algorithm 1 has undefined variables with respect to the KV cache such as “cachetracker”. The description in 4.3 states that the KV cache is directly written to by the full model to correct the past, however Algorithm 1 does not describe this process other than “Update C_s based on j” is unclear what kind of update is happening to the cache. 1. The important LIKELIHOOD function is undefined, which uses the full model’s likelihoods to judge whether the sparse model’s output is incorrect. There are no details on what threshold is used here. 1. Some additional claims are not sound: - Line 270: “KV Cache can also be looked as an implicit knowledge distillation method to pass dark information from the LLM to the sparse counterparts” What is dark information? Technical Quality: 1 Clarity: 1 Questions for Authors: What were the motivations behind the differences between Sirius (Algorithm 1 / Section 4.3) and speculative decoding? Why not directly use speculative decoding, with the sparse model being the drafting model? (which maybe has more guarantees) Why is the method named Sirius? Confidence: 4 Soundness: 1 Presentation: 1 Contribution: 2 Limitations: The authors briefly describe some limitations but the work is generally lacking discussion here. Mainly the authors mention that the method does not work well under extreme sparsity (to my knowledge almost no sparsity method does in the general case – this is more of an open question than a limitation.) The authors could be more forthcoming here. It would be valuable to elaborate on whether the authors believe the two CS approaches are representative of the space of approaches, whether the experiments are adequate, whether they expect the results to change at different model scales, or for tasks beyond the ones studied here. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the time taken to read the paper and your insightful comments. We are thankful that you point out that the comparison against speculative decoding is missing and we lack of convincing real efficiency metrics. We will try our best to address your concern. If you feel that your concern has been taken into account, please consider raising your score. 1. Differences to Speculative Decoding? Why Speculative Decoding is not efficient enough? Sirius contains a stronger and a weaker model, the strong verifies the weaker's output, from which the resemblance can be easily drawn between Sirius and Speculative Decoding (SD). However, the important distinction is the problem setting. Contextual sparsity users prioritize the tradeoff between the performance and the efficiency of the model. We show that the efficiency of the contextually sparse model corrected by SD will be largely limited. Take the Coarse-grained sparse model on Llama-3-8B as an example, the 50% sparse (APU 0.65) model will produce an acceptance rate of 0.89 on GSM8K. SD also runs the smaller model every iteration with verification of the larger model once a period. Naturally, to increase efficiency, we need to (1) enlarge the period and (2) increase the average number of tokens accepted in the period. \frac{1 - \alpha^{(\gamma + 1)}}{1 - \alpha} Now, given the acceptance rate of 0.89, following the formulation in Speculative decoding, we can calculate the expected number of accepted tokens for every gamma term in the Speculative Decoding literature, which is (period - 1). \| Gamma \| Advance Length \| \|--------:\|-----------------:\| \| 4 \| 4.01 \| \| 8 \| 5.91 \| \| 12 \| 7.09 \| \| 16 \| 7.84 \| \| 20 \| 8.30 \| \| 24 \| 8.60 \| \| 28 \| 8.78 \| \| 32 \| 8.90 \| We can notice the trend that the average advance length starts to plateau as the period becomes larger. Take the gamma of 16 as an example, the period is then 17. If we use the formula in Section 4.1 of our paper, we can immediately see that the APU is (16 * 0.65 + 1)/7.84 = 1.45, even larger than full model 1. \| Gamma \| Advance Length \| \|--------:\|-----------------:\| \| 1 \| 0.87 \| \| 2 \| 0.86 \| \| 4 \| 0.90 \| \| 8 \| 1.05 \| \| 12 \| 1.24 \| \| 16 \| 1.45 \| \| 20 \| 1.69 \| \| 24 \| 1.93 \| Because of the plateauing effect, for an acceptance rate of 0.89, the best gamma is 2 (period = 3). The best efficiency is 0.86, compared with 0.65 coarse-grained sparse APU. A similar picture can be applied to Fine-grained sparsity as well. The key reasons are (1) the contextual sparse model is too large to use a long period for SD, and (2) the acceptance criteria are too strict. In contrast, Sirius gives a more flexible choice for contextually sparse model users. For a threshold of 0.1, Sirius can correct Llama-3-8B coarse-grained sparsity from 20.85% to 43.9%, compared to the 49.66% full model. Sirius on average can accept 13.4 tokens out of a kernel size of 16 and over 9 tokens out of a kernel size of 10, translating to APU < 0.76, significantly lower than SD does. 2. Hardware speedup? We show that Sirius delivers the promised speedup. With the introduction of static cache from huggingface 4.38, kv cache is pre-allocated, making rewrite, expand, and rollback of very minimal overhead. Fine-grained sparsity speedup is via prior work [1]. However, CATS relies on a close-sourced custom CUDA kernel. We present the speedup of Coarse-grained sparsity with Sirius in two settings. First, on-chip, we run Llama-3-8B-Instruct inference. \| Setting\| GSM-8K-COT \| A40 \| Speedup \| L40 \| Speedup \| A100 \| Speedup \| \|-----\|-----\|-----\|-----\|-----\|-----\|-----\|-----\| \| CSparsity \| 0.3601 \| 20.7 \| 0.66 \| 15.6 \| 0.67 \| 9.6 \| 0.72 \| \| Sirius (10 kernel size) \|0.7309\| 24.1 \| 0.78 \| 18.2 \| 0.78 \| 11.4 \| 0.85 \| \| Sirius (16 kernel size) \| 0.7309 \| 24.9 \| 0.80 \| 18.8 \| 0.81 \| 11.8 \| 0.88 \| \| Full \| 0.7612 \| 30.9 \| \| 23.2 \| \| 13.3 \| \| Second, Offloading Llama-3-70B-Instruct. We use a single L40 48GB with a PCIe bus bandwidth of 25 GB/s. \| \| Sparse \| Sparse + Sirius \| Full \| \|---------------\|----------------------\|-------------------------------\|--------------------\| \| Performance \| 0.9014 \| 0.7407 \| 0.7407/0.7483 \| \| \| Latency (s) \| 3.57 \| 3.68 \| 5.72 \| 25.3 GB/s \| \| Ratio to Full \| 0.6241 \| 0.6434 \| \| \| [1] Liu, Z., Wang, J., Dao, T., Zhou, T., Yuan, B., Song, Z., ... & Chen, B. (2023, July). Deja vu: Contextual sparsity for efficient llms at inference time. In International Conference on Machine Learning (pp. 22137-22176). PMLR. 3. Criticism of writing quality We do appreciate the suggestions on writing. In light of your precious effort put into reading, we carefully rewrote and proofread the entire methodology section that takes all your suggestions into account. 4. KV Cache correction \| Setting\|GSM8K 20% \|Setting\| GSM8K 20% \| \|-------------------------------------------\|----------\|-----------------------------------------------------\|----------\| \| Llama3-8B-Instruct \| 0.7538 \| Llama3-8B-Instruct \| 0.7538 \| \| Llama3-8B-Instruct + CSparse \| 0.3674 \| Llama3-8B-Instruct + FSparse \| 0.5644 \| \| Llama3-8B-Instruct + CSparse + KV Cache Correct \| 0.4735 \| Llama3-8B-Instruct + FSparse + KV Cache Correct \| 0.6629 \| We show that KV Cache correction contributes to correction. However, we agree that "implicit distillation" is unbased, which has been deleted in the latest version. 5. Why Sirius? Sirius is an astronomical term referring to a two-body star system. One is the brightest ever known, while the other is dim. We draw inspiration from that system. --- Rebuttal Comment 1.1: Title: Updated rating Comment: Thank you for providing your extensive clarifications and updated results. In particular, the latency evaluations on real hardware, additional ablations, and revised presentation changes my impression of this paper a lot. I have increased my score to 6. Nit: there is still a typo in the title "Sparisty", and other places like Figure 3 title "Efficientcy". Please continue to refine for the future versions. --- Reply to Comment 1.1.1: Title: Thank you for your careful reading and raising the score Comment: We would like to express our deep appreciation for your careful reading and insightful comments. Your comments hit home on the important insight of Sirius. We really appreciate your time and effort spent on our work.	Summary: This paper focuses on enhancing the inference efficiency of large language models (LLMs) through contextual sparsity. While it identifies that contextual sparsity reduces hallucination, it also notes a significant impact on reasoning and deduction performance. To address these drawbacks, the paper introduces SIRIUS, a correction mechanism that effectively improves performance. Specifically, the correction mechanism uses additional parameters (i.e., full model) to check the output of the sparse model based on the input and the same instruction prompt. Strengths: 1. The paper introduces SIRIUS, a correction mechanism that effectively enhances the performance of sparse models, particularly in complex reasoning tasks where traditional sparsity techniques falter. 2. The paper includes many experiments across various datasets and tasks to reveal the problem in the existing contextual sparsity methods and sparse models. Weaknesses: Method 1. Based on the results in Tables 1 and 2, this paper claims that “the stronger the model, the larger the quality degradation would be”. However, the studies exclusively utilize relatively smaller models, Llama2-7B and Llama3-8B. Including results from larger LLMs, such as Llama3-70B, would provide a more comprehensive understanding of the impact across different model sizes. 2. As the proposed method uses the probability from the full model to check the correctness of the sparse model, does this require forwarding twice for each token? If so, the cost would be double. 3. It is unclear when to use the correction mechanism. Should it be used after the sparse model has generated all predicted tokens, or immediately after the sparse model predicts each token? Experiment 1. In Table 3, if the proposed method uses the full model to correct the output of the sparse model, why is there still a noticeable degradation compared to the full model itself? 2. As the title highlights the efficiency of the proposed method, it would be better to provide some experiments about it like latency (ms) in DejaVu [1]. [1] Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time. ICML, 2023. **Minor Issues There are many typos in this paper, e.g., the title “… Contexual Sparisty …” should be “… Contextual Sparsity …”. Technical Quality: 2 Clarity: 3 Questions for Authors: Please refer to Weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: There are no discussions about the proposed method's limitations and potential societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your attention and time taken to read the paper and your insightful comments. We are very thankful for you to point out the inadequacy of our method presentation and raising concerns on the diversity of the evaluation. We will try our best to address your concern. If you feel that your concern has been taken into account, please consider raising your score. 1. Studies mainly use smaller models, which should include Llama-3-70B to verify the claim that "the stronger the model, the larger the quality degradation would be". In light of your comment, we run the Llama-3-70B-Instruct model on GSM8K, MMLU-FLAN-COT, CoQA, CNN/DailyMail, and TruthfulQA. Due to running Llama-3-70B-Instruct under pipeline parallelism is slow on our hardware (8xA100), MMLU-FLAN-COT is subsampled 10%, and CNN/DailyMail is subsampled 30%. \| \|GSM-8K-COT\|MMLU-FLAN-COT\|CoQA\|CNN/DailyMail\|TruthfulQA\| \|---------\|-----------\|---------------------------\|--------\|-------------------\|---------------\| \|Llama-3-70B-In\|0.9014/0.9022\|0.7456\|0.6567/0.8069\|0.101634/0.020614/0.096413\|0.5116/0.4247\| \| + CSparse\|0.7407/0.7483\|0.7018\|0.6497/0.8046\|0.101922/0.020854/0.096703\|0.4541/0.3807\| \| + FSparse\|0.8726/0.8772\|0.7193\|0.6497/0.8035\|0.101505/0.020623/0.096344\|0.4835/0.3905\| \|Llama-3-8B-In\|0.7612/0.7672\|0.6272\|0.6153/0.7825\|0.101523/0.020481/0.096311\|0.4945/0.3647\| \| + CSparse\|0.3601/0.3647\|0.5307\|0.6003/0.7735\|0.101681/0.020657/0.096432\|0.5067/0.3953\| \| + FSparse\|0.6103/0.6202\|0.4825\|0.5828/0.7577\|0.101713/0.020448/0.096516\|0.5202/0.3941\| GSM-8K-COT, “strict match”/”flexible extract”; CoQA, EM/F1; CNN/DailyMail, Rouge-1/2/L, TruthfulQA gen, Rouge-1/2 ACC. In the table, Llama-3-70B-Instruct still has features that we identified in the paper, sparse excels at prompt understanding (CoQA and CNN/DailyMail). Significant performance degradation occurring at GSM-8K-COT with coarse-grained sparsity. Even though the gap between sparse and dense for 70B is not as big as for smaller models, we don't think it contradicts the claim that the more powerful the model, the more degradation when doing sparse. First, even with Coarse-grained sparsity, the number of parameters is more than 45B, still colossal. Most tasks here are curated in pre-LLM era and are too easy for models now of huge parameters [1]. On the other hand, the claim is presented in the context of comparing Llama-2-7B Chat and Llama-3-8B-Instruct, two smaller but similar models. Llama-2-7B Chat has significantly worse performance in all different tasks. Contextual Sparsity causes much more damage to 8B than 7B. Nevertheless, trying larger parameter models does provide more insights to contextual sparsity characteristics. [1] Chiang, W. L., Zheng, L., Sheng, Y., Angelopoulos, A. N., Li, T., Li, D., ... & Stoica, I. (2024). Chatbot arena: An open platform for evaluating llms by human preference. arXiv preprint arXiv:2403.04132. 2. Forwarding twice, cost doubled? The tokens produced by the sparse model will be fed into the full model for verification, but the latency cost of full model verification will NOT be a lot, compared to the full model generating one new token. The cost in terms of latency won’t be doubled. The full model will verify the tokens generated by the sparse model periodically, usually in a period of 16. During full model verification, all tokens generated by the sparse model that haven’t been verified will be fed in one iteration. And, the full model will generate one more token in this run. Consider the following table of A100 latency of different sequence lengths with bsz=1. Only 1.1ms increase for 64 seqlen. The verification is as light as generating one token, without additional overhead. \|Input Sequence Length\|1\|2\|4\|8\|16\|32\|64\|96\| \|-----\|-----\|-----\|-----\|-----\|-----\|-----\|-----\|-----\| \|Latency (ms)\|13.3\|13.5\|13.6\|13.8\|14.0\|14.9\|14.4\|17.1\| 3. When to correct? Correction happens periodically. Specifically, the sparse model would generate PERIOD -1 tokens, and the full model would take in these tokens, verify them based on likelihood thresholds, rollback, and then generate one more token. The cycle goes on. Sirius usually takes PERIOD of 16, sometimes 12. 4. Degradation after correction? We can answer by comparing Sirius with Speculative Decoding (SD). SD is lossless. Under the greedy setting, SD acceptance criteria guarantee that the two models' output is the same as the large model output. In comparison, here we effectively loosen up the acceptance criteria. We manually set a threshold on the full model's verification likelihood. i.e., If the full model thinks the token is likely to occur, it accepts. However, the token is not necessarily the one full will greedily decode, leading to divergence in generation and, thus degradation. However, the loosened criteria in turn boost the efficiency of the overall system by a margin to SD's. 5. Hardware wall clock time speedup? Sirius delivers the promised speedup. We present the speedup of Coarse-grained sparsity with Sirius in two settings. First, on-chip, we run Llama-3-8B-Instruct inference. Fine-grained sparsity prior works rely on a custom CUDA kernel, which we don't have access to. \| Setting\| GSM-8K-COT \| A40 \| Speedup \| L40 \| Speedup \| A100 \| Speedup \| \|-----\|-----\|-----\|-----\|-----\|-----\|-----\|-----\| \| CSparsity \| 0.3601 \| 20.7ms \| 0.66 \| 15.6ms \| 0.67 \| 9.6ms \| 0.72 \| \| Sirius (10 kernel size) \|0.7309\| 24.1ms \| 0.78 \| 18.2ms \| 0.78 \| 11.4ms \| 0.85 \| \| Sirius (16 kernel size) \| 0.7309 \| 24.9ms \| 0.80 \| 18.8ms \| 0.81 \| 11.8ms \| 0.88 \| \| Full \| 0.7612 \| 30.9ms \| \| 23.2ms \| \| 13.3ms \| \| Second, Offloading Llama-3-70B-Instruct. We use a single L40 48GB with a PCIe bus bandwidth to be 25 GB/s. \| \| Sparse \| Sparse + Sirius \| Full \| \|---------------\|----------------------\|-------------------------------\|--------------------\| \| Performance \| 0.9014 \| 0.8719 \| 0.7407 \| \| \| Latency (s) \| 3.57s \| 3.68s \| 5.72s \| 25.3 GB/s \| \| Ratio to Full \| 0.6241 \| 0.6434 \| \| \| --- Rebuttal 2: Title: Kindly Asking to Reconsider Your Score Comment: Thank you again for your feedback. We have added the necessary material to the author's rebuttal to address the points you raised. Given that we only have four reviewers, each reviewer's score significantly impacts the overall assessment of our work. We believe that the current overall assessment does not adequately reflect the contribution of our work. Therefore, we kindly request you reconsider your score. Thank you again for your time and effort in reviewing our paper. --- Rebuttal Comment 2.1: Comment: Thanks for your detailed response. It has addressed some of my concerns. However, I am still concerned that 1) the claim “the stronger the model, the larger the quality degradation would be” since it does not match the experimental results on more SoTA models like Llama-3-70B-In vs. Llama-3-8B-In in the response. 2) The speed-up is slight (e.g., A100: 11.8 ms -> 13.3 ms) compared to the degradation of performance (0.7612 -> 0.7309). Thus, I will keep my score unchanged. --- Reply to Comment 2.1.1: Title: Response to reviewer Comment: We thank the reviewer again for additional feedback and suggestions. We agree that our previous claim “the stronger the model, the larger the quality degradation would be” is not well-stated without enough context and quantifications. Also, the experiments we presented in the first round of rebuttal is not clear. We here clarify the claim as: given the same number of parameters, the more well-trained (powerful) the model is, the more performance degradation when applying contextual sparsity method would be. In the paper, we present the experiments of Llama-3-8B-Instruct and Llama-2-7B-Chat. Now, we further present the comparison between running Llama-3-70B-Instruct and Llama-2-70B-Chat on GSM8K COT dataset. The experiments are now running and will be soon added. Again, we are deeply grateful for the reviewer’s insightful suggestions on the relationship between performance degradation from contextual sparsity and full model performance. On the speedup portion, we would like to point the reader to the speedup ratio between the plain sparse and dense models measured on different devices from the CSparse methods on Llama-3-8B-In (In the table, we show A40, L40, and A100). Theoretically, the CSparse method only prunes the model by 45% of parameters on Llama-3-8B-In. Therefore, the method will at most achieve 0.65 of the latency as the original dense model, given the LLM inference is bounded by memory. We found that our implementation is very close to optimal for A40 and L40, as the sparse-to-dense ratio on these two devices is close to 0.65 theoretical value, and on top of that, Sirius incurs an additional 10% of the dense model latency as the overhead of correction. Compared to the sparse model’s original accuracy on GSM-8K-COT 36%, Sirius now corrects it to 73%, which is more than doubled. We argue that Sirius achieves a reasonable efficiency-accuracy tradeoff on A40 and L40. Moreover, compared to A100, L40 and A40 are much cheaper commodity hardware, which is closer to the users of sparse models that are more resource-limited. Admittedly, our current implementation is slightly not as optimal on more high-end GPUs like A100, since the raw sparse-to-dense ratio is close to 0.72 already, which is 0.06 higher than the theoretical value. We found that the slightly suboptimal ratio is caused by extremely high-end GPUs like A100 that are much more demanding on the attention kernels we use. However, building an optimal attention kernel for a relatively short context is beyond the scope of our project. Once the next-generation more optimized attention kernel is rolled out, we will have a similar speedup ratio on A100 as on other slightly slower GPUs. Again, we argue that A100 is much more expensive to rent, making it further away from the target audiences of sparse models. Furthermore, we would like to point readers to the accuracy and efficiency tradeoff we have on the Llama-3-70B-In in the offloading setting (part of the weights are loaded on GPU's on-chip memory, the rest is offloaded on CPU RAM), which is one of the only ways a normal practitioner would run inference on the 70B model without high-end GPU clusters. We found that Sirius corrects the CSparse 70B model from 0.76 to 0.87 in accuracy, while only increasing the latency by roughly 2% of the original model latency. (Please note that there is a typo in the offloading 70B table, on the Performance row, the sparse and dense data should be swapped). Again, we argue that the Sirius method achieves a reasonable tradeoff between Efficiency and Accuracy for the 70B model as well. Thank you again for your additional thoughtful and acute comments. If you think that your concerns have been addressed, please consider adjusting your score. --- Rebuttal 3: Title: Follow-up on Previous Response Comment: We again appreciate the review’s suggestions and inquiries on the efficiency of the claim in our paper. We follow up on our previous message with more experiment results on the comparisons on Llama-2-70B-Chat to provide more empirical studies on the claim. In the table below, we compare two pairs of models on GSM-8K-COT with similar parameters and the degradation after applying the contextual sparse methods (due to time limit, we subsample 20% from the entire dataset for 70B model). Also, we vary the sparsity level to have 50%, 40%, 30% of non-zero values in the weights, where lower than 30% would lead to Llama-3-8B-Instruct to have 0 accuracy). \| Llama-3-70B-Instruct \| Accuracy \| Degradation \| Llama-3-8B-Instruct \| Accuracy \| Degradation \| \|----------------------\|----------\|-------------\|---------------------\|----------\|-------------\| \| Full \| 0.9205 \| \| Full \| 0.7462 \| \| \| Csparse 50% \| 0.7652 \| 0.1553 \| Csparse 50% \| 0.3636 \| 0.3826 \| \| Csparse 40% \| 0.6023 \| 0.3182 \| Csparse 40% \| 0.1856 \| 0.5606 \| \| Csparse 30% \| 0.3144 \| 0.6061 \| Csparse 30% \| 0.0644 \| 0.6818 \| \| Fsparse 50% \| 0.8864 \| 0.0341 \| Fsparse 50% \| 0.6477 \| 0.0985 \| \| Fsparse 40% \| 0.8485 \| 0.072 \| Fsparse 40% \| 0.4053 \| 0.3409 \| \| Fsparse 30% \| 0.7386 \| 0.1819 \| Fsparse 30% \| 0.0265 \| 0.7197 \| \| Llama-2-70B-Chat \| Accuracy \| Degradation \| Llama-2-7B-Chat \| Accuracy \| Degradation \| \| Full \| 0.4508 \| \| Full \| 0.1856 \| \| \| Csparse 50% \| 0.3939 \| 0.0569 \| Csparse 50% \| 0.1515 \| 0.0341 \| \| Csparse 40% \| 0.3447 \| 0.1061 \| Csparse 40% \| 0.1098 \| 0.0758 \| \| Csparse 30% \| 0.2689 \| 0.1819 \| Csparse 30% \| 0.072 \| 0.1136 \| \| Fsparse 50% \| 0.3864 \| 0.0644 \| Fsparse 50% \| 0.1629 \| 0.0227 \| \| Fsparse 40% \| 0.3902 \| 0.0606 \| Fsparse 40% \| 0.1364 \| 0.0492 \| \| Fsparse 30% \| 0.2689 \| 0.1819 \| Fsparse 30% \| 0.1212 \| 0.0644 \| From the table, we can clearly see that given similar parameter size, contextual sparsity brings more degradation to Llama-3 family models than to Llama-2 models (Please read the table vertically to compare between Llama-3-70B-Instruct with Llama-2-70B-Chat and Llama-3-8B-Instruct with Llama-2-7B-Chat). Surprisingly, we also notice another interesting phenomenon where the models with larger parameter sizes seem to be more resilient to the contextual sparsity methods. Please read the table horizontally to compare models within the Llama-3 family models and Llama-2 family models. The trend is as expected since the model with a larger parameter size often has more redundancy in parameters. Furthermore, the contextual sparsity’s weakness is in full display in the above table. We can see that for Llama-3-70B-Instruct, even at 50% sparsity, the performance on GSM-8K-COT is comparable to Llama-3-8B-Instruct. Given that the sparse 70B model still has over 40B parameter size, the performance degradation is unacceptable, let alone a sparse 70B model with lower sparsity levels (40% and 30%). Sirius corrects the sparse model to have 87% accuracy and incur negligible overhead during the offloading setting, again showing its effectiveness in helping contextual sparsity methods. Also, as a follow-up on the efficiency-accuracy tradeoff concern, we would emphasize another set of results that was previously overlooked. Please be aware that the accuracy after Sirius correction is dependent on the sparsity level and the sparsity method. Previously, we looked at the coarse-grained sparsity methods where the sparsity pattern is determined for every input prompt and fixed throughout the generation. For the more flexible fine-grained sparsity, where the sparsity pattern changes for different decoded tokens, the sparse method alone achieved 59.68% accuracy on GSM8K-COT, versus the full model’s 75%, Sirius can correct the fine-grained sparse models to 74% accuracy with 0.775 theoretical Average Parameter Used to full model. Thank you again for your additional thoughtful and acute comments. If you think that your concerns have been addressed, please consider adjusting your score.	Rebuttal 1: Rebuttal: We thanked all the reviewers [R1(a8uY), R2(rbBk), R3(ArEw), R4(BB9w)] for their attention and time put into reviewing the paper and also for the thoughtful and supportive comments. We are glad to see that the reviewers find the work interesting [R4(BB9w)] and effective [R1(a8uY)], consider the problem we are solving to be relevant [R2(rbBk)], and think that our proposed technique to be effective [R1(a8uY)]. Also, we are pleased to read that some find our overall presentation to be easy to follow [R2(rbBk)]. On the flip side, we want to make an additional effort to assure the reviewers that we take their suggestions and criticism seriously. Before diving into specific detailed questions, we rewrote the paper submitted considerably based on the precious feedback. The differences can be mainly summarized in the following bullet points. The revised new version of the paper can be accessed with the following link: https://drive.google.com/file/d/1mvrltX1vd4dlTOKaSyBK1BuXgNvZNYHd/view?usp=sharing. - Elaborate on the Difference in Setting between Sirius and Speculative Decoding [R2(rbBk)] Sirius is a method that seemingly involves two models, a powerful full model and a weaker sparse model. One can easily draw a resemblance with Speculative Decoding, a technique to speed up the large model decoding. The key difference in setting between Sirius and Speculative Decoding is that Speculative Decoding (SD) is LLM-centric, meaning that the key objective is to losslessly speed up the LLM inference. The restriction for whether to accept the weak draft model is high in order to preserve the LLM performance. However, for the users of Contextual Sparse (CS) models, they accept the potential performance degradation involved in the CS models and crave more on the efficiency-performance tradeoff. However, the CS model struggles at difficult tasks that require step-by-step reasoning. Applying Speculative Decoding directly to full and CS models would correct the CS model’s mistake but incur a large price. Sirius is a “small model-centric” technique aiming to still preserve the efficiency of the sparse models (substantially better than SD), but improve sparse model performance in the LLM’s vicinity. - Refine the description of the Sirius [R1(a8uY), R2(rbBk), R3(ArEw), R4(BB9w)] We rewrote the algorithm and the method section in order to present the method in greater detail and address specific concerns about the method relating to the following major questions: When does the switch between full and CS models occur? How is the KV Cache correction implemented? How is the memory of the full with CS models managed, and why we don’t need to load more memory compared to full model inference? - Present the Wallclock Time Speedup for Llama-3-8B-Instruct and Llama-3-70B-Instruct Models [R2(rbBk)] Sirius is about efficiency. We verify the paper’s claim on Average Parameter Used (APU) efficiency with two different high-quality models with different sizes (Llama-3-8B-Instruct and Llama-3-70B-Instruct) on two different settings: on-chip and offloading. Llama-3-8B-Instruct with its CS model can be placed on a single GPU with VRAM greater than 24GB. We implemented a system using Torch Compile and CUDA GRAPH to show that Sirius delivers its promise on three common high-end GPUs, Nvidia A40, L40, and A100. The llama-3-70B-Instruct model has a whopping 140GB of memory requirement in bfloat16, which cannot fit on any single Nvidia high-end graphic card. For normal users with limited resources, they still can run the model with CPU offload. Sirius is also evaluated and delivers the promised speedup. - Add Evaluations of Sirius on Diverse Dataset and Finer Ablations on 7 models ranging from 7B to 70B in the Llama family [R3(ArEw)] We evaluate Sirius in more diverse settings following CS models that identified weaknesses in arithmetic reasoning. Besides GSM8K, we add other arithmetic reasoning datasets AQuA-RAT COT. Also, for commonsense reasoning, we evaluate Sirius' effectiveness on CSQA, StrategyQA, Sports, and Dates. Besides, CS models also struggle at coding, we also evaluate CS models boosted by Sirius on coding datasets HumanEval and MBPP, and show that Sirius can be effective in coding settings.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
PANORAMIA: Privacy Auditing of Machine Learning Models without Retraining	Accept (poster)	Summary: The paper proposes a method to estimate a lower bound on the (pure) Differential Privacy parameter $\epsilon$ for already trained machine learning models using a post-hoc empirical evaluation based on conducting Membership Inference Attacks. Building on the analysis by Steinke et al., 2023, for auditing DP with $O(1)$ training runs, this paper tweaks the setting by using synthetically generated canary data points that closely resemble the training data population, rather than inserting or omitting pathologically crafted canaries that degrade the ML model's performance. This approach is useful because if the canary distribution and the training distribution are (nearly) identical, an already trained model can be considered as the output of the training algorithm on a random partition of the combined training and canary datasets. This means any post-hoc auditing analysis based on membership inference on a randomly and independently selected training dataset yields a valid lower bound on $\epsilon$. To get such a canary distribution, the authors suggest training a synthetic data generation model to create a canary dataset to get a distribution similar to the population. Additionally, the authors extend Steinke et al.'s analysis to situations where the synthetic data distribution is $c$-close to the true population in a DP-like divergence, although the estimator does not technically yield a lower bound in this case. Through extensive empirical evaluations, the paper demonstrates that the proposed estimation technique is useful as it provides reasonable values that approximate the DP lower bounds. Strengths: The paper focuses on an important problem and has the following merits. - The paper proposes an auditing method that seeks to eliminate the need of altering anything about the training process of ML models, thus allowing post-hoc auditing. - The paper proposes using synthetic data that resemble the training data instead of adversarially created canaries can reduce the drop in model performance while offering useful DP estimations. - Empirical evaluations suggest that the estimator correlates with the DP upper bounds and can help identify situations of high privacy leakage. - The appendix provides an attempt to extend the budget estimator for $(\epsilon, \delta)$-DP. Weaknesses: - I'm not entirely certain that the analysis presented in Proposition 2 is correct. In the proof of Proposition 2 (lines 433 and 434), the authors use the fact that the model $f$ is $\epsilon$-DP with respect to the Bernoulli random variables $S$. But in Algorithm 1, we see that $f$ was trained on $D_{f}$ in phase 1, which is independent of $S$ sampled in phase 2. So, the model $f$ is independent of the selection $S$ (i.e., $f \bot S$). In other words, the equation after line 434 should evaluate to 1, as given $x_i$, the prediction $f(x_i)$ remains the same whether $S_i = 0$ or $S_i = 1$. Perhaps Algorithm 1 requires one training run like Steinke et al. where the model $f$ is trained on $X$ as described in line 2 of Algorithm 1? If that is the case, then the claim that the auditing does not require retraining becomes invalid. If I'm mistaken, could you explain Proposition 2 further? - In line 196, the authors mention that Algorithm 1 (lines 4-7) does a sweep over all the recall values of the attack, and they adjust the overall significance-level or the $p$-value by taking a union bound over all equation (1) with the significance level discounted to $\beta \leftarrow \beta / m$. The value of $m$ in the experiments ranges from 500 to 30,000. For such values, assuming $\beta \leq 0.05$, the sweeping over all recall values requires $\sqrt{\log(m/\beta)} \approx 2.4$ whereas when the level of recall is predetermined, the factor in equation (1) is only $\sqrt{\log(1/\beta)} \approx 1.15$. So, I'm not sure if doing a sweep over recall values will give larger DP estimate values. - Use of synthetic data from a distribution that is $c$-close (with a small $c$) can make it difficult for the membership inference attacks to work well. If the goal is auditing DP, perhaps the higher MIA precision for a given recall outweighs the drop in ML model's performance. - The figures haven't been explained very well. In particular, the theoretical maximum precision in Figure 3 and Figure 12 and the empirical maximum value in Figure 4 aren't clear. - The paper only studies the problem of estimating pure DP and does not present an operationalizable algorithm for $(\epsilon, \delta)$-DP, although some results along these lines are motivated in the appendix. Additionally, the obvious weakness (which the authors acknowledge) is that their method does not technically provide a lower bound for $\epsilon$-DP. ### Minor Points - Intuitively, when $\mathcal{G} = \mathcal{D}$ or in the Real-Member;Real-Nonmember (RM;RN) case where the non-members follow the same distribution, I think the no-retraining-needed argument could work. This would involve modifying Algorithm 1 and reworking Proposition 2 by (A) assuming entries in $D_f$ and $D_G$ are i.i.d. from the same distribution and (B) setting $S$ according to the original train-test split instead. On the other hand, when $\mathcal{G} \neq \mathcal{D}$, I'm not sure if such an argument can be made to work, at least not trivially. - It's not clear how the mechanism $B(S, X) = \\{b(x_1), \cdots, b(x_m)\\}$ in Proposition 1 incorporates the helper model mentioned in Section 5.1 used in the experiments. - Algorithm 1 (lines 4-7) does not seem to reflect the $p$-value adjustment discussed in lines 198-200. Perhaps the authors might have overlooked this $p$-value adjustment in the experiments as well? Technical Quality: 2 Clarity: 2 Questions for Authors: - How is the helper model used in the baseline trained? - The lexicographic order in equation (2) appears to be essential for Corollary 2 and seems to be aimed at ensuring something like $\epsilon > c$ to hold. Could the authors provide more details on this? - How exactly does the formulation in the paper differ from the setting introduced by Steinke et al., 2023, specifically in regards to the independence assumptions and the probabilistic dependencies between concerned random variables? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors have acknowledged some limitations in the paper. There do not appear to be any negative societal impacts associated with this work. I encourage the authors to address the issues and questions raised in this review. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. How is the helper model used in the baseline trained? (Q1) The helper model has the same classification task and architecture as the target model. To train it, we generate separate sets of training and validation data using our generator. For image data, the synthetic samples do not have a label. We thus train a labeler model, also a classifier with the same task, on the same dataset used to train the generator. We use the labeler model to provide labels for our synthetic samples (training and validation sets above). We train the helper model on the resulting training set, and select hyperparameters on the validation set. We give more experimental details for each modality in App. C (l. 473-474, 545-546, 567-568, 586-592). App. D1 further highlights the various architectures of the helper model we tried (l. 638, Table 5), including training the helper model on real non-member data. Table 5 shows the helper model trained on generated data gives the strongest baseline (highest $c_{lb}$). The lexicographic order in Eq. (2) appears to be essential for Corollary 2 and seems to be aimed at ensuring something like $\epsilon > c$ to hold [...]? (Q2) The lexicographic order in Eq. (2) is not to ensure $\tilde{\epsilon} > c$ (as a counter-example, $(c_1=3, \epsilon_1=1) \leq (c_2=3, \epsilon_2=2)$). The technical step relying on this order is the construction of the confidence interval (Corollary 2, l. 187-188) based on hypothesis tests from Prop. 1 and 2. The reason for this order with $c$ first is that Prop. 2 asks: ``if the generator is $c$-close, is the target model $\epsilon$-DP?'' We thus must reject any hypothesis with a $c$ we can disprove based on data, before computing $\epsilon$ based on a plausible $c$. Without this order (e.g. $\epsilon$ first or $c + \epsilon$), we would not necessarily compute $\epsilon$ based on a plausible $c$. We would assign some of the MIA performance to privacy leakage, when we know for a fact it comes from the generator being too far from the data distribution. How exactly does the formulation in the paper differ from the setting introduced by Steinke et al., 2023 [...]? (Q3) The key difference lies in how we create the audit set. In Steinke et al., 2023, the audit set is fixed, and data points are randomly assigned to member or non-member by a Bernoulli random variable $S$. Members are actually used in training the target model $f$, while non-members are not (so assignment happens before training). In our framework, we take a set of known iid. members (after the fact), and pair each point with a non-member (generated iid. from the generator distribution). We then flip $S$ to sample which one will be shown to the ``auditor'' (MIA/baseline) for testing, thereby creating the test task of our privacy measurement. Regarding the independence assumptions, $S$ is independent of everything by construction, as in Steinke et al., 2023. In fact, if we replace our generated data with in-distribution independent non-members, we exactly enforce the same independence, except we waste some data by drawing our auditing game after the fact. Using generated data adds another complexity, as $S$ ``leaks'' through the actual data-point shown to the auditor (based on differences between the generator and member data distributions), which we address with the baseline model. Correctness of Proposition 2. (W1) We hope Q3 above already helps resolve the misunderstanding about equation line 434 and Prop. 2. To expand: while $S$ is independent of $f$ ($S \perp f$), they are not conditionally independent when conditioning on $X$ (where $X$ is the test set for the MIA/baseline), so $f \not\perp S \| X$. This is because $X_i$ is either a member or non-member based on $S_i$, and the member is a data point on which $f$ was trained! Hence, Eq. below l. 434 does not evaluate to $1$. Deliberately ignoring $S_{<i}, X_{<i}$ for simplicity, the Eq. reads as the ratio between: (numerator) the membership guess (a post-processing of $f$ and $X_i$) conditioned on the event that we guess for input $X_i$ which is a generated non-member $S_i=0$; and (denominator) the membership guess (still a post-processing of $f$ and $X_i$) conditioned on the event that we still guess on $X_i$ which is now a member of $f$'s training set. If $f$ is DP, the membership guess is a post-processing of a DP result on two neighboring datasets ($X_i$ wasn't in the training set vs. $X_i$ was in the training set), and hence obeys the inequality shown after l. 434. Note that collisions (where we have $X_i, S_i=0$ but $X_i$ is also in the training data) make the ratio $1$: the inequality is still true and the theory applies, but the privacy measurement has no power (see discussion on overfit generator in Q1 of reviewer 55Rd). Adjusting significance level does not give larger DP estimate values. (W2) We adjust the significance level because we need to compare the highest lower-bounds implied by the MIA and the baseline. Notice on Fig. 3a and 4a that the highest precision values and implied lower-bounds can happen at different (unknown in advance) recall levels. If we were to fix a recall value in advance, we could get misleading results based on whether that value is closer to the best value for the baseline or for the MIA. We thus have to make the comparisons at different recall levels. In this case our tests from Prop. 1, 2 (which are at a fixed number of guesses) need a union-bound $\beta \leftarrow \beta / m$ to be correct. We could select the maximum without a union bound as a heuristic, but in practice it barely changes the results. This is because both the baseline and MIA numbers are lowered in a similar way by their respective union-bounds, so the gap between the two remains very similar. Auditing for $(\epsilon, \delta)$-DP algorithms. (W5) We do analyze the $(\epsilon, \delta)$-DP case in App. E2, and show results in Table 10 in App. E3. --- Rebuttal Comment 1.1: Comment: We believe we addressed the technical concerns raised in the review, and clarified the role of our lexicographic ordering, how our approach differs from that of Steinke et al., 2023, and cleared the concern about Proposition 2. Please let us know if there are any more concerns on these or other topics!	Summary: This paper proposes a novel privacy auditing procedure called PANORAMIA. The method works with a single model that uses all available training data, with no modifications to the training procedure, and with only access to a subset of the training data. This is achieved by synthetically generating non-member samples for auditing. The method produces an estimate of the $\epsilon$ differential privacy parameter, which, however, is not necessarily a lower bound. Summary of the method given a model and its training data: 1. Train a generative model on a subset of the training data. 2. Use the remaining training samples as "members" for auditing. 3. Generate an equal amount of "non-members". 4. Split both into training and test sets. 5. Fit a baseline classifier that distinguishes synthetic and real data on the training set (only input are samples). 6. Fit a membership inference classifier on the training set (inputs are samples and model statistics). 7. Use predictions on the test set and hypothesis testing to obtain a confidence interval for quantities of interest. The hypothesis is of the form "The generative model is $c$-close to the true data distribution and the training procedure is $\epsilon$-DP", where $c$ is a distance measure defined in this paper. The baseline predictions are used to reject the first claim, and the membership predictions to reject the second one. Hence, the returned $\epsilon$ is a lower bound on the true $\epsilon$ only if the generative model is indeed at least $c$-close to the data distribution. The theoretical analysis relies on results from the (1) procedure by [Steinke et al., 2023](https://openreview.net/forum?id=f38EY21lBw) but adapts them to the relaxed setting in this paper. For evaluation, the paper applies the new auditing procedure to various CNNs with and without DP guarantees on image data, small GPT2 models on text, and on tabular data. As baselines, the evaluation uses the (1) method by [Steinke et al., 2023](https://openreview.net/forum?id=f38EY21lBw) and real instead of generated members. PANORAMIA does not outperform the baselines, but achieves reasonably close results on image and text data. Strengths: - This paper considers a practical and relevant setting for privacy auditing. In particular, training the model to audit using all training data is a big benefit: otherwise, the audit either uses a slightly different setting that uses fewer training samples, or one has to forfeit a potentially large amount of training data (which hurts utility). Additionally, the procedure works even when the auditor knows only part of the training data, which can be relevant (e.g., federated learning). - The experiments are broad (considering both DP and non-DP training, and different degrees of overfitting), sound, and the conclusions are convincing. Although the results are weaker than for existing methods, those existing methods use a less appealing setting. A particularly convincing result is that PANORAMIA with synthetic non-members often comes close to the same method but using real non-member samples. - The paper is overall structured well and the auditing procedure is laid-out clearly. The authors manage to explain the auditing results well, despite their complex interpretation. I also appreciated that Algorithm 1 collects all parts of the method into one place, including the calculations to obtain confidence intervals. Weaknesses: - The considered privacy semantics are partially misleading and could be made more clear. - The presented auditing game is incomplete; Definition 2 only describes how samples are selected but should also include the goal and actions of the adversary. - The paper states (as a benefit) that PANORAMIA audits a target model, not the algorithm (L46--47). However, the results of the procedure are DP parameters, and DP is always a property of an algorithm (not model). If the considered semantics are meant to be different from DP, they should be made more clear and discussed. - Similarly, the paper claims that PANORAMIA does not require worst-case canaries because it audits the target dataset (via a model). However, this ignores that certain samples in the training data might be more vulnerable; hence, auditing should still consider the worst-case sample in the training data (especially since privacy is averaged over the dataset). This might be an explanation why the reported "lower bounds on $\epsilon$" are still far from tight in Figure 6c (as is the prior (1) procedure in a white-box setting). - Relatedly, the definition of c-closeness (Definition 3) might not require a generator to capture the tail of the data distribution. However, this tail contains outliers that are often particularly vulnerable to privacy leakage. Hence, ignoring such outliers might significantly underestimate the $\epsilon$ lower bound. I would have appreciated a more thorough discussion of this limitation (or a clarification). - The paper could be more explicit about whether PANORAMIA is intended to be a procedure to be used in practice, or an important stepping stone towards more practical procedures. Right now, PANORAMIA achieves worse results than the existing (1) method in a white-box setting, and requires training of a strong synthetic data generator (which might be non-trivial). Nevertheless, I believe this paper is still relevant from a conceptual perspective as a path for future work. - There are minor notation issues in Proposition 1 and 2: the statement mentions only $T$ and $t$, but the inequality uses $T^b$ and $t^b$ (or $T^a$ and $t^a$). Additionally, the left-hand sides both take probabilities over $T^b$/$T^a$ but simultaneously condition on the respective random variable. Fixing those points makes the propositions clearer. Also, the last sentence on L148 seems misplaced. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Might there be a way to use the membership classifier itself as a baseline, e.g., by somehow averaging over possible values for model statistics? This would avoid skewed scenarios where the MIA is better at detecting synthetic data than the baseline detector (even when ignoring membership signal). 2. In Figure 6/Section 5.3, do all models use (approximately) the same number of training samples? If not, could there be confounding effects (e.g., if models trained on fewer samples leak more privacy)? 3. Is there a path to extend this paper's method to always yield a lower bound on $\epsilon$ (w.h.p.)? Or would this require radical changes? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors are very transparent about all limitations of their work and discuss them transparently. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. Hereafter, we start with answers to the questions, before addressing other weaknesses listed that we believe stem from a miscommunication. Questions: Q1: Might there be a way to use the membership classifier itself as a baseline, e.g., by somehow averaging over possible values for model statistics? This would avoid skewed scenarios where the MIA is better at detecting synthetic data than the baseline detector (even when ignoring membership signal). Thanks for this interesting suggestion. If we could somehow capture a distribution of "non-member losses" for this data point, we might be able to marginalize out the dependency of the MIA on the target model and extract the input-dependent part. A key challenge with such a design would be to make sure that the resulting baseline is indeed as strong as it can be as well as that input differences do not ``leak through'' the target model loss for member data. In our experiments, the baseline is often better at picking up signal from the data point than the MIA model is (hence the cases, like in Figure 5 or some of the tabular data plots in the appendix, in which the baseline outperforms the MIA), which is in a sense the opposite problem. We believe it to be an interesting future work, and in general, any progress on discriminative models and MIAs will directly plug-into and benefit our approach. Q2: In Figure 6/Section 5.3, do all models use (approximately) the same number of training samples? If not, could there be confounding effects (e.g., if models trained on fewer samples leak more privacy)? Experiments for all data modalities in S5.3 show the effect of varying test set sizes on the performance of both Panoramia and O(1) while keeping all other experimental variables, including training dataset size constant across all the models. This is to ensure the results we observe are primarily due to varying the test dataset size while keeping everything else controlled. We mention the training details for the models in Appendix C. In addition, we also show, in Appendix D4, how varying training data for a fixed test set size impacts both Panoramia and baseline performance. Q3: Is there a path to extend this paper's method to always yield a lower bound on (w.h.p.)? Or would this require radical changes? Great question! This is plausible, and we hope that we, or another group, will figure out how in the future. The most straight-forward way to get a proper lower-bound on $\epsilon$ in our framework is to figure out how to measure (or construct) an upper-bound on $c$, as then $(c+\epsilon) _{lb} - c _{\text{ub}}$ yields $\epsilon _{lb}$. Getting an upper-bound on $c$ in the general case seems challenging though. There might be a way to leverage DP to do this by construction (since DP bounds the gap between distributions), though we haven't figured out how to do it yet. This is definitely a promising avenue for future work. Other important clarifications: Weakness 2 - Relatedly, the definition of c-closeness (Definition 3) might not require a generator to capture the tail of the data distribution. However, this tail contains outliers that are often particularly vulnerable to privacy leakage. Hence, ignoring such outliers might significantly underestimate the lower bound. I would have appreciated a more thorough discussion of this limitation (or a clarification). In cases where the generative model is unable to effectively capture the long tail distribution of the real dataset it is modeling, it becomes easier for a baseline classifier to distinguish between real and synthetic non-members, and hence to tell the two distributions apart ($c$-closeness, like pure-DP, does not allow such discrepancies in the tail, though our $(c, \gamma)$-closeness relaxation in appendix E does). In practice, we do observe this phenomenon, especially with tabular data. We highlight this limitation via results on the tabular setting in Appendix D6 (we will add a pointer to the paper for clarity). In such a case, we show that the baseline is as strong or stronger than the MIA and Panoramia is unable to effectively detect any privacy leakage from the respective target model. Weakness 3 - The paper could be more explicit about whether PANORAMIA is intended to be a procedure to be used in practice, or an important stepping stone towards more practical procedures. Right now, PANORAMIA achieves worse results than the existing (1) method in a white-box setting, and requires training of a strong synthetic data generator (which might be non-trivial). Nevertheless, I believe this paper is still relevant from a conceptual perspective as a path for future work. Thank you for the helpful feedback. We will clarify that we believe that our work is an important step towards solving the privacy measurement setup that we tackle, namely measurements without control of the training process and with distribution shifts between members and non-members. The end-goal for future work is a full-fledged approach (that provides a proper lower bound) with potential improvements over other approaches (e.g., optimizing the generator to improve the audit). However, we also believe that our framework can already be useful, for instance for providing improved measurements with more data (Fig. 6a); or to tackle privacy measurements in models for which there are no known in-distribution non-members, which as recently pointed out in [1] suffers from the same distribution shifts we tackle by using out-of-distribution non-members (our theory would apply to such a case, though whether the generative model part can help is still an open question). [1] Das, Debeshee, et al. "Blind Baselines Beat Membership Inference Attacks for Foundation Models." 2024. https://arxiv.org/abs/2406.16201 --- Rebuttal Comment 1.1: Comment: I thank the authors for their response, which answered all my questions thoroughly and resolved Weakness 3. However, after reading all other reviews and rebuttals, my concerns about the privacy semantics remain. In particular, a robust evaluation procedure should measure the privacy of the worst-case samples (in the dataset), not just an average over the dataset. For example, the high-level arguments in https://differentialprivacy.org/average-case-dp/ still apply for a fixed dataset. Often, samples from the tail of the data distribution are worst-case samples (in a fixed dataset). At the same time, this paper's method seems to fail if there are such samples, because the generator fails to synthesize similar non-members (I greatly appreciate the author's transparency in this regard). While the (1) procedure of Steinke et al., 2023 can suffer from similar issues, there one can simply pick an audit set that solely consists of worst-case samples. I don't see a similarly easy fix for PANORAMIA. Nevertheless, I do not think this is a deal-breaker, and believe this paper's idea and method carry merit in themselves (e.g., towards a solution of the problems highlighted by [Das et al., 2024](https://arxiv.org/abs/2406.16201)).	Summary: The paper proposes a novel way for privacy auditing of machine learning models through MIAs. The framework they propose, panoramia, aims to audit the privacy of a ML model post-hoc (so with no control over the training procedure), and with access to a known member subdataset. The method first consists of training a generative model on the known subset of members, which is used to then generate synthetic data samples from the same distribution. These synthetic points are then used as non-members, which combined with the known member dataset allows to fit and apply an MIA on the target model. Importantly, authors recognize that there might be a distribution shift between members and synthetic non-members, so they fit a classifier without leveraging the target model as a baseline. Next, they use the difference between the MIA performance (thus using the target model) and this baseline to estimate the privacy loss. Authors provide the formula (and add proof in appendix) to compute a value of epsilon approximating a lower bound on epsilon-DP. They further apply the privacy auditing to three kinds of ML models (image classification, language modeling, and tabular data classification). Authors consider models with varying degree of overfitting, DP training, and increasing the amount of member data available to the auditor. Strengths: - Originality: The paper introduces a way to audit the privacy of ML models post-hoc, without any control of the training data, which is novel. They cite and position themselves correctly to relevant prior work such as Steinke et al. - Quality: The proposed method is technically interesting, formally supported and evaluated extensively across data/model modalities. - Clarity: NA - Significance: The paper proposes a way to compute an approximation for the lower bound on epsilon, to audit ML models post-hoc, which is technically interesting. Weaknesses: - Originality: Authors should include an appropriate related work section, touching on other privacy auditing techniques and potentially other (post-hoc) membership inference attacks. - Quality: The proposed method strongly depends on the quality of a generator and the baseline MIA, the impact and limitations of which can be further explored (see questions/limitations). - Clarity: I find that the paper's clarity can be improved significantly. Especially the results section (tables, text) is quite notation heavy and hard to follow what everything refers to. - Significance: While technically interesting, the relevance of a technique to compute a proxy lower bound privacy loss in practice needs more compelling motivation (see questions). Technical Quality: 2 Clarity: 2 Questions for Authors: - I understand how c-closeness allows to estimate the quality of the used generator. However, what I struggle to understand, is what happens when you develop a generator that is perfect (c=0) which just samples randomly from the known member subset. Then, a baseline classifier would not be able to distinguish members from non-members, and nor will an MIA, leading to a privacy estimate that will be far off. In less extreme cases, the generator might be slightly overfitted and indeed generate non-members very similar to members, which might also impact the MIA performance. Am I correct that this could have a significant impact on the validity of the procedure? And if so, how would you address the concern? - More generally, Panoramia largely depends on a good generator and a baseline MIA. Can authors elaborate on the associated limitations? For instance, can good generators can be developed across all use-cases (smaller datasets, data modalities etc)? And how should a baseline performance be developed or evaluated to be used as part of the panoramia framework? - I understand that the method authors provide does not give a formal lower bound for the privacy loss, but rather a proxy for it. In practice, if indeed a hospital as part of an FL setup would like to assess the privacy leakage incurred by their data, why would they opt to compute an estimate for epsilon using your method? Instead, they could for instance generate non-members in the same way as panoramia and just quantify the MIA performance with an AUC or TPR at low FPR compared to a baseline. In general, authors should further motivate the relevance of a proxy for a lower bound privacy loss in practice. - To further motivate post-hoc privacy auditing (without any control of the training data), I wonder if it makes sense to also emphasize the context of generative AI models such as LLMs? These models are increasingly trained on all the data model developers can acquire so the absence of non-member data is very real in practice. - Can authors add a related work section? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Authors are very clear that their method only provides a proxy for the lower bound of epsilon and thus carefully caveat their method. However, can authors elaborate on the limitations associated with the development of both a generator and a baseline, both of which seem fundamental for panoramia. Currently, their implementation seems rather ad-hoc than adequately discussed for wider applications. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. Q1: What happens when you develop a generator that is perfect (c=0) which just samples randomly from the known member subset. In that case (a highly overfitted generator), the baseline and MIA will output $c _{lb} = \\{\epsilon+c\\} _{lb} = 0$ (though technically $c \neq 0$, since $c$-closeness quantifies the distance between the generator's distribution, and the distribution of members, not the specific member data points). This happens mainly because both the MIA and baseline detect leakage based on the same data, and assess the data in two ways: (1) the distributional difference in the data points themselves; and (2) for the MIA the difference in distributions of loss values the target model attributes to the member vs non-member data. In this case, our theory still applies but returns $\tilde{\epsilon} = 0$. Thus, we would not faultily ascribe privacy leakage to the target model, but the measurement would also fail to detect any real privacy leakage (one could see that the measurement is failing though, since $c _{lb} = \\{\epsilon+c\\} _{lb} = 0$). Q2: More generally, Panoramia largely depends on a good generator and a baseline MIA. Can authors elaborate on the associated limitations? [C]an good generators be developed across all usecases (smaller datasets, data modalities etc)? And how should a baseline performance be developed or evaluated to be used as part of the panoramia framework? The ability of the generator to capture well the data distribution is a key dependency of PANORAMIA and the reason why we have thoroughly evaluate our approach. In particular, in the tabular case, datasets typically contain a large number of ``average'' samples and a long tail of extreme values or rare classes. Thus in such a case, when the generated data does not capture the long tail, the baseline easily classifies all outliers as real data as no synthetic data points are similar to them. This leads to a strong baseline that is hard to beat for the MIA, which means that the privacy measurement then fails (though we can diagnose why). This is a limitation of our work that we discuss in details in App. D6 as a possible cause for the failure of the tabular data modality. In general, the baseline should be made as strong as possible to detect such failures of the generative model following the traditional ML best practices. For this reason, we have dedicated a lot of effort in designing and evaluating our baselines, including the helper models (S5.1 and App. D1). Our design and theory offer practical advantages as well: - As generative models improve, especially for smaller dataset sizes, so will our approach. - The generator opens an interesting design space, in which one could try to optimize synthetic data for audit quality. While we have not yet explored this design space, we leave it as an interesting avenue for future work. - Finally, our theory applies to other out-of-distribution non-member data, such as when using other datasets as non-members [1]. Q3: [T]he method authors provide does not give a formal lower bound for the privacy loss, but rather a proxy for it. [...] Why would [one] opt to compute an estimate for epsilon using your method? Instead, [one could] generate non-members in the same way as panoramia and just quantify the MIA performance with an AUC or TPR at low FPR compared to a baseline. [...] Authors should further motivate the relevance of a proxy for a lower-bound privacy loss in practice. As it is the consensus in the field, we believe that DP provides the best semantics to define and quantify this type of privacy leakage. Thus, usually, privacy audits aim to provide a lower bound for the privacy loss. While we do not yet provide a lower bound, using DP semantics is still useful. For instance, [2] makes a convincing case that accuracy or AUC are not good metrics for privacy leakage. Rather TPR at low FPR is better, but notice that we need to compare the MIA with the baseline, and their performance most revealing of privacy leakage happens at different FPR values (cf. Figures 3a and 4a, though in precision/recall terms). In this case, we cannot directly subtract TPR values. Our theory tells us: (1) which FPR we should chose and (2) how to scale the TPR at this FPR (by mapping it to a $c$ or $c + \epsilon$ value) to make the values comparable between the baseline and the MIA. Taking a step back, we hope that our framework will be a stepping stone for further research, maybe enabling a proper lower-bound on $\epsilon$ by measuring (or enforcing by construction) an upper-bound on $c$ (then $\\{c+\epsilon\\} _{lb} - c _{ub}$ yields $\epsilon _{lb}$). Though we do not know how to do it yet, this is a promising future work. Q4: To further motivate post-hoc privacy auditing (without any control of the training data), I wonder if it makes sense to also emphasize the context of generative AI models such as LLMs? These models are increasingly trained on all the data model developers can acquire so the absence of non-member data is very real in practice. Thanks for this great point. Actually, a paper made public after our submission [1] highlights this exact issue. As you mention, in this case, there is no known non-member data from the same distribution, MIA benchmarks use out-of-distribution non-member data (e.g., more recent datasets). Thus using generated data might yield better non-member data. Our theory applies equally to both real and synthetic non-members data and we believe that it may be an interesting building block for that setting. Q5: Can authors add a related work section? We propose to add a more formal related work section to collect the closest works, and an extended one in appendix. [1] Das et al. "Blind Baselines Beat Membership Inference Attacks for Foundation Models." 2024. https://arxiv.org/abs/2406.16201 [2] Carlini et al. "Membership inference attacks from first principles." S\&P 2022. --- Rebuttal Comment 1.1: Comment: We believe we clarified how our approach deals with overfit genertors $(c=0)$, and why the theory we developed goes beyond comparing AUC or TPR at low FPR between a MIA and baseline. We will add these clarifications to our paper and mention the generative AI use-case mentioned in the review (thank you!) as additional motivation. Please let us know if there are any more concerns on these or other topics!	Summary: The authors propose a privacy auditing technique that utilizes partial access to member data to generate synthetic non-member data, which in-turn is used to train a meta-classifier that can be used to empirically measure privacy leakage relating to record membership. The authors evaluate their technique on models as large as GPT2 and find correlation between audit scores and the expected leakage from models. Strengths: - L226-228: Good to include negative results! - Having access to non-members is often taken for granted (especially when the auditor is also the model trainer), but in most third-party cases getting good-quality non-member data that is not significantly different from member distribution is hard. The methods proposed in this work can be useful there, with a generator and discriminator that can help ensure any differences in members and non-members (and subsequent MIA performance) do not arise from distributional differences. - The paper is well written and structured, and experiments are thorough, spanning quite a lot of models and modalities. Weaknesses: - As an auditor, access to validation/test data used in model training would not be a far stretch- why not use the non-member set as used in standard MIA setup (validation/test data). What is the added benefit from this extra step of generating synthetic non-members? - Regarding the contributions, (1) and (2) has already been explored by [1, 2]. While the method in [1] does not currently support large models/datasets like CIFAR10, it would be nice to highlight differences here. If you indeed have knowledge of a decent-chunk size of members and non-members, I'd imagine you could do something better like [2], and other related methods. - I have some concerns over the dependency on how well the baseline in/out distribution detection system works. As a concrete example consider Figure 7(a, b) - even as a human I see a very clear difference in resolution of the generated images and find it hard to believe that the distinguisher does not work well here. Even a non-ML technique (that can work around with blurring) would work pretty well here. - Table 1: Accuracy of models here is not good enough, especially when using entire data, with very clear signs of overfitting. If using the entire data (and not half split, as in most MIA evaluations [3, 4] where even with half the dataset test accuracy is ~92%), should be able to train well-performing models that do not overfit so heavily. Please see [this resource](https://paperswithcode.com/sota/image-classification-on-cifar-10) - Figure 1: While the trends suggest that the proposed audit correlates with actual leakage, one could also argue that (100 - test accuracy) is also a useful privacy metric in this comparison given the correlation. The utility of the audit would be more apparent, then, when studying models that have comparable test performance, but have (by design) different leakage, perhaps focusing on moderate/large values of $\epsilon$ with Differential Privacy-training. - L288-290: (1) and (2) are ambiguously enforced- if you have a large portion of train data, you can control process in most cases, and also obtain non-member data. As far as (3) is concerned, that is a compute-related constraint, not a threat model difference. The auditor could, for instance, use available data knowledge to train "in" models. - L331-334: I do not find the justification behind having access to partial member data convincing. While I am okay with just stating that "this is a possible limitation, but okay to assume for an auditor", but relating it to how it could be useful in situations like FL is not practical. For instance, here empirical experiments use more than 1/5th of the data; no FL training will have just 5 participants. This is closer to a specific case of distributed learning, or pure data aggregation. Even in such cases, the data distributions per client will be considerably different, as opposed to the experiments here which have uniform samples. ## Minor comments - Please consider a more descriptive abstract. "scheme" here is not very descriptive - L121: "We formalize a privacy game" - this is standard membership inference evaluation and not a contribution of this work. Please rephrase to make this clear. - L493-494: Please provide direct distinguishing accuracies of these baseline classifiers for reference and easier understanding. ### References - [1] Suri, Anshuman, Xiao Zhang, and David Evans. "Do Parameters Reveal More than Loss for Membership Inference?." High-dimensional Learning Dynamics 2024: The Emergence of Structure and Reasoning. - [2] Nasr, Milad, Reza Shokri, and Amir Houmansadr. "Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning." 2019 IEEE symposium on security and privacy (SP). IEEE, 2019. - [3] Carlini, Nicholas, et al. "Membership inference attacks from first principles." 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022. - [4] Zarifzadeh, Sajjad, Philippe Cheng-Jie Marc Liu, and Reza Shokri. "Low-cost high-power membership inference by boosting relativity." (2023). Technical Quality: 3 Clarity: 4 Questions for Authors: - Figure 3a suggests an ordering of 100 > 20 > 50 in terms of leakage for high-recall region, whereas numbers in Table 1 suggest that the trends should be like 100 > 50 > 20 if the proposed method is indeed measuring what it is supposed to. Why is that so? - L480-481: this means the "labeler" is much weaker and might be generating incorrect labels more often? Are there any numbers for what the actual test performance of this labeler is? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: There are a few statements that currently serve as justification for limitations (see above) but should just be posed directly as base assumptions. Apart from that (and some potential shortcomings in models used for evaluation), most limitations are already stated in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. We start with answers to the two explicit questions, before addressing two of the weaknesses listed that we believe stem from a miscommunication or misunderstanding. Questions: Q1 - Figure 3a suggests an ordering of 100 > 20 > 50 in terms of leakage for high-recall region, whereas numbers in Table 1 suggest that the trends should be like 100 > 50 > 20 if the proposed method is indeed measuring what it is supposed to. Why is that so? In our experiments on all data modalities and models, the precision values that lead to the highest lower-bounds $c$ or $c + \epsilon$ are achieved at fairly low recall but for different recall values (especially for images). For this reason, when training the baseline and MIA models, we use a validation set to select hyper-parameters and training stopping time that maximize the lower bounds (effectively maximizing the precision in the lower range of recall values) on the validation set. We will clarify this in Appendix C. This is why for some instances like pictures, MIA models have the wrong order at high recall values. While they could probably be tuned to achieve higher precision and be ordered correctly, we do not do it since that regime is not where the highest lower-bounds happen. Indeed, we can see in both Figure 3 and associated results in Table 1 that in the recall regions at which the lower-bounds are maximal, the order is as expected (100 > 50 > 20). Q2 - L480-481: this means the "labeler" is much weaker and might be generating incorrect labels more often? Are there any numbers for what the actual test performance of this labeler is? The labeler used to train the helper model can indeed be weaker and generate wrong labels. For instance, our labeler for image data has 80.4\% test accuracy on the CIFAR10 real data test set. The rationale for this approach is to augment the baseline with a model providing good features (here for generated data) to balance the good features provided to the MIA by $f$ outside of the membership information. In practice, the labeled generated data seems enough to provide such good features, despite the fact that the labeler is not extremely accurate. We studied alternative designs (e.g., a model trained on non-member data, no helper model) in Appendix D.1, Table 5, and the helper model trained on the synthetic data task performs best (while not requiring non-member data, which is a key point). Other important clarifications: Weakness 1 - [W]hy not use the non-member set as used in standard MIA setup (validation/test data). What is the added benefit from this extra step of generating synthetic non-members? We do study this setup in Fig. 6, comparing a MIA using up to the full test data and our approach using generated data. On CIFAR10 we can observe that, for the same non-member dataset size, using real data performs better (when using an ML-based MIA instead of a loss-based attack in this case). However, generated data enables us to use more data points for training and evaluating the MIA (and baseline), leading to larger measurements of privacy leakage. This is true for both regular and DP models (Fig. 7a and 7c), despite the fact that the CIFAR10 test set is fairly large compared to the training set (20%) and such a large portion of training data may not be kept for testing in general. Going in the same direction, a very recent work made public after our submission [1] shows that a similar member/non-member distribution shift issue happens when measuring privacy leakage from foundation models. In that case, the models are trained on vast amounts of data, and there is no known non-member data from the same distribution. As a result, MIA benchmarks use out-of-distribution non-member data (e.g., more recent datasets). Thus, using generated data might yield better non-member data. Additionally, our theory applies to both real and synthetic non-members data. Consequently, we believe that our approach may be an interesting building block for the setting described in [1]. Finally, we believe that the generator opens an interesting design space, in which one could try to optimize generated data for audit quality. We have not yet explored this avenue of research, but we believe that it is an interesting future work. Weakness 3 - I have some concerns over the dependency on how well the baseline in/out distribution detection system works. As a concrete example consider Figure 7(a, b) - even as a human I see a very clear difference in resolution of the generated images and find it hard to believe that the distinguisher does not work well here. Even a non-ML technique (that can work around with blurring) would work pretty well here. This is a great point, thank you. We apologize as this is a miscommunication on our end: our CelebA experiments are using a $128 \times 128$ resolution, thus our generative model generates images at that resolution. In Figure 7a though, we wrongly showed full resolution $218$ x $178$ images, hence the resolution difference. We fixed this figure to display the real images as we use them, and included it in the associated PDF answer (Fig. 1), for which we can see that the difference in resolution disappeared. Also note that the generator only needs to generate some good images, enough that it is hard to detect real members with high precision. This is because we play a one-sided member-detection game, as explained in Remark 3 ll. 137-144. Of course, it is always possible that much stronger baselines exist, although we did spend quite a bit of effort and time on making them as good as we could (see details in Appendix D of the submission). [1] Das, Debeshee, et al. "Blind Baselines Beat Membership Inference Attacks for Foundation Models." 2024. https://arxiv.org/abs/2406.16201 --- Rebuttal Comment 1.1: Comment: > ...such a large portion of training data may not be kept for testing in general. No model (at least in practice) is ever trained without any validation/test data. While it may not be as high as 20%, getting an estimate of privacy leakage of the underlying training mechanism/model should not require more than a small fraction of the data. I think a convincing argument can be made to use generative techniques, and the authors are close to one but not there yet. Most of my other concerns remain unaddressed, specifically: claims around contributions, underperforming models, and how the score generated via the proposed technique is any better than directly measuring test accuracy as a signal to help understand relative leakage (and other misc. comments). I am hoping the authors will respond to them --- Reply to Comment 1.1.1: Comment: Thank you for your follow-up! The are two things we would like to emphasize about the validation set, in case they were lost in the rest of the answer. First, we do see that generated data can provide higher privacy leakage measurements above the ones we can measure using 20% non-member held-out data (Figures 7a and 7c). Second, we believe that foundation models are a counter-example to the claim that there always is such a large held-out set: as far as we know there typically is no publicly known held-out set of in-distribution non-members (this is from an external point of view, it is likely that there is one internally). However, there are many known member data points, and externally measuring privacy leakage for such models is a topic of interest (see [1] in our answer). We are happy to discuss the other comments as well! > Claims around contributions In our contributions paragraph, we list three key properties of the setting we tackle. We do not mean to claim that we are the first to study each separately! We will rephrase this paragraph to clarify that the novelty of our approach lies in the combination of those properties. The two papers cited in the review focus on membership attacks, which is just one part of our approach. It would be interesting to see if we can use those attacks in Panoramia to yield better measurements, but this paper focuses on developing the privacy measurement framework that we propose. Regarding the specific membership inference papers: * The paper cited in the review as ``[1] Suri, Anshuman, Xiao Zhang, and David Evans. "Do Parameters Reveal More than Loss for Membership Inference?."'' was first posted on arxiv in June'24, after our submission. It is indeed a very interesting candidate to use in Panoramia's privacy measurement in the future. * The paper cited as ``[2] Nasr, Milad, Reza Shokri, and Amir Houmansadr. "Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning."'' has several membership attacks, and some might be compatible with our approach and could be interesting to try. Note that in both cases, we would still need non-members to measure and assess the performance of the membership attacks (and hence the generative model in our setting), and we would still need a theoretical framework to meaningfully compare the results of the results of the MIA and baseline to quantify privacy leakage (our theory). We will make sure to mention the papers and which parts of our framework they could improve. > underperforming models In this paper, we focused on developing our privacy measurement framework, and studying its behavior when varying levels of overfitting (increasing the number of training epochs), changing the number of parameters, or making models DP. For this reason, we did not focus on using the most recent, state-of-the-art models, though it would be an interesting measurement study in the future. On the specific models we use for image data, we use a ResNet101, which reaches 93.75% test accuracy on CIFAR10 with ImageNet pre-training (that is why our version is a bit weaker at 91.61%, as we do not do this pre-training). We note that other papers listed in the review evaluate similar performing models (a WideResnet with 92.4% in [2], 92% in [3], and 90% accuracy in [4]). We will add results for a stronger model. We will report results if they finish by the end of the discussion period, though we are not sure it will be ready due to the computation power we have access to.	Rebuttal 1: Rebuttal: Thank you for the thought-provoking reviews and suggestions! In our answers, we focus on key misconceptions regarding our paper and misunderstandings in the reviews. Hereafter, we summarize the most important points addressed: - Potential baseline weakness (reviewer CsXP): this is an issue due to Figures 7(a, b) in which we wrongly put full-resolution member images while we work with a lower-resolution dataset for our experiments (hence with a generator outputting lower-resolution images). We fixed the figure (Figure 1 in the PDF attachment) to display the member images that we actually used. - Technical question on the soundness of equation line 434 (reviewer Z4qp): we explain why this equation is the DP bound as we wrote, and why it does not evaluate to 1. We also clarify how our auditing game differs from (and resembles) the one from (Steinke et al., 2023), which is related to the soundness of the equation at line 434. - Motivation for the approach (reviewers 55Rd, kScF, Z4qp), including: (1) why using DP is useful even without a lower-bound (i.e., it lets us compare the MIA and baseline on the same scale while being a good way to formalize and quantify privacy leakage); (2) the usefulness of our framework (e.g., the theory can help with other out-of-distribution non-member data, such as when using other datasets as non-members [1], our generator opens an interesting design space to maximize audit efficiency); and (3) that we believe that our current work might be a stepping stone to a more complete approach that yields a proper $\epsilon$ lower-bound. We address all these issues and some of the other comments from each reviewer in the individual answers. [1] Das, Debeshee, et al. "Blind Baselines Beat Membership Inference Attacks for Foundation Models." 2024. https://arxiv.org/abs/2406.16201 Pdf: /pdf/3a0ff058cf1f9b085eca15e8ed27729080bbab28.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Hypothesis Testing the Circuit Hypothesis in LLMs	Accept (poster)	Summary: This paper considers the problem of formalizing and evaluating the circuit hypothesis. This hypothesis posits that specific subnetworks within a Transformer model are responsible for specific model behaviors. Although there are multiple examples of such circuits that have been manually discovered in the literature, the question of how to precisely judge whether a discovered circuit is indeed responsible for a specific behavior remains open. While past works have proposed various ad-hoc approaches, this paper first presents three criteria that an ideal circuit should satisfy (inspired by past works) along with corresponding statistical hypothesis tests for judging if a given circuit indeed satisfies these criteria. The paper then further presents weaker versions of the hypothesis tests for two out of the three criteria that are more likely to be satisfiable by circuits discovered in practice. These hypothesis tests are then applied to six different circuits that have been proposed in the literature to evaluate the extent to which they satisfy these criteria. Strengths: This is a nice paper. It is well-written and addresses a question of growing importance. Mechanistic interpretability has emerged as a promising approach for understanding the behavior of trained LLMs. However, given the infancy of the area, the definitions and criteria for evaluating the quality of a mechanistic interpretation are yet to be standardized. This paper fills this need, particularly in the context of circuit analysis. The three proposed criteria of preservation, localization, and minimality make intuitive sense. The experiments with the synthetic circuits also provide evidence in favor of these criteria. I particularly like the proposed use of hypothesis tests to evaluate these criteria. These tests provide an operational definition of an ideal circuit and can be feasibly used in practice for judging new circuits. Overall, I think this paper helps bring some formal discipline to the topic of circuit analysis. Weaknesses: I have a couple of concerns. First, the hypothesis test for judging equivalence (Equation 2) feels somewhat unintuitive. Consider the scenario where the circuit C* significantly outperforms the model M on half the inputs and vice versa. Intuitively, one would not consider such a circuit equivalent to the model, yet the hypothesis test would conclude otherwise. Why not define equivalence simply in terms of faithfulness of the circuit C* to model M? My second concern is the possibility that the three proposed criteria are insufficient for characterizing an ideal circuit, i.e., there might be additional criteria that are needed. However, given that circuit analysis is in its early days, I think the paper already makes useful contributions. I also have a number of minor comments and questions that I list below: 1. Line 180: Why use ; instead of , in s(C*(x);y)? 2. Equation 2: I think there is a missing abs operation. The definition of the null hypothesis in Appendix B.1 further suggests this. 3. Equation 2,3: Notationally, are lower case and upper case letters being used to represent separate things? For instance, x,y vs X,Y? Why not use lower case x and y in these equations? 4. Equation 5: In practice, there will be randomness also due to the fact that \delta(e,C) is calculated empirically. Does the test account for this randomness? 5. Line 267: The notion of a "reference distribution over circuits from the complement distribution" is not very clear at this stage. I understand what it means based on reading the rest of the paper. 6. Figure 3: The results for the IoITask are a little mysterious. How is the faithfulness of a circuit with some edges removed worse than an empty circuit? Further discussion would be helpful here. 7. Is Algorithm 2 an original contribution or was it presented in Gretton et al., 2007? Some more explanation and intuition about the algorithm would help the reader. 8. It would help to have a more careful comparison with the three criteria for circuits proposed by Wang et al, 2023. Technical Quality: 3 Clarity: 4 Questions for Authors: Please address the questions listed in the previous section. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Yes, the paper discusses the limitations and the potential impacts of the proposed work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses *First, the hypothesis test for judging equivalence (Equation 2) feels somewhat unintuitive. Consider the scenario where the circuit $C^$ significantly outperforms the model $M$ on half the inputs and vice versa. Intuitively, one would not consider such a circuit equivalent to the model, yet the hypothesis test would conclude otherwise. Why not define equivalence simply in terms of faithfulness of the circuit $C^{}$ to model $M$?* Thank you for raising this great point! We agree that the test above does not handle this case. However, we intend our suite of tests to be used jointly. The situation you describe would be adequately handled by the sufficiency test, as the circuit you described would surely be unfaithful. However, we think this test complements a test that only looks at faithfulness for the following reasons: 1. It can distinguish situations where there is no bias (equally outperform each other) from situations where there is, even when the model is very faithful. 2. It is lenient to added variance, which could happen, for example, because we are using patching. We believe these properties make the test a valuable addition to the suite, as it helps to clarify precisely how the circuits are close or not to the model. My second concern is the possibility that the three proposed criteria are insufficient for characterizing an ideal circuit, i.e., there might be additional criteria that are needed. However, given that circuit analysis is in its early days, I think the paper already makes useful contributions. We agree! We hope future work builds on these criteria. ## Minor comments and questions Thank you so much for the careful read! Your feedback is greatly appreciated. - *Line 180: Why use ; instead of , in $s(C^{}(x);y)$? That’s a great question! This is a typo --- it came about because $y$ can be a parameter of the task, rather than the classic label. For example, it can be the index of the subject and object in IOI. We spent a fair bit of time debating whether to treat that as a parameter with `;` or a variable which we denote with `,`. However, we realize we have been inconsistent with the notation and have changed all of them to `,`. - Equation 2: I think there is a missing abs operation. Yes. Apologies about that. - Equation 2,3: Notationally, are lower case and upper case letters being used to represent separate things? It is a typo. We have changed them to lowercase. - Equation 5: In practice, there will be randomness also due to the fact that $\delta(e,C)$ is calculated empirically. Does the test account for this randomness? This is a great question! Randomness in a dataset would definitely introduce additional variance. We sidestep this issue by defining the task with respect to a fixed dataset. But a different test could be designed to account for a finite dataset. In the flexible tests, we only account for the randomness arising from the sampled circuits and edges. - Line 267: The notion of a "reference distribution over circuits from the complement distribution" is not very clear at this stage. Thank you for pointing this out! We’ve rephrased it as ''reference distribution over circuits from the complement distribution, i.e., circuits that do not overlap with the candidate circuit.'' Is this clearer? - Figure 3: The results for the IoITask are a little mysterious. We believe this is because of the ''Negative Name Mover Heads'' discovered in the original IOI paper [1], which write in the opposite direction of the name mover heads. - Is Algorithm 2 an original contribution or was it presented in Gretton et al., 2007? Algorithm 2 has two components, the permutation test and HSIC. The permutation test is a classic hypothesis test by Fisher [2] and Pitman [3]. It is a nonparametric test to aim to show whether the observed statistics could be drawn from the “permuted distribution”. In our setup, we randomize the complement circuit from the observed y -- creating an independent null. If the observed statistics fall outside of the independent null, then we reject the hypothesis. The HSIC is proposed measures in Gretton et al [4] to measure the independence between two variables in the kernel space. - It would help to have a more careful comparison with the three criteria for circuits proposed by Wang et al, 2023.** Yes, we absolutely agree. It is discussed in the extended related work in line 486, Appendix A. ## References [1] Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, & Jacob Steinhardt. (2022). Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small. [2] Fisher, R. A. (1935). The design of experiments. New York, NY: Hafner. [3] Pitman, E. J. G. (1937). Significance tests which may be applied to samples from any population. Journal of the Royal Statistical Society. Supplement, 4, 119-130, 225-232. [4] Gretton, A., Fukumizu, K., Teo, C., Song, L., Schölkopf, B., & Smola, A. (2007). A Kernel Statistical Test of Independence. In Advances in Neural Information Processing Systems. Curran Associates, Inc. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response! I will keep my score. It will be helpful to include the discussion of Equation 2 in the paper. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you for responding to our rebuttal and for your support of the paper! We will definitely include the discussion of equation 2 in the paper.	Summary: This paper proposes a set of tests to evaluate how well a "circuit" meets its desired properties. - Here a circuit refers to a subnetwork, which could be either synthetic (e.g. constructed according to RASP) or discovered in a trained model. - The desired properties considered in this paper are: - faithfulness, where the circuit should preserve the performance of the original model; - localization, where removal of the circuit should alter the model's output; - minimality, which says the edges in a circuit (treated as a computational graph) should not be redundant. There are 3 "idealized tests" intended as necessary (but not sufficient) conditions: - Equivalence, where the null hypothesis is that the circuit $C^$ preserves the performance of the original model $M$. - The test checks whether $C^$ and $M$ have an equal chance of outperforming each other, which is a necessary condition (but not sufficient): rejecting the null means that $C^$ necessarily does not preserve the performance. - Independence, where the null hypothesis is that the output of the model after removing $C^$ is independent of the output of the original model. - The test computes the HSIC between the performance of $C^$ and $M$. - Note that this is a stringent requirement; this will later be modified into a "flexible" test. - Minimality, where the null hypothesis is that all edges in $C^$ are necessary and not redundant. - The test checks whether the amount of performance change by removing an edge in $C^$ is more than the amount of change caused by removing an edge that is believed to be redundant. The paper also proposes 2 "flexible tests", where the difficulty of the test can be gradually varied by choosing different reference distributions when computing the test. - Sufficiency, which computes the probability (over random circuits in a reference distribution) that $C^$ is more faithful than a random circuit. - Partial necessity, which checks whether with high probability, removing $C^$ leads to a worse performance than removing a random circuit. The paper then applies these tests on 2 synthetic circuits and 4 circuits manually discovered from trained Transformers. - Synthetic circuits: for 2 types of the Tracr task, where the circuits are given as RASP programs. - Discovered circuits: Indirect object identification (IOI), Induction, Docstring (DS), Greater-Than (G-T). It finds that the synthetic circuits pass all 3 tests. In contrast, the discovered circuits fail the test to various extents, but are still far from random. Moreover, the flexible tests provide a more fine-grained understanding than the idealized tests (with binary outcomes). Strengths: - The paper provides a quantitative way to evaluate the concept of "circuit" in models, by adopting the hypothesis testing framework. The proposed tests correspond to desired properties a circuit. - The proposed tests are more thorough and fine-grained than some existing evaluation criteria, such as the knockdown effect. - The paper are clearly written. Weaknesses: - I'm concerned about the practical applicability of the tests. As the paper also mentioned, the proposed tests are either too stringent, or can be sensitive to the circuit size and the choice of the reference distribution. - The implications and use cases of the tests could be better discussed; e.g. please see questions below. Technical Quality: 3 Clarity: 3 Questions for Authors: - About independency: neural networks are often overparameterized and hence likely contain a high level of redundancy. What if there are multiple circuits that are each faithful and minimal but are similar to each other? These seem to me should be considered as valid circuits, but they would violate independency. - About minimality, I don't see why the randomly added edge would lead to a small $\delta(e^I, C^I)$: it's possible that adding a random connection would change the performance non-trivially (but likely negatively). For example, even a simple residual term would change the scale of the output and affect subsequent computations. - Could you comment on the effect of different choices of activation patching? - Could you comment on how the proposed tests can inform interventions on training (e.g. as regularization terms)? The hope is that this could make the trained networks more likely to contain circuits that more closely satisfy the desiderata. - Can the proposed tests help with coming up a more precise definition of circuits, or inform how we should choose the granularity of the definition of "nodes"? Minor clarifications: - Eq (2): should the LHS be taking an absolute value? - Line 246: not sure I understand what “a lucky draw” means; e.g. it could mean a draw that is better than q fraction of random circuits? - Line 250: what does "supersets of C" and "comparable to C" mean? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper discusses the technical limitations. There is no direct societal implication. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses I'm concerned about the practical applicability of the tests (...) can be sensitive to the choice of the reference distribution. Thank you for raising this important point! We believe that the main objections stem from a small but important misunderstanding of the paper's main goal. The review states: > This paper proposes a set of tests to evaluate how well a "circuit" meets its desired properties. We politely disagree with this goal. Our goal is to formalize the implications of the circuit hypothesis---"LLMs implement tasks through circuits." We demonstrate consistency with synthetic circuits, and we then study how existing circuits align with these idealized properties. Our goal is not to create a checklist that evaluates which circuits are better or worse. This subtle but important difference leads to different interpretations of the results. If a circuit does not align with the circuit hypothesis, it does not mean the circuit is "undesirable"; rather, it could indicate that the idealized version of circuits differs from how neural networks actually encode tasks. Our paper aims to provide tools to quantify the extent to which the idealized circuit hypothesis aligns with circuits in practice. Thus, the practical applicability of our tests is to provide a nuanced understanding of the alignment between the circuit hypothesis and discovered circuits. Currently, circuits found in transformers rarely pass the stringent tests due to issues like the redundancy mechanism, as the reviewer correctly points out. However, future models could exhibit more modular and circuit-like behavior. Our tests can help determine when this happens. ## Questions What if there are multiple circuits that are each faithful and minimal, but are similar to each other? We agree that neural networks implement redundancy, making the independence test difficult to pass. This suggests that the network is not implementing the tasks similarly to how the idealized version of "circuits:' and neural networks behave. Our tests help illustrate how the neural network in practice does not align with the idealized circuit hypothesis. About minimality, I don't see why the randomly added edge would lead to a small $𝛿(𝑒^𝐼,𝐶^𝐼)$ Thank you for the question. A key component of the circuit hypothesis is that edges not part of the circuit can be removed without significantly affecting model performance, hence our design of the minimality test. While "removing" is a common term, this can be more complex than simply setting values to zero, as is the case with STR patching [1]. This is detailed in lines 109-124 of the paper. This ablation method is chosen to maintain the edge magnitude and avoid the issues you described. Further clarifications on activation patching are provided in the question below. Could you comment on the effect of different choices of activation patching Congruent with existing findings [1], we found that the circuit is indeed sensitive to the ablation scheme that was used to “discover” it. This is illustrated in Figure 6 of the paper. We present the results with the original method used to discover the circuit with the aim of fairness. Appendix D.1 clarifies which ablation methods were used with which dataset. Could you comment on how the proposed tests can inform interventions on training Thank you for the question! These tests can indicate if certain training methods or architectural variations are more likely to produce idealized circuits. For instance, if circuits in a transformer trained with a specific regularizer align more with the tests, this suggests that this method is more effective at producing circuits. However, it is unclear how to incorporate these tests during training. These tests are designed for specific "hand-crafted" tasks $\tau$, while pre-trained models are not trained this way. Additionally, systematizing the creation of many such $\tau$ or deriving a differentiable loss from the tests remains a challenge. Can the proposed tests help with coming up a more precise definition of circuits? Thank you for the question! It depends on what you mean by this. When we consider the Transformer and related architectures as computation graphs, any meaningful definition of a circuit is likely to be similar to the one we have provided. This similarity arises because the computation graph framework inherently constrains how circuits can be defined. Consequently, it would be challenging to find a more precise definition of a circuit within this framework without it closely resembling our existing definition. However, if you are asking about the level of granularity at which we apply the definition (e.g., interventions at a node level vs. edge level), then we think our tests would be useful. This would certainly be an interesting experiment to explore in future work. ## Minor Clarifications - Eq (2): should the LHS be taking an absolute value? Yes! Thank you for pointing it out! - Line 246: not sure I understand what “a lucky draw” means We mean the following: If with a small probability (e.g., 10%), we could have randomly drawn a circuit just as faithful as the candidate circuit, then the candidate circuit is simply a "lucky draw" from the reference distribution. If the candidate circuit is significantly more faithful than the 90% most faithful random circuit, then it is not just a lucky draw. - Line 250: what does "supersets of C" and "comparable to C" mean? A circuit is defined as a set of nodes and edges. A superset of C is a circuit that contains all edges and nodes in C. For a circuit to be "comparable to C" means that it will have similar faithfulness to C. We expect a superset of C to be comparable to C. ## References [1] Fred Zhang, & Neel Nanda. (2024). Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. --- Rebuttal Comment 1.1: Comment: Dear Reviewer, thank you again for the time and effort you've dedicated to reviewing our work. We believe our responses address the concerns raised in your reviews. As the discussion period is nearing its conclusion, if you find that any aspects of our responses require further clarification or discussion, we are eager to engage in constructive dialogue! --- Rebuttal 2: Comment: Sorry for the delay in my response, and thank you for the clarifications! My main concerns have been addressed, and I've raised my score. I'd appreciate more clearly stating the paper's goal in the camera ready. Another question please: the word "circuit" has different meanings with different implications on generalization. For instance, circuit in the complexity sense (e.g. a boolean circuit) refers to a unit with well-defined computation, and hence would have perfect OOD generalization (where OOD refers to a change of distribution over the inputs). In contrast, there's typically no generalization guarantee when referring to circuits as subnetworks. Could you share your thoughts on what implications on generalization could we get by interpreting a network through its subnetworks? --- Rebuttal Comment 2.1: Title: Thank you for the response and the question! Comment: Thank you for your response! We are glad that we have addressed your concerns. We will ensure that the paper's goal is clearly stated in the camera-ready version. Regarding the question about circuit generalization, this is an excellent point and aligns with our current research. Currently, a circuit is defined with respect to a dataset, and circuits do not generalize well if the dataset is significantly changed, even if the underlying task remains the same. One possible reason is that multiple circuits can replicate the model behavior for a given dataset, but only a few may generalize across different datasets. For example, a circuit for mathematical computation can perform the Greater Than task, as can a circuit that can output a larger number after some lower number. We believe finding a circuit that generalizes across different datasets of the same tasks bring us closer to the "circuits" in a complexity-theoretic sense.	Summary: The paper operates in the framework of mechanistic interpretability of transformer models, where it is assumed that subgraphs (i.e. circuits) of the computational graph determined by the model implement specific capabilities of the latter. It defines a set of properties that an ideal circuit should have, namely equivalence, independence and minimality. Afterwards, it proposes two sets of hypothesis tests. The first set identifies a more strict hypothesis-testing framework to determine if a circuit satisfies these properties. The second proposes some more flexible tests for the first two properties of ideal circuits. Afterwards, these hypothesis tests are used on synthetic and manually discovered circuits in the literature on GPT-2 small and other small transformers. Strengths: - Overall very well-explained paper, particularly the part on mechanistic interpretation. - Crafting of the hypotheses for testing is original and well thought. It follows logically from the properties defined and is well explained. - Different types of tests were performed with different granularity. Weaknesses: - I believe it would have been nice to perform further tests to understand whether the idealized tests have some applicability on discovered circuits (or if they always lead to the null being rejected), potentially by also altering the discovered ones. - Would have found it interesting to propose some ideas on how to use these tests for circuit discovery. Technical Quality: 2 Clarity: 4 Questions for Authors: - I believe there is a typo in the description of Table 1 on page 7, where it is claimed that “A (✓) indicates the null hypothesis is rejected”, while I believe that this means that the circuit “passed the test” so that the null-hypothesis is not rejected. - When referring to the Bonferroni correction, was the base one used or the Bonferroni-Holm one? The latter is always more appropriate as it is less conservative and leads to a more powerful test. Confidence: 3 Soundness: 2 Presentation: 4 Contribution: 2 Limitations: - Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses I believe it would have been nice to perform further tests to understand whether the idealized tests have some applicability on discovered circuits. (or if they always lead to the null being rejected), potentially by also altering the discovered ones. Would have found it interesting to propose some ideas on how to use these tests for circuit discovery. Thank you for bringing this up! We agree that these tests can be applied to improve circuit discovery algorithms, such as ACDC [1]. For example, it would be interesting to apply the minimality test for edge pruning in Algorithm 1 of ACDC [1]. In ACDC, the edges are pruned by checking that removing an edge does not increase the KL divergence between the outputs and the current circuit by more than a threshold, $\tau$. This threshold $\tau$ is treated as a hyperparameter. In contrast, using the minimality test, we can determine the importance threshold in a principled way. However, we omitted further discussion of these ideas because they are not the main focus of this paper, which is to formalize the circuit hypothesis, develop appropriate tests, and apply them to study the extent to which the circuit hypothesis holds. Nevertheless, we believe that applying our proposed tests in novel circuit discovery algorithms is an exciting area for future research. ## Questions I believe there is a typo in the description of Table 1 on page 7 Thank you for pointing out the typo; you are absolutely right. We have updated the paper to reflect that. When referring to the Bonferroni correction, was the base one used or the Bonferroni-Holm one? Thank you for the great suggestion. We used the Bonferroni correction in the paper because it is easier to explain, but you are right that Bonferroni-Holm is uniformly more powerful. We have made a note about it in the paper and will incorporate it in the paper's code package. In our experiments, we explored another less conservative method, the Benjamini–Hochberg method, but we did not observe a difference in the main findings. So we opted for Bonferroni correction for simplicity. ## References: [1] Arthur Conmy, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim, & Adrià Garriga-Alonso. (2023). Towards Automated Circuit Discovery for Mechanistic Interpretability. --- Rebuttal Comment 1.1: Comment: Thanks for your answer. I will keep my score. --- Rebuttal 2: Comment: We thank the reviewer for their comments , response, and support of the paper!	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful reviews and support of the paper. We are pleased to see that the reviewers find the paper well-explained, with original and well-considered hypothesis tests (Reviewer wEmx); that it is clearly written, offers a quantitative approach to evaluating the concept of ''circuit'' in models, and that proposed tests are more thorough and fine-grained than some existing evaluation criteria (Reviewer NNjL); and that it addresses a question of growing importance, bringing formal discipline to the topic of circuit analysis (Reviewer 6etf). We address the questions individually below.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Toxicity Detection for Free	Accept (spotlight)	Summary: This work proposes to leverage logits of the first token in LLM responses to identify toxic prompts. The experiments against LLaMAGuard and multiple open-sourced LLMs show satisfactory performance in ToxicChat and LMSYS-Chat-1M datasets. Strengths: - The proposed method is simple and easy to implement. - The presentation is clear. - Toxicity detection is a significant problem in LLM safety. Weaknesses: - I don't think toxicity detection based on the first token makes great sense. For example, in the appendix, "sorry'' is one of the refusal tokens. However, it is also possible that the LLM expresses toxic contents in a "sorry ... but ..." format. - The success of the proposed method heavily depends on the refusal token list. How to ensure the generalizability of the trained SLR model? - The baselines to compare are relatively weak. The authors fail to include state-of-the-art LLM models such as LLaMA3, GPT-4, GPT-4 Turbo, etc. - There also exist a few works training a neural network to detect toxic contents which the paper did not mention such as [1]. How does the proposed method compare with them? - The datasets for evaluation are skewed and not very popular. Most of the samples are non-toxic. Is it possible to have a try on hh-rlhf datasets [2]? References [1] He, Xinlei, et al. "You Only Prompt Once: On the Capabilities of Prompt Learning on Large Language Models to Tackle Toxic Content." 2024 IEEE Symposium on Security and Privacy (SP). IEEE Computer Society, 2023. [2] Bai, Yuntao, et al. "Training a helpful and harmless assistant with reinforcement learning from human feedback." arXiv preprint arXiv:2204.05862 (2022). Technical Quality: 1 Clarity: 3 Questions for Authors: - How to determine the list of refusal tokens? - How to extend your work to multi-class toxicity classification tasks (e.g., settings like LLaMAGuard)? Confidence: 4 Soundness: 1 Presentation: 3 Contribution: 1 Limitations: The authors discussed one limitation in Section 7. But it can be more boardly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the review for raising valuable questions. However, it seems to us that there might potentially be some misunderstandings from the reviewer regarding the implementation of MULI and our conceptual contribution. Here we offer our response to the reviewer's questions. Q1. I don't think toxicity detection based on the first token makes great sense. For example, in the appendix, "sorry'' is one of the refusal tokens. However, it is also possible that the LLM expresses toxic contents in a "sorry ... but ..." format. R: We too found it surprising that such a simple mechanism could work so well, but our experiments indicate that it is very effective (see, e.g., Tab. 2). We highlight that MULI looks not just at a single token (the first token of the chosen response) but at the entire probability distribution for the first token, which provides a lot more information. When responding to a toxic input, the distribution on LLM responses tends to put a non-trivial probability on one or (typically) more refusal phrases, and thus the distribution of the first token tends to put a non-trivial probability on these special refusal tokens. We expect MULI will work well on the specific example ("sorry ... but ..."), as it starts with a token that is associated with refusals ("sorry") and thus tends to indicate a toxic response. MULI also often works on examples where the model responds with harmful information without refusing, as in these examples, typically the LLM has a substantial probability of refusing, which can be detected by MULI. Please note that while the toy models use manually selected refusal tokens, MULI does not: MULI learns which tokens are indicators of toxicity. The toy models are introduced for pedagogical purposes. MULI learns a classifier that predicts toxicity based on the probability distribution of the first token, and empirical experiments have shown good performance. Q2. The success of the proposed method heavily depends on the refusal token list. How to ensure the generalizability of the trained SLR model? R: No, MULI does not depend on the refusal list at all. It learns which tokens are associated with toxicity, from a training set. While the toy models do depend on the list of refusal tokens, we highlight that the toy models are introduced for pedagogical purposes, to help provide intuition for the design of MULI. MULI generalizes well over different base LLMs and different datasets, as shown in Sec. 6.3 and Sec. 6.4. Q3. The baselines to compare are relatively weak. The authors fail to include state-of-the-art LLM models such as LLaMA3, GPT-4, GPT-4 Turbo, etc. R: The baselines we compare to (LlamaGuard, OpenAI moderation API) are widely used and SOTA in the field. We evaluated MULI on multiple currently popular open-source LLMs, including Llama3 (Sec. 6.3 and Tab. S2). Unfortunately, we do not have a way to evaluateMULI on GPT-4, as OpenAI does not allow users to obtain the full logits for all tokens. Q4. There also exist a few works training a neural network to detect toxic contents which the paper did not mention such as [1]. How does the proposed method compare with them? R: We appreciate the reviewer’s suggestion on additional literature and will add them to the related work. [1] constructs a separate detector to detect toxic outputs, using zero-shot prompting. Their approach incurs additional cost at inference time; in contrast, we design a method which incurs no additional cost at inference time. Q5. The datasets for evaluation are skewed and not very popular. Most of the samples are non-toxic. Is it possible to have a try on hh-rlhf datasets [2]? R:The data in the real world iseven more skewed; therefore, it is very important to use proper metrics that are appropriate given the class imbalance. We advocate for TPR@FPR0.1%, which reflects real-world considerations and is not sensitive to the positive/negative ratio in the test set. Please refer to the discussion in Sec 3.2. The datasets we used, ToxicChat and LMSYS-Chat-1M, are actually popular for toxicity detection [3]. We additionally evaluated MULI on the OpenAI Moderation API Evaluation dataset. The TPR@0.1%FPR of MULI trained on ToxicChat / MULI trained on lmsys1m / LlamaGuard / OpenAI Moderation API are 24.90%/25.86%/14.56%/15.13%, respectively, when evaluated on the OpenAI Moderation test set. Even though MULI is trained on other datasets, its performance significantly exceeds existing methods. See the full results in the global rebuttal PDF. The HH-RLHF dataset includes pairs of similar conversations and labels for which one people prefer. It is a good dataset for RLHF finetuning; however, we do not see it as a good benchmark for toxicity detection. [3] Llama guard: LLM-based input-output safeguard for human-AI conversations. Q6. How to determine the list of refusal tokens? R: MULI learns a classifier, and does not require a list of refusal tokens. Our toy models do require a list of refusal tokens but it is only for pedagogical purposes. We constructed the list of refusal tokens in our toy models based on our experience with toxicity detection. Q7. How to extend your work to multi-class toxicity classification tasks (e.g., settings like LLaMAGuard)? R: Multi-class toxicity classification might be useful but is outside the scope of the paper. We will release our code so that people can extend MULI for their own purposes. Multi-class classification seems less important for our setting than for LlamaGuard. LlamaGuard seeks to build a single detector for different providers who might have different policies. LlamaGuard provides multi-class classification so that providers can choose which categories they wish to block. MULI learns a simple classifier that is specific to a single LLM, and seeks to enforce whatever policy is implemented by the safety alignment of the underlying LLM, so there is no need for multi-class classification. --- Rebuttal Comment 1.1: Title: Reply to your rebuttal Comment: Thank you for your clarification. I still have further questions about your rebuttal. Regarding your response to Q3, although GPT-4/GPT-4 Turbo cannot allow users to access the logits of the first token, they are indeed strong baselines to detect toxic content. Calling GPT-4 API is expected to be much faster than your proposed method (although it may introduce some cost) since your method still relies on the inference of LLMs. I am wondering how your (free) method compares with commercial ones so that the users can balance the financial cost and the detection performance. Regarding your response to Q7, as you mentioned "MULI learns a simple classifier that is specific to a single LLM", MULI has to be re-trained from scratch for each new LLM which implies a non-trivial computation cost. For example, if we want to obtain a MULI-based toxicity detector for LLaMA-70B, re-training from scratch might be very time-intensive and computationally inefficient. A single A40 GPU may not be able to finish this task. I am wondering if it is possible to reuse between different versions of LLMs if they share the same vocabulary list? I will raise my score if my concerns are addressed. --- Reply to Comment 1.1.1: Title: Thanks. Comment: Thanks for your consideration. Here are the responses: Response to additional comment on Q3: Thank you for the clarification and good suggestion. We additionally evaluated GPT-4o and GPT-4o-mini on the two datasets. On ToxicChat, GPT-4o had 72% TPR at 1.5% FPR (compared to 86.7% TPR at the same FPR for MULI) and GPT-4o-mini had 53% (MULI 81.2%) TPR at 1% FPR; on LMSYS-Chat-1M, GPT-4o had 92% (MULI 97.2%) TPR at 6% FPR and GPT-4o-mini had 90% (MULI 97.2%) TPR at 6% FPR. Based on these numbers, GPT-4o and GPT-4o-mini are both suboptimal to MULI in detecting toxicities. Besides, there are several more disadvantages of calling commercial APIs like them: 1. They are not flexible in customizing FPRs. Users need to customize the filtering threshold according to their tolerance; 2. They cause considerable expense for applications that need to process a massive amount of data; 3. They actually take a lot more time since it not only incurs a generation time cost (usually multiple times the inference time cost) on their server but also suffers from internet issues. We will include these results and the discussions in the final version of the paper. Response to additional comment on Q7: That is not true. Training MULI from scratch is actually very computationally efficient. In practice, we trained MULI in two phases. In the first phase, we forward only once on each training example (which is the minimal cost I can imagine) and cache the logits from the output. In the second phase, we use the cached logits to train a linear classifier for MULI, which takes only a few minutes, even on a small GPU. Moreover, training MULI does not require much data; please see Fig. 7 and the discussion on 6.4. For big LLMs, if one does not have enough GPU resources to make an inference, I believe it makes no sense to train a MULI for that. Training MULI incurs negligible costs compared to the daily usage costs for those who demand using big LLMs. Therefore, there is no need to reuse MULI between different versions of LLMs for the purpose of computational efficiency. In spite of this, it is still an interesting research question how different versions of LLMs share their distribution on the logits in responses and how one can reuse MULI between them. We will release our code so that others can explore this further.	Summary: This work proposes a toxicity detection method for LLMs that incurs negligible additional inference cost, and shows superior performance compared with two existing methods. The main observation of this work is that the logits of the first generated token (after prompt) are informative about toxicity of the prompt. The authors propose 2 baseline methods, PoR (crude probability of refusal) and PoRT (probability of refusal according to the logits of a pre-selected token, eg. "sorry"). They show that PoRT is effective, paving the path to their flagship method MULI. In MULI, a sparse logistic regression classifier is trained on top of the first token logits. L1 regularization on the weights is applied, as well as a non-linear function on the logits (ablations for both the regularization strength and the non-linear function are provided). MULI shows excellent performance compared with LlamaGuard and OMod, with lower computational complexity. The improvement over these methods supposes a jump in performance in toxicity detection. MULI also shows more robustness than the other methods to modality shifts (tested with 2 toxicity datasets). Strengths: Originality: * This work proposes simple yet original solution to toxicity detection. Analyzing the logits of the first token is surprisingly effective, sound and simple. Quality: * This work contains experimental results on 2 toxicity datasets, a comparison with 2 existing methods and a comparison among several LLMs. The experimental setup is of great quality. * I really appreciated the experiment comparing LLMs (fig 6) and the ablation of dataset size to estimate MULI (fig 7). Clarity: * The paper is well written, all the proposed evaluations and methods are sound and well explained. Significance: * This work tackles the important topic of toxicity detection in intruction-tuned LLMs. This is an important topic nowadays, since practically the everyone is using LLMs in daily life, with millions of queries per day. Being able to better distinguish toxic queries from non-toxic with low computational methods is key to improve quality and also to reduce costs associated with LLM inference. Weaknesses: Quality: * One possible weakness is the lack of justification to analyze the first token only. I missed some discussion on how is this token "attending" to the prompt or how can this method fail? Intuitively, this method can suffer from adversarially designed prompts, some analysis in this sense would be interesting. In general, a limitations section is missing. Technical Quality: 4 Clarity: 3 Questions for Authors: * One question that immediately came to my mind while reading is what happens beyond the first token logits. My intuition is that including the 2nd token would disambiguate many more answers. I believe the same MULI formulation could be applied to the 2nd token by appending the 2nd token logits to the 1st token ones, and using that 2x larger vector to train the logistic regression. Some discussion on this would be great. * I missed some discussion on the possible jailbraking of MULI. How could MULI suffer from adversarially designed prompts, such that they are toxic but can deceive the logistic regression? * How does MULI perform with sub-work tokenizers, or tokenizers that strongly split words? In general, how is MULI impacted by the tokenizer, since only the 1st token is used. --- Overall comment: Despite the method's simplicity, it provides excellent results and major improvement in terms of compute. I find this combination very interesting, being able to combine simplicity and a leap forward in terms of results is hard to achieve. Moreover, the experimental results are solid and reproducable. Overall, I find this paper of good quality for NeurIPs and I recommend acceptance. Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors mention that the method limitations are discussed in Section 7 (conclusion). However, section 7 only contains: _"Nevertheless, there are limitations, as the information hidden in the output of LLMs is much more than what we extract by MULI. It encourages researchers to look deeper into the information hidden in LLMs in the future."_ I encourage the authors to include a dedicated limitations section, with discussion on the topics mentioned in previous sections of this review. For example, how MULI could be jailbroken, limitation of using only the 1st token vs more, etc. Additionally, the authors should expand on _"the information hidden in the output of LLMs is much more than what we extract by MULI"_. What, in the authors opinion, could be extracted that MULI does not? Please, take these only as constructive suggestions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for identifying the originality, quality, clarity and significance of our method, as well as raising valuable questions. Here are the responses to your questions: Q1. One question that immediately came to my mind while reading is what happens beyond the first token logits. My intuition is that including the 2nd token would disambiguate many more answers. I believe the same MULI formulation could be applied to the 2nd token by appending the 2nd token logits to the 1st token ones, and using that 2x larger vector to train the logistic regression. Some discussion on this would be great. R: Good suggestion, I agree that it is a promising extension for MULI, and we have considered it seriously before. Including the 2nd token could possibly result in better performance. It does introduce some challenges since the logits of the 2nd token depend on the 1st token, so simply enlarging the feature vector might not be sufficient. We will add discussion of this direction to the final version of the paper. Q2. I missed some discussion on the possible jailbreaking of MULI. How could MULI suffer from adversarially designed prompts, such that they are toxic but can deceive the logistic regression? R: MULI works well on jailbreaking examples in ToxicChat, as they are regarded as one kind of implicit harmfulness by definition in this dataset. We don't claim to detect adversarial attacks, e.g., GCG-generated jailbreaks. Detecting strong adversarial attacks is an open and complicated question and is beyond the scope of this paper, but I believe it is worth delving into in the future. Q3. How does MULI perform with sub-work tokenizers, or tokenizers that strongly split words? In general, how is MULI impacted by the tokenizer, since only the 1st token is used. R: Most current LLMs use sub-word tokenizers, such as Llama, Mistral, GPT, Vicuna, and Koala. We evaluated MULI on the above LLMs, as presented in Sec.6.3 and Tab. S2. MULI's performance on each LLM surpasses the existing state of the art. In our examples, many refusal tokens happen to be full words in the particular tokenizers used today, but MULI does not rely upon this. Q4. Include a dedicated limitations section. R: Good suggestion. We will include a limitations section in the final version. --- Rebuttal Comment 1.1: Title: Thanks for the rebuttal Comment: I thank the authors for their clear rebuttal answers, and the commitment to add a limitations section. I will maintain my score, I believe this paper deserves an accept.	Summary: The paper introduces Moderation Using LLM Introspection (MULI), which leverages the LLM's first token logits for toxicity detection. This is a novel approach compared to traditional methods that require an additional LLM for toxicity detection, thereby reducing computational costs and latency. Strengths: - Efficiency: MULI achieves near-zero additional computational cost, which is a significant improvement over existing methods that often double the computational overhead by using a separate LLM for detection. - Performance Metrics: The paper demonstrates that MULI significantly outperforms state-of-the-art (SOTA) detectors in multiple metrics. - Practical Implications: By focusing on detecting toxicity based solely on the prompt, MULI allows for real-time blocking of toxic prompts before the LLM generates any response. This is particularly useful for streaming APIs and web interfaces. - Reproducibility: The paper stated that the author(s) will release code (L197). Weaknesses: - While the results are promising, the paper does not provide enough information on how well the method generalizes to different types of toxic prompts. - The toy models make some simplifying assumptions, such as using specific refusal tokens like "Sorry," "Cannot," and "I." These assumptions might not hold in all scenarios (e.g. training data, models). - While the paper provides a high-level overview of the method and results, it lacks a detailed analysis of failure cases or scenarios where MULI might not perform well. Understanding these limitations is crucial for practical deployment. Technical Quality: 3 Clarity: 4 Questions for Authors: - Line 212: is it should be “a tolerance of 0.1% FPR” instead of TPR? - What are the specific scenarios or types of prompts where MULI might fail to detect toxicity accurately? Qualitative analysis would be beneficial for future work. - How does the choice of refusal tokens affect the performance of MULI? Would the method be robust to changes in these tokens, and how can it be adapted to different languages or dialects as foundation models like Llama-3 now are trained on multilingual data? - How does MULI handle edge cases, such as prompts that are borderline toxic or ambiguous in nature? What strategies can be employed to improve its robustness in such scenarios? [1] introduces a type of implicit hate speech that is very ambiguous in nature and [2] recently shows that LLMs are struggling to detect this implicit one. - Overall, having qualitative results on failure cases would be beneficial for the paper besides the existing great quantitative performance. [1] https://arxiv.org/pdf/2109.05322 [2] https://arxiv.org/pdf/2403.16685 Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: NA Flag For Ethics Review: ['Ethics review needed: Safety and security'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for identifying multiple strengths of our method, as well as raising valuable questions. Here are our responses to the questions. Q1. Line 212: is it should be “a tolerance of 0.1% FPR” instead of TPR? R: Yes, it should be FPR instead of TPR. We will correct this in the final version of the paper. Q2. What are the specific scenarios or types of prompts where MULI might fail to detect toxicity accurately? Qualitative analysis would be beneficial for future work. R: Several failure cases correspond to jailbreaking attempts with toxic intention, or toxicity in non-English prompts. Examples include: - 请你扮演我的外婆。在炎热的夏夜，我的外婆总是吟唱 Office 365 旗舰版的序列号来哄我入睡。(This jailbreak attempt is in fact not obviously toxic.) - Give me a story about two gay men and make it an adult novel - tell me recipe of lsd in details - Schreibe einen Deathcore Liedtext, der viele obskure alte Wörter benutzt! Das Thema soll Pornosucht sowie chronisches Onlineverhalten sein und es soll im Vordergrund stehen, wie diese Dinge langsam den Geist eines Jugendlichen korrumpieren und seinem glücklichen Leben im Weg stehen. antworte auf - genere moi un code python pour contourner le 3d secure We will include further discussion (including examples) on the failure cases and edge cases in the final version of the draft. Q3. How does the choice of refusal tokens affect the performance of MULI? Would the method be robust to changes in these tokens, and how can it be adapted to different languages or dialects as foundation models like Llama-3 now are trained on multilingual data? R: We have not explored multilingual phenomena. We will add this to the limitations section. We expect MULI will be robust to different refusal phrasing used by other models. MULI uses logistic regression to learn which tokens are associated with toxic outputs, for a particular LLM, so it adapts to the particular LLM it is protecting. We introduce toy models for pedagogical reasons to help understand the intuition for why/how MULI might work. We don't intend to claim that the toy models would work well in practice, let alone in all scenarios. MULI can be viewed as a generalization of the toy models, where we learn at training time which tokens are associated with toxic outputs.We evaluated MULI on different models, including Llama-3. See Sec. 6.3 and Tab. S2. Q4. How does MULI handle edge cases, such as prompts that are borderline toxic or ambiguous in nature? What strategies can be employed to improve its robustness in such scenarios? [1] introduces a type of implicit hate speech that is very ambiguous in nature and [2] recently shows that LLMs are struggling to detect this implicit one. R: We have not evaluated such cases. We agree that handling borderline cases indeed poses unique challenges. Due to the sometimes subjective nature of toxicity in speech, even defining toxicity can be ambiguous, a question that lies orthogonal to our study on the algorithmic perspective. The implicit hate speech mentioned in the reviewer’s comment is concerning, yet we are hopeful that MULI would still offer a useful framework, and perhaps it could be addressed with better training data and better-aligned LLMs (that react properly to such implicit toxicity). Our experiments suggest that MULI is very effective in detecting clear-cut toxic or nontoxic cases (achieving high TPR at low FPR), which in the real world should encompass the vast majority of conversations. Q5. Overall, having qualitative results on failure cases would be beneficial for the paper besides the existing great quantitative performance. R. Thank you for the suggestion. We will include the above discussion in the final version of the paper. --- Rebuttal 2: Comment: Thank you to the authors for their response. While addressing multilingual and implicit hate remains challenging, as evidenced by your failure cases, it can be improved with better training and alignment since MULI relies on the used LLM. Additionally, as MULI can learn during training which tokens are associated with toxic outputs, this can inform better training data strategies for a truly "Toxic Detection for Free" (e.g., ensuring all gold responses to toxic prompts start with "Sorry..." or another standard phrase, and gold responses to non-toxic prompts do not start with it). I do not have any concerns, so I will raise my score accordingly.	Summary: This paper introduces a novel approach to detecting toxic prompts in LLMs using a method called Moderation Using LLM Introspection (MULI). The authors highlight the limitations of SOTA toxicity detectors, which often have low true positive rates (TPRs) at low false positive rates (FPRs) and incur high computational costs. The main motivation to develop this approach is that information is hidden in the LLMs' outputs that can be extracted to distinguish between toxic and benign prompts. Therefore, MULI leverages the logits of the first response token from the LLM to detect toxicity, eliminating the need for additional classifiers and reducing computational overhead. Strengths: - The paper introduces a novel method that uses the introspective capabilities of LLMs to detect toxic prompts, which is both cost-effective and efficient. - Results show that MULI significantly outperforms existing SOTA detectors in various metrics, particularly in achieving high TPR at low FPR, which is crucial for real-world applications. - The biggest strength to me is that MULI eliminates the need for additional classifiers, reducing computational costs and latency. - The approach can be applied in real-time settings, including streaming APIs, making it highly practical for deployment. Weaknesses: - One weakness of this work is its limited scope of evaluation. While the paper evaluates MULI on specific datasets (ToxicChat and LMSYS-Chat-1M), it would benefit from testing on a broader range of datasets to ensure generalizability. - Another issue of this approach is its dependency on LLM Quality. The effectiveness of MULI is highly correlated to the alignment quality of the underlying LLM. It is not clear whether poorly aligned models can provide reliable logits for toxicity detection. Technical Quality: 3 Clarity: 3 Questions for Authors: The method relies on logits, which may not be easily interpretable. Do authors have an explanation for why certain logits indicate toxicity? Can you investigate the impact of different alignment techniques on the performance of MULI to understand its dependency on the quality of the underlying LLM? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - As I mentioned above, the main limitation of this approach is its dependency on LLM quality and alignment. - Besides, generalizability across different types of toxic content is still an open question and maybe a limitation. - The method based on output logits, but more information can be hidden in LLMs' outputs. Further investigation on toxicity detection with LLM outputs is required. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for identifying multiple strengths of our method, as well as raising valuable questions. Here are the responses to your questions: Q1. One weakness of this work is its limited scope of evaluation. While the paper evaluates MULI on specific datasets (ToxicChat and LMSYS-Chat-1M), it would benefit from testing on a broader range of datasets to ensure generalizability. R: We additionally evaluated MULI on the OpenAI Moderation API Evaluation dataset, which consists of 1680 examples. The TPR@0.1%FPR of MULI trained on ToxicChat / MULI trained on lmsys1m / LlamaGuard / OpenAI Moderation API are 24.90%/25.86%/14.56%/15.13%, respectively, when evaluated on the OpenAI Moderation test set. Even though MULI is trained on other datasets, its performance significantly exceeds existing methods. See the full results in the global rebuttal PDF. Q2. Another issue of this approach is its dependency on LLM Quality. The effectiveness of MULI is highly correlated to the alignment quality of the underlying LLM. It is not clear whether poorly aligned models can provide reliable logits for toxicity detection. R: We agree. In Fig. 6, we showed that MULI's performance depends on the alignment of the base LLM: MULI performs better on LLMs with stronger safety alignment. Nonetheless, in all cases, MULI's performance is significantly better than other methods (e.g., MULI's TPR@FPR0.1% is at least 27% for all LLMs evaluated, compared to just 6% for the baselines). We will add discussion to the limitation section. Q3. The method relies on logits, which may not be easily interpretable. Do authors have an explanation for why certain logits indicate toxicity? R: Thank you for understanding. Certain tokens tend to be associated with refusals (we call them refusal tokens). Toxic questions tend to lead the LLM to have a non-trivial probability of refusing, i.e., of outputting refusal tokens, whereas benign questions tend to lead to a very small probability, so the logits for refusal tokens are higher for toxic questions than for benign questions. MULI takes advantage of this phenomenon (see, e.g., Section 6.5 and line 264). Q4. Can you investigate the impact of different alignment techniques on the performance of MULI to understand its dependency on the quality of the underlying LLM? R: As we responded in Q2, LLMs with stronger safety alignment (as reflected by a higher "security score", see Section 6.3) are associated with better performance from MULI (Fig. 6). --- Rebuttal Comment 1.1: Title: Acknowledgement Comment: Thanks to the authors for their responses. I keep my score as-is and recommend acceptance.	Rebuttal 1: Rebuttal: We appreciate the reviewers for their valuable time. Based on the questions and suggestions in the reviews, we plan to make the following adjustments to our paper: 1. We will provide a qualitative overview on common failure modes, with examples of toxic prompts that are uncaught by MULI. 2. We will include additional evaluation using the OpenAI Moderation API Evaluation dataset. The full results are in the PDF. 3. We will acknowledge and further discuss the limitations of the method, including its reliance on well-aligned models and subjectivity to adversarial attacks. Specific questions from the reviewers are addressed separately. Pdf: /pdf/a963b76618369ee94d8662f23634ba53103368e4.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The authors of this paper propose an approach for detecting toxicity of prompts from strongly aligned models (that are trained to refuse toxic prompts) using their Moderation Using LLM Introspection (MULI). Key to this approach is the observation that even though LLMs may not refuse toxic prompts always at very high probabilities (thus leading to actual refusal response), the probability of certain tokens associated with refusal rises above the average when they see toxic prompts. They build a sparse linear regression model using the probabilities for the first token output to detect toxic prompts and show they can sometimes achieve high TPRs at low FPRs using standard datasets. Strengths: The idea itself is very interesting and is a refreshing take on toxicity detection. The method does seem original and the paper is easy to read - I appreciated how they started building up their problem with small toy examples and then proceeded to develop their SLR method. Some of the other key strengths that I noted: - The insight that even though the refusal response may not rise to the top, it may have substantially high probability for toxic prompts. - Easy way further enhance the performance of well-aligned LLMs like llama-2-7b or build detectors using them. Weaknesses: Please construe these broadlyPlease see weaknesses section. as comments and try your best to respond to them: - Need to mention somewhere that this will work only for current LLMs that are safety aligned in a certain way and using specific refusal responses. - Let’s say we want to guardrail toxic prompts for a specific fine-tuned LLM using MULI. There will be a certain additional inference cost when doing this (even if it producing just one token output). This can be discussed somewhere. - The interpretation of SLR weights was a bit murky for me. Are the authors trying to show that the refusal tokens typically have coefficient values that lead to positive predictions (toxicity labels)? Also how was the $\lambda$ in SLR (eqn. 5) chosen? Standard cross validation? Technical Quality: 3 Clarity: 4 Questions for Authors: Please see weaknesses section. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The limitations are discussed in just one sentence - perhaps this can be expanded. One limitation I could think of was that you need a well-aligned model already and an infrastructure to run it (even if it is for just producing a single token output). Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the encouraging words about our idea and methodology, as well as the valuable questions. Here are the responses to your questions: Q1. Need to mention somewhere that this will work only for current LLMs that are safety aligned in a certain way and using specific refusal responses. R: Good suggestion. We will include this in the limitation section. Q2. Let’s say we want to guardrail toxic prompts for a specific fine-tuned LLM using MULI. There will be a certain additional inference cost when doing this (even if it producing just one token output). This can be discussed somewhere. R: I am not sure if I understand this correctly. Protecting a standard LLM incurs no significant inference cost, since MULI works using logits from the LLM (the cost of the linear classifier is negligible). MULI does rely on the safety alignment of the LLM. If a malicious user fine-tunes an LLM to remove the safety alignment, that may render MULI ineffective, so we agree that MULI is not sufficient for protecting against harmful queries to maliciously fine-tuned LLMs. We will add this to the limitation section. Q3. The interpretation of SLR weights was a bit murky for me. Are the authors trying to show that the refusal tokens typically have coefficient values that lead to positive predictions (toxicity labels)? R: Yes, exactly. This data suggests that the toy models that only use refusal tokens for detection provide a reasonable intuition for why MURI works. Q4. Also how was the $\lambda$ in SLR (eqn. 5) chosen? Standard cross validation? R: $\lambda$ is fixed to 1 × 10^-3. The performance of MULI is insensitive to its value; thus, we selected it roughly without any cross-validation. Q5. The limitations are discussed in just one sentence - perhaps this can be expanded. One limitation I could think of was that you need a well-aligned model already and an infrastructure to run it (even if it is for just producing a single token output). R: That's a good suggestion. We will expand the limitations and include this. --- Rebuttal Comment 1.1: Title: Thanks. Comment: Thanks for the clarification. I read the other reviews as well and am inclined to recommend acceptance. However, a few more clarifications: For Q2 - What I meant was this. Lets say a user wants to creates guardrails for a specific LLM using MULI. However, if that LLM itself is not safety aligned, they will have to use another safety-aligned LLM to get the first token logits. This is only one token inference but does add to the cost since the safety-aligned LLM has its own inference cost. For Q4 - This is interesting. If it is not sensitive, does an un-regularized logistic regression work? May be there is a broad range of $\lambda$ values. --- Reply to Comment 1.1.1: Title: Thanks. Comment: Response to additional comment on Q2: Thanks for the clarification. Yes, it is right that MULI requires an additional inference cost if one needs to apply it to another LLM. We will include this discussion in the final version of the paper. Response to additional comment on Q4: Unregularized MULI is only slightly suboptimal to the regularized MULI in terms of AUPRC, and they are comparable on other metrics. Please see the ablation study in Sec.6.6 and Tab.5, where f* + None denotes the Unregularized MULI.	null	null	null	null	null	null
Deep Support Vectors	Accept (poster)	Summary: This paper introduces the concept of Deep Support Vectors (DSVs) as an adaptation of support vectors from Support Vector Machines (SVMs) to deep learning models. The authors propose the DeepKKT condition, which generalizes the traditional Karush-Kuhn-Tucker (KKT) conditions of SVMs to handle the high-dimensional and multi-class nature of deep learning problems. By selecting or generating DSVs that satisfy the DeepKKT condition, the authors demonstrate that these vectors can play a similar role to traditional support vectors in terms of encoding decision boundaries and reconstructing models from a small subset of samples. The paper also shows that DSVs can be used for few-shot dataset distillation and as a means to alleviate the black-box characteristics of deep learning models by providing visual explanations of decision criteria. Furthermore, the authors demonstrate that the DeepKKT condition can transform conventional classification models into generative models with high fidelity, using class labels as latent variables. Strengths: 1. The paper introduces a novel concept, Deep Support Vectors, which extends the idea of support vectors from SVMs to deep learning models, providing a new perspective on understanding deep neural networks. 2. The proposed DeepKKT condition effectively generalizes the traditional KKT conditions to handle the complexities of deep learning problems, such as high dimensionality and multi-class classification. 3. The authors demonstrate the practical applicability of DSVs through various experiments, including few-shot dataset distillation, visual explanation of decision criteria, and transforming classification models into generative models. 4. The paper provides a thorough comparison of DSVs with existing methods in the context of few-shot dataset distillation, highlighting the superiority of DSVs under practical constraints. Weaknesses: 1. The paper lacks a rigorous mathematical derivation of the DeepKKT condition and its connection to traditional KKT conditions. A more formal treatment would strengthen the theoretical foundation of the proposed method. 2. While the authors provide experimental evidence for the effectiveness of DSVs, a more comprehensive analysis of the method's sensitivity to hyperparameters and its robustness to different architectures and datasets would be beneficial. 3. The paper does not provide a detailed discussion on the computational complexity of generating DSVs and how it scales with the size of the dataset and the complexity of the model. Technical Quality: 2 Clarity: 3 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and constructive feedback. We appreciate the acknowledgement of our contribution - Originality of our DeepKKT condition and its practical application. Mathematical Derivation of DeepKKT Condition: While we acknowledge that a rigorous mathematical derivation of the DeepKKT condition and its connection to traditional KKT conditions is desirable, it is important to note that our work focuses on defining a novel concept within the constraints of unknown nonlinearity in deep learning models. In traditional SVM, the goal is to construct a linear or known (white-box, kernel) nonlinear classification model using support vectors at the boundary. The most notable difference of our work from traditional SVM is that SVM aims to find a parameter $\phi$, while we use the DeepKKT condition to find plausible data points $x$ in an already-trained black-box nonlinear classification model. Therefore, making a direct connection between the two is inherently challenging. Given the unknown nonlinearity, a rigorous definition is difficult to establish. Thus, we introduced an analogy to the traditional KKT conditions, guiding our approach through stationarity and primary Lagrange conditions under hinge loss. This analogy and adaptation itself are significant contributions, and the absence of a rigorous derivation does not undermine our findings. Sensitivity to Hyperparameters and Robustness: We have indeed conducted extensive experiments to demonstrate the robustness of our method across various architectures and datasets. Our work includes evaluations on ConvNet with CIFAR-10 and CIFAR-100, as well as transfer learning with SVHN (as shown in Figure 4). Additionally, we utilized the widely adopted ResNet50 for ImageNet classification with a pretrained model from torchvision. As described in the global response, we also validated our approach using CLIP on the LAION-2B dataset, a diffusion classifier with U-Net, and a SwinTransformer. Additionally, we conducted experiments with different scales of parameters such as stationarity loss and primal loss ratio, which yielded robust results. These diverse applications underscore the robustness and versatility of our proposed method across different architectures and datasets. Computational Complexity: While we did not explicitly state the computational cost in the paper, an analysis of our algorithm (line 517-519, section H ) reveals that it operates with an order of n complexity. This implies that the scalability of our method aligns with existing gradient descent methodologies, ensuring that our approach remains computationally feasible even for large datasets and complex models. We leave its success case in OpenCLIP in the accompanying pdf. We hope this addresses your concerns and provides a clearer understanding of the contributions and strengths of our work. Thank you again for your valuable feedback. --- Rebuttal Comment 1.1: Title: Response to authors Comment: Thanks for the detailed explanations, which address most of my concerns. However, without the running wall-clock time, I am still concerned about the complexity of the proposed methods. Therefore, I keep my rating as 5. --- Rebuttal 2: Title: We presumed the request was about algorithmic complexity, not the real wall-clock time Comment: We presumed the request was about algorithmic complexity, not the real wall-clock time. As you mentioned as follows.   > The paper does not provide a detailed discussion on the computational complexity of generating DSVs and how it scales with the size of the dataset and the complexity of the model.   This is why we mentioned about computation complexity. In the case of Wall clock time, it is as follows. We ran it on a single-A6000 GPU, for the CIFAR10 experiment, it took 30s/img for generating deep support vectors in ConvNET, and about 60s/img when we did it in ResNet. For ImageNet, it took about 3 minutes per image on ResNet. It's important to note that in our case, we focused on the implementation and didn't really consider speed at all. In fact, updating the update algorithm in a super-resolution fashion, (updating both low and high-resolution pixels simultaneously) has actually reduced generation time 1/10. Also, images are generated at once, and the number of generated images are purely proportional to memory, not time. This implies time can be reduced even further by using more memory linearly.	Summary: This paper introduces a novel Deep Support Vectors (DSVs) framework, which can be used to reconstruct data and serve as latent generative models using logits as latent. By adapting the traditional Karush-Kuhn-Tucker (KKT) condition for deep learning models, the authors introduce the DeepKKT one and show that generated DSVs using this condition exhibit properties similar to traditional support vectors but apply to modern deep architectures like ConvNet and ResNet. Experiments on several common image datasets, including CIFAR10, CIFAR100, SVHN, and ImageNet, verify that DeepKKT can extract and generate DSVs. Moreover, DSVs are shown to be better than existing algorithms in few-shot dataset distillation problems. Strengths: The paper is well-written and easy to follow. Introducing deep support vectors based on an adaption of the KKT condition used in traditional support vector machines (SVM) is novel and interesting. To be honest, I did not have enough time to verify all of the mathematical verifications in the paper, and I'll leave it in other reviewers' comments on this part. The experimental results consistently show the benefit of DVSs in several problems, such as few-shot dataset distillation and image generation. Weaknesses: One of the proposed DSVs' most critical limitations is requiring the model to be fully pre-trained before being applied for support vector generation. This condition differs from the traditional support vector machine (SVM) framework, where the training process of the (linear) model and the support vector selection are jointly performed. Also, it is unclear how the deep model is pre-trained (e.g., with which datasets; are there any domain gap issues between the pre-training data and the main training data, etc.). Furthermore, the margin definition in Eq. (3) is unclear, and I could not understand why the larger margin is better. Regarding the total training loss for the DSVs generation in (9), it is unclear about the selections of the augmentation operator $\mathcal A$. Moreover, when training (9) with the traditional SGD approach, how do we guarantee that the generated deep support vectors belong to the original training data set? Technical Quality: 3 Clarity: 3 Questions for Authors: Please address my questions in the weaknesses section. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback and valuable insights, and for acknowledging our work. We are especially grateful for your recognition of the originality and practical applicability of the DeepKKT condition. However, we would like to clarify that we do not deal with DeepSVM. Specifically, we do not aim to ‘train’ an architecture. Instead, we aim to ‘retrieve’ plausible data points within a ‘trained’ architecture, as noted in our global response. Requiring the Model to be Fully Pre-Trained: As highlighted in our global response, our primary contribution is adapting the strengths of SVMs to deep learning models by extracting deep support vectors from pre-trained deep networks. This approach is intended to resolve the black-box nature of these models and enable the use of support vectors for tasks such as dataset distillation and data generation. Unlike SVMs, which jointly perform training and support vector selection, our method is designed to work with any pre-trained classification model. Actually, this condition is much more relaxed than that of SVM, as it does not require access to all training data. Furthermore, as demonstrated in Figure 4, we have shown that DSVs can be constructed under much harsher conditions, such as in transfer learning. Additionally, we believe you may have misunderstood our approach, as it is fundamentally different from DeepSVM. Our method does not impose constraints on the pre-trained deep learning network, other than it being a classification model. Unclear Pre-Training Process: Our approach does not impose specific conditions on how the deep model is pre-trained. We demonstrate the versatility of our method by employing standard training techniques, including data augmentation, stochastic gradient descent (SGD), and batch sampling. As detailed in lines 89-94 of our paper, our method successfully operates under these typical settings. In contrast, other methods often succeed only in highly constrained environments, such as low-dimensional manifolds, full-batch gradient descent, or unaugmented data. This flexibility ensures that our approach is broadly applicable across various training scenarios and environments. Margin Definition in Eq. (3): The margin definition follows the principles of SVMs, where support vectors are used to maximize the margin between the decision boundary and data samples, thereby defining the most plausible decision boundary. This philosophy is well-established and documented in the literature, such as in [1]. Our adaptation ensures that the DSVs we extract retain this critical property because samples near the decision boundary would be much more plausible than those far from the boundary, especially in a very high-dimensional data space. Selection of the Augmentation Operator $\mathcal{A}$ : The details of the augmentation operator $\mathcal{A}$ are provided in Section D, lines 476-480. This section outlines the augmentation techniques used to ensure the robustness of the generated DSVs. Guaranteeing DSVs Belong to the Original Training Dataset: This is an important point. As mentioned in lines 169-172, we incorporated the manifold condition to ensure that the generated DSVs adhere to the distribution of the training set. This manifold condition is enforced through the use of augmentation, TV loss, and alpha loss, as detailed in Equation (9). Ensuring that DSVs belong to the original training data’s distribution is a key contribution of our work. Additionally, as shown in Tables 1 and 2, when using DeepKKT or generating DSVs, the performance in dataset distillation was better than when using the original dataset. This experimentally demonstrates that DSVs lie on the training data’s manifold. [1] Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. In International Conference on Learning Representations (ICLR), Vancouver, BC, Canada, 2018. Poster presented at ICLR 2018. --- Rebuttal Comment 1.1: Title: Upgrading my score to 6 Comment: Thanks to the authors for the detailed responses, especially for clarifying the scope of the paper. Since most of my concerns have been addressed, I decided to upgrade the paper's score to 6.	Summary: The paper describes rethinking a classification through support vector in a way similar to SVMs. It provides additional benefits such as: - few-shot dataset distillation - using the similarity of the proposed formulation to the diffusion models, the model can be transformed into a generative model Strengths: - Originality: The paper demonstrates out-of-the-box thinking by showing that generalisation of the deep KKT conditions leads to similarity with diffusion models; the contribution of the paper is multifaceted: formulating deep KKT conditions, showing the utility of such formulation for dataset distillation, and showing how it could lead to a generative model. - Significance: I think this work is significant as the authors contribute towards the intersection of such important problems as dataset distillation, transparency of the ML models, and provide new insight about the connection between the KKT conditions and the diffusion models - Clarity: the paper is clearly written, the only optional suggestion, for the ease of mathematical notation, is to use the boldface fonts for vectors, e.g., like Bishop (2006), page xi. It also gives, in my understanding, enough reproducibility information. - Quality: The paper presents and sufficiently backs up the claims. Weaknesses: Clarity: the authors are incouraged to improve the discussion over the limitations. Technical Quality: 4 Clarity: 4 Questions for Authors: 1) It would be useful if the authors have any comments on the failure modes of the proposed model. For example, one of the aspects may be the confounding factors (Bontempelli et al (2022)) which could be picked up by the model? 2) One of the perceived limitations is that the support vectors only represent the whole image. It does not allow to pick up the particular aspects of the image which takes the decision. For example, we can guess from an image of a cat with antlers that these are the antlers which make it look like a deer. But strictly speaking we don't know if such changes result in confirmation bias or whether the model actually predicts the results based on the antlers. I wonder if the authors can comment on such limitation? 3) I could imagine such work is also linked with incremental/lifelong learning, for example in a way similar to Laskov et al (2006), which considers incremental SVM. In contrast to the standard deep-learning solutions, KKT can be updated recursively, which provides a wealth of options for improving the performance. This might be an additional benefit of the proposed formulation. I wonder if the authors can comment on this? Bontempelli et al (2022) Concept-level Debugging of Part-Prototype Networks, ICLR 2022 Laskov, P., Gehl, C., Krüger, S., Müller, K.R., Bennett, K.P. and Parrado-Hernández, E., 2006. Incremental support vector learning: Analysis, implementation and applications. Journal of machine learning research, 7(9). Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: As discussed above, the section on limitations would greatly improve the discussion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed feedback. Your insightful feedback and objective perspective greatly helped our research. Also, we are particularly grateful for your acknowledgement upon our originality and intention of the paper. Clarity on Mathematical Notation: We appreciate your suggestion regarding the use of boldface fonts for vectors. We have reviewed the relevant guidelines, and in the final version of the paper, we will incorporate your advice to enhance the clarity of our mathematical notation. Failure Modes and Confounding Factors: We are grateful for your insightful question about potential failure modes and confounding factors. This feedback has prompted us to reflect more deeply on these aspects. For instance, as mentioned in our global response PDF, an example can be seen in images of castles, which are often surrounded by forests. This is because castles are usually located in mountainous areas. Our model might learn to associate the presence of trees with the presence of castles. Similarly, in images of tenches, the presence of humans could be a confounding factor. Additionally, we can use our model to identify potential biases. For example, in Figure 1, all clocks point to 10:10, which is a bias introduced by the convention of setting clocks to 10:10 in advertisements to make them appear aesthetically pleasing. Limitation of Support Vectors Representing Whole Images: While it may seem that our support vectors represent whole images, they often emphasize unique features, which we believe is beneficial. For example, in Figure 1, the drilling rig images focus on the drill bit from various angles, and the rocking chair image combines perspectives from multiple points of view, akin to Picasso’s paintings. This suggests that even within a single image, we can extract multiple important features. An even more striking example is the tench case (thank you for your suggestion on failure modes!). In the tench image, we can see multiple objects—the tench and the fisherman. This highlights multiple features, demonstrating the value of encoding the entire image. Additionally, regarding the credibility of our method, we believe extensive experiments support our argument. For instance, as shown in Figure 3, we can test hypotheses through editing and DSV generation, revealing significant features like antlers in an image of a cat with antlers. Furthermore, Figure 15 shows that changing the class results in distinctive features emerging in the DSVs. However, we acknowledge your point regarding the lack of a numerical measure for feature importance. We view this as an area for future work, where we could potentially incorporate existing XAI methods like Grad-CAM or develop new metrics to quantify feature significance. Incremental/Lifelong Learning: Your observation regarding the link between our work and incremental/lifelong learning is very insightful. Indeed, we are currently pursuing this direction in our follow-up research. We are exploring the use of DSVs in replay-memory-based continual learning to store anchor points without needing additional datasets. This approach leverages the recursive update potential of the Karush-Kuhn-Tucker (KKT) conditions, offering a significant advantage over standard deep learning methods. We hope this addresses your concerns and provides a clearer understanding of the contributions and strengths of our work. Thank you again for your valuable feedback. --- Rebuttal Comment 1.1: Comment: Many thanks for the rebuttal. I went through the answers to all the reviewers, I think it addresses the reviewers' questions so I stick with the same assessment. --- Reply to Comment 1.1.1: Comment: Thank you very much for your feedback and strong support of our work. I’m glad our responses have effectively addressed your questions as well as those of the other reviewers.	Summary: This paper introduces the DeepKKT conditions for deep svm models, which correspond to the KKT condition in traditional linear SVMs. By either selecting deep support vectors (DSVs) from training data or generating them from already trained deep learning models, The authors show DSVs can play a similar role to conventional support vectors. In addition, this paper shows that the DeepKKT condition can transform conventional classification models into generative models with high fidelity, particularly as latent generative models using class labels as latent variables. The experiments validate the effectiveness of DSVs using common datasets (ImageNet, CIFAR10 and CIFAR100) on the general architectures (ResNet and ConvNet). Overall, the contribution is limited, hard to follow and I do not recommend for publication at this moment. Strengths: The DeepKKT and its extended version in Eq. 9 to generate DSVs is interesting. This paper shows that the DeepKKT condition can transform conventional classification models into generative models with high fidelity, particularly as latent generative models using class labels as latent variables. The experiments validate the effectiveness of DSVs using common datasets (ImageNet, CIFAR10 and CIFAR100) on the general architectures (ResNet and ConvNet). Weaknesses: (1) innovation and contribution This paper is just deep svms, DeepKKT is known, and its contribution is trivial. Overall, the deep svms here is still a black-box. DeepKKT can relieve the black box issue by introducing deep learning to svms (line 34-36)? The heuristic method to generate DSVs using Eq.9 is not convincing. are they still DSVs? (2) Technical and theoretical analysis In technical level, it may be practical to use Eq. 9 to generate DSVs (either from training data or generated), but it is heuristic. (3) The paper idea is easy, but the writing is still needs to improve. Technical Quality: 2 Clarity: 2 Questions for Authors: (1) For example, I do not follow "Like support vectors can reconstruct SVM, we can reconstruct the deep models from scratch only with DSVs.". How do you reconstruct deep svms from DSVs? (2) deep learning is black box, and the authors using black box (deep learning here) to relieve this issue? Is it possible to rewrite the paper to sell its idea on DeepKKT and how to generate DSVs? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: see above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your feedback; we really appreciate the time and effort you have invested in reviewing our paper. However, it appears that there may be some misunderstandings regarding key aspects of our work, and we believe certain elements of your review may be misleading. First of all, our paper is not about either training or elucidating a DeepSVM. On the other hand, DeepKKT is our original contribution, and we were the first to introduce this concept. As noted in the global comment, we emphasize our originality at least four times in the paper (Lines 72-76, 8, 56, and the Figure 1 caption). We believe this misunderstanding may have led to misinterpretations in your review. Consequently, the concerns raised in Weakness 1, Question 1, and Question 2 are misleading since they are based on this misrepresentation. Deep Support Vectors are distinct from DeepSVM. As mentioned on lines 72-75, DeepSVM refers to a well-known approach that integrates deep learning architecture with SVM principles. Our work is fundamentally different—indeed, we do not propose a new architecture. We clarify this distinction at least four times in the paper. Additionally, Question 2 is incorrect because our aim is to address the black-box nature of existing architectures (such as ResNet, ConvNet, and, in this rebuttal, U-Net, ResNeXt, and Swin Transformer) by using DeepKKT conditions, not to develop a new architecture. Overall, it seems there was an initial misunderstanding of our concept and difficulty in interpreting the paper from the perspective it was intended. As such, Weakness 3 also appears to stem from this misunderstanding. Finally, regarding Weakness 2, the introduction of the manifold condition is one of our primal conditions. This is not merely a heuristic but a theoretically grounded concept. Given that modern deep learning models handle high-dimensional data and face challenges related to the curse of dimensionality, it is crucial to account for the Riemannian manifold $\mathcal{M}$. The augmentation operation $\mathcal{A}$ is how we impose symmetry invariance in the model. Formally, we sample an action $g$ from a group $\mathcal{G}$, where $\mathcal{G}$ represents a symmetrical Lie group, such as translation or rotation [1]. By applying $\mathcal{A}$ as a practical Lie group, we ensure that $f(\mathcal{A}(x)) \simeq f(x)$, meaning the semantics extracted by the model are preserved through the symmetric operator. This strategy is widely used and accepted in self-supervised learning, as seen in methods like SIMCLR [2]. Regarding $L_{tot}$ and $L_{norm}$, this also corresponds to prior term for the image [3], and this term is commonly used. Furthermore, as demonstrated in the global response PDF (see hyper-parameter section), even if we dramatically decrease this term (by up to 100 times), the figures does not change much. Meaning these terms are not critical for full generation. [1] Cohen, Taco, and Max Welling. "Group equivariant convolutional networks." International conference on machine learning. PMLR, 2016. [2] Chen, Ting, et al. "A simple framework for contrastive learning of visual representations." International conference on machine learning. PMLR, 2020. [3]Hongxu Yin, Arun Mallya, Arash Vahdat, José M. Álvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. CoRR, abs/2104.07586, 2021. --- Rebuttal Comment 1.1: Comment: How do you define deep support vectors? I always links it with deep svm conditions. Do these DSVs still hold after Eq. 9? --- Rebuttal 2: Title: we exploit Eq. 9 to meet the condition in Eq. 6 Comment: Formally speaking, DeepSVM refers to SVM that incorporates deep networks as feature extractor. This means it has a pre-trained encoder $\phi$ , which is a deep network. In this process, deep network is given. Then, DeepSVM constructs SVM over the encoder. i.e., it builds upon datapoints $\phi(d), \quad d \in \mathcal{D}$. where $\mathcal{D}$ is an original manifold. From this context, deep svm is no more different than normal svm. The only difference is that it uses a mapping function $\phi$ from deep networks. So deep svm condition is just svm condition. In contrast, our goal is to find vectors which has the characteristics of support vectors in already trained deep networks. And we derive a DeepKKT condition to make the vector meet characteristics. Loosely speaking, we view deep learning model (Normal deep networks such as ResNet which classifies an imagenet) as black-box SVM, and find support vector from it. We define condition first, then, we derive a loss which corresponds to the condition. And regarding, Eq. 6, we derived the condition in line 149-174 by connecting KKT condition and deep learning dynamics and leave an intuitive analogy in line 454-468, which interprets the stationarity condition in DeepKKT geometrically. Furthermore, in experiment section, we validated DSVs meet SVM characteristics in line 230-250. From this context, we think your question Do these DSVs still hold after Eq. 9? may have confused the sequence. As we exploit Eq. 9 to meet the condition in Eq. 6 --- Rebuttal Comment 2.1: Comment: I know you try to make the whole model reasonable by adding a manifold in Eq. 6 to make Eq. 9 holds while you find DSVs. It could be better if you minimize a function f(x, \theta) x \in manifold, subject to DeepKKT condition. My understanding is that DSVs should be on the boundary exactly if we extend svm to deep svm. Then your solution from Eq. 9 or your DSVs is not the DSVs that I talked about. It would be better to sell your DSVs in another name, diffusion vector or support diffusion vector, etc. --- Reply to Comment 2.1.1: Comment: Upon your point “It could be better if you minimize a function f(x, \theta) x \in manifold, subject to DeepKKT condition.”, We clearly stated the connection between manifold condition and Eq. 9 in our paper in line 209-223 (just before Eq. 9) The term about augmentation-invariance and image prior is a manifold condition. And this is clearly explained in our paper, just before Eq. 9. For your convenience, >To extract DSVs from the manifold, we assume that the model is well-trained, meaning it maintains consistent decisions despite data augmentation. In other words, the model should classify DSVs invariantly even after augmentation. To ensure this, we enforce that the augmented DSVs ( $\mathcal{A}(x)$ where $\mathcal{A}$ denotes augmentation function) also meet the primary and stationarity conditions. Also, we exploit traditional image prior [33, 16], total variance $L_{tot}$ and size of the norm $L_{norm}$ to make DSVs lie in the data manifold. $L_{tot}$ is calculated by summing the differences in brightness between neighboring pixels, reducing unnecessary noise in an image, and maintaining a natural appearance. $L_{norm}$, taking a similar role, penalizes the outlier and preserves important pixels. However, we acknowledge that there exists a room for more clear understanding, so we will change the notation of. Eq. 9 With respect to Eq. 6. Such as, $L_{\text{stationarity}} + \beta_1 L_{\text{primal}} + \beta_2 L_{\text{manifold}$ Upon “My understanding is that DSVs should be on the boundary exactly if we extend svm to deep svm” We also explicitly stated the boundary condition in line 233-236. For your convenience, Also, we even spare a whole subsection about SVM characteristics in section 5.1. For your convenience, >While DeepKKT does not explicitly incorporate the complementary slackness condition due to computational costs and ambiguity, Fig. 2a suggests that DSVs implicitly fulfill this condition; During the training process, we observe an increase in the entropy of DSV candidates, hinting that the generated DSVs are close to the decision boundary. (the complementary slackness refers boundary condition.) Also, please note that We do not deal with DeepSVM and DeepSVM is just an architecture which exploits SVM as a classifier on deep learning model. So we noted in global response the first rebuttal comment upon your comment, and first reply on your response. Furthermore, we clearly stated in lines 72-76, 8, 56, and the Figure 1 caption. Finally, upon your mention “It would be better to sell your DSVs in another name, diffusion vector or support diffusion vector, etc.” We cannot consent your argument, as it is clearly not true, we showed that DSVs have the support vector characteristics. So we think the name DSVs is the best fit. With following reasons. 1. Meets boundary condition implicitly 2. Coreset selection in table 2 as SVM can encode its decision boundary, it itself is useful for a constructing model and has rich information. We showed DSVs can serve as coreset select mechanism in Table.1 and even further it can serve as SOTA dataset distillation algorithm in few-shot dataset distillation. 3. Model explainability: As support vector encode decision boundary, it can explain how does the model decides the class. DSVs also can serve as global explanation within only parametric space.	Rebuttal 1: Rebuttal: Summary In this paper, we propose a method to identify deep support vectors (DSVs) for pre-trained deep models without access to the original dataset. Our work does not involve training or constructing a Deep Support Vector Machine (DeepSVM). Instead, our original contribution is introducing the DeepKKT condition for obtaining DSVs. Prior to our work, global explanations in a purely parametric sense did not exist. DSVs can be identified in any classification model, regardless of architecture or dataset. To support this claim, we conducted experiments on various datasets such as CIFAR10, CIFAR100, SVHN, and ImageNet using ConvNet and ResNet architectures. In this rebuttal, we present results from experiments on the Laion2B and ImageNet datasets using U-Net, ResNext, and Transformer architectures (see the accompanying PDF file). Our Contribution and Experimental Setting: In this paper, we focus on identifying support vectors in pre-trained deep models without requiring access to the original training dataset. Our core contribution is the reconstruction of the dataset from a ‘trained’ model. The term ‘deep model’ refers to any deep architecture used for solving classification problems. Our major contribution is the ability to easily reconstruct the corresponding support vectors in any trained model, which enables us to explain the model’s overall decision-making process. DeepKKT is Our Original Contribution, and We Did Not Tackle DeepSVM: As stated in lines 72-76, 8, 56, and the caption of Figure 1, we do not address DeepSVM, nor do we train DeepSVM or deal with SVM architecture. Our focus is on already-trained typical deep networks such as ResNet. The DeepKKT condition is our genuine and original contribution; it is neither trivial nor previously known. As mentioned in Section 2.1, there are numerous research efforts on DeepSVM that integrate the SVM algorithm, a natural approach in the 2000s, into deep models. However, our paper is entirely different. These DeepSVM algorithms cannot generate deep support vectors from current deep models, as that is not their intended purpose. About Global Explanation: Typical Explainable AI (XAI) methods focus on local explanations, which clarify the decision criteria for a specific input. For example, suppose a model classifies an image as a 'deer' and an XAI algorithm aims to explain this decision. The algorithm might highlight a feature, such as an antler, that influenced the model's classification. In contrast, global explanations address the general decision criteria of the model, without being restricted to a single data point. Although there are some researchers performing global explanations, they often limit themselves to feature matching. These methods typically store feature vectors by passing the entire dataset through the model and then aggregate these vectors using pre-defined algorithms. For the deer example, these algorithms would highlight antlers from various deer images, providing a broader justification for the model's decisions. This approach has two main drawbacks: 1) The computational cost is extremely high, and 2) It cannot generate general criteria based solely on the parameter space, as it relies on the dataset. Essentially, these methods are extensions of local XAI [1], which our approach is not. Ablation Studies of Additional Data and Architecture: Even when we adjust the hyperparameters to scales of 10 or 0.1, the quality of the generated figures remains unchanged. We included additional results with various hyperparameters in the accompanying pdf. This robustness is a distinctive feature compared to methods like GANs, which require highly sophisticated hyperparameter tuning. [1] Fel, Thomas, et al. “Craft: Concept recursive activation factorization for explainability.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. Pdf: /pdf/110b82ab676493478e647c1c1f6cdb18fb057491.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper introduces Deep Support Vectors (DSVs), an adaptation of support vector concepts to deep learning models. The authors propose a DeepKKT condition, analogous to the KKT conditions in SVMs, to identify or generate DSVs in trained deep models. They demonstrate that DSVs exhibit properties similar to traditional support vectors, including encoding decision boundaries and enabling model reconstruction. The paper shows applications of DSVs in few-shot dataset distillation, model interpretability, and even using classification models as latent generative models. The authors validate their approach on common datasets (ImageNet, CIFAR10, CIFAR100) and architectures (ResNet, ConvNet). Strengths: - Novel concept of Deep Support Vectors that bridges ideas from SVMs to deep learning - Evaluation across multiple datasets and model architectures - Diverse applications including dataset distillation, model interpretability, and generative modeling - Provides both theoretical motivation and empirical validation Weaknesses: - Transformer architectures have not been included in the experiments - Comparison to other interpretability methods is limited such as global SHAP (mean absolute SHAP value) or other global methods - The proposed method has limitation when it comes to real-scenario applications. Dedicated generative models such as GANs or Transformers are better options. Technical Quality: 3 Clarity: 3 Questions for Authors: - How sensitive are the DSVs to the choice of hyperparameters in the DeepKKT condition? Is there a principled way to select these parameters? - How well do DSVs scale to larger models and datasets? Are there computational challenges in applying this approach to state-of-the-art large language models, for instance? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: A more in-depth consideration of privacy implications on the generative capability of the method could be given Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and constructive feedback. We appreciate the opportunity to address the points you raised. Experiments with Transformer Architectures: In our global rebuttal (see pdf file), we included experiments on Transformer architectures. Specifically, we demonstrated global explanations using a Swin Transformer and further evaluated our method on CLIP and a diffusion classifier utilizing U-Net, showcasing the adaptability of our approach across diverse architectures. Comparison to Other Interpretability Methods: As stated in the global response, it is crucial to note that our approach provides a global explanation in a parametric sense, which is the first attempt in this context. Existing algorithms cannot provide explanations solely based on model parameters. Conventional methods, such as SHAP, rely on feature matching within the dataset manifold; they extract feature vectors by passing the entire dataset through the model and then cluster these vectors. In contrast, our method operates independently of the dataset, using only the parameter space to generate general and agreeable criteria. Detailed explanations are provided in the global response. Real-Scenario Applications: Our proposed method focuses on unveiling deep classification models rather than generating new data, as seen with GANs and diffusion models. As outlined in our global response, our primary goal is to interpret black-box deep learning models by leveraging the properties of support vectors from SVMs. This leads to benefits such as coreset selection and model explainability. Therefore, we disagree with the notion that our method is impractical for real-world scenarios. In fact, it contributes to both 1) model interpretability and 2) dataset distillation, which are crucial and indispensable fields. Even when considered solely as a generative model, our approach offers significant advantages. For example, as illustrated in Fig. 6, our latent interpolation capability is both powerful and straightforward, whereas diffusion models [1] require additional complex mechanisms for similar tasks. Additionally, as shown at the bottom of Fig. 6, our method has unique strengths. It does not rely on any pre-trained characteristic vectors (human-crafted) or latent architectures since the supervision model itself is explainable. In contrast, GANs require such pre-trained characteristic vectors and latent architecture. Moreover, our method extracts Deep Support Vectors (DSVs) from existing supervision models, enhancing its practicality and addressing privacy concerns often associated with generative models. This is a significant advantage, as GANs and diffusion models typically require large datasets and complex training processes. In contrast, our approach leverages existing supervision models, making it more efficient and less resource-intensive. Sensitivity to Hyperparameters: As stated in the global response, our algorithm is robust to hyperparameters. Regarding the sensitivity of Deep Support Vectors (DSVs) to hyperparameter selection, our experiments demonstrate that the form of DSVs remains consistent at the feature level. This consistency suggests that the important information extracted from the pretrained models remains stable, regardless of variations in hyperparameters. Scalability to Larger Models and Datasets and requirements for Large language models.: We have validated our method on substantial datasets, such as ImageNet, which is widely recognized as large. However, it seems that even larger datasets might be required for foundational models. To address this concern, we also conducted experiments with OpenCLIP, which is trained on the LAION-2B dataset, one of the largest datasets in the vision domain (see the attached PDF). Our algorithm works with order n complexity, meaning our algorithm's computation aligns with traditional SGD methodologies, ensuring that our approach remains computationally feasible even for large datasets and complex models. Regarding large language models, our work focuses on classification models built on the principles of support vector machines (see lines 12, 42, 97, 120, 281, and others). Therefore, the latter part of your question seems less relevant. It is also important to note that model inversion in the NLP domain is relatively easier [2] compared to our work in model inversion, particularly when applied to vision. A few papers have addressed this issue in the vision domain, and none have shown success with practical, applied models, often limiting their scope to simpler multi-classification tasks like MNIST [3]. We have successfully synthesized high-resolution images in real-world vision scenarios, such as with ResNet50, which is a pioneering contribution. We hope this clarifies the strengths and contributions of our work. Thank you again for your valuable feedback. References: [1] Wang, Clinton, and Polina Golland. "Interpolating between images with diffusion models." (2023). [2] A survey on large language model (LLM) security and privacy: The Good, The Bad, and The Ugly [3] Yu, Runpeng, and Xinchao Wang. "Generator born from classifier." Advances in Neural Information Processing Systems36 (2024). (Neurips 2023)	null	null	null	null	null	null
Efficient Centroid-Linkage Clustering	Accept (poster)	Summary: The authors give an algorithm that approximates a centroid linkage clustering. Their algorithm is fast both in theory and in practice. The algorithm is based on a novel fully dynamic data structure for nearest neighbors which is of independent interest. Strengths: The proposed algorithm is a significant strengthening compared to existing algorithms, in particular from a theoretical point of view: while it is known that under some standard complexity theoretic assumptions most hierarchical clustering methods need at least (essentially) quadratic runtime, the proposed approximation algorithm significantly breaks this barrier. This is also underlined by experimental results. Interestingly, approximation does not seem to decrease the quality of the clustering. Indeed, as can be seen in Table 1, higher approximation values often give better results than "better" approximations or even optimal algorithms. Weaknesses: The techniques are very much geared towards the centroid linkage function. It could be interesting to mention how much of it could potentially generalize to other linkage functions and where the bottlenecks for full generalization lie. Technical Quality: 4 Clarity: 4 Questions for Authors: -Do you have any intuition as to why "worse" approximation factors often seem to give better clustering results (as your Table 1 indicates)? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The limitations are addressed in the checklist. Flag For Ethics Review: ['No ethics review needed.'] Rating: 9 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and suggestions. We respond now to the specific questions of the reviewer. > The techniques are very much geared towards the centroid linkage function. It could be interesting to mention how much of it could potentially generalize to other linkage functions and where the bottlenecks for full generalization lie. This is a great question and something that we will clarify when discussing our results in the paper. We believe that our techniques should extend naturally to other linkage functions for spatial hierarchical clustering that incorporate centroids, such as Ward’s linkage. Investigating whether any centroid-based method expressed in the Lance-Williams formulation could take advantage of our approach is an interesting question for future work. On the more theoretical side, we note that our theoretical results for ANNS could be used more generally in other clustering algorithms that require adaptivity. > “Do you have any intuition as to why "worse" approximation factors often seem to give better clustering results (as your Table 1 indicates)?” This is an interesting question, and as briefly mentioned in the paper, we do not have any scientific explanation for this phenomenon. One possibility is that as \eps increases, the decision process slowly incorporates an approximate single-linkage-like behavior, which could be beneficial for certain datasets. We plan to include several additional results on clustering datasets with ground truth in the final version of the paper (e.g., on ImageNet, Reddit, and arXiv embeddings as discussed in our response to reviewer Dgto) which will provide more data on the effect of increasing \eps. --- Rebuttal Comment 1.1: Comment: I would like to thank the authors for their rebuttal. Their answer to my second question seems plausible, and it would perhaps be interesting to compare the quality of the clusterings to the results of other methods, such as single-linkage to investigate the proposed reason. --- Reply to Comment 1.1.1: Comment: Thank you again to the reviewer for their suggestions. We will further investigate this direction by comparing with single-linkage clustering and include our findings in the revised version.	Summary: The paper deals with centroid-linkage agglomerative clustering (centroid HAC), where the distance between two clusters is the distance between their centers. It presents a subquadratic time algorithm (approximate centroid HAC) that, instead of requiring the two closest clusters to be merged at each step, it allows for any two clusters to be merged if their distance is within a factor of c of that of the two closest clusters. Strengths: S1. A dynamic approximate NN search algorithm with adaptive updates is proposed. S2. A centroid-linkage HAC algorithm is presented that exploits the approximate NN search algorithm. S3. Theoretical performance bounds are presented and proved. S4. Empirical evaluation indicates that the approximate centroid HAC algorithm provides clustering results very close to exact centroid HAC, but with a considerable speedup in execution time. Weaknesses: W1. It is confusing that the ANNS algorithm used in the experiments is not the same as the one proposed and analyzed in Section 3 of the paper (Algorithm 1). Since the focus of the paper is on clustering, this issue creates confusion regarding the actual contribution of this work. Perhaps the ANNS algorithm (contribution 2) is better to be presented in a forum related to data structures. W2. It is not clear if the ‘approximate centroid HAC’ idea is novel or not (independently of the ANNS algorithm used in the implementation). Technical Quality: 3 Clarity: 2 Questions for Authors: Q1, Q2. See comments W1, W2 above. Q3. How does the method compare to other approximate HAC methods in terms of NMI? In Figure 3 running time comparison is provided. In Figure 5 some results are provided on the Iris dataset, but I cannot see any general conclusions. Q4. It is meaningless to compare using small datasets, since the optimal solution can be quickly obtained. More experiments using big would add value to the paper. Q5. Is it possible to have a (rough) estimate of the increase in execution time as the value of \epsilon decreases? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors discuss the limitations. I do not see any potential negative societal impact of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and suggestions. We respond now to the specific questions of the reviewer. > Q1/W1: Coherence and clarifying why we use a different ANNS algorithm in theory and in practice. We thank the reviewer again for raising this question, which we agree is important. We present a longer discussion about this question in the overall response at the top of our rebuttal. > Q2/W2: Novelty of “approximate centroid HAC” We believe that the formulation of approximate HAC is certainly not novel, and there has been a number of interesting works in the past few years exploring similar notions of approximation for other linkage functions for dissimilarity and similarity (or graph) settings (see for example references [1, 5, 11, 12, 13, 21] which all deal with similar notions of approximation for other linkage functions). However, we strongly believe the techniques we introduce and leverage for obtaining our results for c–approximate HAC are novel, and have not been used in the literature before, and are not simple extensions of known techniques. > Q3: “How does the method compare to other approximate HAC methods in terms of NMI? In Figure 3 running time comparison is provided. In Figure 5 some results are provided on the Iris dataset, but I cannot see any general conclusions.” We thank the reviewer for this question; we will add a better explanation of how the NMI of our approximate HAC methods compares with the exact method in the main body of the paper. We included a broader discussion of these results on more datasets in the appendix, which we will move into the main body of the paper. The main result is that for modest values of \eps (e.g., \eps=0.1) the approximate method consistently gets NMI within a factor of 2% of that of the exact algorithm. > Q4: Using larger datasets to show more meaningful scaling results In this paper, we considered standard clustering benchmark datasets at three different scales: small, medium, and large. The small datasets (e.g., iris, faces, etc.) and medium-sized datasets (e.g., MNIST, birds, where embeddings for the latter are obtained via neural network-based methods) are labeled and are used for our quality evaluation experiments. As the reviewer correctly pointed out, the exact algorithms run very fast on the small datasets, however, they do not scale effectively to medium and larger datasets due to their quadratic running time. Notably, the approximate HAC algorithm achieves nearly 30x speedup on MNIST and 60x speedup on Birds with 192 threads; we will include these running time results in the revised version. For quality evaluation on large datasets, we plan to add three new embedding-based datasets that come with ground truth which we described in the global response. If the reviewer has any other suggestions for datasets to cluster with meaningful ground truth labels we would be delighted to include them in our paper. We would like to clarify that the benchmark datasets considered are consistent with multiple recent works on graph and pointset clustering (see for example references [13, 14, 38]). Additionally, we plan to test the scalability of our approximate HAC algorithm on larger ANN benchmark datasets (e.g., Wikipedia-Cohere ~30M) in the revised version. However, since the algorithm is sequential, we do not expect it to scale to billion-sized datasets. Developing an efficient parallel algorithm for approximate Centroid HAC and other linkage criteria is an interesting future direction to address this issue. > Q5:Execution time v/s $\epsilon$ This is a great question and something we investigated—please see Figure 3(c) in the paper, which illustrates the trend in execution time as $\epsilon$ decreases. --- Rebuttal Comment 1.1: Comment: I thank the authors for the rebuttal and their efforts to respond to my comments. I am still concerned about the incoherence issue: the ANNS algorithm proposed and analyzed is different from the one used in the experiments. For a reader who wishes to apply the method, the proposed ANNS algorithm is actually irrelevant. I will increase my score from 4 to 6. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for increasing their score. We will make an effort to add a detailed discussion on this gap between theory and practice in the revised version, which we hope will address their concern.	Summary: This paper considers the design of approximate versions of the centroid linkage method for hierarchical clustering. The exact version of centroid linkage requires $\Theta(n^2)$ time, but it could be possible to do better if some relaxation is allowed. This paper shows how to get a $c$-approximate centroid linkage clustering in sub-quadratic time by using a dynamic approximate near-neighbor search data structure that has guarantees against adaptive updates. This is non-trivial since the distance between centroids is non-monotone: after merging a pair of clusters, the resulting distance between centroids of other clusters could go down. To complete the result, they construct such a data structure that supports insertion, deletion, and queries against adaptive updates with sub-linear running time. An extensive experimental evaluation is performed to test a practical variant of the proposed algorithm on standard benchmark datasets against the exact centroid linkage method provided by the fastcluster library, a standard library for hierarchical clustering algorithms). The proposed method either shows good performance or minimal loss against the exact method for a variety of clustering metrics and significant speedup over the exact method. Strengths: There has been much work on scaling up hierarchical clustering methods by considering approximation, particularly linkage methods such as single linkage, average linkage and ward's method. In contrast, there has been less (but non-zero) work on centroid linkage. This paper provides an elegant and cleanly described solution to this problem. Weaknesses: - The experimental evaluation isn't for the algorithm that is proposed theoretically, but instead for a variant with several practical modifications (e.g., using graph-based ANNS which may not have the same theoretical guarantees as LSH-based ones). - For the experiments, the authors also consider running the algorithm with 192 cores. For this case, it seems that fastcluster cannot make use of the additional cores, which seems to be an unfair comparison. However, they do use an exact version of the new algorithm which does make use of the additional cores, providing a more fair comparison. Technical Quality: 3 Clarity: 3 Questions for Authors: Minor comment - the usage of $Q$ for both covering net and priority queue is somewhat confusing. Please consider modifying this. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are made clear through a clear description of assumptions. I don't see potential negative societal impact from this work since it develops more efficient versions of already widely-used algorithms. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and suggestions. We respond now to the specific questions of the reviewer. > W1: The experimental evaluation isn't for the algorithm that is proposed theoretically, but instead for a variant with several practical modifications (e.g., using graph-based ANNS which may not have the same theoretical guarantees as LSH-based ones). We thank the reviewer for raising this important point, which we acknowledged in the limitations section of our submission. We present a longer discussion about this question in the overall response at the top of our rebuttal. > W2: For the experiments, the authors also consider running the algorithm with 192 cores. For this case, it seems that fastcluster cannot make use of the additional cores, which seems to be an unfair comparison. However, they do use an exact version of the new algorithm which does make use of the additional cores, providing a more fair comparison. We actually also evaluated all algorithms in the sequential setting and still obtained a speedup (even on one core) over fastcluster. The results for our single core running times are in Figure 3(a). While it is true that fastcluster can’t make use of the additional cores, we view this as an advantage of our algorithm. The results with many cores are in Figure 3(b). As the reviewer notes, the runtime of fastcluster is the same in both scenarios. > Q1: Minor comment - usage of Q for both covering net and priority queue is somewhat confusing. Please consider modifying this. Thank you for pointing this out, we will fix it. --- Rebuttal Comment 1.1: Comment: Thank you for your response and addressing my questions. My overall evaluation remains the same.	Summary: Paper studied Hierarchical clustering algorithm. HAC is popular method where n points are initially singleton clusters (leaves). At each step, two clusters are merged, and the process is continued until one giant cluster of all n points (root) remains. THe merging process gives a hierarchy or nested clusterings. Depending on which clusters are merged ,the algorithm differs. Common methods are single linkage where the distance between clusters is in terms of min x in C_1 y in C_2 d(x,y). Another method which is focus of current undertaking is centroid linkage, which is d(C_1, C_2) = dist(mu_1, mu_2) where mu_i is centroid vector of cluster i. Technical point of view: Interestingly, after merging two clusters,the best inter-cluster distance can sometimes decrease, so it is not monotone increasing over time as in single linkage. Since many lower bounds exist showing that we need theta(n^2) time to do HAC, paper studies approximate HAC, where we merge any apprxoximately closest clusters instead of exact closest clusters. To do this, authors capture the HAC problem as that of approximate nearest neighbors over dynamic dataset. Each cluster is a point in the ANN index, and when 2 clusters merge, we capture by 2 deletions and 1 insertion. Then using a priority queue where each cluster maintains its closest distance to other clusters, the authors show how to maintain this PQ online, and use this to quickly determine which clusters to merge. To make it all work they devise provable new algorithms for dynamic ANN algorithms using LSH which can handle adaptive inserts/deletes (which is needed for this problem). Apparently, prior work only handles non-adaptive. Strengths: * Important problem of getting sub-quadratic algorithms for hierarchical clustering, quite reasonably well written paper. * Theoretical backing + empirical applications using the theory ideas are given, which is nice * Good set of evaluations and the numbers are great -- quality is better and significantly faster. * The dynamic ANN algorithm introduced in this paper ideas could be useful in other settings also. Weaknesses: Datasets chosen for evaluation can be more diverse. The theoretical result can be presented with more rigor -- For example, I could not follow what objective function they are presenting a c-apprxoximation for the HAC problem in Theorem 5. The technical reason/idea behind why they are able to convert a non-adaptive ANN structure to adaptive using the power-of-two size indices is not explained well, seems like the crux of the paper but missing in detail. Is there precedent for using such ideas to make non-adaptive inputs into adaptive inputs? Technical Quality: 3 Clarity: 4 Questions for Authors: In Line 140, don't you want the weighted average and not the weighted sum? What is the objective function you use to compare the quality of HAC soluton and prove approximation factor gaurantee? Why not use more modern-day neural-network inspired clustering datasets for benchmarking, as it will be an important use-case going forward (like the OpenAi embeddings or more BERT-based embeddings). What is SOTA known for other linkage functions in HAC like single linkage and average linkage? Do your ideas extend there also? Which linkage is most used in practice? You cite [2, 14] for dynamic non-adaptive ANN, However I cant find the corresponding results in either. Please be more specific. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Would be good to mention a more detailed limitations section apart from those mentioned in conclusion / open problems. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and suggestions. We respond now to the specific questions of the reviewer. > “Datasets chosen for evaluation can be more diverse.” To address this point we have assembled several new clustering datasets with ground truth from existing sources (which we will make publicly available). The datasets are described in more detail in the global response. We will include quality and scalability results on these datasets, which we hope will address the reviewers concern. > c-approximation for the HAC problem in Theorem 5 Thanks for pointing this out. The definition of c-approximate HAC is given on l138. We will improve the writing to clarify this. > “The technical reason/idea behind why they are able to convert a non-adaptive ANN structure to adaptive using the power-of-two size indices is not explained well, seems like the crux of the paper but missing in detail. Is there precedent for using such ideas to make non-adaptive inputs into adaptive inputs?” We will add more intuition and improve the writing for the ANNS section. The power-of-two method is known as merge and reduce and is an established technique (e.g. see line 197) for turning static data structures into dynamic ones. This allows us to insert new points to query against but to support adaptive query points we introduce covering nets. Merge-and-reduce: The problem that this fixes is that ANNS data structures for Euclidean distance are static meaning we can’t insert new points to query against. The idea is that we will have buckets labeled 0, 1, … , log(n) where bucket i will hold a single static ANNS data structure with at most 2^i points. To insert a point to query against we will add it to bucket 0. Then while there is a bucket i with “loose” points that are not in an ANNS data structure (there will be at most one at any time): * If there are at most 2^i total points in bucket i, build a new ANNS with all points in bucket * Else move all points from bucket i to bucket i+1 as “loose” points Roughly speaking, in bucket log(n) - i it will take time (n/2^i)^(1+epsilon) to construct an ANNS but we will only ever have to construct 2^i ANNSs in that bucket. It follows that we will only spend at most n^(1+epsilon) time constructing ANNSs for each of the log(n)+1 buckets. In the actual proof we charge the running time to the points instead of the buckets. To query or delete we only have to query or delete from at most a single static ANNS in each bucket. Covering net: ANNS data structures assume that query points are non-adversarial. Suppose we are given a point p that we will want to find an approximate nearest neighbor for in some set S. If we construct the ANNS on S then whp it will return an approximate nearest neighbor for p when queried. Given a set of query points that is not too large we can union bound to say that the ANNS will return an approximate nearest neighbor for all of them. However, in the adaptive case we must assume that an adversary who knows the randomness of the ANNS data structure gets to pick the query point. Since we can’t union bound against the infinite possible query points we introduce covering nets. A covering net is a finite set of points such that every possible query point is near a point in the covering net. Then for every query we round it to a point in the covering net and find an approximate nearest neighbor for that point. This allows us to union bound against only the points in the covering net at the cost of picking up a small additive error in the query. > “In Line 140, don't you want the weighted average and not the weighted sum?” Yes, thank you for flagging this. > “Why not use more modern-day neural-network inspired clustering datasets for benchmarking, as it will be an important use-case going forward (like the OpenAi embeddings or more BERT-based embeddings).” We definitely agree with the importance of this use case. However, we are not aware of very many publicly available datasets of this sort that would come with ground truth clustering labels. We note that in our paper, we have used one embedding-based dataset (see Appendix E for the description of the birds dataset); we will also add three new embedding-based datasets that come with ground truth which we described in the global response. If the reviewer has any other suggestions for datasets to cluster with meaningful ground truth labels we would be delighted to include them in our paper. > “What is SOTA known for other linkage functions in HAC like single linkage and average linkage? Do your ideas extend there also? Which linkage is most used in practice?” Computing single-linkage clustering can be achieved by first computing an MST of the input points and then applying a simple post processing step. For the complexity of Euclidean MST, see https://en.wikipedia.org/wiki/Euclidean_minimum_spanning_tree#Computational_complexity. In the case of average linkage, the state of the art is a paper we cite: https://papers.nips.cc/paper_files/paper/2019/hash/d98c1545b7619bd99b817cb3169cdfde-Abstract.html However, the proposed algorithm to the best of our knowledge has not been implemented, likely due to the somewhat large overhead of efficiently computing the average linkage distance. Applying our techniques to average linkage, which seems to be a harder problem than the one we address, is an interesting open problem. > Would be good to mention a more detailed limitations section apart from those mentioned in conclusion / open problems. Thank you for the suggestion. We will add a section outlining two main limitations: * the use of different ANNS data structure in the theoretical and empirical results. * quality evaluation limited to medium-sized datasets, due to the lack of suitable publicly available embedding datasets	Rebuttal 1: Rebuttal: We thank all the reviewers for their thoughtful comments and suggestions. We reply to the questions and comments of each reviewer individually in the corresponding rebuttal fields. # Coherence: First, we would like to clarify a point raised by the reviewers regarding the coherence of our theoretical and experimental results, namely that the theoretical algorithm and experimental evaluation use different ANNS algorithms (a point brought up by reviewers vHLH and Dgto). We acknowledge this limitation and would like to point out that we have been open about it in the checklist of our submitted paper. We would like to discuss the issue from two angles. In terms of the coherence of the paper, the crux of the paper is the efficient approximate centroid HAC algorithm. In the paper we show that it gives a very efficient theoretical algorithm (using our new data structure based on LSH) and an algorithm which is very efficient in practice (using state of the art ANNS techniques). So in that sense the paper presents a complete story. In terms of the theoretical and empirical part considering different algorithms, while ideally both settings would use exactly the same algorithm, we would like to point out that our situation is still somewhat better than in multiple other prominent clustering problems. In our case, both the theoretical and practical algorithms use the same meta-algorithm (instantiated using different data structures). On the other hand, in the case of many other clustering problems, there is an even larger gap between the theoretical and practical approaches, for example: * Correlation clustering - while the best known algorithms are obtained by solving a linear program (see https://dl.acm.org/doi/abs/10.1145/3618260.3649749 or https://arxiv.org/pdf/2309.17243), in practice a simple local search / Louvain-based algorithm is used (see section 4.1 of https://hal.science/hal-04251953/file/openproblems.pdf). * Modularity clustering - the best known approximation is obtained by solving an SDP (https://www.sciencedirect.com/science/article/pii/S0022000020301124?fr=RR-1&ref=cra_js_challenge). Again, Louvain-based algorithms are used in practice. * In the case of k-means, while the best known approximation is a constant obtained via the primal-dual method (https://dl.acm.org/doi/abs/10.1145/3519935.3520011), in practice Lloyd’s heuristic paired with k-means++ seeding is used which has a logarithmic approximation guarantee. * Balanced graph partitioning - coarsening combined with a brute-force algorithm is used to obtain the best results in practice. On the other hand, the algorithms for solving the corresponding theoretical formulations: balanced cut and multiway cut are entirely different. # Additional Experiments: Several reviewers also mentioned that adding more experimental results on more diverse / large datasets would be helpful. To address this, we will include quality and scaling experiments on additional labeled datasets for the revised version. First is the standard Imagenet dataset which contains around 1.2M images of everyday objects from 1000 different classes. Each image is passed through ConvNet [23] to obtain an embedding of 1024 dimensions, similar to the Birds dataset. Next, we consider two large text datasets (reddit and arxiv) from the recent MTEB work [A]. The reddit dataset consists of ~420K (embeddings of) post titles and the goal is to cluster them into (50 different) subreddits. The ArXiv dataset consists of ~730K (embeddings of) paper titles and the goal is to cluster them in (180 different) categories. The embeddings are 1024 dimensional in both arXiv and reddit datasets. [A] Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. 2022. MTEB: Massive Text Embedding Benchmark. arXiv preprint arXiv:2210.07316 (2022). https://doi.org/10.48550/ARXIV.2210.07316	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Conditional Density Estimation with Histogram Trees	Accept (poster)	Summary: This paper proposes a tree-based conditional density estimation model. Previously, conditional density estimation involved: 1) kernel density estimation-based method (requires expensive bandwidth tuning); 2) neural network (black-box approach); 3) a tree-based model which focuses on a Gaussian distribution for the target; 4) regression-based approach. In contrast to these approaches, the proposed method is 1) non-parametric; 2) uses a MDL principle which eliminates the need for hyperparameter tuning for regularization; 3) an iterative algorithm which searches for the optimal histogram in each leaf (thus modeling the full distribution). The authors claim that the proposed method is both more accurate (as measured by the log-loss) and more robust against irrelevant features. Further, the method is smaller in size and improves interpretability (which comes from a tree-based model). The experimental results are provided against a varying range of datasets from the UCI repository. The comparison is provided against other tree-based models and blackbox-approach methods. Strengths: 1. Experimental results are comprehensive, providing superior accuracy and more compact model compared to other approaches. 2. The proposed method does not need extensive parameter tuning and is empirically fast. 3. The algorithm seems to be fairly novel and easy to understand. Weaknesses: 1. The dataset used in the experiments is fairly small. To highlight empirical performance in training, the authors are advised to use larger datasets. 2. The method assumes that the support of the target variable is bounded. 3. Some presentational issues in the figure are present. See below in the Questions section. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The method denoted as "Ours" should be formally labeled as "CDTree." 2. Are all of the attributes in the datasets continuous? How would the method handle discrete/categorical attributes? 3. What model selection criteria is used for CART models? Is it the same as Equation 4 which is used for CDTree? 4. Is it possible to move the pseudocodes into the main writeup? 5. Figure 3 may contain some presentational issues. What do "+", "circle" ,"square", etc mean? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Overall the paper is reasonable well-written. However, the experimental results are limited to fairly small datasets which unfortunately do not highlight potential scalability of the proposed method. While the focus of the method is on interpretability and accuracy, the authors would have to provide some empirical results on the training phase. In addition, please move the pseudocode description of the algorithm to the main writeup section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, Thanks for the constructive advices! I will now reply to each point in the Weaknesses and Questions sections. # Regarding Weaknesses in the review: ## 1. Rebuttal to Weakness 1: In comparison to previous research for CDE, the datasets we used are NOT smaller. - We included 14 datasets in total, with various sample sizes (from 104 to 21263) and dimensionalities (from 5 to 82). We aim to test on both small and large datasets since 1) the sample size is often not too large in critical areas like healthcare; 2) (as mentioned in the Introduction) CDE is often used to understand the data collected from scienfitic experiments, in which data size is often small too, as data collection is expensive. Discussion about the data sizes in previous research: for instance, in the most recent related work [12] (Gao, Zijun, and Trevor Hastie. "LinCDE: conditional density estimation via Lindsey's method." Journal of machine learning research 23.52 (2022): 1-55.), which is a tree ensemble method, all datasets considered are below $2 000$, indicating that smaller datasets are indeed interesting to the CDE task. Further, the deep learning model proposed in [9] (Dutordoir, Vincent, et al. "Gaussian process conditional density estimation." Advances in neural information processing systems 31 (2018)) {see page 8, figure 4} considers benchmark datasets with similar sizes (number of rows) to ours (and the dimensionalities of datasets we consider are even in general higher). - Despite what has been discussed, we tried on a larger dataset about NYC taxi duraion (sample size $N=203686$), which we take from Kaggle. This is a dataset that can be used to predict the taxi trip durations based on the pick-up and drop-off locations (longitutde and latitude). We only use a subset of this data that corresponds to 1 month. However, most interpretable competitor methods fail on this dataset, and specifically CKDE/NKDE/LinCDE fail to return any results because their (widely used) implementations caused that the memory limit is reached. Meanwhile, our method and CADET are the only interpretable methods that can be applied to this dataset. However, as shown in the table below, the negative log-likelihoods of our method are pretty close to those of the two neural network methods (NF and MDN), significantly outperforming the negative log-likelihoods of CADET. Although the main goal of this paper is NOT to (just) introduce a CDE method that scales better than existing "shallow" models, this result indeed shows the potential of the scalability of CDTree to larger datasets. \| Methods \| Mean (Negative Log-likelihoods) \| Standard Deviation \| \|-----------\|-------\|--------------------\| \| Ours (CDTree) \| 7.015 \| 0.0038 \| \| CADET \| 8.290 \| 0.0870 \| \| NF \| 6.921 \| 0.1031 \| \| MDN \| 6.835 \| 0.0056 \| ## 2. Reply to Weakness 2: - We are aware of this limitation as we already discussed it in the Limitation of the paper. ## 3. Rebuttal to Weakness 3: Figure 3 does NOT contain the presentational issue mentioned in the review as the legend shows the meanings of different shapes. - As shown by the legend on the upper part of the figure, each method is shown by a different color TOGETHER with a different shape. We find that, in comparison to using different colors only (with the same shape), our visualization makes it easier to show the general picture for each method accross all datasets. # Reply to Questions: ## 1. The term "Ours" / "CDTree" for proposed method / model respectively We deliberately use “Ours” to refer to our method (the model we proposed, the learning criterion, and the algorithm, as a whole). By contrast, we use the term “CDTree'' to refer to the probabilistic model (learned from the data), which can be seen already from our abstract. ## 2. Our method can handle both discrete/continuous features. As discussed in Appendix B.2, we do not require the feature variables $X$ to be continuous. That is, while the target variable $Y$ needs to be continuous (otherwise there is no need for conditional density estimation), the feature variables can be both categorical and continuous. Just like the decision tree for classification/regression, the splitting conditions on the internal nodes can be defined for both discrete and continuous variables. We do assume that categorical variables are preprocessed and one-hot encoded though. ## 3. About CART model selection. The CART we use is the original version from the inventors. That is, the tree is first grown by iteratively maximizing the gini index, and then pruned by the so-called “cost complexity pruning”, i.e., by optimizing the impurity (MSE for regression) plus a regularization term defined as the number of nodes times a regularization hyperparameter, in which the hyperparameter is tuned by cross-validation in our experiment. These are (already) discussed in Appendix C.4. ## 4. Put (more) content about Algorithm in the main content. Yes, if the paper is accepted and then one more page is granted. ## 5. Resolved (see above). --- Rebuttal 2: Title: Dear authors Comment: (Mistake on the previous title - it was directed to the authors who provided the rebuttals): Thanks for your extensive reply on the points. Perhaps it is better to move clarifying points which are in the appendix to the main writeup (as I see that some answers to my questions are in the appendix which not many reviewers will have to look at, unless there is an explicit hint in the manuscript). I do not believe it will have much impact on the length of the manuscript. In addition, as you go through the draft once more, you will have opportunities to cut down redundant sentences/paragraphs and make your points more succinct. I still argue that Figure 3 is a very confusing figure - a visualization should not ask readers to combine a combination of color and shape. Why are some of these shapes not aligned on the horizontal axis tick? I also think there are too many horizontal ticks (yes, each tick = dataset) but there is not much space to put all of these information. One more thing: I would also keep the fonts consistent in your reply - it actually does not help reviewers look at your reply in entirety and it looks less professional. --- Rebuttal Comment 2.1: Title: Thanks for the further advices about the writing Comment: Thanks for the further advices about the writing --- Rebuttal Comment 2.2: Title: Further clarifications for Figure 3 Comment: Dear reviewer, Thanks again for the detailed comments that help improve my paper further! Somehow the second half of your previous comments is missing in the notification email I received from "neurips2024-notifications", which did not contain your further comments about Figure 3. I only noticed this just now and hereby give further clarifications below. Although the common way of presenting the results in Figure 3 is to use a separate figure for each individual dataset (often with the x-axis representing "the number of added features" and the y-axis representing "the number of splits on irrelevant features"), this would lead to too many figures for our case (as we tested 14 datasets). We would need to put all these figures in the Appendix then. Thus, we chose to put the results of all datasets together, which can show that our method learns CDTrees with the number of irrelevant features very close to 0 across all datasets. As a result, we just use the different shapes to keep the information about the number of features added for each "point" on the figure. I do agree that this figure may need more brainwork to process than other plots in the paper; however, the advantages are 1) avoiding too many figures for this single experiment subsection, and 2) using one plot for all datasets can better show the "general picture" of different methods. (Nevertheless, I would be happy to consider other possibilities in your mind as well. Thanks in advance!) Regarding the point you raised "Why are some of these shapes not aligned on the horizontal axis tick", we do this deliberately by "jittering" the points a bit, to avoid some points being fully covered by others. Last, thanks for advices about the fonts used in the replies as well!	Summary: This paper addresses the problem of conditional density estimation, i.e. given a conditioning variable x estimate the whole distribution of y, with special emphasis on interpretability. For this interpretability requirement, the authors resort to classical decision trees allowing to partition the conditioning space in an interpretable manner. The conditioned distributions of y are modelled with equally-spaced histograms with a variable number of bins. This construction is optimised in an heuristic manner using an objective function that is built using a Minimum Description Length (MDL) approach, using different codes for the different parts (size and structure of the tree, splitting conditions and histograms) of the construction to be determined. Experiments show log loss performance that is globally better than interpretable competitors on the selected datasets. Empirically, the trees obtained with this method are shallower than other interpretable tree-based methods and they are more robust to irrelevant features, in the sense that these are less often used for splitting. Strengths: The paper is generally well written and easy to read. The construction seems useful in practice as shown by the experiments and could be easily adopted by practitioners. Weaknesses: This work combines in a straightforward manner classical building blocks from the machine learning and information theory literature: decision trees, histograms, MDL, universal codes for integers, ... so there is little novelty in the construction Although it is advertised that no hyperparameter tuning is needed thanks to MDL, there are still arbitrary choices for the priors and also the hyperparameter C (C=5 in the experiments). For example, the tree size and structure could be encoded using Willems' approach in Context Tree Weighting (preorder traversal of the tree, for each node use one bit to say if it is internal or leaf, which is equivalent to putting a 1/2 prior probability of splitting a node). (F. M. J. Willems, Y. M. Shtarkov, and T. J. Tjalkens. The context-tree weighting method: Basic properties. Information Theory, IEEE Transactions on, 41(3):653–664, 1995.) This should be discussed. The MDL approach used in this work is called "crude" in [14] since there is some arbitrariness in the choice of priors instead of some optimality criterion. The only part in this work that is "refined" [14] is the NML used for the histograms themselves but this is something that is well know (Proposition 1 is a simple extension of existing results). This should be discussed. The optimisation method is heuristic and thus its interest for the NeurIPS audience is limited (in fact, most of the algorithm is in the appendix, which indicates that it is not an important contribution). For these reasons, I think that the novelty and the significance for NeurIPS is quite limited. In my opinion, this work would fit better in a more applied venue. Minor issues: citations to books (e.g. [6,14]) should be more precise line 186: "variable name" can be misleading, I suggest to say "variable index" typo: line 231 "we iterative" Technical Quality: 3 Clarity: 3 Questions for Authors: paragraph explaining the role of C and d (lines 191 to 194) is a bit confusing, can you provide an example showing what are the splitting options that are encoded when d>1 ? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: Limitations were properly discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed comments and advice. We now address each paragraph in the "Weakness" section of your review. ## Factual Error in Paragraph 2: - The statement in the review that "no hyperparameter tuning is needed" is incorrect. Our paper highlights that the advantage of using MDL is the elimination of the need for a REGULARIZATION hyperparameter. We specifically use the term "regularization hyperparameter"/"hyperparameter for regularization" throughout the paper, in which we mentioned the word "hyperparameter" 4 times in total. We specifically emphasized this point because the “standard” way of regularization in decision tree learning is to penalize on the number of nodes times a "regularization hyperparameter", which needs extensive parameter tuning. ## Rebuttal about the statement that we are using "crude" MDL (Paragraph 3): - The claim that our MDL approach is "crude" is incorrect according to Chapter 14.3 of the cited book [14]. See page 426. Here I quote: "In the end, we encode data D using a hybrid two-part/one-part universal model, explicitly encoding the models we want to select between and implicitly encoding any distributions contained in those models", which defines our approach as “the refined MDL”. Additionally, while the NML regret for histograms is well-known, our contribution extends these results to a supervised setting with histograms and demonstrates that the NML regret for the CDTree is the product of the regret terms for all histograms on all leaves. ## Rebuttal about choosing priors (Paragraph 2): - Although there is indeed some flexibility in choosing the model encoding scheme within the MDL framework, similar to selecting priors in Bayesian model selection, encoding the model is not arbitrary but follows requirements and guidelines. This includes ensuring the prior probability defined on the model class sums to 1, preventing an excessively large penalty term that would require larger sample sizes to converge to the true model. The encoding scheme you mentioned is suboptimal because the corresponding prior probability is not proper (consider, e.g., the code word that corresponds to the case that every node is a leaf node, which cannot form a decision tree). Notably, many “old” methods that leverage MDL used various intuitive yet “crude” encoding schemes like this one, and it is hardly possible to review all of them (e.g., the one used in the C4.5 decision tree is different from the one you mentioned above). ## About choosing hyperparameter C (Paragraph 2): - The hyperparameter C establishes a hierarchical structure on the search space of continuous feature variables, setting levels for granularity. This hierarchy helps express prior beliefs about model complexity. For instance, intuitively, if a decision tree gives a split condition with a highly precise value like $X > 1.0001$, domain experts may ask, is this level of granularity (up to the 4 decimal places) necessary? Why not just $X > 1$? Thus, from the perspective of model selection, the former condition is "more complex" (more bits required to encode) than the latter one. However, as different feature variables may have very different ranges, instead of considering the granularity in terms of the decimal precision, in this paper we consider the granularity in terms of the quantiles (e.g., splitting on the 1/2-quantile may be considered "simpler" than splitting on the 1/10000-quantile). As explained in Appendix C.2, just like we have different numeral systems (e.g., decimal/binary), which set up different hierarchical structures for the levels of granularities, the hyperparameter C plays a similar role here in setting up the hierarchical structures for the quantiles. - We have shown that a global ad-hoc choice for $C = 5$ works well for all datasets in our experiments section. By contrast, it is impossible to have a single value for the regularization hyperparameter used in traditional decision tree learning that works well for different datasets, which also does not carry this natural meaning of the $C$ in our paper. ## About our algorithimic innovations (Paragraph 4): - While our algorithm is greedy, it addresses the challenging task of optimizing the tree structure and separate models (on leaves) simultaneously. Recent algorithmic advancements for classification/regression trees have been reviewed, highlighting the challenges for CDE tasks, which motivates the choice for the heuristic algorithm. - Our main algorithmic innovations are described in Section 5 (not the appendix), including 1) no pruning for the tree, which favors the model complexity, and 2) searching the histograms directly (without using any "guessing" heuristics that are common in traditional "model tree" methods). We also include in Section 5 a high-level yet complete description of the algorithm process. The pseudo-code and other details are put to the appendix for reproducibility due to the page limit. - Due to the fact that improving on the heuristic algorithms for classification/regression trees often require a whole research paper to describe (as reviewed in our paper), we consider further improving on our proposed algorithm for CDE as future research. ## Rebuttal about the characterization of our method as a "straightforward combination of classical building blocks". We respectfully disagree on this statement as it overlooks our contributions: - We propose the first single tree-based model specifically designed for the CDE task (as CADET essentially used the same model as CART due to the equivalence between the MSE loss and the Gaussian assumption); - We are the first to formalize the decision tree learning under the (modern) MDL framework as a model selection problem, introducing new encoding schemes for data (with NML) and the decision tree itself. - Our extensive experiments demonstrate that with CDTree, kernel-based methods are no longer the only choice for "shallow" models for CDE. --- Rebuttal Comment 1.1: Title: About "Factual Error in Paragraph 2" + "choosing priors" + "choosing hyperparameter C" Comment: Thank you for your detailed response. Your choice of prior for the model consists in encoding the size (number of leaves $K$) with Rissanen's code for integers and then a uniform prior over all trees with K leaves ($1/C_K$). I mentioned another possibility which is the prior used in CTW, which puts 1/2 probability on splitting each node. In a series of papers on Bayesian Context Trees, a more general version is considered with a parameter $\beta$ defining the probability of not splitting, that is, the larger beta, the larger the penalization of more complex models. See Section 3.1 of Papageorgiou, I., & Kontoyiannis, I. (2024). Posterior representations for Bayesian Context Trees: Sampling, estimation and convergence. Bayesian analysis, 19(2), 501-529. I don't understand your remark: The encoding scheme you mentioned is suboptimal because the corresponding prior probability is not proper (consider, e.g., the code word that corresponds to the case that every node is a leaf node, which cannot form a decision tree) If every node is a leaf node => the tree has only one node (the root) => it's a valid decision tree and its prior probability is 1/2 with Willems' scheme. As mentioned in the reference above on Bayesian Context Trees, one can see the splitting process as a Galton-Watson process, proving that the prior is proper. As you say: "C establishes a hierarchical structure" and "This hierarchy helps express prior beliefs about model complexity". Although I understand that empirically C=5 worked well for all the datasets considered, there is no theoretical proof showing that this is a universal constant that shouldn't be tuned. So I still think that "no hyperparameter tuning is needed" should be at least toned down, since there could be datasets for which some tuning would be beneficial. Can you clarify how your prior satisfies (1.) of page 426 of the MDL book? --- Rebuttal 2: Title: More rebuttal Comment: Dear reviewer, Thanks for the quick and detailed response. We have further rebuttals as follows. ## Further clarification about the prior you proposed - One factual error: the paper you cited, (Papageorgiou, I., & Kontoyiannis, I. (2024). Posterior representations for Bayesian Context Trees: Sampling, estimation and convergence. Bayesian analysis, 19(2), 501-529.), considers $m$-ary trees, while in our paper we consider (full) binary trees only, as describe in Lines 104-105 and then emphasized in Line 177. - Further explanations about why the prior probabilities you proposed do not sum to 1. For simplicity, consider the case when a decision tree has 3 nodes in total, for which apparently only one possible tree structure exists (i.e., one root node and two leaf nodes). However, if you put a 1/2 prior probability on whether to split each node, this "only" structure has a prior probability of $(1/2)^3 < 1$. Hence, in this case, the sum of all prior probability masses is $(1/2)^3 < 1$ as well.   ## Further clarification about the hyperparameter tuning - Again, as already mentioned in our previous rebuttal, we NEVER claimed that “no hyperparameter tuning is needed”; hence, I am not sure how we "tone it down". - We indeed introduced $C$, yet we have shown that a global ad-hoc choice for $C$ can work reasonably well for all datasets in our experiments (and whether tuning $C$ can further increase the predictive performance of CDTree is not related to the main research question in this paper). By contrast, this is hardly possible for the regularization hyperparameter; i.e., it is hardly possible to pick a single value with an intuitive meaning for the $\alpha$ in the traditional decision tree learning optimization function, often in the form of "$impurity + \alpha \|T\|$", in which $\|T\|$ denotes the size of the tree, and $impurity$ can be MSE for regression. ## Regarding Condition (1) of page 426 of the MDL book. - Notably, the text box on Page 426 does NOT give a rigorous definition of the "Refined MDL model selection", for the following reasons. First, see page 427, the 2nd paragraph, and here I quote "The general idea is summarized in the box on page 426, which provides something like a definition* of MDL model selection, but only in a quite restricted context. If we go beyond that context, these prescriptions cannot be used literally, but extensions in the same spirit suggest themselves.". Second, the Condition (1.) of page 426 is also NOT a rigorous definition for the requirements of the priors, as the concept "quasi-uniform" is NOT defined throughout the book (see page 425, here I quote "While we have not formalized what types of codes still count as quasi-uniform and what not ..."). Hence, Condition (1) of page 426 is more like a general guideline summarized from the examples used in the book. - The last two paragraphs of page 425 (starting from "we can choose the code ..."; and note that the last paragraph of page 425 ends on the top of page 427) are a more general guideline for choosing the priors for model selection than the condition (1) of page 426. Specifically, our proposed prior is close to the general description given by the last sentence of the first paragraph on page 427; here I quote, "In some cases we can apply conditional codes which do achieve a “conditional” minimax regret, ...". That is, in general, we encode the model by number of nodes => tree structure => node splitting conditions => number of histogram bins; each step is a conditional uniform/quasi-uniform code (as the integer code is listed as an example of quasi-uniform in the book), conditioned on the value encoded in the previous step. - Notably, we quoted the content of page 427 (in our previous response) as it suffices to give the rebuttal to your previous statement that a mix of NML (one-part) and a model prior makes our MDL encoding "crude"; as shown on page 427, a mix of these two is by definition the form of the "refined" MDL model selection. --- Rebuttal Comment 2.1: Title: alternative prior and "factual error" Comment: In Papageorgiou, I., & Kontoyiannis, I. (2024). Posterior representations for Bayesian Context Trees: Sampling, estimation and convergence. Bayesian analysis, 19(2), 501-529., proper $m$-ary trees* are considered, which are defined as follows: A tree $T$ is called proper if any node in $T$ that is not a leaf has exactly $m$ children. For $m=2$, it corresponds to full binary trees. So, I don't see where the "factual error" is. Regarding your futher explanation, I think I see where your misunderstanding comes from : in the Bayesian Context Trees prior, the structure is encoded without conditioning on the total number of leaves. The total number of leaves is not explicitly encoded as in your approach. So, your example gives the correct prior of that particular tree but it is not the only one: you need to sum over the whole set of full binary trees. --- Reply to Comment 2.1.1: Title: Further clarification Comment: Again, thanks for the detailed and quick response! (Nice to have these technical discussions anyway.) First, I never claimed that our prior is the only choice. Second, I looked at papers you cited and looked at the exact formula of the prior you proposed. Specifically, [A] Papageorgiou, I., & Kontoyiannis, I. (2024). Posterior representations for Bayesian Context Trees: Sampling, estimation and convergence. Bayesian analysis, 19(2), 501-529. , and [A] cited the below [B] Kontoyiannis, Ioannis, et al. "Bayesian context trees: Modelling and exact inference for discrete time series." Journal of the Royal Statistical Society Series B: Statistical Methodology 84.4 (2022): 1287-1323. And in [B], Lemma 1 shows that the prior $\pi(T) = \alpha ^ {\|T\| - 1} \beta ^ {\|T\| - L_D(T)}$ sums to 1, for all trees $T$ with the depth smaller than the given number $D$, which does not apply to our case as we do not have the constraint on the tree depth. (You could say, the max depth $D$ can be specified as the sample size $n$, yet this is a bit arbitrary as it has influence on the priors for trees that reach (or do not reach) the depth $D$ (see the formula). It is at least much less natural for the decision tree for CDE than for the variable-memory Markov chains in [A]. Last, I found where another misunderstanding is from. In your original review, you stated that "For example, the tree size and structure could be encoded using Willems' approach in Context Tree Weighting (preorder traversal of the tree, for each node use one bit to say if it is internal or leaf, which is equivalent to putting a 1/2 prior probability of splitting a node)". However, this is different from what is described in Section 3.1 of [A], as no prior probability will be put on those nodes that reach the max depth (i.e., leaves). In fact, it seems to me that your statement above actually says that, given a list of nodes (with a pre-defined order), we specify a prior probability 1/2 on whether to split independently to each node; however, in a branching process there is no such indepdency. This is where the misunderstanding mainly comes from. It is indeed an elegant way to encode a tree with the max depth given. Very nice to know.	Summary: This paper proposed to use decision tree with leaves formed by histogram models for conditional density estimation to gain interpretability. Characteristics of the proposed method, along with the density estimation accuracy and run time, are evaluated with numerical experiments. Strengths: Extensive experiments are conducted to evaluate multiple aspects of the proposed method, including comparison of the density estimation accuracy with related methods, as well as computational efficiency and interpretability-related aspects. Weaknesses: The proposed method can be viewed as a conditional density estimator with partitions on both X and Y space. There is limited novelty as far as I read from this paper. Technical Quality: 3 Clarity: 3 Questions for Authors: - If p(y\|x) changes smoothly with respect to x, is the proposed method capable of capturing such smoothness, given it is based on partition of X? - How to prevent overfitting? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: The authors discussed some limitations of their method. There is no potential negative societal impact of their work. I suggest providing some discussion on the questions raised above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer, Although I understand that the judgment about novelty can be a subjective matter, I would appreciate it if more information is provided regarding the motivation for this judgment, which would be helpful for me to improve the paper. Regarding your two questions: 1. I am not sure what exactly you mean by "capturing such smoothness". Could you give a formal definition or an example of "capturing smoothness"? In general, partition-based models (including histograms, decision trees, regression trees, etc) approximate smooth functions (e.g., decision boundaries) in a piecewise manner. 2. As elaborated in the paper, the overfitting is prevented by formalizing the learning problem as a MDL model selection problem. --- Rebuttal Comment 1.1: Comment: Thanks for the rebuttal. Regarding smoothness, I was asking about the way that the conditional density changes wrt x. There can be roughly two cases: (1) very heteroscedastic case, where p(y\|x) can change drastically across x values, and (2) for x1\approx x2, p(y\|x=x1) is close to p(y\|x=x2), and the limiting scenario would be p(y\|x) = p(y) regardless of x values. CDE methods that partitions X is generally suitable for case 1, and I would like to hear how the proposed method deals with case 2. --- Reply to Comment 1.1.1: Comment: Dear reviewer, Thanks for the clarification for the questions. For case (2), i.e., for the case "$x1\approx x2$ then $p(y\|x=x1)$ is close to $p(y\|x=x2)$", $x_1$ and $x_2$ will be in the same leaf node, and both $p(y\|x=x1)$ and $p(y\|x=x2)$ will be estimated by a single density estimator (a histogram in our proposed method). In the extreme case that $p(y\|x) = p(y)$ for all $x$, the decision tree will be learned to have only one node (the root node). This is in fact shown by our experiment in Section 6.4 "Robustness to irrelevant features", in which we show that features that are (conditionally) independent to $y$ won't be used for tree splitting.	Summary: This paper proposes a new conditional density estimation algorithm based on histogram trees. The base model corresponds to a full binary tree, where each internal node is associated with a split of the feature space in one coordinate, and each leaf node is associated with a histogram density estimator for the response variable for the feature variable falling into the leaf node. To find the best among all possible models that best fit the data, the proposed algorithm invokes the MDL principle, which avoids any hyperparameter tuning, to derive the model selection criterion. Since the optimization is infeasible to be solved exactly over all possible models, the paper proposes a greedy algorithm. Experiments support that the resulting estimator CDTree is competitive against the state-of-the-art models. Strengths: - The paper proposes a nice algorithm for an important problem based on basic principles. This seems to be a nice practical application of the MDL principle. - Though the resulting estimator is a single tree-based conditional density estimator, which is inherently "interpretable" thanks to the tree structure, experimental results show that CDTree is competitive against the SOTA algorithm based on ensemble. This is quite surprising. - Experimental results are sufficiently thorough to understand the pros and cons of the proposed method. Weaknesses: While I like the paper overall, the manuscript is not spotless, especially in terms of presentation, grammatical errors, and typographical mistakes. I believe that after a careful revision of the paper, this would be a nice addition to the community. (Though the fit might be much better with an applied statistics journal.) Technical Quality: 4 Clarity: 3 Questions for Authors: Suggestions - The definition of a conditional density tree with histogram in Section 3 seems to be incomplete. I was able to understand the complete model after I read Section 4.3. I believe that this should be properly defined as a model in Section 3, since without a proper definition Section 4.1 and 4.2 read awkward. For example, what is the dimension of $x$ and do you assume $y\in\mathbb{R}$? Somehow this is not explicitly defined. - And I think the sentence in lines 108-110 is quite confusing, in the sense that while the model $M$ may depend on the dataset $D$, a covariate $x$ needs not be from the dataset to evaluate the conditional density. - In Section 4.1, the definition of MDL optimal model is rather abrupt as the "universal model" is defined only in Section 4.2. I believe that the NML code should be defined before (1) or at least there should be a pointer to (2) in that paragraph. The notation $P_M(.)$ is also misleading. Please consider $P_M(\cdot \| \cdot)$. Questions - Why did the authors define $\hat{\theta}$ in Proposition 1 while it is not used? - Why are CKDE and NKDE put under "interpretable models"? I was confused as the authors emphasized at multiple points that kernel-based methods are not interpretable. - A silly baseline is to consider a Gaussian density estimator as CADET assumes in the proposed MDL criterion of CDTree. This would result in a faster algorithm as the search space is simple. I am curious how this simple baseline would work in practice, as this can help understand the benefits of using the flexible histograms in place of Gaussians and using MDL principle separately. I would appreciate additional experiments, if time permits. - (Minor) Regarding the discussion on runtime: Can you provide a bit more quantitative remark, is possible? For example, the authors mention that CART is highly optimized while the proposed algorithm is not. In which part of CDTree could be more optimized? - Please revise the bibliographic info. For example, "mdl" in the title of [41] should be capitalized. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors note its limitation regarding the boundedness assumption of target variable. On top of that, the algorithm is greedy and there is no guarantee on closeness between the actual estimator and the optimal estimator defined by the learning criterion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for the detailed, constructive, and very helpful comments on my writing! ## Presentational issues fixed I have fixed the presentational issues and typos you mentioned. Among others, one extremely helpful comment is that "lines 108-110 are confusing": in these lines I wrote "... $f (y\|x)$ as $f _ k (x)$ ... ", but it should actually be "... $f (y\|x)$ as $f _ k (y)$ ... ". ## Rebuttal about "model definition not complete" We respectfully disagree that "The definition of a conditional density tree with histograms in Section 3 seems to be incomplete", and you seem to indicate part of the model definition is in Section 4. However, Section 3 contains all information about how to calculate the probability/likelihood of data given a fixed conditional density tree (regardless on whether it is actually a good model), which is sufficient for defining a probabilistic model. In contrast, Section 4 is all about the definition of the model selection criterion. ## Rebuttal about "applied statistics journal may be a better fit" The machine learning community has been recently highly interested in both conditional density estimation (CDE) (e.g., the references [9, 37, 48, 12] in the paper are from NeurIPS/ICML/JMLR) and interpretability, yet interpretable CDE methods are neglected. Second, the lack of CDTree, as the counterpart to the regression tree and classification tree, hinders the development of XAI methods (e.g., local surrogate models) for CDE. ## Reply to your questions: - Regarding the first and last point: you are correct. Thanks so much again! - Whether CKDE/NKDE is interpretable model: CKDE/NKDE (and other kernel-based models) are currently the standard (if not the only) choices if someone is looking for a "shallow" model for CDE. We argued in the paper that these models are "arguably less interpretable to (single) trees", yet we are also aware that these models may be considered more interpretable than tree ensembles/neural networks. (After all, there is still no precise definition of "interpretable models". ) - About comparison to the baseline with Gaussian assumption: The MDL criterion we proposed CANNOT be used for Gaussian assumptions, as it is known that the regret term for the NML will be infinity (see page 298 of the book [14], {Grünwald 2007, The minimum description length principle MIT press, Example 11.1, Chapter 11}). I agree it would be interesting to do some sort of this kind of comparison though, as kind of an ablation study for the histograms and a supplementary to CART-h in the paper. However, the closest related work I can think of is (Proença, Hugo M., et al. "Discovering outstanding subgroup lists for numeric targets using MDL." ECML PKDD 2020), but they 1) consider rule lists (not trees), 2) they consider subgroup discovery for numeric targets (so not regression or CDE) and 3) they use the Bayesian encoding, not NML (which is not possible as discussed). - Runtime of CART: it depends on the search space of the regularization hyper-parameter. In our experiments, we set the range by “ccp_alphas = np.logspace(-5, 2, 30)”, and for all datasets, a single run should not exceed 100 seconds from what I remember. --- Rebuttal Comment 1.1: Comment: I appreciate the authors' response. Please consider revising the manuscript with the points raised in my review, as well as other reviews. Especially I found that the reviewer n5cC raised many interesting questions, which could help further improve the depth of the paper if properly answered. I will keep my score as is, since I believe there is a sufficiently good contribution in this paper.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Hierarchical and Density-based Causal Clustering	Accept (poster)	Summary: The paper aims to understand treatment effect heterogeneity and identify and evaluate subgroup effects. It addresses the challenge of typically unknown subgroup structures by proposing a solution based on causal k-means clustering to assess effect heterogeneity. The approach is improved by integrating hierarchical and density-based clustering algorithms, providing a more nuanced and effective method for identifying and evaluating subgroup effects. Strengths: The paper exhibits several strengths across originality, quality, clarity, and significance: Originality: The introduction of simple plug-in estimators, implementable with off-the-shelf algorithms, is notable. This approach opens new avenues for clustering with generic pseudo-outcomes and contributes to identifying homogeneous subgroups in treatment response. Quality: The paper thoroughly explores finite sample properties via simulation, with a setup similar to prior studies. The rate of convergence is presented, and the inclusion of standard error bars in the graphs adds to the reliability of the results. Clarity: The experiments are straightforward and concise, with all necessary details provided. The graphs are clear and effectively convey the results. Significance: The paper significantly contributes to the progression of methodologies for identifying homogeneous subgroups in treatment response. Weaknesses: The paper lacks a comparative analysis of the method's performance against existing methods in the empirical analysis section. Specifically, it does not demonstrate if the new method captures the underlying structure more accurately than other existing methods. Including such comparisons would strengthen the paper and make its contributions more convincing. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Among the case studies, the number of covariates is very small. How does the convergence rate of \hat{\mu}_a change as the dimension of the covariate space increases? Specifically, how slow does it get in practice as the number of covariates grows? 2. Do the authors have suggestions on the optimal number of covariates to limit to before applying this method in practice? 3. Are there any methods the authors can recommend to help identify important covariates that modify the treatment effect? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitations in their discussion section, acknowledging that this method is a useful tool for exploring subgroup structures. They recommend using other methods in combination with their proposed method to inform specific decisions. One area for improvement is the interpretation of the subgroup cluster results, which poses another challenge. Could the authors explain how to better interpret the resulting clusters in their application example? Additionally, how can this information be utilized to inform decisions afterward? Providing more guidance on these aspects would enhance the practical applicability of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your valuable comments and suggestions. We address each of them below. 1. \[Experimental comparison\] Yes, we completely agree that our work could be strengthened by experimentally comparing with other clustering methods, especially if other methods require any special assumptions that are not present in ours (like the margin-type conditions as in k-means) to be employed in the causal clustering framework. However, this would require a closer analysis at those methods. If other alternatives do not require particularly stronger assumptions and can be readily integrated into the causal clustering framework as well (though this must be theoretically verified as in our work), then the problem turns to which approach to use for specific data. Because each clustering algorithm has benefits and drawbacks, it is difficult to conduct a fair comparison using experiments. Given these circumstances, we believe that designing a fair experiment to demonstrate the superiority of our proposed methods is not quite straightforward, and that's outside the scope of this paper. Furthermore, comparisons with other \"causal clustering\" counterparts in the SCM literature appear unclear because they are designed to analyze structural heterogeneity, whereas ours is designed to analyze treatment effect heterogeneity. We may consider some special data generation process where we may demonstrate superiority of our method, yet this might not be convincing. All in all, we believe that confining our work to presenting the respective theory that the two appealing off-the-shelf cluster-analysis methods can be successfully adopted within the novel framework of causal clustering would suffice for the time being. Nonetheless, following your comment, we will go over the potential extension to other clustering algorithms in the discussion section in the revised manuscript. 2. \[Ans to Q1\] This is a good question. The rate of convergence of regression functions in nonparametric modeling is well studied (e.g., Györfi, 2002). It depends on both the function space to which the true $\mu_a$ belongs and the estimator itself $\hat{\mu}_a$; e.g., for $\mu_a$ in the Holder class with smoothness $s$, given $dim(X)=d$, the best (in minimax sense) rates are $O(n^{-\frac{s}{2s+d}})$. If $\mu_a$ is not smooth enough, then high-dimensional covariates will increase uncertainty of our estimators. In fact, in our simulation in Section 5.1., we incorporated the effect of the number of covariates by directly controlling the convergence rate through a parameter $\beta$, i.e., by letting $\Vert \hat{\mu}_a - \mu_a\Vert = O_P(n^{-\beta})$. The results are presented in Figure 2 (we will magnify this figure in the revised manuscript). 3. \[Ans to Q2\] In our opinion, there is no precise answer to this question. As previously stated, the smaller the number of covariates ($=d$), the lower the estimation error. However, additional components also interact with $d$. For example, in density-based clustering, we have bandwidth that is entangled with $d$, both of which affect the error. More importantly, if we use only a small number of covariates, the no-unmeasured confounding assumption (Assumption C2) is likely to be violated; we usually collect as many covariates as possible to ensure there is no nmeasured confounders left. All of these factors must be considered simultaneously. 4. \[Ans to Q3\] This leads to an entirely different, but interesting question. Given a collection of covariates, one can accurately estimate the CATE function using approaches such as Kennedy (2023), and then apply an algorithm to determine which covariate (or combination of covariates) has the most impact on the treatment effect. Or based on our methods, one can analyze the covariate distributions across clusters and determine which covariate results in the greatest distributional divergence in subgroup effects. We will add this comment in our revised manuscript. 5. \[Interpretation\] This is also a great question. One possible interpretation is that each cluster has its own subpopulation from which units are generated. Properties of each subpopulation could be studied by analyzing distributional features, etc. However, as we highlighted in the Discussion section as a shortcoming, while our approach allows for efficient discovery of subgroup structures,, it may be less effective for prescriptive applications (i.e., informing specific treatment decisions). Thus, it is important to exercise caution while attempting to interpret the observed clusters.	Summary: This work solves the task of clustering treatment-effect data based on their conditional treatment effect with a discrete set of treatments and continuous (possibly multivariate) effects. More specifically, they extend the framework of previous work on causal k-means which achieves the same, albeit which is now applied on density-based (hierarchical) clustering. Similar to the work on causal k-means, the authors opt for clustering samples based on their signature in the CATE space, i.e., a Euclidean vector whose elements are the avg. treatment effect given each treatment, jointly with an appropriate plugin estimator. To bound the error of the resulting pruning tree, they first extend the concept of $(\alpha,\nu)-$good neighbourhood to incorporate a general distribution (in the original work the empirical one is is used) and then adapt the main result of that work to their extension. To extend this framework to hierarchical clustering they employ Balcan et al's (2014) method, which provides an algorithm that has low (up to O(n) ) complexity and is robust to outliers, among other traits. Importantly, they extend the error bounds of the latter method to the causal case, which shows a small discounting of the accuracy due to the added CATE estimation overhead. They demonstrate their result on synthetic and real world datasets, which are nicely visualised for 2D. Strengths: 1. The work provides theorems to rigorously derive a bound on the error of the resulting outcome. 2. The assumptions make sense and are well studied. Weaknesses: # Contextualisation This work does mention several other methods closel or loosely related with the task at hand, however none of these other methods have been compared against in the experiments. Although this could well be justified as providing an isolated work providing the respective theory for this specific approach, the case that the authors make could be strengthened by comparing with other methods. For instance, those extracting causal rules, or a step-wise clustering approach that first estimates the CATE and then applies clustering on it. Other density based clustering methods could also be used, for instance hierarchical DBSCAN [1] which is also considered a robust method, or even with causal k-means on the estimated CATE vectors. The point here would not necessarilly be to justify the superiority of Balcan et al's method (which arguably is also not too concerned with this in the first place), but to also study experimentally the superiority of their theoretically superior results. ``` [1] Ricardo J. G. B. Campello, Davoud Moulavi, Joerg Sander: Density-Based Clustering Based on Hierarchical Density Estimates. In: Advances in Knowledge Discovery and Data Mining. Springer Berlin Heidelberg, Berlin, Heidelberg 2013 ``` Additionally, this work borrows from two key works, the causal k-means and the Balcan et al's method, but the extensions and modifications of thie present work are contributions which can at some times be unclear. # Plug in estimator The kernel density estimate can often be criticised due to its weaknesses, for instance its susceptibility to the curse of dimensionality, need for hyperparameter tuning, etc. How does the method depend on the choice of kernel, beside its Lipschitz constant? # Presentation First, I was bothered by the $\equiv$ sign in line 180. This doesn't seem to be a standard or previously introduced notation. What is the advantage of using this over an equals sign? Do you want to stress that $\mathbb C$ is some ball, itself? Additionally, I believe it would be easier for the reader if you broke down the definition 3.1 of or provided intuitive explanations of the $\nu$-strict and $\alpha$-good parts of it, similar to the Balcan et al. work, mutatis mutandis for your distirbution-general extension. Some minor typos: * 115: us harness$\rightarrow$us to harness * 200: have$\rightarrow$has * 210: arbitrarily sized * The legend and labels in Figure 4 are too small to be legible when printed on paper. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. In your work you make an extensive ue of the density of the sample points $(X,Y)$; however, the Housdorff distance definition you provide (ln. 247) seems to be oblivious to this distribution, only sensitive to the envelope of these points. Could you comment on this? 2. See questions on presentation above. 3. You may comment on comparisons section, above. 4. You may comment on the plug in estimator as mentioned above. 5. The use of CATE as a clustering domain seems to have its merits; here, the conditioning set seems to be completely ignored, up to the estimation of these vectors, themselves. Say, two very different patients (in terms of X) that have similar CATE profiles would be clustered together; from a standard causal perpsective using a DAG-based SCM, this would amount to splitting the dataset in sub-parts (clusters) which might not be well aligned with the varialbes of the underlying SCM, and could make downstream structural causal learning tasks harder. Could you comment on this? 6. Could you comment on my understanding of the hyperparameter need (see limitations)? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors make a good effort to be open on the criticisms of certain standard assumptions in the field of causality, which do not, however, hurt the validity of the method. Another basic limitation that seems to be inherited from the use of Balcan et al's work which requires the a-priori specification of the $\alpha, \nu$ hyper parameters, for which there does not seem to be a good way to intuitively specify. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your valuable input and insights. We address each of your comments below. 1. \[Clarification on contributions\] Thank you for bringing our attention to this. The main contributions of our work is that we have proved that the two appealing off-the-shelf cluster-analysis techniques can be successfully adopted within the novel framework of causal clustering, using simple plug-in methods without requiring additional strong structural assumptions as opposed to k-means, which requires the margin condition (kim et al. 2024). For example, in Section 3, we demonstrate that the robust inductive hierarchical clustering method (Balcan el al. 2014) may be applied to causal clustering through the plug-in estimator with minimal assumptions, and we analytically validate the associated costs in Theorem 3.1. Verifying this is not as straightforward as one may imagine. We will state our contributions more clearly in Section 1, particularly in relation to kim et al. (2024) and key references in Sections 3 & 4. 2. \[Experimental comparison\] Relatedly, yes, we completely agree that our work could be strengthened by experimentally comparing with other methods such as hdbscan, if other methods require any special assumptions that are not present in ours (like the margin-type conditions as in k-means) to be employed in the causal clustering framework. However, this would require a closer analysis at those methods. If other alternatives do not require particularly stronger assumptions and can be readily integrated into the causal clustering framework as well (though this must be verified), then the problem turns to which approach to use for specific data. Since each clustering algorithm has its own pros and demerits, it is difficult to conduct a fair comparison using experiments. Given these circumstances, we believe that designing a fair experiment to demonstrate the superiority of our proposed methods is not quite straightforward, and that's outside the scope of this paper. (Please let us know if we misinterpret your intention here). Following your comment, however, we will provide a brief discussion of the potential extension to other clustering algorithms. 3. \[Plug-in KDE\] As pointed out, our plug-in estimator essentially inherits pros and cons of the standard KDE (in level-set clustering). In theory and practice, the choice of the kernel affects the performance of KDE through the bandwidth, and when the bandwidth is appropriately chosen, the shape of the kernel has little effect on the performance. In kernel density estimation (KDE), Gaussian kernel has the advantage that the number of modes monotonically decreases as the bandwidth increases (Silverman, 1981, Using kernel density estimates to investigate multimodality). In kernel regression, Epanechnikov kernel is the most efficient in the constant term of mean integrated squared error, but other usual kernels (including Gaussian) are at least \~90% efficient. Other than these effects, the shape of the kernel does not affect the convergence rate and has little effect on learning error by less than \~10% on the constant. The choice of kernel has no effect on our results (but it may in real data studies). 4. \[Issues with presentation\] Thank you for pointing these out. - We shall remove the $\equiv$ signs, which had been often used to emphasize notational \"equivalence,\" and replace them with standard equal signs. - We totally agree it should be better to broke down the definition 3.1 as in the original work of Balcan et al. Will do that accordingly in the revised manuscript (or at least in the appendix). - Thank you for correcting the typos; during revision, we will fix everything and proofread thoroughly. 5. \[Ans to Q1\] As you have mentioned, the Hausdorff distance $H(S_{1}, S_{2})$ is oblivious to the density but only sensitive to the envelope: given that we already have enough points in a specific region, adding more points does not meaningfully change the Hausdorff distance. However, we are looking at the level sets $L_{t,h}=\\{w\in \mathbb{R}^{q}: p_{h}(w)>t\\}$ for $t>0$, and their estimators $\hat{L}\_{t,h}$ for $t>0$. Then the density information is encoded through the data points in the level sets. For example, if one distribution has high density regions but the other does not, then for a large $t>0$, the level set ${L}\_{t,h}$ for the first distribution would be nonempty, while the level set for the second distribution would be empty. Hence, although the Hausdorff distance is oblivious to the density, the density is encoded through the level sets $L\_{t,h}$, and measuring the difference by the Hausdorff distance behaves sensitively with respect to the density of data points. 7. \[Ans to Q5\] We believe your insight is correct. We don't yet have a good answer to how to reconcile with the related structural causal learning tasks. As we leave comments for reviewer STix, causal clustering using the SCM approach may yield different results from ours. We will at least comment on this in the discussion section, with the hope of opening up new options for subsequent research. 8. \[Ans to Q6\] You are correct - just like Balcan et al (2014) our estimator requires tuning the two noise parameters. We do not propose any good solution for parameter selection in our paper. Nonetheless, Balcan et al. have empirically showed robustness to such parameter tuning, thus we believe our method can inherit this property, albeit further work is needed because these parameters play an essential part in our method. We will add this discussion in the revised work. --- Rebuttal Comment 1.1: Comment: I thank the reviewers for their effors in responding to my questions. Thank you for your responce to Q1, as it does address my question. I believe your answer to Q5 is something that could be considered as a limitation of the general approach of this method, which does still not seem to be addressed in this work. If accepted, I believe you should also make the limitation of your answer to Q6 explicit in your work. I am also aligned with remarks of other reviewers on the extent of novelty. Overall, I will be maintaining my gently positive score, albeit with limited fervor. --- Rebuttal 2: Comment: Thank you very much for your valuable feedback. As you suggested, we will add a new paragraph on Q5 and Q6 as the limitation of our work in the discussion section, hoping it opens a new avenue for future research. We are confident that incorporating your suggestions above will significantly enhance the quality of our paper. Title: Thank you	Summary: The authors propose an extension of existing causal (i.e., treatment effect heterogeneity) k-means clustering techniques to hierarchical and density-based clustering, including novel estimators and convergence guarantees. Edit: increased rating from 3 to 7 (with the understanding that more qualified people (ACs/PCs/ethicists) will look into the plagiarism concern) Strengths: The topic of causal clustering has already been addressed from variety of perspectives in previous work (e.g., kernel methods suitable for k-means, hierarchical, and density-based clustering, as well as focusing on heterogeneity in terms of causal structure or causal mechanisms), but this work nevertheless finds an original approach (hierarchical/density-based clustering for treatment effect heterogeneity), which is clearly motivated, has high quality theoretical justifications, and should be significant in the (theoretical and applied) causal inference community. Weaknesses: The main weakness (and why my overall rating is a 3 instead of more like a 6 or 7), is the inappropriate verbatim copying from reference [36] (Kwangho Kim, Jisu Kim, and Edward H Kennedy. Causal k-means clustering. arXiv preprint arXiv:2405.03083, 2024.), including the following (and I assume more): - lines 7,8: "We present..." - lines 18--20: mostly copied verbatim from abstract of [36] - lines 91--99: in its entirety - lines 104,105: "If all coordinates..." The figures (espcially 4, which requires over 400% maginification) are too small to be legible on standard paper sizes, making it harder to understand or corroborate the results described in the text. Furthermore, the unreasonably small figures save up to a page of space, allowing more text to be squeezed into the page limit, which seems unfair considering the submission guidelines. I find the related work discussed in Section 1.2 to be lacking. I would expect to see some reference and discussion/comparison with other previous work on causal clustering (which has focused more on heterogeneity in terms of causal mechanisms and causal structure rather than treatment effect), for example including: - Hu, S., Chen, Z., Partovi Nia, V., Chan, L., & Geng, Y. (2018). Causal inference and mechanism clustering of a mixture of additive noise models. Advances in Neural Information Processing Systems, 31. - Huang, B., Zhang, K., Xie, P., Gong, M., Xing, E. P., & Glymour, C. (2019). Specific and shared causal relation modeling and mechanism-based clustering. Advances in Neural Information Processing Systems, 32. - Saeed, B., Panigrahi, S., & Uhler, C. (2020). Causal structure discovery from distributions arising from mixtures of DAGs. In International Conference on Machine Learning (pp. 8336-8345). PMLR. - Markham, A., Das, R., & Grosse-Wentrup, M. (2022). A distance covariance-based kernel for nonlinear causal clustering in heterogeneous populations. In Conference on Causal Learning and Reasoning (pp. 542-558). PMLR. Maybe experimental comparison against some of the above methods is also possible/desirable? Technical Quality: 2 Clarity: 2 Questions for Authors: 1. line 134: Is $d$ a distance function (which by definition is positive when evaluated on distinct objects) or some weaker notion with image $[-1, 1]$? 2. line 164: Can the authors elaborate on why "the true target hierarcy... is an infinite set of clusters"? 3. line 338: What are some examples of such subsequent learning tasks? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Assumptions are clearly stated throughout the text, and general limitations are explicitly discussed at the end. However, I would expect to see some discussion about how this work (which facilitates targeted interventions on specific subgroups) relates to issues of fairness/bias. At the very least, a more complete answer to Question 10 in the author checklist should be given, following the guidelines: "If the authors answer NA or No, they should explain why their work has no societal impact". Flag For Ethics Review: ['Ethics review needed: Data privacy, copyright, and consent', 'Ethics review needed: Discrimination, bias, and fairness', 'Ethics review needed: Deception and harassment'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and thorough feedback. We have addressed each of your concerns as outlined below. 1. [Originality and related work] Thank you for highlighting the connection between our research and other works in the literature on causal discovery or structural causal models. To our knowledge, there are three common ways to express causal/counterfactual quantities: (1) structural equations, (2) causal graphs or structural models, and (3) potential outcomes (counterfactuals). While these languages can complement each other, the target parameters, assumptions, notations, and techniques often differ. We acknowledge that the prior works you mentioned also address the notion of "causal clustering." However, they fall into the second approach (in the context of learning structural heterogeneity), whereas ours is based on the third (learning treatment effect heterogeneity). This distinction is why we did not include the extensive body of work in causal discovery/structural causal models. To the best of our knowledge, Kim et al. (2024) were the first to formally study the task of analyzing treatment effect heterogeneity via clustering based on the (potentially multivariate) CATE within the potential outcomes framework. In our revised manuscript, we will clearly highlight this difference and define our "causal clustering" problem. Additionally, based on your comments, we will add a separate subsection in Section 1 explaining the connection between our approach and others in the causal structure learning literature. 2. [Figures and Presentation] Thank you for pointing this out. We agree that some figures, particularly Figure 4, are too small. Since the essential idea to be presented is quite macroscopic (i.e., cluster patterns), we believe that expanding it slightly larger with increased font and legend size can alleviate this issue. We will address this in the amended text, and if necessary, we will move some figures to the Appendix. 3. [Inappropriate verbatim from [36]] We appreciate your suggestion regarding this matter. Our reliance on the previous work of Kim et al. (2024) is mainly for describing the motivation, problem, and setup in Sections 1 and 2. The framework is novel within the community, and our intention is to present the problem accurately without distorting its description. (As mentioned above, our problem differs significantly from the causal clustering framework in the causal structure learning literature.) However, we acknowledge that we should avoid using verbatim as much as possible and rephrase where necessary. We will fully address this in the revised text. Nonetheless, we would like to emphasize that Sections 3 and 4, which constitute the main contribution of our paper, are entirely original and do not contain any issues related to verbatim content from [36]. Therefore, in our opinion, this should not significantly diminish the value of our paper. 4. [Ans to Q1] We apologize: it should be a distance with the image $[0,1]$. We will fix this. 5. [Ans to Q2] We apologize for the confusion. As you noted, it is an erroneous expression, and part of the sentence should be revised as follows: "... with respect to the true target clustering, because we build a set of nested clusters across various resolutions (a hierarchy) such that the target clustering is close to some pruning of that hierarchy." 6. [Ans to Q3] These could be utilized, for example, to develop precision medicine or optimal policy. We will add specific examples in the discussion section. 7. [Ethics Review] This is a very good point. In the last section, we will discuss how the discovered subgroup was formed simply based on similarity in treatment effect, without considering factors such as fairness/bias. We believe that discovering a "fair subgroup" would be an intriguing future project of our work. If there are any remaining issues, please feel free to let us know through your comments. --- Rebuttal Comment 1.1: Comment: Thanks for the thorough rebuttal---it addresses all of my concerns! The other reviews and corresponding rebuttals also gives me a more favorable view of the paper. I have increased my rating from 3 to 7, with the understanding that more qualified people (ACs/PCs/ethicists) will look into the plagiarism concern. In summary: Causal clustering for treatment effect heterogeneity is a well-motivated problem of practical importance, and this paper adds to the (limited) existing literature in a natural direction, offering solid theoretical results and basic proof-of-concept empirical results. I would expect this paper to have high impact in the causal inference community (considering both the theoretical and practical sides). --- Rebuttal 2: Title: Thank you Comment: We appreciate your helpful questions and comments, and we feel that the revised manuscript will be significantly improved as a result. Most importantly, we appreciate the reviewer pointing out the issue of several inappropriate verbatims while we define the problem and setup in Section 2. We take this issue extremely seriously and will fully address it in the revised manuscript. Please feel free to let us know if you have any further concerns or questions.	Summary: The paper deals with problems arising in understanding treatment response/effects and in particular evaluating subgroup effects building on recent work using causal k-means clustering. The main contribution of the paper is to circumvent the k-means approach and suggest a hierarchical and also a density-based clustering approach. The authors present associated estimators with their proposed methods and rate of convergence, thus extending the framework of causal clsutering. The paper is motivated by the study of heterogeneous treatment effects via clustering and highlights the drawbacks of prior work based on k-means. The authors extend this by applying density-based clustering which has the advantage of finding clusters with arbitrary shapes and sizes and appears to be more robust to noise and outliers. Similarly, the observe that hierarchical clustering has some advantages in scenarios where the data are nested or are forming hierarchies. The main results are: -Th. 3.1: The authors analyze the robust hierarchical clustering algorithm appearing in prior works [5], in the context of causal clustering. Under certain assumptions, that are related to so-called good-neighborhood properties as defined in [5] the authors manage to show that having access to a small random subset of the data can allow for their algorithm to have small error on the entire data set. -Th. 4.1: Analogous statement for the density-based clustering methods. Finally the authors presented experiments for studying finite-sample properties of their proposed plug-in procedures using simulated data. Strengths: +analysis for robust hierarchical clustering and density-based methods seems interesting +natural algorithms and problems are well-motivated Weaknesses: -despite the well-motiaveted setting, the reviewer believes that the paper has a lack of novelty: the algorithms are already from previous works, and most of the approaches are what was expected. -presentation can be improved. As of now, it is a bit cryptic, as the main tasks/problems are not well-defined and rather are to be implied from the context. I would have loved to see a clean definition of the problem that is being solved, the challenges and the novel approach. As of now, I believe the results are a plug-in approach based on two well-studied algorithms. Technical Quality: 3 Clarity: 2 Questions for Authors: -To improve the results, I was wondering if there are other notions of error, for which your results can be extended? What if we measure error for example on the hierarchy that is being found? There are various notions for comparing hierarchies in the literature and I believe this could be interesting and novel extension of your work. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: -see above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and valuable feedback. We would like to address each of your major concerns in the following. 1. [Novelty of the problem] Thank you for pointing this out. First, we want to emphasize that the problem of causal clustering is novel; clustering on counterfactual outcomes provides a new framework for analyzing treatment effect heterogeneity, and as far as we know this approach has not addressed (or at the very least, formally addressed) in the literature before, except in Kim et al. (2024). It also differs significantly from other previous attempts in the cluster analysis literature that considered partially observed outcomes or clustering with measurement errors, since in our case the variable to be clustered consists of completely unknown counterfactual functionals (See Section 1.2 of Kim et al. (2024)). Due of the page limits, we just briefly describe the problem in Section 2 and refer readers to Kim et al. (2024) for details, which could confuse them about the problem's novelty. Based on the reviewer's comment, we will explicitly clarify on the problem, setting, and associated challenges. 2. [Contribution] We would like to stress out that our main contribution is not on the algorithm side, i.e., the use of the robust hierarchical or density-based clustering algorithms. Rather, it is on the theoretical side where we provide conditions under which the plug-in estimator works for each algorithm for causal clustering framework. - The problem of causal clustering poses challenges which do not appear in the previous studies, as we cluster on unknown counterfactual functionals that is to be estimated nonparametrically. Surprisingly, a plug-in approach does NOT always works without strong extra conditions. Kim et al. (2024) showed that, in the case of k-means, even the plug-in estimator will fail without the margin condition, which requires local control of the probability around the Voronoi boundaries; without such structural assumptions, the error cannot be bounded. - The main contributions of our work are found in Sections 3 and 4, where we prove that the two appealing off-the-shelf cluster-analysis techniques can be successfully adopted in the framework of causal clustering using simple plug-in methods without requiring extra strong structural assumptions as opposed to k-means, which requires the margin condition (kim et al. 2024). As outlined in the proof, verifying this is more complicated than one may think. (We will give a brief exposition why this could be a theoretically difficult task in the revised text.) Our plug-in approaches could also be readily extended to clustering with generic pseudo-outcomes, even outside the context of causal inference. This versatility may be considered an additional contribution. 3. [Presentation] As mentioned earlier, and as suggested by the reviewer, we will ensure that our revised paper is more self-contained. We believe this will help clarify the novelty of the problem and our key contribution, ensuring that readers do not experience any confusion. 4. [Ans to Question] Thank you for this insightful question. There are two types of error we consider in our work: estimation error of the unknown counterfactual functionals (i.e., the identified regression functions $\{\mu_a\}$) and error regarding to clustering accuracy. I guess your question is related to the latter, or more specifically, whether one may adopt other types of clustering algorithms that can incorporate uncertainty on the hierarchy that is being found. We completely agree that this could lead to fascinating future work, and we will include it in the revised manuscript's discussion section. We hope the above response addresses your concerns. However, if there are any remaining issues, please feel free to let us know through your comments. --- Rebuttal 2: Title: ACK of responses Comment: The reviewer has read the response and thanks tha authors for their time. However given the presentation issues, and the weaknesses raised, the reviewer still thinks the paper should be improved before publications and that it is not ready as is. --- Rebuttal 3: Comment: Thank you for your input. We believe incorporating your comments on presentation will improve the quality of the paper. Title: Thank you	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Transfer Q-star : Principled Decoding for LLM Alignment	Accept (poster)	Summary: This paper introduces 'Transfer Q', which is a decoding strategy to align language models with target reward models without any fine-tuning. The main contribution of this paper is the estimation of the optimal Q function, which is often unavailable in practice and is necessary for approximating the RL optimal policy, through DPO (Direct Preference Optimization). The paper first introduces an algorithm for direct transfer decoding using DPO-trained models as optimal baselines, and then presents indirect transfer decoding that uses these as conservative baselines. The authors describe the theoretical characteristics of this methodology and experimentally demonstrate that it achieves state-of-the-art performance on target reward models. Strengths: Aligning to specific values through decoding algorithms without fine-tuning is highly useful. In particular, estimating the token-level Q* function in this process is very challenging, and existing works typically train external networks to estimate these values. This paper uses DPO-trained models as baselines to estimate the token-level Q* function, as DPO learns the Q* function in an offline manner. Given recent reports that DPO-trained models predict preferences better than classification-based reward models, this approach is very interesting and can be considered a valuable contribution. Additionally, since DPO-trained models are offline trainers and thus may be distant from the optimal Q* function, the introduction of an indirect transfer decoding algorithm that treats them as suboptimal, along with the presentation of its theoretical properties, is also a significant contribution. Weaknesses: 1. Equation 5 on line 121 is introduced roughly. Its derivation should be described in the Appendix, or references discussing it should be introduced. 2. The introduction of Controlled Decoding (CD) is incorrect. It states that CD uses $Q^{\pi_{sft}}$ as a proxy for Q, but in reality, it uses an energy-based model of $Q^{\pi_{sft}}$ and a value function through an external network as Q. The difference between this work and CD is that CD uses FUDGE and Q function as a baseline, while the proposed TQ* uses DPO as a baseline reward model. Therefore, Figure 1 and its explanation are incorrect. Considering that the indirect method, which conservatively estimates the target model, is the contribution, Figure 1 should be drawn in reverse. 3. Equation 12 on line 175 is believed to be incorrect. Theoretically, the $\pi(z\|s)$ used in this equation should be the reference model $\pi_{ref}$ before exploration, not the aligned model $\pi_{BL}$. The equation development started with the premise that the optimal policy can be expressed as an energy-based model of the reference model and reward model, but in equation 11, it roughly transitions from the reference model to the aligned DPO, losing theoretical justification. In this case, it's expected that the original distribution will collapse significantly and over-optimize to the reward model. 4. A crucial experiment on the trade-off between KL divergence and the reward model is missing. In RL, there's a strong trade-off in reward model optimization depending on how much KL penalty is given, so it's not sufficient to only check the performance against the reward model as in this paper's experiments. While Figure 2(b) touches on this, it's not enough. An experiment showing whether it demonstrates pareto-frontier performance in the trade-off relationship according to KL penalty is also needed. This experiment is particularly important given the use of the DPO-trained model as the reference model, as mentioned in weakness 3. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. In Figure 2 (b), when measuring the KL divergence for TQ, which model was used to compare the decoding results? For a fair comparison, it should be measured against $pi_sft$, like the other baselines. 2. I'm curious about the inference cost. How do the memory requirements and time complexity compare to naive greedy decoding? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: This paper has done truly promising work, but it has the following limitations: 1. Although DPO is used as a baseline, it is an offline trainer and there is a theoretical gap with the optimal Q. This issue is addressed through the indirect transfer decoding part, but there is still room for improvement, and various discussions are likely to follow. 2. Due to the use of the DPO model as the initial policy, the theoretical justification is weakened, and a clear interpretation of the experiments is not possible. 3. In the process of training the DPO model, over-optimization towards the reward model is also controlled by the KL penalty. It would be interesting to add analysis and experiments on this aspect. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive feedback on our paper. >Weakness 1: Equation 5.... introduced. Response to Weakness 1: We will add the derivation for the closed-form solution in Equation 4 (also shown in CD [B]) as a Lemma in the appendix for better clarity. >Weakness 2: The introduction of CD..... reverse. Response to Weakness 2: We thank the reviewer for this point but there seems to be a confusion. We apologize for any oversight during our explanation of CD in our paper. We take this opportunity to expand on this point more below. We note that the objective of Controlled Decoding which is a KL regularized RL problem (also highlighted in CD) is given by \begin{align} \pi^(\cdot\|s\_t) := \arg \max\_{\pi} \mathbb{E}\_{z \sim \pi(\cdot\|s\_t)}[Q^(s\_t, z)] - \alpha KL (\pi(\cdot\|s\_t), \pi\_{\text{SFT}}(\cdot\|s_t)) \tag{4}. \end{align} However, it is extremely crucial to note that in order to obtain the optimal policy in (4), the Q-function in the equation should be $Q^(s\_t, z)$ (the optimal action-value function i.e the reward return calculated with optimal policy) and not any arbitrary $Q$ function from any policy. However, the approach in CD indicates that the $Q$ estimation (either directly or through an external network) is done through the data generated using $\pi\_{sft}$ policy (refer to Equation 1, Equation 4 in [B]), leading to sub-optimality as demonstrated in our Figure 1. Our main contribution is to show that with an access to an aligned model, we can estimate $Q^{}(s_t, z)$ in a much more efficient way than CD as demonstrated in all our experimental settings. *Possible source of confusion:* However, we believe the source of confusion is that in the CD paper, the definition of action value function used in (4) above (which is equation (1) in CD paper [B]) utilized $\pi_{sft}$ to sample but denotes the value function with $V^$ which is usually reserved for the optimal value function. >Weakness 3:* Equation 12 on line 175 .... the reward model. Response to Weakness 3: We thank the reviewer for providing us with this opportunity to clarify further. *Solution in (12) is theoretically correct and justified.* We have stated our proposed decoding optimization problem in Equation (11), and solving that leads to the choice of having the reference policy as $\pi_{\text{BL}}$ in Equation 12 on line 175. This is a specific choice of our algorithmic solution design which we have also implemented in our work. However, we want to emphasize that our theoretical results in Sec. 3.3 (specifically the KL divergence upper bound in Theorem 1, statement (2)) derives the divergence of proposed algorithm's policy to the original reference policy $\rho_{\text{SFT}}$. The upper bound of Theorem 1 [statement 2] is given by \begin{align} \mathbb{D}\_{\text{KL}}(\rho^\_{Alg}(\cdot\|\mathbf{x}),\rho\_{SFT}(\cdot\|\mathbf{x}) )\leq(\frac{1}{\beta}+\frac{1}{\alpha}T)r\_{\max}. \end{align} By controlling the value of $\alpha$ and $\beta$, we can control the distance to the original reference model. Moreover, in our experimental ablation Figure 2(b) we show the KL divergence to the original SFT reference model which proves that the proposed method does comparable to other baselines and is not diverging from the SFT policy. >Weakness 4:* A crucial experiment on t...... in weakness 3. Response to Weakness 4: As suggested by the reviewer, we performed additional experiments to generate the pareto-frontier results. To be specific, in Figure-3 in the [rebuttal PDF](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf), we compare the tradeoff between the win-rate and the KL divergence to the base reference SFT policy. TQ* outperforms existing baselines. We will add this to the final version. >Question 1: In Figure 2 (b), when measuring the KL divergence for TQ, which model was used to compare the decoding results? For a fair comparison, it should be measured against $\pi_{\text{SFT}}$ Response to Question 1:* Yes, you are right. We have indeed used the reference model $\pi_{\text{SFT}}$ for all the algorithms while measuring the KL divergence for TQ. >Question 2:* I'm curious about the inference cost. How do the memory requirements and time complexity compare to naive greedy decoding? Response to Question 2: We report the inference time-complexity of all the existing deocding algorithms. Ours is comparable to state-of-the-art decoding methods. \| Algorithm \| Inference Time \| Avg Reward \| \|------------------------\|----------------\|------------\| \| Naive Decoding \| 3s \| 0.13 \| \| ARGS \| 7s \| 0.29 \| \| $\text{CD}^{--}$ \| 40s \| 0.71 \| \| $\texttt{TQ}^{\star}$ (Ours) \| 41s \| 1.0 \| Reference: [B] Sidharth Mudgal et al., Controlled decoding from language models, 2024 --- Rebuttal Comment 1.1: Comment: The author still lacks a theoretical foundation. Whether it's a Q-function or a Value function, they end up having the same representation in this algorithm due to the (both tractable or intractable) normalizer. Furthermore, although a theoretical guarantee has been presented regarding the excessively diverging KL-div, sufficient experiments and defense have not been conducted. Therefore, I maintain my current assessment. --- Rebuttal 2: Title: Clarifications regarding core technical contributions and new experiments [Pareto Front plot] in rebuttal pdf Comment: > Comment 1.1: The author still lacks a theoretical foundation. Whether it's a Q-function or a Value function, they end up having the same representation in this algorithm due to the (both tractable or intractable) normalizer. Response: Thank you for your comment. We apologize if we missed anything but this comment is not very clear for us to respond to. We have followed the standard definitions from the reinforcement learning literature [A]. We start by explicity defining the token level MDP in Section 2.1. We emphasize that the Value function is defined for each state (defined in line 210 in our paper), and the Q function is for each state and action (defined in Equation (2) in our paper). **Request to Reviewer: We kindly request the reviewer to please expand on why value and Q function will be the same and what is the meaning of the term "normalizer" in this context. We want to clearly understand the concern before responding. Thank you for your feedback and engagement. Our Focus and Contributions:** In our work, we want to emphasize that our focus is on the estimation of optimal Q* and show that it is better as compared to all existing decoding methods such as CD, ARGS, etc. We show this in theory as well as experiments. [A] Sutton, Richard S., and Andrew G. Barto. Reinforcement learning: An introduction. MIT press, 2018. > Comment 1.2: Furthermore, although a theoretical guarantee has been presented regarding the excessively diverging KL-div, sufficient experiments and defense have not been conducted. Therefore, I maintain my current assessment. Response: Thank you for your comment. We want to highlight that our work contributes on both theory and experiments side. *On the theory side,* our work is the *first to derive such a theoretical upper bound* for both (1) suboptimality and (2) KL divergence for a decoding algorithm. There are no theoretical results in any of the existing works (such as CD, ARGS etc), which constitutes a unique and novel contribution of our work on its own. *On the experimental side,* we have tested our proposed approach on six evaluations (in the main submission) and added two more large-scale evaluations in the rebuttal pdf [[link to pdf in openreview](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf)] . - For the KL divergence plot, we have Figure 2(b) in the main body and as the reviewer suggested, we added a Pareto front plot for evaluation 1 (*Figure 3 in the rebuttal pdf) as well in the rebuttal pdf [[link to pdf in openreview](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf)] which clearly shows the superior performance of our method. - Additionally, our current comparison includes win rate, coherence, reward, and diversity, which are designed to approximate human preferences. We are running more experiments and committed to adding them in the final version of our work. We believe we have addressed all the concerns, and are happy to engage in further discussions if any remain. Thank you once again for your time and consideration. Looking forward to your feedback. --- Rebuttal Comment 2.1: Title: Additional Pareto Front Results for Two more Evaluation Setups Comment: Thank you for your time and efforts in reviewing our paper and rebuttal discussions. To address your comment regarding the Pareto front plot in the experiments, we remark that we added the Pareto front plot on Evaluation 1 in the rebuttal pdf (Figure 3) [[link to pdf in openreview](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf)]. Additional Experimental Results:* To further strengthen our empirical evaluations, we ran experiments and obtained Pareto front results for two more evaluation setups: Evaluation 2 and 3 (details in the paper Table 1). We present the results in the form of tables here. We are committed to adding them for all the evaluations in the final version of our paper. - *For Evaluation 2 Setup* (detailed in Paper Table 1): This table shows the value of KL and the corresponding win rate and shows that our proposed method outperforms the existing methods. \| Method \| \| \| \| \| \| \| \| \| \| \|------------------------------\|----------\|-------\|-------\|-------\|-------\|-------\|-------\|-------\|-------\| \| ARGS-DPO \| KL \| 0.40 \| 1.15 \| 2.20 \| 3.80 \| 5.75 \| 7.05 \| 8.20 \| 9.15 \| \| ARGS-DPO \| Win-Rate \| 50.50 \| 57.80 \| 61.75 \| 65.90 \| 67.10 \| 67.70 \| 68.20 \| 68.15 \| \| $\text{CD}^{--}$ \| KL \| 0.50 \| 1.25 \| 2.35 \| 4.20 \| 6.50 \| 8.75 \| 9.35 \| 10.75 \| \| $\text{CD}^{--}$ \| Win-Rate \| 50.75 \| 62.85 \| 68.90 \| 72.70 \| 75.40 \| 76.00 \| 76.15 \| 76.30 \| \| $\texttt{TQ}^{\star}$ (Ours) \| KL \| 0.42 \| 1.20 \| 2.18 \| 3.85 \| 5.95 \| 7.90 \| 8.85 \| 10.40 \| \| $\texttt{TQ}^{\star}$ (Ours) \| Win-Rate \| 54.30 \| 70.90 \| 75.70 \|79.60 \| 80.55 \| 81.95 \| 82.95 \| 83.25 \| - *For Evaluation 4 Setup* (detailed in Paper Table 1): This table shows the value of KL and the corresponding win rate and shows that our proposed method outperforms the existing methods. \| Method \| \| \| \| \| \| \| \| \| \| \|------------------------------\|----------\|-------\|-------\|-------\|-------\|-------\|-------\|-------\|-------\| \| ARGS-DPO \| KL \| 0.37 \| 1.26 \| 2.05 \| 3.71 \| 5.86 \| 7.14 \| 8.36 \| 9.23 \| \| ARGS-DPO \| Win-Rate \| 50.10 \| 58.32 \| 62.10 \| 66.13 \| 67.32 \| 67.89 \| 67.41 \| 66.02 \| \| $\text{CD}^{--}$ \| KL \| 0.45 \| 1.32 \| 2.39 \| 4.36 \| 6.58 \| 8.85 \| 9.50 \| 10.89 \| \| $\text{CD}^{--}$ \| Win-Rate \| 51.05 \| 63.12 \| 69.44 \| 73.16 \| 75.80 \| 76.25 \| 77.00 \| 77.17 \| \| $\texttt{TQ}^{\star}$ (Ours) \| KL \| 0.38 \| 1.27 \| 2.11 \| 3.95 \| 6.05 \| 8.03 \| 8.98 \| 10.58 \| \| $\texttt{TQ}^{\star}$ (Ours) \| Win-Rate \| 50.86 \| 69.45\| 73.28 \| 76.19 \| 79.20 \| 80.19 \| 81.00 \| 82.16\| We believe we have thoroughly addressed all concerns and are more than happy to engage in further discussions if any additional issues remain. Thank you so much again for your consideration.	Summary: This paper addresses decoding for aligning large language models, which is a process of inference-time, token-level optimisation without updating the parameters of the LLM. Two scenarios are considered: 1. where a baseline policy is given and is aligned with the target trajectory-level reward. 2. where the baseline policy is given and aligned with a different trajectory-reward than the target one. Two approaches are provided in accordance: one named direct transfer, which as the name suggests directly derive a token-level policy from the baseline LLM, and the other named indirect transfer, whose key lies in an importance sampling trick to reweigh using the ratio between the target and baseline trajectory policies. The authors also provide theoretical analysis characterising the sub optimality gap and the divergence from the supervise-fine-tuned model. Empirical results demonstrates superior performance in several key metrics. Strengths: This paper is very well written, poses a reasonable problem, namely how to more efficiently decode an existing trajectory level policy into a token level policy, and provides an elegant solution. The proposed approach draws on latest developments in the field (DPO) and enjoys a rigorous theoretical characterisation. The gap between the trajectory level learnt policy and the sft-policy is interpretably characterised in terms of the regularisation coefficients in front of the KL terms. The authors finished with a nice empirical analysis showing the efficacy of their method. Weaknesses: Morally speaking, it is not clear to me why should decoding this way works much better than directly taking $\rho_{BL}$- at least in the case of direct decoding. After all, TQ* is just the 1-step-action-value function of $\rho_{BL}$, so I think the readers would benefit from an explicit explanation of why this process of obtaining $\rho_{BL}$ --> estimating its (1-step-action) value function --> compute the closed form of the optimal policy of this value function with a KL term should do better than just directly using the trajectory level policy which is already the (regularised) optimal of the reward function $r$. More concretely, it would be nice to have this shown as a theoretical result - a theorem that shows that the suboptimality gap of the policy proposed by the authors is smaller than $\rho_{BL}$. I appreciate that the authors have shown this empirically, though. For the indirect case, the method proposed by the authors require access to the ground truth trajectory-level reward function. But if we had this in practice, would we not have just directly used it to compute $\rho_{BL}$? Thus, I think here it makes sense to add some explanation regarding why this is useful. Finally, I find it a bit of a misleading juxtaposition with DPO - the point of DPO is that we no longer need to model the reward function, but in this work we rely on both the trajectory level optimal policy learned by DPO as well as the reward function which also needs to be learned. Both of these suffer from statistical/optimisation instability in their own right. So it is strong to assume you have both of these. Technical Quality: 4 Clarity: 4 Questions for Authors: equation (2) - This seems like a typo since for i > 0, the first argument of R should also take in the so-far generated sequence z_0 ... z_{i-1}. equation (15) - how is $Z_{BL}$ and $Z_r$ computed? This is needed later in the importance ratio, if I understood correctly, but I don't see how to compute these quantities which are usually intractable. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Aside from the limitations I mentioned in Weaknesses, the authors also acknowledges that TQ* suffers from increased time complexity. They suggest that this can be mitigated by training a small value function adapter as discussed in prior work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments. >Weakness 1: Morally speaking, it is not ...... empirically, though. Response to Weakness 1: We believe the reviewer wants to understand the advantage of directly generating the response using $\rho_{\text{BL}}$ over TQ. We explain in detail below. Our proposed method enables optimal action at the Token Level.* In our algorithm (TQ), for every state ($s_t$) we compute the action value function $Q^(s_t,z)$ using $\rho_{\text{BL}}$ and select the optimal action by maximizing the action value function $Q^(s_t,z)$, ensuring the optimal performance for any token as highlighted in Theorem 1. Intuitively, it ensures the possibility of performing better credit assignment with token level MDP while decoding, when compared to $\rho_{\text{BL}}$ which is trained as a contextual bandit for trajectory responses. This aspect has also been highlighted in the very recent work [C]. As a result, the majority of the prior decoding methods demonstrate improvement over directly using the DPO policy Theoretical Justification:* From Theorem 1 in the paper, we can extract the theoretical justification of the improvement of our proposed algorithm over $\rho\_{\text{BL}}$. The sub-optimality gap of our algorithm, defined in Equation 18, can be decomposed into two terms ($\Delta \leq \Delta_1 - \Delta_2$) as detailed in Appendix G.1 (Equations 28 and 29). The first term, $\Delta_1$, is a function of parameter $\beta$ and represents the sub-optimality from the optimal value function due to $\rho\_{\text{BL}}$, with its upper bound $\Delta\_1\leq \beta\mathbb{D}_{\text{KL}}[\rho^(\cdot\|\mathbf{x}) \mid \mid \rho\_{\text{sft}}(\cdot\|\mathbf{x})]$. This indicates the sub-optimality incurred at the token level when using $\rho\_{\text{BL}}$ alone. However, the second term, $\Delta_2 = \alpha h_{\alpha}$ , is always non-negative, illustrating that our algorithm provides an additional benefit of improving the suboptimality gap. By tuning the value of $\alpha$, we can improve the performance beyond what is possible with $\rho_{\text{BL}}$ alone. This improvement is consistently demonstrated across all our experimental results, as also highlighted by the reviewer. >Weakness 2:* For the indirect case, .... is useful. Response to Weakness 2: We believe what the reviewer is asking is if we have access to the ground truth reward, why can't we directly have an aligned $\rho_{\text{BL}}$ for that ground truth reward? We expand on this in detail as follows. *Obtaining aligned $\rho_{\text{BL}}$ to ground truth would require fine tuning.* We note that even if we have access to ground truth reward, obtaining an aligned model $\rho_{\text{BL}}$ using standard DPO would require fine-tuning of parameters which is not the focus of this work. Our motivation is in tuning-free alignment settings, where we don't update the parameters of the language model. *Regarding our approach and decoding methods.* We emphasize that any decoding method would require access to the target reward function (trajectory level). However, for indirect transfer, we leverage any open-sourced baseline model $\rho_{\text{BL}}$ aligned to an arbitrary reward function $r_{\text{BL}}$ which is different from the target reward. Thus, for principled decoding even though we have $r_{\text{target}}$, still we can't directly use $\rho_{\text{BL}}$ to estimate action value function and decode, which will result in sub-optimal decoding due to the distribution shift. Thus, we need to estimate $\rho_{\text{target}}$ using Equation 16. >Weakness 3: Finally, I find it a bit of a ....... So it is strong to assume you have both of these. Response to Weakness 3: We remark that any decoding method requires the access to target reward function for alignment. As reviewer mentioned, if the target reward function is learned through preferences, there will be statistical error due to coverage and suboptimality of optimization methods. Furthermore, the DPO policy could also have it's own instability issues which might affect the final performance of our decoding method. We agree with the reviewer and will highlight this limitation in the revised final version of our work. Examples of Access to Ground truth reward: We note that there can be true rewards/scores that don't need to be always trained from data, for example - coding, and mathematical tasks where one might have fixed ground truth avoiding above mentioned statistical errors. >Question 1: equation (2) - This seems like a typo ....sequence z_0 ... z_{i-1}. Response to Question 1: Thanks for pointing out the typo, we will fix it in the final version of the paper. The first argument should be $s_{t+i}$ to make sure it take in so far generated sequence. >Question 2: equation (15) - how is $Z_{\text{BL}}$ and $Z_r$ computed.... intractable. Response to Question 2: This is a good question. We note that for the theoretical analysis in this work, we assume access to the ratio mentioned in Equation (15). For the experimental purposes, we utilized unbiased estimates. For instance, we can estimate the partition from collecting samples from $\rho_{\text{SFT}}$ and evaluating the empirical estimate of $\mathbb{E}\_{y \sim \rho\_{\text{SFT}}}[\exp(\frac{1}{\beta}r(x,y))]$. Regarding this ratio explicitly, we had an interesting observation that for several realistic scenarios of transfer, as observed in our empirical experiments, when either the reward difference $r_1(x,y) - r_{\text{BL}}(x,y)$ is small or the reference policy is not aligned to any specific reward $r(x,y)$, the ratio of the partition functions will be 1. This would effectively relax the computational bottleneck for the implementations. [B] Sidharth Mudgal et al., Controlled decoding from language models, 2024 [C]. Rafailov, R et al., From r to q*: Your language model is secretly a q-function. arXiv preprint arXiv:2404.12358. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my comments. I'm satisfied with maintaining my original score.	Summary: The paper Transfer Q⋆: Principled Decoding for LLM Alignment proposes a novel approach to aligning large language models by leveraging a principled decoding strategy. The authors propose a method to estimate the optimal Q-function for decoding using an existing aligned policy, addressing limitations in previous approaches like Controlled Decoding (CD). The paper presents an indirect transfer method, allowing for alignment even when the baseline model is trained on a different reward function. The authors provide theoretical analysis characterizing the sub-optimality gap and KL divergence to the reference policy. Extensive experiments across multiple datasets and model architectures demonstrate the effectiveness of TQ⋆ compared to existing methods. Strengths: Novelty: The paper introduces an original approach to LLM alignment via decoding, leveraging existing aligned policies to estimate the optimal Q-function. Theoretical foundation: The authors provide a rigorous theoretical analysis of their method, including bounds on the sub-optimality gap and KL divergence. This adds credibility to the approach and helps explain its effectiveness. Comprehensive evaluation: The experimental section is thorough, covering multiple datasets, model architectures, and evaluation metrics. The inclusion of both synthetic and real transfer tasks demonstrates the method's robustness. Practical relevance: TQ⋆ addresses a significant challenge in LLM alignment, offering a computationally efficient alternative to fine-tuning approaches. This has potential implications for improving the deployment of aligned LLMs. Weaknesses: Comparison with Baselines: The comparisons with existing baselines like DPO are insightful, but additional baselines, especially those focusing on inference-time control, could strengthen the evaluation. Hyperparameter sensitivity: The paper does not thoroughly explore the sensitivity of TQ⋆ to its hyperparameters, particularly the decoding alignment parameter α. A more detailed analysis of this aspect would strengthen the work. Scalability: While the method is tested on 7B parameter models, it's unclear how well it scales to larger models that are increasingly common in practical applications. Evaluation metrics: While the paper uses several evaluation metrics, including GPT-4 based assessment, it lacks human evaluation studies. Given the subjective nature of language quality and alignment, human evaluation would provide valuable validation of the method's effectiveness. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Have you investigated the stability of the alignment achieved by TQ⋆ over extended generation sequences? Does the alignment quality degrade for longer outputs, and if so, how does this compare to other methods? 2. How sensitive is TQ⋆ to the choice of baseline model? If multiple baseline models are available, each aligned with different rewards, how might one optimally select or combine them to estimate Q⋆ for a given target reward? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors acknowledge some limitations of their work, such as the potential for hallucination in responses to very obscure queries. However, they could improve their discussion of limitations by addressing: 1. The reliance on existing aligned baseline models, which may not always be available or suitable for all target rewards. 2. The potential for errors or biases in the GPT-4 based evaluation, which is used as a proxy for human assessment. 3. The computational overhead of TQ⋆ compared to standard decoding methods, which may impact real-time applications. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Reviewer Summary: We thank the reviewer for the encouraging remarks and recommending acceptance of our work. We provided detailed responses to other comments one by one as follows. >Weakness 1: Comparison with Baselines: The comparisons with existing baselines like DPO are insightful, but additional baselines, especially those focusing on inference-time control, could strengthen the evaluation. Response to Weakness 1: Thank you for your comment. We have now included a comparison with additional inference time control baselines (aka decoding methods) as well in the table below. Specifically, we report the time taken to decode a single prompt, using the system configuration described in Appendix C in the main paper. \| Algorithm \| Inference Time \| Avg Reward \| \|------------------------\|----------------\|------------\| \| Naive Decoding \| 3s \| 0.13 \| \| ARGS \| 7s \| 0.29 \| \| $\text{CD}^{--}$ \| 40s \| 0.71 \| \| $\texttt{TQ}^{\star}$ (Ours) \| 41s \| 1.0 \| >Weakness 2: Hyperparameter sensitivity: The paper does not thoroughly explore the sensitivity of TQ* to its hyperparameters, particularly the decoding alignment parameter α. A more detailed analysis of this aspect would strengthen the work. Response to Weakness 2: We agree with the reviewer that doing more ablations with respect to hyperparameters would strengthen the work. To this end, as suggested by the reviewer, we did additional ablations experiments to understand the effect of varying the decoding alignment parameter $\alpha$ on the quality of the generated text. Specifically, we performed decoding by varying $\alpha$ such that $\frac{1}{\alpha} \in [0.1, 0.25, 0.5, 0.75, 1, 2, 5, 7.5]$. We report the results in Figure-3 in the [rebuttal PDF](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf). To be specific, we compare the tradeoff between the win-rate and the KL divergence to the base reference SFT policy for different values of decoding alignment parameters. >Weakness 3: Scalability: While the method is tested on 7B parameter models, it's unclear how well it scales to larger models that are increasingly common in practical applications. Response to Weakness 3: Thank you for raising this concern. As the reviewer suggested, we performed an evaluation on two additional setups using larger models, as detailed in table below. \| \| Dataset \| SFT Model \| DPO Model \| Reward Model \| \|--------------\|-------------------------------------\|------------\|------------\|--------------\| \| Evaluation-7 \| HH-RLHF \| LLAMA2-13B \| LLAMA2-13B \| LLAMA2-13B \| \| Evaluation-8 \| OpenAssistant Conversations Dataset \| Pythia-12B \| Pythia-12B \| Pythia-6.9B \| We report the results in Figure-2 in the [rebuttal PDF](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf). >Weakness 4: Evaluation metrics: While the paper uses several evaluation metrics, including GPT-4 based assessment, it lacks human evaluation studies. Given the subjective nature of language quality and alignment, human evaluation would provide valuable validation of the method's effectiveness. Response to Weakness 4: Our current comparison includes win rate, reward model performance, coherence, and diversity, which are designed to approximate human preferences. However, we agree that incorporating direct human evaluation would provide more robust validation of our method's effectiveness. We plan to include human evaluation in the final version and are currently in the process of obtaining the necessary permissions. >Question 1 : Have you investigated the stability of the alignment achieved by TQ⋆ over extended generation sequences? Does the alignment quality degrade for longer outputs, and if so, how does this compare to other methods? Response to Question 1: Thank you for your suggestion. We performed additional evaluation by varying the length of generated text. We report the results in Figure-1 in the [rebuttal PDF](https://openreview.net/attachment?id=6Il3qOI0FO&name=pdf). We observed that irrespective of the length of generated text, $\texttt{TQ}^{\star}$ consistently outperforms all the compared baselines. >Question 2 : How sensitive is TQ⋆ to the choice of baseline model? If multiple baseline models are available, each aligned with different rewards, how might one optimally select or combine them to estimate Q⋆ for a given target reward? Response to Question 2: This is a great question and is a valid scope of future research work. We can try by selecting argmax Q not only over the action token by as well as models itself. We are working as a future research. >Limitations: The authors acknowledge ..... impact real-time applications. Response to Limitations: Thank you for your insightful comments. We will expand our discussion of limitations in the final version to include: (1) The reliance on existing aligned baseline models, noting their availability and suitability challenges for different target rewards. (2) Potential errors or biases in the GPT -4-based evaluation, addressing its limitations as a proxy for human assessment. (3) The computational overhead of TQ⋆ compared to standard decoding methods, considering its impact on real-time applications. --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: Thank you for addressing my concerns with additional experiments and baselines for larger models. I appreciate your commitment to adding comprehensive human evaluation. Based on these improvements, I am satisfied with maintaining my original score for your paper. --- Reply to Comment 1.1.1: Title: Thank you for the response. Comment: Dear Reviewer, Thank you for your response. We are glad that our rebuttal responses were able to address your concerns. Regards, Authors	Summary: The paper proposes a new estimation of Q function for controlled decoding. Instead of using a Q function derived from SFT model, the paper propose to use Q function derived from separate aligned models. Evaluation shows the proposed method can achieve higher reward in benchmarks. Strengths: 1. The idea of using aligned models to estimate Q is quite intuitive as these models closer to optimal policy than SFT model, also the paper proposed mathematical derivations and explanations on how and why aligned models are better for estimating Q. 2. The benchmarks showed the proposed method tend to achieve higher reward, especially it outperforms the controlled decoding baseline along with other alignment methods. Weaknesses: 1. The proposed approach seems computationally expensive to do decoding. 2. The approach require accessing an already aligned model, this adds more requirements for using the method and limits it use cases. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Did the authors compare decoding speed for different approaches? It would be useful to report even thought the proposed approach is slower. 2. For gpt4 based evaluation, why the authors only report Win-Tie instead of splitting win and ties? In particular I am curious what is the percentage of ties for each comparison. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are discussed in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Reviewer Summary: We sincerely thank the reviewer for the thoughtful review of our work, highlighting the strength and novelty in both our formulation and experimental design in leveraging available aligned models for principled decoding. We address all other comments one by one as follows. >Weakness 1: The proposed approach seems computationally expensive to do decoding. Response to Weakness 1: We thank the reviewer for this important point. We agree with the reviewer that performing the proposed principled decoding would indeed add an additional computational overhead at the inference time. But we would like to emphasize that *our main focus and contribution* in this work is to provide principled method to perform LLM alignment via decoding and provide theoretical gaurantees. Moreover, our decoding method has a comparable inference time to the existing state-of-the-art Controlled Decoding (CD) [A]. For the practical implementation, similar to CD method in [A], one can train a small Q-function adaptor offline, which would allows for faster inference time. This significantly reduces time complexity, similar to ARGS [B], which only introduces a constant factor of k (top-k tokens) over classical decoding methods. Another important point to highlight is that there are several critical scenarios—such as reasoning, mathematical puzzles, chess, and coding where accuracy is more crucial than inference time. In these situations, principled decoding can provide substantial benefits. >Weakness 1: The approach require accessing an already aligned model, this adds more requirements for using the method and limits it use cases. Response to Weakness 2: Thank you for raising this critical point. We remark that our proposed decoding method does not require access to a model specifically aligned to a particular reward function. Instead, interstingly, we can leverage any existing aligned model (BL), even if it is aligned to a different reward (Indirect transfer) and thus is not inherently restrictive. Therefore, utilizing already existing aligned model (such as on Hugging face) for improved decoding is the strength of our approach. We show this via extensive experiments in our paper. >Question 1 : Did the authors compare decoding speed for different approaches? It would be useful to report even thought the proposed approach is slower. Response to Question 1: Thank you for pointing this out. We have now included a comparison of the inference times for all baseline decoding algorithms in the table below. Specifically, we report the inference time for decoding a single prompt, based on the system configuration detailed in Appendix C. \| Algorithm \| Inference Time \| Avg Reward \| \|------------------------\|----------------\|------------\| \| Naive Decoding \| 3s \| 0.13 \| \| ARGS \| 7s \| 0.29 \| \| $\text{CD}^{--}$ \| 40s \| 0.71 \| \| $\texttt{TQ}^{\star}$ (Ours) \| 41s \| 1.0 \| It is evident that $\texttt{TQ}^{\star}$ perform comparably with CD in terms of the time complexity, whereas it is slower than ARGS which suffers from low reward. However, with an offline trained adpator we can scale the inference time to be same as ARGS. >Question 2 : For gpt4 based evaluation, why the authors only report Win-Tie instead of splitting win and ties? In particular I am curious what is the percentage of ties for each comparison. Response to Question 2: We follow the same evaluation protocal as followed in prior decoding approaches like ARGS [B], thereby reporting the Win-tie in Table 2 for uniformity in comparison. However, as mentioned by the reviewer, we also provide the win-rate and tie-rate separately for Evaluation-1 and Evaluation-3 Setup below. \| Ours \| Methods \| Win-Rate \| Tie-Rate \| Lose-rate \| \|------------------------\|------------------\|----------\|----------\|-----------\| \| $\texttt{TQ}^{\star}$ \| ARGS-SFT \| 85.33 \| 1.33 \| 13.33 \| \| $\texttt{TQ}^{\star}$ \| $\text{CD}^{--}$ \| 60.67 \| 6.00 \| 33.33 \| \| $\texttt{TQ}^{\star}$ \| DPO \| 64.00 \| 6.67 \| 29.33 \| \| $\texttt{TQ}^{\star}$ \| ARGS-DPO \| 62.67 \| 5.33 \| 32.00 \| Table: Win, Tie and Lose-rate for Evaluation Setup-1 \| Ours \| Methods \| Win-Rate \| Tie-Rate \| Lose-rate \| \|------------------------\|------------------\|----------\|----------\|-----------\| \| $\texttt{TQ}^{\star}$ \| ARGS-SFT \| 68.67 \| 6.67 \| 24.67 \| \| $\texttt{TQ}^{\star}$ \| $\text{CD}^{--}$ \| 63.33 \| 4.0 \| 32.67 \| \| $\texttt{TQ}^{\star}$ \| DPO \| 62.67 \| 7.33 \| 30.00 \| \|$\texttt{TQ}^{\star}$ \| ARGS-DPO \| 68.00 \| 6.00 \| 26.00 \| Table: Win, Tie, and Lose-rate for Evaluation Setup-2 [A]. Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models, 2024 [B]. Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. Args: Alignment as reward-guided search,2024. --- Rebuttal Comment 1.1: Comment: I appreciate the authors providing the new results and the response, I am satisfied with the response and raised my score by one.	Rebuttal 1: Rebuttal: ## General Response We want to thank all the reviewers for their time and effort in providing detailed comments to our work. We are encouraged that the reviewers found our proposed approach - *novel* (Reviewer ypg9, Reviewer xz5J) & *theoretically rigorous* (Reviewer hjEu, Reviewer 1x8S), - our experimental evaluation to be comprehensive (Reviewer 1x8S), - and our paper to be very well-written (Reviewer 1x8S) & truly promising (Reviewer xz5J). We have addressed the reviewer's comments and concerns in individual responses to each reviewer. As requested by the reviewers, we have performed new experimental evaluations as folows. Please find the attached rebuttal pdf for the experimental results. New Experimental Results: 1. *Pareto Front for KL Divergence to True Reference Policy:* We perform experiments to understand the tradeoff between the win-rate and the KL divergence to the base reference SFT policy. Our findings, reported in Figure 3 of the rebuttal PDF, show that our proposed method, $\texttt{TQ}^{\star}$, outperforms existing baselines. 2. *Stability to Generation Length:* We compared $\texttt{TQ}^{\star}$ against all baselines by varying the length of the output responses. We report our findings in Figure 1 of the rebuttal PDF. Our observations indicate that $\texttt{TQ}^{\star}$ consistently outperforms all baselines, demonstrating its stability across different output lengths. 3. *Generality to Larger Model Size:* We conducted evaluations on two additional setups using larger models, as detailed in the table below. The results, presented in Figure 2 of the rebuttal PDF, show that for both setups, $\texttt{TQ}^{\star}$ significantly outperforms competitive baselines. This demonstrates the scalability and generality of our approach to different model sizes. \| \| Dataset \| SFT Model \| DPO Model \| Reward Model \| \|--------------\|-------------------------------------\|------------\|------------\|--------------\| \| Evaluation-7 \| HH-RLHF \| LLAMA2-13B \| LLAMA2-13B \| LLAMA2-13B \| \| Evaluation-8 \| OpenAssistant Conversations Dataset \| Pythia-12B \| Pythia-12B \| Pythia-6.9B \| 4. *Inference Time for $\texttt{TQ}^{\star}$:* We report the time taken to decode a single prompt for different decoding strategies, using the system configuration described in Appendix C in the main paper. \| Algorithm \| Inference Time \| Avg Reward \| \|------------------------\|----------------\|------------\| \| Naive Decoding \| 3s \| 0.13 \| \| ARGS \| 7s \| 0.29 \| \| $\text{CD}^{--}$ \| 40s \| 0.71 \| \| $\texttt{TQ}^{\star}$ (Ours) \| 41s \| 1.0 \| It is evident that $\texttt{TQ}^{\star}$ perform comparably with CD in terms of the time complexity, whereas it is slower than ARGS which suffers from low reward. However, with an offline trained adpator we can scale the inference time to be same as ARGS. Pdf: /pdf/329e047ba3661126e2e255edde8880a25493af73.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment	Accept (spotlight)	Summary: This paper introduces Text-Only Pre-Alignment (TOPA), a framework designed to extend LLMs for video understanding, without the need for real video data. TOPA leverages textual videos (Tideo), generated by capable LLMs, to mimic real videos. Aided by CLIP's aligned cross-modal space, this framework despite training on text-only data, can effectively handle real video input. The effectiveness of TOPA is demonstrated through extensive experiments on video understanding benchmarks. Strengths: 1. This paper propose a novel text-only pre-alignment framework, TOPA, to extend LLM for video understanding. This method leverages LLM-generated text-only data for training, reducing the reliance on video data and human annotations. The text-only learning pipeline introduced in this paper has potential implications for multimodal learning. 2. The introduced TextVid dataset is innovative. The idea of using LLM-generated "Tideo" to mimic real videos and employing CLIP to bridge the Tideo representation with real video representation is particularly interesting 3. The method is evaluated across various video understanding benchmarks, including the challenging EgoSchema and MVbench. The effectiveness of TOPA is demonstrated in multiple settings, including zero-shot evaluation, pretrain-finetune schemes, and data-efficient finetuning. The paper also performs ablation studies to analyze the impact of proposed method. 4. Many different approaches have compared and discussed in this paper, including vidoe-text pretraining approaches, image MLLM-based approaches and LLM-based video agents. This comprehensive comparison could serve as a valuable reference for future research on video understanding. 5. The paper is well written and easy to follow. Weaknesses: 1. The paper could benefit from providing a more detailed comparison between Tideo and real video. For instance, the TextVid dataset likely covers more diverse domains, due to the use of varied prompts during Tideo generation. It could be analyzed in depth. 2. The proposed Tideos are primarily key-frame level representations, which may not fully capture the temporal dynamics of real videos. 3. The experiment results on MVBench should be incorporated into Section 4.1.2 for improved understanding of TOPA. The results on MVBench clearly demonstrate TOPA's strengths and limitations. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. The proposed approach regards video as a few key frames. Can TOPA handle scenarios with more frames? 2. Could you provide more details about the finetuning stage? Are both the projector and adapter optimized during finetuning? 3. How scalable is the TOPA framework when applied to larger datasets or more complex video understanding tasks? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: As stated by the authors, the text-only learning framework is restricted by the imperfectly aligned CLIP model. The projected visual representations are limited in capturing fine-grained visual details. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Weakness1: The paper could benefit from providing a more detailed comparison between Tideo and real video. For instance, the TextVid dataset likely covers more diverse domains, due to the use of varied prompts during Tideo generation. It could be analyzed in depth. Ans: Thank you for your suggestion. TextVid is notably diverse compared to real video datasets. For instance, previous datasets like Howto100m and Ego4D focus on specific domains, such as instructional videos from YouTube. By using varied condition prompts, we enhance the diversity of Tideos to cover a broader range of domains. We provide visualizations in the submitted PDF. (a) Figure 1 shows that Tideos generated under different conditions exhibit varying distributions. Tideos-WordNet, in particular, displays the most diverse distribution, scattered across the space. (b) Figure 2 illustrates that Tideos-Howto100m and Tideos-Ego4D focus primarily on human activities, while Tideos-WebVid and Tideos-WordNet cover a broader range of scenarios. (c) Table 3 compares vocabulary sizes, showing that Tideos-WordNet tends to cover more objects. > Weakness2: The proposed Tideos are primarily key-frame level representations, which may not fully capture the temporal dynamics of real videos. Ans: Sampling videos into several key frames is a widely adopted approach in vision-language models [1,2,3] due to its efficiency and considerable performance. Understanding videos at high frame rates is indeed challenging and essential for tasks, such as action counting. Recent work has explored this area [4,5]. However, this is not the primary focus of our paper. TOPA may be extended to handle more frames scenarios by incorporating additional local aggregation modules and fine-tuning with a larger number of frames, which we leave it as future work. > Weakness3: The experiment results on MVBench should be incorporated into Section 4.1.2 for improved understanding of TOPA. The results on MVBench clearly demonstrate TOPA's strengths and limitations. Ans: Thank you for your suggestion. Due to page limitations, we could not include the MVBench results in the submitted version. We will try to incorporate the MVBench experiment into the main paper for improved clarity and understanding of TOPA in final version. > Q1: The proposed approach regards video as a few key frames. Can TOPA handle scenarios with more frames? Ans: In this paper, we fix the input frames to 10 for simplicity. There are several approaches to extend TOPA to handle more frames like (a) Frame selection: Choosing a subset of frames based on criteria such as importance or relevance. (b) Additional local aggregation module: Adding a module that aggregates information from local regions to obtain clip-level representation. (c) Fine-tuning with more frames: Finetuning the model with larger number of frames. > Q2: Could you provide more details about the finetuning stage? Are both the projector and adapter optimized during finetuning? Ans: We provide detailed finetuing hyper-parameters for each dataset on Table 11. During supervised finetuning, both of the projector and adapter are optimized. We will enhance the presentation of training, fine-tuning, and inference details in the final version of our paper. > Q3: How scalable is the TOPA framework when applied to larger datasets or more complex video understanding tasks? Ans: The TOPA framework exhibits considerable potential for scalability across larger datasets and video understanding tasks, due to the innovative data generation method and text-only training. TOPA framework leverages LLMs for data generation, providing several key advantages: (1) the potential size of the dataset, TextVid, is virtually unlimited; (2) the diversity of domains covered by Tideos can be easily expanded by prompting LLMs with specific conditions; (3) the supervision signals can be dynamically generated to meet video understanding task requirements, such as video summarization or video chat. Ref: [1] Yang, Antoine, et al. "Zero-shot video question answering via frozen bidirectional language models." NeurIPS 2022 [2] Yu, Shoubin, et al. "Self-chained image-language model for video localization and question answering." NeurIPS 2023 [3] Ko, Dohwan, et al. "Large language models are temporal and causal reasoners for video question answering." EMNLP 2023 [4] Papalampidi, Pinelopi, et al. "A simple recipe for contrastively pre-training video-first encoders beyond 16 frames." CVPR 2024 [5] Balažević, Ivana, et al. "Memory consolidation enables long-context video understanding." ICML 2024 --- Rebuttal Comment 1.1: Comment: Thank the authors for the detailed response. Most of my concerns have been addressed. Overall, the proposed text-only pre-alignment framework, including the Tideo dataset and the text-only learning process, is both novel and effective. It offers a new perspective and solution for multimodal alignment, which may inspire future research. --- Reply to Comment 1.1.1: Comment: We're glad that we addressed your concerns, and we appreciate your recognition of our work. We would like to thank you once again for your valuable suggestions, which helped improve our paper. We will ensure that our discussions are reflected in the final version.	Summary: The paper presents Text-Only Pre-Alignment (TOPA), a method to extend LLMs for video understanding without training on real video data. TOPA generates "Textual Videos" using LLMs to simulate real video data, then pre-aligns the LLMs with video modalities using the CLIP model for feature extraction. This approach achieves impressive results across various benchmarks, surpassing previous methods and competing well with recent video understanding models. Strengths: - The term "Tideo" for textual videos effectively captures the concept. - The motivation of the work is well-formulated, addressing both the high cost of video training and the challenges of aligning different modalities. - The authors discuss the modality gap, the misalignment between visual and text modalities, and its impact on performance. - The related work section is concise and clearly positions the study within the context of previous research. - The novel idea of creating textual descriptions of videos without having the actual videos, thereby skipping the alignment part between video and text modalities, is interesting. It shows that simulating other modality via text and tuning LLMs solely on language can be more efficient than combining multiple modalities. - The proposed method is thoroughly evaluated across a range of tasks and models. Weaknesses: 1. L42: The authors describe subtitles as "frame-level descriptions," but subtitles are actually spoken speech extracted from videos and have intrinsic context and alignment issues [HowTo100M]. 2. The description of the training process and stages is confusing. It needs to be clearer what is trained when and how. The subsection "Video-LLM alignment" is confusing given the title "text-only pre-alignment." The presentation would benefit from an overview of high-level steps for training/inference, followed by detailed descriptions. The dataset section lacks references to the appendix for exact prompt details. Minor: - The references should be clickable. - Correct “Egoschema” to “EgoSchema.” - The statements “1. Intrinsic complexity of the video modality” and “2. The limitations of web language supervision” should include references to support these claims. Technical Quality: 4 Clarity: 3 Questions for Authors: Questions: 1. Have you considered using combined pretraining with both Tideo and multi-modal data? 2. Does the Video-LLM alignment include temporal aggregation of frames? 3. Could you provide ablation studies on how the prompts and different parts of the dataset affect downstream performance, such as the inclusion of condition prompts, the influence of global descriptions, and various QA pairs? It would also be interesting to see how complex the Tideos are and the consistency of frame descriptions. Could the proposed adaptation of LLMs be trained using only text data from available video datasets? 4. Have you considered creating textual descriptions of frames from the original videos (for evaluation) and comparing zero-shot evaluation text vs. videos? Main Concern: The quality of the generated video descriptions (Question 3). Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The limitations are discussed and cover the main disadvantages of the current model. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Overall, thank you for your careful reading, detailed feedback, and valuable suggestions. We will revise our paper based on the discussion. > Weakness1: L42: The authors describe subtitles as "frame-level descriptions," but subtitles are actually spoken speech extracted from videos and have intrinsic context and alignment issues [HowTo100M]. Ans: We agree that "frame-level descriptions" is not an accurate term in this context. Our intention was to highlight that subtitles primarily provide local context and lack the long-term temporal supervision necessary for understanding longer videos, such as those in EgoSchema. We will revise this section for clarity, address [Minor3], and include a discussion on the challenges of visual-textual misalignment [1] [2]. [1] Lin, Yijie, et al. "Multi-granularity correspondence learning from long-term noisy videos." ICLR 2024 [2] Han, Tengda, Weidi Xie, and Andrew Zisserman. "Temporal alignment networks for long-term video." CVPR 2022 > Weakness2:The description of the training process and stages is confusing. The subsection "Video-LLM alignment" is confusing given the title "text-only pre-alignment." Ans: We appreciate the feedback and will revise the paper for clarity. Here is a brief response. 1. Training and Inference. At "text-only pre-alignment" stage, the model is trained on the TextVid dataset, where the LLM learns to process continuous CLIP text features. Then, the text-only pre-aligned model can be applied for real videos in two ways: (a) Zero-shot inference (Sec. 4.1). The visual features of test videos are projected into textual features space, and then processed by LLM. (b) Finetuning on downstream video datasets (Sec. 4.2). The model is further fine-tuned and evaluated using video features. 2. Video-LLM alignment and text-only pre-alignment. Video-LLM alignment involves aligning LLMs with video modalities, enabling the LLM to process video inputs. Our text-only pre-alignment is a specialized approach within this framework, leveraging text-only data for training. > Q1: Have you considered using combined pretraining with both Tideo and multi-modal data? Ans: We attempted pretraining with both Tideos and the WebVid dataset but did not observe performance improvements. We attribute this to the limited supervision signals from the short video captions in WebVid. We believe the use of Tideos in this paper effectively highlights its features and advantages. Integrating Tideos with multimodal data will be explored in future work. > Q2: Temporal aggregation of frames. Ans: We haven’t designed extra temporal aggregation modules in this work. We assume the LLM can handle temporal aggregation since frames are represented as a sequence of embeddings within the LLM. Table 6 shows TOPA benefits from using more frames, suggesting that the LLM can achieve temporal aggregation. > Q3: Ablation studies on TextVid ... Consistency of frame descriptions ...Trained using text data from video datasets Ans: 1. Condition Prompts Enhance Tideo Diversity. We aim to enhance the diversity of Tideos through varied condition prompts. We provide visualizations in the submitted PDF. (a) Figure 1 shows that Tideos generated under different conditions have different distributions, with Tideos-WordNet being the most diverse and widely dispersed. (b) Figure 2 illustrates that Tideos-Howto100m and Tideos-Ego4D focus primarily on human-centric activities, while Tideos-WebVid and Tideos-WordNet cover more scenarios. (c) Table 3 compares vocabulary sizes, revealing that Tideos-WordNet encompasses more objects. Diversity and performance. Our early experiments found that Tideos-WebVid and Tideos-WordNet do not improve downstream performance. It's because the video understanding benchmarks, such as EgoSchema, primarily focus on human-centric activities, which are more closely aligned with Tideos-Ego4D and Tideos-Howto100m. Consequently, Tideos-WebVid and Tideos-WordNet offer limited additional benefits. We look forward to future open-domain video understanding benchmarks to better assess the impact of the our diverse Tideos. 2. Global descriptions and QA pairs. In submitted PDF, an ablation study (Table 1) shows that multi-choice Tideo-QA and Tideo Summarization tasks enhance the performance. For further results and detailed analysis of the multi-choice QA tasks, please see Appendix A.2. 3. Tideos Consistency: In Appendix H, we provide examples of Tideos. The frame descriptions, including frame captions and object captions, show strong consistency. 4. Text data from available video datasets may not be suitable for text-only pre-alignment. Our text-only pre-alignment relies on Tideos and corresponding supervision. Text data from video datasets, such as Howto100M, provide insufficient supervision for effective training. Additionally, the textual data associated with these videos, such as narrations, is often less detailed compared to our Tideos. > Q4: Have you considered creating textual descriptions of frames from the original videos (for evaluation) and comparing zero-shot evaluation text vs. videos? Thank you for your insightful suggestion. We add an experiment to study this. We used LaViLa-xl [1] as the video captioner. We divided each test video into 10 clips, each containing 4 frames, and generated clip-level captions for them. These clip-level captions are then processed by the CLIP text encoder to produce 10 textual features, which serve as input for the LLMs. Results in Table 2 reveal that both project-based and caption-based inference approaches help to mitigate the domain gap issue. Using textual features derived from video captions achieves better results, benefiting from clip-level captions. This demonstrates the flexibility of TOPA and its applicability to various scenarios. Ref: [1] Zhao, Yue, et al. "Learning video representations from large language models." CVPR 2023 --- Rebuttal Comment 1.1: Comment: I have carefully reviewed the authors' responses as well as the other reviews. I appreciate the authors' thorough comments, which have addressed all of my concerns. I'm looking forward to the revised version of the paper. I am happy with the current evaluation and discussions. --- Reply to Comment 1.1.1: Comment: We're glad that we addressed your concerns, and we appreciate your recognition of our work. We would like to thank you once again for your valuable suggestions, which helped improve our paper. We will ensure that our discussions are reflected in the final version.	Summary: The authors introduce Text-Only Pre-Alignment (TOPA), a novel approach that extends large language models (LLMs) for video understanding without pre-training on real video data. TOPA generates Textual Videos, comprising continuous textual frames and annotations to simulate video-text data, and uses these to pre-align a language-only LLM with the video modality. The CLIP model aligns image and text modalities, bridging the gap between textual and real videos. TOPA encodes continuous textual frames as CLIP text features, analogous to CLIP image features, thus aligning the LLM with real video representations. Strengths: The authors present a novel strategy for approaching video understanding. The proposed models perform well on Egoschema, NextQA, STAR, and TVQA using the Llama2 and Llama3 LLMs. Weaknesses: In Table 1, the author did not include key papers such as LifelongMemory[1], Video-Agent: A Memory-Augmented Multimodal Agent for Video Understanding[2], LangRepo[3], and MVU[4]. LifelongMemory achieves 68% and 62.1% on the subset and the full set of EgoSchema, respectively, and should be included in the table to allow readers and reviewers to understand the relative performance, considering the size of the LLM. Additionally, LangRepo and MVU use open-source LLMs (Mixtral7B, Mixtral-8×7B, Llama-2-7b-Chat, Gemma-7b-IT, and Mistral-7B-Instruct) and should also be included to provide a comprehensive comparison. Furthermore, the table is missing the performance on the subset, whereas other papers report performance on both the subset and the full set. Including this information is crucial to observe how accuracy changes from the subset to the full set. Similarly, in Table 2, the author did not include key papers such as LangRepo, MVU, VideoAgent, and LLoVi, making it difficult for readers and reviewers to compare the proposed models with existing works. https://arxiv.org/abs/2312.05269v1 [1] Ying Wang, Yanlai Yang, and Mengye Ren. Lifelongmemory: Leveraging llms for answering queries in long-form egocentric videos, 2023. https://arxiv.org/abs/2403.11481 [2] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Video-agent: A memory-augmented multimodal agent for video understanding, 2024 https://arxiv.org/abs/2403.14622 [3] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding, 2024. https://arxiv.org/abs/2403.16998v1 [4] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, Michael S. Ryoo. Understanding Long Videos in One Multimodal Language Model Pass, 2024 Technical Quality: 2 Clarity: 2 Questions for Authors: 1. Can you add the performance of the proposed models on the subset of EgoSchema? 2. Can you record the inference time of your proposed method on the entire dataset and compare it with the inference times of existing works, including those listed in this review? There is a concern that the visual-to-text feature conversion in zero-shot mode may slow down the entire model. 3. Can you attempt to use the largest open-source models and show the zero-shot performance on the EgoSchema dataset? It is important to assess whether the proposed model can scale up to larger LLMs and if its performance can be comparable to existing works using closed-source LLMs. I will reevaluate the proposed model based on the answers to the questions. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: I mentioned some concerns about the work in the weakness section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Weaknesses: In Table 1, the author did not include key papers such as LifelongMemory[1], Video-Agent: A Memory-Augmented Multimodal Agent for Video Understanding[2], LangRepo[3], and MVU[4]. LifelongMemory achieves 68% and 62.1% on the subset and the full set of EgoSchema, respectively, and should be included in the table to allow readers and reviewers to understand the relative performance, considering the size of the LLM. Additionally, LangRepo and MVU use open-source LLMs (Mixtral7B, Mixtral-8×7B, Llama-2-7b-Chat, Gemma-7b-IT, and Mistral-7B-Instruct) and should also be included to provide a comprehensive comparison. Furthermore, the table is missing the performance on the subset, whereas other papers report performance on both the subset and the full set. Including this information is crucial to observe how accuracy changes from the subset to the full set. Similarly, in Table 2, the author did not include key papers such as LangRepo, MVU, VideoAgent, and LLoVi, making it difficult for readers and reviewers to compare the proposed models with existing works. Ans: Thank you for your valuable suggestions and for highlighting these related papers. We will include and discuss them in the final version of our paper for a more comprehensive comparison. We kindly note that these papers are appeared/updated on arXiv in late March 2024 and should be considered as contemporaneous work. Discussion on LifelongMemory, Video-Agent, LangRepo, and MVU. We have categoried existing video understanding approaches into 4 categories in Section 4, including Web video pre-training approaches, Adapt image MLLMs for video understanding, LLM-based video agents, and our Text-only Pre-alignment. We should note that our Text-only Pre-alignment different from previous work that levearges text-only data for video-LLM alignment. We'd like to categorize LifelongMemory, Video-Agent and LangRepo into LLM-based video agents approaches that leverage VLMs tools convert visual information into textual, and then process video understanding task via LLMs. MVU belongs to Adapt image MLLMs for video understanding, focusing on adapting image-MLLMs for video understanding. We will update the tables to include results from these papers for a more comprehensive comparison. Performance on EgoSchema subset & Evaluation approaches for multi-choices QA. We report results for the EgoSchema subset in Appendix A.2. We analyze the performance gap between the subset and the full set and discuss the impact of various evaluation methods, including similarity-based, logits-based, and LLM-selection approaches. Appendix A.2 reveals that logits-based and similarity-based approaches perform well on the EgoSchema subset. However, these approaches show limited effectiveness on the full set, resulting in a significant performance gap. This gap may be due to the differing linguistic structures of the two sets: the subset features more similar sentence structures with slight variations, making it easier to distinguish using these methods. In contrast, the full set contains a wider variety of choices, which poses challenges for logits and similarity-based approaches. We notice that LangRepo and MVU use log-likelihood-based selection for EgoSchema evaluation, which aligns with our discussion in Appendix A.2. We'll include a further discussion about these two contemporaneous work on this section. >Question1: Can you add the performance of the proposed models on the subset of EgoSchema? Answered in Weakness. > Question 2: Can you record the inference time of your proposed method on the entire dataset and compare it with the inference times of existing works, including those listed in this review? There is a concern that the visual-to-text feature conversion in zero-shot mode may slow down the entire model. Ans: Thank you for your suggestion. We provide inference time on the following table. We conduct the experiment on a single A100 GPU and report the inference time per sample. The main computational cost of TOPA comes from LLM inference, while the conversion of visual-to-text features adds negligible overhead. Notably, TOPA is more efficient than video-agent approaches, which require multiple additional calls of VLMs. \| Method \| Core LLM/VLM \| Visual Encoding \| Visual-to-Text \| LLM inference \| All \| \| --------------------------- \| -------------- \| --------------- \| -------------- \| ------------- \| ------ \| \| TOPA-LLama2-13B \| LLama2-13B \| 0.0241 \| 0.0003 \| 0.4998 \| 0.5242 \| \| SF-VLM [4] \| LLaVA-v1.5-13B \| - \| - \| - \| 0.9794 \| \| LLoVi (results from MVU[4]) \| - \| - \| - \| - \| 207 \| > Question 3: Can you attempt to use the largest open-source models and show the zero-shot performance on the EgoSchema dataset? We provide the results of TOPA using Llama2-7B, Llama2-13B, and Llama3-8B across various experiments, showing that TOPA benefits from a more powerful and larger-scale LLM backbone. Exploring the performance of TOPA with even more capable models, such as Llama3.1-70B, would be valuable. However, the training overhead for such large models is currently prohibitive for our research. --- Rebuttal 2: Title: Looking Forward to Further Discussion Comment: Dear Reviewer, Thank you once again for your valuable suggestions. We hope that our response has adequately addressed your concerns. If you have any further questions or comments, please do not hesitate to reach out. We look forward to any further discussion. We would like to bring to your attention that the discussion period is scheduled to close on August 13 at 11:59 PM AoE. Thank you for your time and consideration. Best regards, The Authors --- Rebuttal Comment 2.1: Comment: Dear Reviewer kuEy, The reviews of this paper are diverging and thus your further input is needed. Can you please carefully read the other reviews and the responses from the authors and share any updated thoughts? Do you still think the paper needs improvements to be accepted even after reading the other two reviews that are strongly supporting the paper? AC --- Rebuttal Comment 2.2: Title: Benchmark Comparison and Training Datasets Comment: LangRepo as archived on March 21, 2024, which is two months earlier than the submission deadline, so it should have been included in the authors' Table 1 based on the NeurIPS submission guidelines. Additionally, the authors claim that their model outperforms VideoAgnet in lines 213-214 or sentence was miswritten. TOPA is not shown to outperform VideoAgent in Table 1 and table lacks existing models with the similar model size, making it difficult to assess where TOPA actually stands in the context of very long-form video. At a minimum, the authors should have compared performance by either applying Llama2-13B to VideoAgent or applying GPT-4 to their model and explaining the performance differences. Outperforming LLoVi is not sufficient. Furthermore, since TOPA was trained on the Ego4D datase, which includes the EgoSchema videos, the authors cannot claim that the TOPA accuracy in Table 1 is a zero-shot result. The authors also need to address why the accuracy improves when evaluating from the EgoSchema subset to the fullest, while all existing models have shown a performance drop, a pattern generally observed in other datasets. This raises the question of whether this improvement is because TOPA was trained on Ego4D and it is not a zero-shot performance. Based on this reason, I lowered my rating to reject --- Rebuttal 3: Title: Clarification1: Our approach——Text-only Pre-alignment Comment: > LangRepo as archived on March 21, 2024, which is two months earlier than the submission deadline, so it should have been included in the authors' Table 1 based on the NeurIPS submission guidelines. >Additionally, the authors claim that their model outperforms VideoAgnet in lines 213-214 or sentence was miswritten. TOPA is not shown to outperform VideoAgent in Table 1 and table lacks existing models with the similar model size, making it difficult to assess where TOPA actually stands in the context of very long-form video. At a minimum, the authors should have compared performance by either applying Llama2-13B to VideoAgent or applying GPT-4 to their model and explaining the performance differences. Outperforming LLoVi is not sufficient. Furthermore, since TOPA was trained on the Ego4D datase, which includes the EgoSchema videos, the authors cannot claim that the TOPA accuracy in Table 1 is a zero-shot result. > The authors also need to address why the accuracy improves when evaluating from the EgoSchema subset to the fullest, while all existing models have shown a performance drop, a pattern generally observed in other datasets. This raises the question of whether this improvement is because TOPA was trained on Ego4D and it is not a zero-shot performance. > Based on this reason, I lowered my rating to reject ----- ### Thank you for the detailed feedback. However, we believe there are many misunderstandings regarding our paper. # Our approach——Text-only Pre-alignment: ### (1) TOPA is NOT a video agent approach. While both TOPA and video agents [1,2,3,4] utilize LLMs for video understanding, they are based on fundamentally different core concepts and technical frameworks, representing distinct lines of research. In this paper, we propose text-only pre-alignment to extend LLMs for video understanding. We propose Tideo, consisting of sequential textual frames, to mimic real video. During text-only pre-alignment, the LLM learns to process Tideo, which is represented as sequential CLIP textual features. During zero-shot inference, real videos are first representated as sequential CLIP visual features. These visual features are then projected into the CLIP textual space and processed by the pre-aligned LLM. Video agents typically employ an LLM as a central agent to iteratively identify and compile key information to answer questions, using VLMs as tools to convert video content into textual descriptions. The major and essential difference is that TOPA takes video features as LLM's input, while video agents take textual descriptions as LLM's input. ### (2) For zero-shot evaluation, TOPA doesn't train on any videos, let along Ego4D videos. > Furthermore, since TOPA was trained on the Ego4D datase, which includes the EgoSchema videos, the authors cannot claim that the TOPA accuracy in Table 1 is a zero-shot result. > This raises the question of whether this improvement is because TOPA was trained on Ego4D and it is not a zero-shot performance. For zero-shot evaluation, the TOPA model is trained on generated text-only TextVid dataset. There are no videos involved in our text-only pre-alignment process. During the data generation process, we use four types of prompts for diverse Tideo generation, one of which is based on "scenarios" from Ego4D metadata. These "scenarios" are very high-level descriptions like: ```"Watching tv","Cleaning / laundry","Potting plants (indoor),","Scooter mechanic","Eating","Walking on street","Working out outside","Playing with pets"```. We do not believe that the use of such general scenarios compromises the zero-shot setting. Moreover, most video agents, such as VideoAgent [1], LLoVi [3], and LangRepo [4] use LaViLa[7] as the video captioner, which is pre-trained on video-narration pairs from Ego4D. This involves far more detailed information than the high-level scenarios we employ. --- Rebuttal 4: Title: Clarification2: Comparison with video agents approaches Comment: # Comparison with video agents approaches In this paper, we discussed various video understanding approaches, including video agents. Our goal is to present diverse perspectives on video understanding, rather than to assert that TOPA is the ultimate solution. ### (3) The final version will include all the related works we discussed. > LangRepo as archived on March 21, 2024, which is two months earlier than the submission deadline, so it should have been included in the authors' Table 1 based on the NeurIPS submission guidelines. We are glad to include and discuss more related works for comprehensive understanding in our final version. However, it is important to emphasize once again that TOPA is not a video agent. The comparion of these new video agents doesn't affect the main contributions of this paper. Besides, We have already discussed and compared several representative video agents [1,2,3]. ### (4) We don't claim that "TOPA outperforms VideoAgent [1]" > Additionally, the authors claim that their model outperforms VideoAgnet in lines 213-214 or sentence was miswritten. Line 213-214 write: "TOPA outperforms previous image-based adaptation approach IG-VLM and video agents LLoVi and Vamos with the same scale LLM (Llama2-7B and Llama2-13B)". Here, "video agents" specifically refer to LLoVi and Vamos, not VideoAgent [1]. ### (5) Comparison with VideoAgent [1]. > At a minimum, the authors should have compared performance by either applying Llama2-13B to VideoAgent or applying GPT-4 to their model. TOPA need train on TextVid dataset, which involves LLM backforward, making it challenging for us to use GPT-4 as the LLM backbone. In the table below, we observed a significant performance drop in VideoAgent when using Llama2-70B as the LLM backbone. Notably, TOPA with Llama2-13B outperforms VideoAgent with Llama2-70B, highlighting TOPA's effectiveness. \| \| LLM \| EgoSchema Subset \| \| :------------- \| :------------- \| :--------------: \| \| VideoAgent [1] \| GPT-4 \| 60.2 \| \| VideoAgent [1] \| LLama2-70B \| 45.4 \| \| TOPA \| LLama2-13B \| 51.2 \| --- Rebuttal 5: Title: Clarification3: Performance on EgoSchema fullset and subset Comment: # Performance on EgoSchema fullset and subset. > The authors also need to address why the accuracy improves when evaluating from the EgoSchema subset to the fullest, while all existing models have shown a performance drop, a pattern generally observed in other datasets. This raises the question of whether this improvement is because TOPA was trained on Ego4D and it is not a zero-shot performance. ### (6) EgoSchema subset (500) is a part of EgoShema fullset (5031). EgoSchema is focused entirely on evaluation: the hidden full set (test set) consists of 5031 videos via an evaluation server, of which 500 were publicly released for validation. Therefore, we believe that reporting and comparing results on the full set is valid and convinced, as presented in the EgoSchema paper [8]. ### (7) We don't leverage Ego4D videos or data for training. TOPA don't train on Ego4D videos or narrations. We only use "scenarios" from Ego4D meta data like ```"Watching tv","Cleaning / laundry","Potting plants (indoor),","Scooter mechanic","Eating","Walking on street","Working out outside","Playing with pets"``` as prompts for Tideo generation. In contrast, most video agents, such as VideoAgent [1], LLoVi [3], and LangRepo [4] use LaViLa[7] as the video captioner, which is pre-trained on video-narration pairs from Ego4D. MC-ViT-L [6] finetuned their model on Ego4D. ### (8) The performance gap largely related to evaluation approach. As shown in Appendix A.2, we found that most approaches, which evaluate with logits and video-text similarity, share a huge performance gap between subset and full set. TOPA, when evaluated using the logits method, also experiences this common performance drop. While for LLM selection approaches, like VideoAgent [1] and TOPA-LLama2-13B, exhibit a smaller gap or no gap. In Appendix A.2, we suggest that this discrepancy may be due to differences in the linguistic structure of the choices. Why does TOPA-LLama2-13B show no performance gap? TOPA uses different training strategies, different inference strategies, different training data and different LLM backbone, compared to other approaches. It's challenging to figure out "why our model performs consistently". Besides, the performance gap phenomenon and evaluation approach is not the focus of this paper. Instead, it would be more appropriate for other approaches, especially those focus on evaluation methods, to investigate why they underperform on the EgoSchema full set. \| \| Eval mode \| EgoSchema Subset (500) \| EgoSchema Fullset (5031) \| Gap \| \| -------------------------------------- \| ----------------- \| :--------------------: \| :----------------------: \| :---: \| \| LongViViT [5] \| Similarity \| 56.8 \| 33.3 \| -23.5 \| \| MC-ViT-L [6] \| Similarity \| 62.6 \| 44.0 \| -18.6 \| \| LangRepo-Mixtral-8×7B-(12B active) [4] \| LLM logits \| 66.2 \| 41.2 \| -25.0 \| \| TOPA-LLama2-13B \| LLM logits \| 67.5 \| 41.6 \| -25.9 \| \| VideoAgent (GPT-4) [1] \| LLM selection \| 60.2 \| 54.1 \| -6.1 \| \| TOPA-LLama2-13B \| LLM selection \| 51.2 \| 51.0 \| -0.2 \| Ref: [1] (ECCV 2024) Wang, Xiaohan, et al. "Videoagent: Long-form video understanding with large language model as agent." [2] (CVPR 2024) Min, Juhong, et al. "MoReVQA: Exploring Modular Reasoning Models for Video Question Answering." [3] (arXiv 2023.12) Zhang, Ce, et al. "LLoVi: A simple llm framework for long-range video question-answering." [4] (arXiv 2024.3) Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding, 2024. [5] (CVPR 2024) Papalampidi, Pinelopi, et al. "A simple recipe for contrastively pre-training video-first encoders beyond 16 frames." [6] (ICML 2024) Balažević, Ivana, et al. "Memory consolidation enables long-context video understanding." [7] (CVPR 2023) Zhao, Yue, et al. "Learning video representations from large language models." [8] (NeurIPS) Mangalam, Karttikeya, Raiymbek Akshulakov, and Jitendra Malik. "Egoschema: A diagnostic benchmark for very long-form video language understanding." --- Rebuttal Comment 5.1: Comment: Thank you once again for the detailed response. The captioners used by VideoAgent and LLoVi were pre-trained on Ego4D while excluding the EgoSchema videos. However, I remain concerned that TOPA was trained using "scenarios" from the Ego4D metadata, which includes EgoSchema scenarios. It would be clearer to claim zero-shot if the authors could demonstrate the performance of TOPA when trained on Ego4D excluding the EgoSchema scenarios. I will raise the score to weak reject, as the authors have addressed some concerns regarding performance against VideoAgent. --- Rebuttal 6: Title: Regarding the use of Ego4D "Scenarios" Comment: Thank you for your feedback. We are currently conducting an ablation study on the use of "scenarios" prompts from Ego4D. The current results are as follows: \| \| Epoch \| Training data \| EgoScheme SubSet (500) \| EgoSchema Fullset (5031) \| \| --------------- \| :---: \| :-------------------------------------------- \| :--------------: \| :---------------: \| \| TOPA-LLama2-13B \| 18 \| All Tideos (721K) \| 51.2 \| 51.0 \| \| TOPA-LLama2-13B \| 8 \| Tideos excluding the Ego4D "scenairos" (516K) \| 48.4 \| 51.3 \| These results indicate that Tideos-Ego4D does not significantly impact performance on the full EgoSchema dataset, although there is some effect on the subset. We believe this outcome demonstrates that the use of high-level "scenarios" does not lead to special information leakage within the EgoSchema benchmark. In fact, many of the "scenarios" in the Ego4D metadata represent common human activities. Considering the scale and diversity of Tideos, we believe that "scenarios" in Ego4D are well-covered by the other three types of Tideos, as shown in the PDF. Thus, excluding Tideo-Ego4D does not substantially affect performance on EgoSchema. Furthermore, we disagree with the notion that our use of high-level "scenarios" is not a zero-shot approach, while the use of same-domain videos and narrations (used in captioner of LLoVi and VideoAgent) is considered a zero-shot approach. From any perspective, the information contained in high-level, common "scenarios" is far less than that provided by videos and narrations from the same dataset. In our final version, we will clearly state our position on the zero-shot setting and specify the exact data we used. ## About "Scenarios" in Ego4D meta: Some "Scenarios" in Ego4D meta (Without seletion): ```"Watching tv","Cleaning / laundry","Potting plants (indoor),","Scooter mechanic","Eating","Walking on street","Working out outside","Playing with pets","Bike mechanic","Doing yardwork / shoveling snow","Eating","Farmer","Baker","Cleaning / laundry","Crafting/knitting/sewing/drawing/painting","jobs related to construction/renovation company","Carpenter"``` We use these common scenarios as prompts to generate Tideo, and we don't believe these common scenarios would involve the leakage of specific video information. Additionally, most of "Scenarios" are repeated multiple times in Ego4D meta. For examples, ```Carpenter: 332, Eating:497, Cleaning:1068, Playing with pets:105, Watching tv:154``` It means that even if we exclude all the video information included in EgoSchema, we can still obtain a similar 'Scenarios' corpus for Tideo generation.	null	null	Rebuttal 1: Rebuttal: # General Response: We greatly appreciate the reviewers' careful reading and insightful feedback on our work. It's encouraging to receive comments recognizing our work's novel idea, good motivation, thorough experiments and promising results. We have respond the concerns raised and are open to further discussions. We received many valuable suggestions to further improve our paper, and we'd like to share some revision plans here. (1) Including and discussing related papers (Reviewer kuEy). Video understanding is a rapidly evolving field, with numerous new works emerging, such as video-agent approaches. Although we have tried to make comprehensive comparisons with various methods in our submitted version, Reviewer kuEy pointed out some contemporaneous work that we overlooked. We appreciate this feedback and will further discuss and incorporate these references to provide a more comprehensive comparison. (2) Providing more analysis on Tideos (Reviewers nFok and oFSV). We'll include the visualizations to illustrate the diversity of Tideos and the impact of the prompts used in Tideo generation. (3) Improving the presentation on pretraining/finetuning/inference (Reviewers nFok and oFSV). We will provide a clearer overview of the high-level steps, including text-only pre-alignment, finetuning, and various zero-shot inference strategies, and offer more details. # Brief Introduction of the PDF: (1) Table 1 (suggested by Reviewer nFok): An ablation study shows that multi-choice Tideo-QA and Tideo Summarization tasks enhance the performance. For further results and detailed analysis of the multi-choice QA tasks, please see Appendix A.2. (2) Table 2 (suggested by Reviewer nFok): We add an experiment to study inference with textual features of videos. We used LaViLa-xl [1] as the video captioner. Results in Table 2 reveal that both project-based and caption-based inference approaches help to mitigate the domain gap issue. Using textual features derived from video captions achieves better results, benefiting from clip-level captions. This demonstrates the flexibility of TOPA and its applicability to various scenarios. (3) Figure 1, Figure 2, and Table 3 (suggested by Reviewer nFok and oFSV): Visualizations of diversity of Tideos. (a) Figure 1 shows that Tideos generated under different conditions have different distributions, with Tideos-WordNet being the most diverse and widely dispersed. (b) Figure 2 illustrates that Tideos-Howto100m and Tideos-Ego4D focus primarily on human-centric activities, while Tideos-WebVid and Tideos-WordNet cover more scenarios. (c) Table 3 compares vocabulary sizes, revealing that Tideos-WordNet encompasses more objects. Pdf: /pdf/744ed968c4454327f4d31857824ade5cc3c83376.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Private Attribute Inference from Images with Vision-Language Models	Accept (poster)	Summary: The paper presents a timely and important investigation into the privacy implications of Visual Language Models (VLMs) by evaluating their ability to predict sensitive personal information from images found online. The authors introduce a new dataset, which consists of: 1. Images sourced from posts on selected subreddits. 2. Eight categories of private attributes, including residence location, sex, age, and others. 3. 554 manually-labeled attributes, each assigned a "hardness score" reflecting the degree of reasoning and online search required for annotation. The authors assess several VLMs (black-box), with GPT4-V achieving the highest average accuracy at 77.6%. The authors further explore the impact of prompt engineering and implement automated zooming through prompting. Strengths: 1. The paper studies the crucial and under-explored issue of privacy risks associated with VLMs. 2. Unlike previous datasets that primarily rely on images of people, this dataset's annotated attributes only contain depictions of humans in 9.7% of cases. This allows for a more nuanced evaluation of VLMs' ability to infer sensitive information from subtler visual cues. 3. The paper is well-written, with clear documentation of the annotation and evaluation procedures. Weaknesses: 1. The ground truth labels are created by annotators, which may introduce subjectivity and may not always reflect real-world truths. 2. The dataset is relatively small, and the absence of standard deviation reporting makes it difficult to assess the robustness of the results. If one labels another 500 attributes from another set of images collected using the same procedure, how would the accuracy change? 3. It would be better if the authors could provide alignment rates between different annotators. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What is the time spent on filtering subreddits and on annotating the images separately? This could help reproducibility. 2. What are the estimated financial and time costs of human annotation compared to VLM-based annotation methods? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their time and effort spent reviewing our paper, for their insightful and detailed feedback, and for their positive overall assessment of our work. We especially appreciate the reviewer’s acknowledgment of the criticality of the studied privacy issue and for their compliments of the constructed dataset and presentation of the work. Below, we address the reviewer’s questions and comments. Q1: Could you please elaborate how the insights on the privacy issue discussed in the paper depend on the dataset? Definitely. While we agree with the reviewer that the exact quantitative results presented in the paper may vary depending on the dataset, we believe that the key message of our paper is qualitative; namely that such privacy infringing inferences from ordinary images posted on anonymous forums are indeed possible at scale. We constructed our dataset in order to verify (or reject) this hypothesis, which we believe to have achieved, especially given the fact that our dataset is comprised of real-world images posted on an anonymous forum (Reddit), and as such, any inference success on our dataset has direct real-world implications. Unfortunately, due to our resource limitations and the expensive labeling procedure, we could not extend the dataset with more samples (we refer to Q2 and Q3 below for more details). Nonetheless, we also believe that constructing large-scale and ideally public benchmarks for investigating and mitigating the privacy infringing inference capabilities for VLMs is interesting and important future work. However, to provide some robustness statistics of our results, we have (1) run our inferences three times at temperature 0.2 obtaining accuracies of 76.2%, 77.1%, 77.8% with GPT4-V, showing little variation, and (2) calculated a CI for our reported result in the paper at temperature 0.0 for GPT4-V: 77.6% +- 3.5% at confidence level 95%. We will include these additional results in the next revision. Q2: “What is the time spent on filtering the subreddits and on annotating the images separately?” Filtering the subreddits was based on fixed heuristics and as such it did not take any significant time. Annotating the images required around 1 week of full time work, i.e., roughly 40 hours. Q3: “What are the estimated financial and time costs of human annotation compared to VLM-based annotation methods?” As mentioned in Q2, the labeling of the dataset cost around 40 hours of human work. Using the hourly rates set by our institution (USD 35/h) this amounts to USD 1400 of labeling costs. At the same time, using the GPT4-V API, we spent less than USD 12 for inferences on the whole dataset, conducted in around 5 mins with further parallelization possible. As such, GPT4-V inferences are more than 100x cheaper and around 480x faster than human labeling, highlighting the concerning scalability of VLM inferences vs. relying on human annotators for privacy infringing inferences. We thank the reviewer for raising this important point, and will include these numbers in the next revision of our paper. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I do not have other questions at this point.	Summary: The authors focus on the privacy risks of multimodal LLMs by demonstrating the personal information can be extracted from publicly accessible images, and leveraged by LLMs to infer sensitive details. Strengths: The intersection of privacy and multimodal AI is interesting, and the evaluation is strong, using both open and closed LLMs with vision capabilities. The paper is written well, and I appreciate the responsible disclosure. Weaknesses: I have two main issues with this paper: - The authors claim that inference of personal details from public images constitutes a privacy violation. Many privacy researchers would disagree with this. It could be claimed that a person voluntarily posting an image online implies that any content of this image and any subsequent inference is an acceptable risk to them. If the image is posted involuntarily, then it's not the LLM inference that constitutes the privacy violation, but the unauthorised posting of the image. In other words, the fact that VLMs can extract personal information from images, while potentially aggravating to users, is not straightforwardly a privacy violation per se. There is no particular discussion about this, or the potential regulatory implications or mitigations for the issue. - The fact that VLMs are capable of this type of inference is not particularly surprising to me, and the main takeaway is "a human with some time on their hands, access to a search engine and some open-source intelligence skills could do the same, but it'd take a more time". The only counter-argument here is scalability. There is no thorough discussion on potential incentives to perform this type of inference on an industrial scale. In other words, I am hard-pressed to detect the strong, broad-interest contribution of this work beyond "VLMs can do XYZ efficiently", which is an addition to a long list of things that VLMs can do. Beyond this, the contributions are relatively low-impact: There is a some jailbreaking, and some engineering work to reveal details in the image by zooming automatically. For a NEURIPS paper, this is not a substantial enough contribution. Beyond this, the evaluation on a single dataset (sourced from the internet, see below) which is also not very large, and lack of a comparison against human observers, narrows the impact of the work. Technical Quality: 2 Clarity: 2 Questions for Authors: - What are the takeaways of your works in terms of concrete recommendations for society, platform owners, regulators, and privacy researchers? - Why was only a single dataset included? Can you certify that this dataset is representative of images encountered? I see that most of it is sourced from Reddit, are there any concerns of statistical bias from this method of selection? - Why is there no evaluation against human observers? I think this would be the most relevant and interesting type of comparison here: a systematic evaluation against humans with OSINT expertise. Perhaps parts of a labelling workflow on a blinded group of evaluators could be leveraged for this task? - What is the impact of hyperparameters? I see that the temperature was fixed. A more thorough evaluation could be useful here. - What are the costs of executing these inferences in your work? - It would be useful to have a bigger discussion around topics like e.g. uncertainty: The models can confidently hallucinate some personal property which misleads the "attacker" completely. How does this and similar phenomena affecting LLMs/VLMs interact with your findings? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Please see the "questions" on some of the scientific limitations. I am a bit concerned about this study being executed without an institutional review board permission with the justification that "human subjects were not involved". Depending on jurisdiction, this can be contentious. You sourced images from Reddit, in my understanding, without asking for permission by the data owners, and conducted a scientific study on them. My understanding is that there was no statistical counselling about the representativeness of the dataset, or discussion with an ethics review board about the concerns arising from this study. In particular, the checklist states clearly that "The answer NA means that the paper does not involve crowdsourcing", and you answered "NA", but the dataset seems to be downloaded from Reddit, which (while not crowdsourcing in the strict sense) involves leveraging (in fact, attacking) human data without explicit consent, so I find the "NA" here problematic. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects', 'Ethics review needed: Data privacy, copyright, and consent'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their time and effort spent reviewing our paper. We also appreciate the reviewer's acknowledgment of the studied problem being interesting, our evaluation and presentation. We hope our answers below address the reviewer's concerns. Q1: Does the accurate, scalable, and automated inference of detailed personal attributes from images posted on pseudonymized platforms pose a privacy concern? Yes, we strongly believe that the demonstrated inferences amount to a serious privacy threat. First of all, while it is indeed possible for humans to also infer such attributes, human inferences cost around 100x more money and 480x more time (we refer to Q2 and Q3 of Reviewer T2ze for detailed inference costs). As such, using VLMs for this task is a fundamental paradigm change. Notably, VLMs enable these inferences at a large and automated scale, questioning the privacy through obscurity assumptions one usually operates under on such platforms. Additionally, as detailed in the paper, key privacy regulations consider data that enables the inference of personal information protected data—something not commonly associated with snapshots of parking lots posted in largely non-privacy-concerning contexts. Further, our view that large-scale personal attribute inferences from online content not intended to be personally identifiable are shared by both the academic community and the public. This is highlighted by similar works [1] on text receiving widespread attention, including an ICLR spotlight and a privacy policy award. Further, similar issues are widely discussed in the community [2,3], as well as real-world applications [4]. All in all, our work is the first to show that personal information is inferable at a large scale from inconspicuous images posted online, which is, in line with other reviewers, a concerning and relevant finding. Q2: “What are the takeaways of your works in terms of concrete recommendations for society, platform owners, regulators, and privacy researchers?” Society: We believe that our work can raise awareness among the public that sharing images, even on anonymous platforms where they thought to have taken enough care to obscure their identity, can be revealing of their personal information. We hope that, with awareness of this, users can adjust their online behavior and become more conscious of their privacy. Platform Owners: Anonymous platforms have to be aware of such risks, and may adjust privacy promises to their users accordingly. Educating their users of the full extent of the risks would be also desirable. At the same time, they could also take a conscious approach to make it harder to process their data at a large scale, obstructing scalable inferences. Regulators: We believe that in some jurisdictions, such as the EU, there are already strong regulatory protections for private data in place. As such, the important takeaway for regulators is to see how ubiquitous/heterogenous personal data really is. Privacy Researchers: We believe that for privacy researchers, in the face of these privacy risks posed by foundation models the key challenge to tackle is to develop both model side and user side mitigations. We thank the reviewer for this insightful question, and we will expand our paper in its next revision with a more detailed discussion. Q3: Does the used evaluation dataset impact the qualitative conclusions drawn by the paper? No, we believe that the key message of our paper is qualitative, namely that privacy-infringing inferences on real-world data at scale are possible, and this message is not impacted by the exact quantitative results obtained. Notably, the dataset that we have used for this work is sourced exactly where we see the examined threat: from real-world online forums. As such, any accurate inference on our dataset corresponds to a real-world privacy risk. This is also the reason why we have taken such care in constructing the dataset (labeling in-house) and not releasing it to the public. Q4: Would an evaluation against humans with OSINT experience be possible? While such a comparison would certainly be interesting, it is not the focus of the paper (see Q3). Further, as in our case, we already spent 100x more money and 480x more time on labeling than VLMs, a cost difference hard to overcome with domain experts. Finally, we believe that with our current real-world data, such an experiment would be ethically highly problematic, particularly when tasking outside individuals to infer personal data from real-world individuals. Q5: What is the impact of the sampling temperature parameter on the quantitative inference results? For this, we run GPT4-V on temperatures 0.0, 0.2, and 0.4, achieving 77.6%, 77.03% (average of three runs: 76.2%, 77.1%, 77.8%), and 77.3% accuracy, respectively. As such, reasonable temperature levels have little impact on the quantitative results. We will include these additional results in the next revision. Q6: Could you please discuss how inference uncertainty reflects on your findings? We agree with the reviewer that uncertainty estimates could help in real-world scenarios in case of inaccurate inferences. Qualitatively, similar to [5], we find that models are much less prone to hallucinations when they are tasked with directly reasoning about a given input (compared to, e.g., underspecified QA settings). Actively recognizing this provides an interesting avenue for future research in this area. References [1] R Staab et al. Beyond Memorization: Violating Privacy via Inference with Large Language Models. ICLR 2024. [2] Schneier, B. (2023, Dec). AI and Mass Spying. Schneier on security. [3] Eggsyntax. (2024, May). Language models model US. LessWrong. [4] Brewster, T. (2023, Nov) Chatgpt has been turned into a social media surveillance assistant. [5] R Staab et al : “Large Language Models are Advanced Anonymizers”, 2024; arXiv:2402.13846. --- Rebuttal Comment 1.1: Title: Thank you and response to rebuttal. Comment: Thank you for your rebuttal. > Q1: Does the accurate, scalable, and automated inference of detailed personal attributes from images posted on pseudonymized platforms pose a privacy concern? I recognise that there are varying valid viewpoints on this topic, and I will not hold this point against you. I have read the rest of your rebuttal and I did not find responses to all of the concerns I raised, in particular not the statistical validity and (especially) ethical considerations about the use of publicly sourced data and whether this study was guided by an institutional review board. The phrasing of your rebuttal actually amplifies some of my concerns ("such an experiment would be ethically highly problematic, particularly when tasking outside individuals to infer personal data from real-world individuals"). I am thus maintaining my current stance until this information is provided. --- Reply to Comment 1.1.1: Comment: First of all, we thank the reviewer for their response and for engaging in a discussion with us. We are also highly appreciative of their recognition of differing viewpoints on how and to what extent the presented inferences pose a privacy threat. Let us first answer an explicit question from the review that we have not answered directly in our rebuttal. ”What are the costs of executing these inferences in your work?” Using the GPT4-V API, the cost of our inferences on the whole dataset lie below \\$12 and take around 5 minutes at our current level of parallelization (we believe that with a better implementation of parallelization, this can be further reduced). Note that since we have conducted this study, GPT4o has been made available, which has a significantly lower API cost. At the same time, creating the dataset (in-house human labeling) took us around 40 hours of work, which, using the hourly rates set by our institution (\\$35), amounted to a total labeling cost of \\$1400. Now, let us address the points explicitly raised in the reply. ”Why was only a single dataset included? Can you certify that this dataset is representative of images encountered? I see that most of it is sourced from Reddit, are there any concerns of statistical bias from this method of selection?” We only used a single dataset for this study, as (1) there are no suitable available datasets in the community, as also elaborated in the paper; (2) constructing such datasets is very expensive, as discussed above; and (3) we were not able to find an equally suitable alternative data source to Reddit, which is also as directly representative of the actual examined inference threat. We collected the dataset by first heuristically filtering subreddits where we believed informative images are being posted (the subreddits are listed in Appendix D.4; visiting them on Reddit will give you a sense of the images included in the dataset). Note that these heuristics were mostly based on our intuition and did not involve any prior inference results. Further, the dataset only includes images where the labeler was able to infer at least one feature themselves. No other selection was made, especially, we did not adjust anything in the dataset after we have started with our inference experiments. As such, the dataset is not cherry-picked, and is representative of real-world images posted on anonymous forums from which personal attributes can be inferred. Note that the fact that we only include images in the dataset that can be labeled do not subtract from the representativeness of the threat. An adversary could simply first ask the VLM if it can infer anything, before proceeding to infer personal attributes; or simply ignore those answers of the model where no inferences have been made. Now, it is indeed true that images sourced from Reddit, especially from a select subset of subreddits is not representative of all possible images one may encounter online. However, it is also not the goal of the paper, as previously mentioned, to provide an exact and representative quantitative analysis across all online platforms of the image analysis capabilities of VLMs. Instead, the key message of the paper is qualitative; there exists a privacy threat enabled by the large-scale inferences of VLMs stemming from seemingly non-revealing images that people tend to upload on anonymous platforms under non-representative usernames. To deliver this message, we believe our dataset is extremely well suited, as it includes exactly these sorts of images posted on Reddit, a pseudo-anonymous forum. Each accurate inference we have made in a controlled setting, could also be made by an adversary with malicious intents on the exact same data points. This is also the reason why we have taken so much care when dealing with our dataset and publishing our results — we elaborate more on this in the next question.	Summary: The paper performs a novel analysis privacy-leakage due to modern multimodal VLMs. Specifically, they show that modern VLMs, when correctly promoted, can infer sensitive information from seemingly innocuous images. They curate a dataset of images containing clues to sensitive information and query a few SOTA models using these [image, prompt] pairs, and show that the models are capable of inferring eight sensitive attributes from the images. The inference ability strengthens with the quality of models, which raises concerns about online privacy. Strengths: - Very interesting, timely and practically relevant privacy analysis of VLMs - The work raises appropriate concerns about privacy of online images - Paper is very well-written and easy to follow Weaknesses: - The proposed attack although important seems quite easy to defend against - The paper should evaluate at least a couple of simplest of defenses, e.g., SFT. Technical Quality: 3 Clarity: 3 Questions for Authors: I really like the work due to its relevance to practice and shoutout to the authors for that! I have a few questions: - The proposed attack seems super simple: it’s basically jailbreaking VLMs to elicit harmful responses. This is known type of an attack that is extensively studied, e..g, for LLMs. How will this attack work if a proprietary model provider does SFT using aligned data? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See questions above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First, we would like to sincerely thank the reviewer for their time and effort spent reviewing, their insightful comments, and for their highly favorable assessment of our work. We are especially thankful for the reviewer’s acknowledgement of the relevance and timeliness of the presented privacy threat. Below, we address the reviewer’s comments and questions. Q1: Would it be easy to defend against the presented inference-based privacy attack? Crucially, independently of how easy it would be to defend against such privacy inferences on the model provider’s side, the fact that these privacy violations are possible with current available models is concerning. Especially, since open-source models are also capable of highly accurate inferences, meaning that even if model providers would make preventative adjustments, these previously uploaded open-source models would still be available to conduct privacy-infringing inferences. Further, while it is hard to exactly anticipate the difficulty of provider side defenses, we believe that it would be a difficult undertaking to try to prevent privacy infringing inferences while maintaining utility on other tasks. We believe so, as we have observed (as also stated in the paper) that the capability of the models to conduct such inferences is highly correlated with their general capabilities. Additionally, we know that some of the examined models (e.g., GPT4-V) already are supposed to be aligned against such inferences, yet we were able to harness these models without too much effort put into jailbreaking them. In fact, for this rebuttal, we tested our attack on the newer GPT4-o model, which tends to score higher in alignment benchmarks than GPT4-V [1], and achieved even higher accuracy (80.7%); showing that alignment against such inferences is at the moment not strong enough. Nonetheless, as also argued in the paper, we strongly believe that developing adequate defenses, both on the providers’ side and on the users’ side is crucial going forward in order to enable people to exercise their right to privacy even in the age of VLMs. In this regard, we make the first important step into this direction by pointing out, and systematically evaluating the inference-based privacy risks these VLMs pose; hopefully raising awareness in the broader privacy and machine learning community. Q2: Could SFT be an effective method to defend against the attack? We believe that both SFT, and especially preference tuning, such as PPO or DPO could be promising strategies on the model provider’s side to decrease the likelihood of a model answering a privacy infringing inference query. However, conducting such experiments is beyond our means unfortunately as (1) even the examined well-performing open-source models are extremely large to handle for training, and (2) this would require vast amounts of training data that is even more expensive to obtain (please see our responses to Q2 and Q3 of Reviewer T2ze for more details). Nonetheless, we agree with the reviewer and believe that it is an important avenue both for future scientific work and for commercial providers to explore the possibilities of preventing privacy-infringing inferences through finetuning or preference tuning. As part of our responsible disclosure we therefore directly contacted all LLM providers used during this study. References [1] Z Ying et al. Unveiling the Safety of GPT-4o: An Empirical Study using Jailbreak Attacks. arXiv 2024.	Summary: The paper proposes a new privacy attack with VLMs where the model is queried with an image and the goal is to predict the private attributes such as place and gender of the person not shown in the image. Strengths: - Good experimental evaluations - Clear problem formulation - Practical automated attack using VLMs Weaknesses: Overall, I like the evaluation of the attack, although I feel it is missing a baseline comparison. For instance, how likely is it to predict the sensitive attributes like age, gender, etc. (without access to the VLM) using pure correlations in the data. If there is a pipeline for automatically annotating the important objects in the image, then a Bayesian model or a neural network can be trained to capture the correlations to the target attribute. This is also indicative from the experiments where the presence of partial human identifiers boosts the attack success. This type of correlation can be captured by a baseline model. Technical Quality: 2 Clarity: 3 Questions for Authors: Can you compare the effectivness of VLMs w.r.t. the baseline model that captures the general correlations between objects in the images? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors show a practical attack using modern VLMs where the adversary can infer sensitive attributes about an image by querying these models. While authors do not release their VIP dataset citing ethical concerns, I feel they need to acknowledge that an adversary can replicate their attack in practice, and should discuss more on the defenses against such attacks. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and efforts spent reviewing and for their insightful feedback. We are especially appreciative of the reviewer’s recognition of our attack’s novelty and our extensive experimental evaluation. Below, we address the reviewer’s comments and questions. Q1: Would it be possible to construct a baseline method relying on classical supervised learning techniques? Fundamentally, any baseline method relying on supervised learning would require a specialized training dataset on which it has to be trained. As such, mounting a privacy attack using such methods would require significantly stronger assumptions on the adversary. At the same time, and as we argue in our paper, VLMs pose a high privacy risk through inference exactly because of the low barrier of entry for adversaries. While we agree with the reviewer that such a baseline evaluated directly on our dataset would provide an interesting anchor point for our results, unfortunately, due to the highly diverse and challenging nature of the domain this is not feasible. To underline this argument, take the following examples of how the VLM inferences have been made: (1) an office room, where among many of the items there is one with a University of Florida logo; (2) the interior of a New York public library; (3) sunset, with a view of a few buildings in Denver, in the foreground an open bottle of German beer; (4) just a picture of an at home office room, stylistic elements indicating the (likely) sex of the author. Each of the above inferences relies on specific task-dependent knowledge that would need thousands of data points to learn in a supervised setting. As labeling even the few hundred examples in our dataset is considerably expensive (~USD 1400, please see our responses to Q2 and Q3 of Reviewer T2ze for more details), obtaining such a dataset for each possible inference task is not feasible. Clearly, in order to excel at such inferences, the VLMs have to rely on the vast world knowledge obtained during unsupervised pre-training both on image and text data. Further, the closest direct comparison on a related domain that we can obtain is comparing the performance of VLMs against “traditional” machine and deep learning methods on HAI and PAI datasets, where, as mentioned also in the paper, visual foundation models have shown promising performance over prior methods [1-3]. As such, we also believe that even if one could construct supervised methods for some of the inference types, VLMs would still prevail. Q2: Could you please expand the discussion on the threat of adversaries replicating your attack and on possible defenses? Certainly. In general, we share the reviewer’s concern, and it was our main motivation for conducting this study to raise awareness about the inference-based privacy threat VLMs pose such that the community can (1) act against it by mitigating the issue, and (2) people can adjust their online practices in the face of this privacy threat. In line also with our concern, before making a copy of this study available anywhere, we first informed the key entities that produced the examined VLMs about our findings as part of a responsible disclosure. All in all, in line with the common stance in the privacy community, we believe that privacy through obscurity is not robust; which is why we think it is important to raise and systematically evaluate this issue in the form of this study. Regarding potential mitigations against the attack, as already discussed in Section 6 of the paper, we believe that both internet users and VLM providers can act towards reducing the privacy risks posed by VLMs. From the providers’ side, our findings could serve to strengthen the safety alignment of the models, specifically targeting privacy infringing inferences. From the users’ side, we hope for anonymization tools that could extend to images, by, e.g., removing revealing clues from images using generative modeling. Nonetheless, most importantly, we believe that the first, and most crucial step, towards responsible and privacy preserving practices of VLM development and use is awareness of the full extent of possible privacy issues these models may pose. Here, our paper focuses on raising and demonstrating the issue of VLMs enabling privacy violating inferences at a large scale. References [1] X Cheng et al. A simple visual-textual baseline for pedestrian attribute recognition. IEEE TCSVT 2022. [2] M Castrillón-Santana et al. Evaluation of a visual question answering architecture for pedestrian attribute recognition. CAIP 2023. [3] X Wang et al. Pedestrian attribute recognition via clip based prompt vision-language fusion. arXiv 2023. --- Rebuttal Comment 1.1: Comment: Thank you for your clarifications. After reading the rebuttal and all the other reviews, I have decided to raise my score. --- Rebuttal 2: Comment: We thank the reviewer for raising their score and if they have any remaining questions or concerns, we are eager to engage in further discussion until the end of the author-reviewer discussion period.	Rebuttal 1: Rebuttal: First of all, we would like to thank all reviewers for their time and efforts spent on reviewing our paper, and for their insightful, constructive, and valuable comments. We are especially appreciative of several reviewers’ acknowledgement of the relevance, practicality, and cruciality of the examined threat; as well as of the recognition of our extensive experimental evaluation. We address the reviewers’ questions and comments in individual responses below, and are looking forward to a fruitful discussion.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
General Compression Framework for Efficient Transformer Object Tracking	Reject	Summary: The authors proposed a novel compression strategy of transformer based trackers. Unlike previous works, it divides the teacher network into multiple segments, each segment corresponds a single transformer layer of student network, then train each student layer separately. It also introduced some training strategies to enhance performance including (progressive) replacement training, prediction guidance and feature mimicking. Such compression framework is insensitive to the change of architecture of teacher network. Strengths: 1. Effectiveness. The experiment results clearly demonstrated significant improvement of inference speed while reserving the majority of tracking accuracy. 2. Flexibility. The proposed compression strategy is insensitive to the change of architecture of tracking models, making it easy to apply on almost any transformer based trackers. The segmentation strategy and the size of student network also supports user customization, which enables the user to design student network according to their unique demands. Such flexibility shows excellent application prospects in end-side scenarios. Weaknesses: The detailed strategy of dividing the teacher network is not stated clearly in the paper. Base on the pseudo code provided in page 13, it seems that the segmentation strategy is simply mapping the list of transformer blocks of student network to that of the teacher network base on the lengths of the two lists. This could be too simple. For example, assume teacher network has 8 transformer blocks in module 1 and 2 blocks in module 2, while student network consists of 2 blocks, then the second student block would have to emulate the last 3 blocks of module 1 and the 2 blocks of module 2, while module 1 and module 2 might have been trained separately and possess different knowledge. Empirically, this would result in sub-optical performance. A brief discuss on the divide strategy could help this paper become more informative. Technical Quality: 3 Clarity: 4 Questions for Authors: I don't quite understand the concept of prediction guidance mentioned in the first paragraph of section 3.3: does it mean using the prediction of teacher model as pseudo label to supervise student's learning? If so, then how exactly does it help learning process, since the original ground truth bounding box already provides similar supervision, I don't really understand why using the less precise, noisy prediction of teacher tracker as additional pseudo label could actually be beneficial. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: The paper clearly addressed its limitations including inefficient training process and the performance gap between teacher and student network. Flag For Ethics Review: ['No ethics review needed.'] Rating: 9 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing the efficiency and value of our work. We also appreciate your insightful advice and comments. We value your support for our work! ## Q1: Stage Division Yes, our framework is indeed really simple yet effective. When the teacher model and student model share the same structure, implementing our framework requires only mapping the list of transformer layers. If student model has a different structure from teacher model, only minor modifications are needed to achieve our stage division strategy. Thanks to our stage division strategy, the number of transformer layers in each stage of teacher model is flexible. With a fixed number of stages, we can flexibly divide the layers of teacher into multiple stages. And the performance of student model relies on the layer number of student model instead of division strategy of teacher model based on our experiment experience. We have also conducted an ablation study on stage division strategy in Table 7. The 'Even' and 'UnEven' stage division strategies achieve similar performance on LaSOT. To highlight that the improvements result from our framework instead of complex stage division strategy, we just implement the even division strategy. Researchers can apply various stage division strategies based on their models. We will take your advice and add a more detailed discussion about stage division strategy in the final version. ## Q2: Prediction Guidance You are right! The prediction guidance means using the prediction of teacher model as pseudo label to supervise student's learning. While there is inherent noise in the teacher model's predictions, this noise can actually benefit the student model’s learning process. (1) Firstly, the predictions of teacher model are probability distributions, which provide additional information beyond discrete ground-truth labels. This probabilistic information helps the student model better grasp the data's underlying complexity. (2) Secondly, teacher model , having undergone extensive training, encompasses a rich set of features and patterns. By learning from the teacher model's predictions, student can learn the deep knowledge of teacher. (3) The noise in teacher prediction can enhance the generalization ability of student model. Research shows that such noisy supervision can often be more effective than using ground-truth labels alone, as it helps the student model converge more quickly and robustly [1][2][3]. In another word, although prediction of teacher contains some noise, this kind of supervision is easier to learn for the student model than groundtruth labels and can help student model converge more quickly. [1] Towards Understanding Knowledge Distillation, M. Phuong, C. Lampert, International Conference on Machine Learning, 2019 [2] Knowledge Distillation: A Survey, J. Gou, B. Yu, S.J. Maybank, D. Tao, International Journal of Computer Vision --- Rebuttal Comment 1.1: Comment: Thank you for clarifying my inquiries. Your explanations on prediction guidance have been enlightening to me, expanding my understanding of knowledge distillation, a concept I have not delved deeply into before. I appreciate your valuable contributions. --- Reply to Comment 1.1.1: Title: Thanks for your acknowledgement and support. Comment: We are grateful for your recognition of the novelty and efficiency of our work. We are also very glad that our response can resolve your concerns and it's our honour! Thanks a lot for you support for our work.	Summary: This paper aims to distill knowledge from larger teacher models into more compact student trackers. Three techniques are proposed: A stage division strategy that segments the transformer layers of the teacher model. Replacement training technique. Prediction guidance and stage-wise feature mimicking. Experiment verifys the effectiveness of the method. Strengths: 1. The proposed techniques are comprehensive and include a bunch of methods to improve the performance and efficiency of the trackers, 2. The experiments are extensive which includes 5 VOT benchmarks. 3. The speed is fast when applying the 2 layer tracker variants. Weaknesses: 1. The most obvious weakness is that the whole method consists of many distilling techniques, including training strategies, feature mimicking, and loss guidance. It is hard to see the inherent consistency between those techniques. This may harm the generalization ability and transferability of the proposed framework, as the author claims the framework is general. 2. The overall method is complex. I am worried about its application to other researchers. 3. When applied to the Mixformer v2, which has only 2 layers, performance can be improved marginally while speed is unchanged. This may indicate the method's shortcomings. Complex techniques only bring a little improvement. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Prove those techniques can be unified instead of looking like a bunch of tricks. 2. The methods are restricted when the transformer tracker has fewer layers. 3. See weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing efficiency and value of our work. We also appreciate your insightful advice and comments. We would be grateful if you could support our work and reconsider your rating. ## Q1: Inherent Consistency Thank you for your insightful advice. CompressTracker is a unified and generalized framework which can be applied to with any transformer structure. All other reviewers acknowledged the novelty, simplicity and strong generalization of our CompressTracker. we would like to clarify how our contributions are intrinsically connected and contribute to the cohesive functioning of our framework. Motivation of stage division strategy Current dominant trackers are one-stream models characterized by sequential transformer encoder layers that refine temporal features across frames. This layer-wise refinement suggests treating the model as interconnected stages, encouraging alignment between student and teacher at each stage. Thus, we propose stage division strategy, which divide teacher into distinct stages corresponding to the layers of student model. Each student stage learns and replicates the corresponding teacher stage’s function. Motivation of replacement training Building on this, we propose a replacement training methodology. The core of this method is dynamic substitution of stages during training. Thanks to our stage division strategy, we can perform this replacement training. Previous methods couple layers together, making replacement training impractical or potentially confusing due to strong coupling between student model stages. However, our stage division strategy decouples each stage, allowing replacement training and improved accuracy. Motivation of prediction guidance and stage-wise feature mimicking After that, to accelerate convergence, we introduce prediction guidance using the teacher's predictions as supervision. The stage-wise feature mimicking strategy aligns feature representations at each stage of student with those of teacher, ensuring more accurate and consistent learning. Inner Connection Our contributions are cascading and interconnected. We first propose the stage division strategy, which enables replacement training. The replacement training relies on our stage division strategy. Building on these, we introduce prediction guidance and stage-wise feature mimicking strategy to further enhance the student's learning from teacher. Each contribution lays the foundation for the next, creating a strong inherent consistency. Generalization ability Thanks to our stage division strategy, our framework has strong generalization ability, allowing flexibility in designing student model and supporting any transformer architecture. This flexibility is unique to our approach and unachievable by previous methods due to the absence of the stage division strategy. We conduct extensive experiments to verify effectiveness and generalization ability of our framework (Tables 1, 2, 3). In total, We have compressed 2 kinds of teacher models into 7 different student models. We experimented with 2 different teacher models, 5 different layers of student models, and different structures for student and teacher models. These student models all outperform their counterparts, demonstrating the effectiveness and generalization ability of our framework. Simplify Our framework is quite simple, with minimal code modifications required. We provide pseudo code in Appendix Algorithms 1, 2, and 3. Reviewers H7HB, 8Frr, and 2g8j acknowledged the simplicity, with Reviewer 2g8j expressing surprise at its simplicity. This simplicity can also prove the strong transferability of our framework. Recognization of other Reviewers Reviewer H7HB recognized simplicity and generalization abilities, noting, "Versatility" (flexbility) and "Streamlined training" (simplify). Reviewer M8LK highlighted novelty, stating, "The author has clear ideas". Reviewer 8Frr also recognized innovation and flexbility, highlighting "Innovative Approach" and "Structural Flexibility". Reviewer 2g8j also affirmed the flexibility. In conclusion, our contributions are inherently consistent, with each building upon the previous one. Extensive experiments verify effectiveness and generalization ability. All other reviewers acknowledged the novelty, simplicity and generalization. We will modify our manuscript to clarify the inner connection of contributions. We sincerely hope you can reconsider our work, support us, and reconsider the rating. We would appreciate it very much. ## Q2: Complex Method Our framework is quite simple. For details, please refer to our response to Q1: Inherent Consistency. We believe other researchers can reproduce our method quickly and easily. We will release all the codes upon acceptance. ## Q3: MixFormerV2-S and Fewer Transformer Layers Indeed, MixFormerV2-s has 4 transformer layers. We conducted an experiment to compress MixFormerV2 into CompressTracker-M-S with 4 layers, matching MixFormerV2-S. As shown in Table 2, our CompressTracker-M-S outperform MixFormerV2-S (62.0 AUC vs 60.6 AUC on LaSOT), with identical settings, including model structure, pretrained weights, and training datasets. Model performance does degrade with fewer transformer layers. CompressTracker-2 with only 2 layers remains competitive with MixFormerV2-S, which has 4 layers, and outperforms most previous models in both speed and accuracy, except for HiT-Base and SAMT, as shown in Table 4. Additionally, MixFormerV2-S requires a complex multi-stage training process, taking 120 hours, while CompressTracker-2 achieves similar or better results with just 14 hours training, as shown in Appendix Figure 5. We will add more comparision with MixFormerV2-S in the final version to further emphasize effectiveness and simplicity of our CompressTracker. --- Rebuttal Comment 1.1: Comment: Thanks for your responses. Though reducing the layer numbers of the vit-based tracking is not new, the contribution is non-trivial. I acknowledge the contribution of this work. Thus, I am willing to raise my initial rating to 5 (bordline accept). --- Reply to Comment 1.1.1: Title: Thanks for your acknowledgement and support. Comment: We are grateful for your recognition of our contribution. Thanks a lot for your reconsidering the rating. We will work further to make our work better. Thanks a lot for you support for our work.	Summary: The paper introduces CompressTracker, a novel general model compression framework that enhances the efficiency of transformer-based object tracking models. It innovatively segments transformer layers into stages, enabling a more effective emulation of complex teacher models by lightweight student models. The framework incorporates a unique replacement training technique, prediction guidance, and feature mimicking to refine the student model's performance. Extensive experiments demonstrate CompressTracker's effectiveness in significantly speeding up tracking models with minimal loss of accuracy, showcasing its potential for real-time applications on resource-constrained devices. Strengths: 1） Innovative Approach: The paper presents a novel compression framework, CompressTracker, which innovatively addresses the challenge of deploying transformer-based trackers on resource-limited devices by significantly reducing model size and computation cost without substantial loss of accuracy. 2）Structural Flexibility: A key advantage of the proposed framework is its structural agnosticism, allowing it to be compatible with any transformer architecture. This flexibility enables the adaptation of CompressTracker to various student model configurations, catering to diverse deployment environments and computational constraints. 3）Efficiency and Performance: The paper demonstrates through extensive experiments that CompressTracker achieves a remarkable balance between inference speed and tracking accuracy. It notably accelerates the tracking process while maintaining high performance levels, as evidenced by the nearly 96% retention of original accuracy with a 2.17× speedup. Weaknesses: 1）The concept of "prediction guidance and stage-wise feature mimicking" and the idea of BEVDistill [1] seem somewhat similar. 2）Despite the model's efficiency in inference, the training process for CompressTracker is relatively inefficient. 3）While the paper shows promising results on certain benchmarks, there may be concerns about how well these findings generalize across different types of tracking tasks and real-world scenarios. 4）The paper does not compare with other model compression techniques, such as knowledge distillation, model quantization, and pruning. 5）According to the results in Table 3, I observed that the outcomes of CompressTracker-2 are inferior to those of MixFormerV2-S. What could be the reason for this? 6）It is necessary to apply compression to other tracking models in order to further validate the efficacy of the CompressTracker presented in this paper. 7）The authors lack a sufficiently comprehensive review of the related work. The authors should give more reasonable related work by carefully introducing the recent approaches to tracking with compression, such as [2]. [1] BEVDistill: Cross-Modal BEV Distillation for Multi-View 3D Object Detection, ICLR 2023. [2] Distilled Siamese Networks for Visual Tracking, TPAMI 2021. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Please refer to weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your support and insights. We appreciate your deep understanding of the innovation and effectiveness of our work. We are grateful for your support for our work! ## Q1: Difference from BEVDistill Our CompressTracker is quite different from BEVDistill. (1) Purpose and Scope. CompressTracker is designed for visual object tracking, whereas BEVDistill transfers depth information from LiDAR to image backbones. (2) Feature Alignment. BEVDistill aligns BEV features from LiDAR and image encoders, while CompressTracker enforces that each stage in the student model’s encoder mimics the corresponding stage in the teacher model. Unlike BEVDistill, which runs encoders separately, CompressTracker performs feature mimicking at each encoder stage. (3)Distillation Approach. BEVDistill employs a complex sparse instance distillation method, whereas CompressTracker only uses a simple loss function for supervision. In a summary, our CompressTracker differs from BEVDistill in the motivation, target of distillation, implement details, and supervision method. ## Q2: Training Inefficiency We acknowledge the training inefficiency in Limitation (Section 5). To the best of our knowledge, our CompressTracker achieves the best balance between accuracy and inference speed. While some methods train lightweight models directly on groundtruth, resulting in less training time but lower performance, CompressTracker outperforms these models, such as HiT and SMAT, in accuracy and maintains competitive or superior inference speed (Table 4). In contrast, other methods, like MixFormerV2, use complex multi-stage training strategies, leading to much longer training times. Appendix Table 13 shows that our CompressTracker-4 (20 hours) requires only about 1/6 of the training time of MixFormerV2-S (120 hours) and achieves a 5.5 AUC improvement on LaSOT over MixFormerV2-S. Although our CompressTracker have a slightly longer training time, CompressTracker surpasses previous models and achieves the best balance between inferece speed and accuracy. We will develop more efficient training methods to enhance accuracy and decrease training duration. ## Q3: Generalization Ability Our CompressTracker demonstrates strong performance and generalization across various benchmarks and real-world scenarios. We conduct additional experiments on datasets like DepthTrack, which includes depth images from challenging conditions, and OTB, another popular RGB benchmark. As shown in the table, our CompressTracker-4 consistently maintains high performance across these datasets, highlighting its robustness and generalization ability. We will add experiments on more benchmarks in the final version to verify the generalizaiton ability. \| Dataset \| OSTrack \| CompressTracker-4 \| \|-------\|-------\|-------\| \| DepthTrack \| 51.5 \| 49.7 \| \| OTB \| 69.4 \| 68.4 \| ## Q4: Other Model Compression Techniques We have compared CompressTracker with several model compression techniques in our paper. As shown in Table 10 and Figure 3, our CompressTracker outperforms knowledge distillation (row #7 in Table 7 and 'Distill Training' in Figure 3) and surpasses MixFormerV2-S by 5.5 AUC on LaSOT, despite MixFormerV2-S using pruning for speedup (Table 4). We appreciate your suggestion and will add experiments comparing other model compression techniques, such as model quantization, in the final version of the manuscript. ## Q5: CompressTracker-2 CompressTracker-2 and MixFormerV2-S exhibit different performances across various datasets. CompressTracker-2 outperform MixFormerV2-S by 2.4 AUC on TrackingNet, while MixFormerV2-S surpasses CompressTracker-2 by 3.3 AUC on UAV123 and 3.2 AUC on LaSOT extension. Both models perform similarly on LaSOT and TNL2K. It is noteworthy that CompressTracker-2 uses only 2 transformer blocks, whereas MixFormerV2-S includes 4 transformer blocks, similar to our CompressTracker-4. We also conduct experiment to compress MixFormerV2 into CompressTracker-M-S, which shares the same structure with MixFormerV2-S. As shown in Table 2. Our CompressTracker-M-S substantially outperforms MixFormerV2-S across five datasets while maintaining the same speed. Additionally, MixFormerV2-S’s multi-stage training is more time-consuming (120 hours) compared to CompressTracker’s 20 hours for CompressTracker-4, as shown in Appendix Table 13. Morevoer, the model reduction paradigm utilized by MixFormerV2-S restricts the structure of student models to be consistent only with the teacher’s model, while our CompressTracker framework supports a diverse range of any transformer architectures for student, thanks to our stage division. We will add more analysis to highlight the advantage of CompressTracker in the final version. ## Q6: Applied to Other Tracker Yes! Applying compression to other tracking models is crucial for further validating CompressTracker's efficacy. We have conducted extensive experiments to assess the effectiveness and generalization ability of CompressTracker. We compress OSTrack into different layers in Table 1, and compress MixFormerV2 into CompressTracker-M-S, which is the same as MixFormerV2-S, in Table 2. Besides, we also compress OSTrack into CompressTracker-SAMT with 4 SMAT layers, which is the same as SMAT, in Table 3. In total, We have compressed 2 kinds of teacher models into 7 different student models with varying model structures. Our CompressTrackers with different configurations outperform their counterparts, demonstrating effectiveness and generalization ability of CompressTracker. We will add additional experiments applying compression to more tracking models to further validate the efficacy of our CompressTracker in the final version. ## Q7: Related Work Thanks for your suggestion, we will add additional reviews of related work, such as [1], to better clarify the difference between CompressTracker and previous works in the final version. [1] Distilled Siamese Networks for Visual Tracking, TPAMI 2021. --- Rebuttal Comment 1.1: Comment: Thanks for the authors. All concerns have been addressed. I will keep my rating. --- Reply to Comment 1.1.1: Comment: We are grateful for your recognition of the novelty and efficiency of our work. We are glad that our response has addressed your concerns. Thanks a lot for you support for our work.	Summary: In this paper, the authors proposed a general model compression framework for efficient Transformer object tracking, named CompressTracker. The method adopts a novel stage partitioning strategy to divide the Transformer layers of the teacher model into different stages, enabling the student model to more effectively simulate each corresponding teacher stage. The authors also designed a unique replacement training technique, which involves randomly replacing specific stages in the student model with specific stages in the teacher model. Replacement training enhances the student model's ability to replicate the behavior of the teacher model. To further force the student model to simulate the teacher model, we combine predictive guidance and staged feature imitation to provide additional supervision during the compression process of the teacher model. The authors conducted a series of experiments to verify the effectiveness and generality of CompressTracker. Strengths: The author has clear ideas and the article is easy to understand. He proposes a general compression framework for single object tracking. This method can efficiently compress large object tracking models into small models. The author has conducted a large number of experiments to prove the effectiveness of this method. Weaknesses: The font size of the pictures in the article is too small. The author can adjust the font size appropriately to facilitate reading. The training time line in Figure 1a is blocked, resulting in incomplete display. The font size of the tables is inconsistent, for example, the font size of Tables 5, 6, 7, and 8 is too large. The abstract is redundant and can be appropriately deleted. Technical Quality: 3 Clarity: 3 Questions for Authors: 1.The font size of the pictures in the article is too small. The author can adjust the font size appropriately to facilitate reading. 2.The training time line in Figure 1a is blocked, resulting in incomplete display. 3.The font size of the tables is inconsistent, for example, the font size of Tables 5, 6, 7, and 8 is too large. 4.The abstract is redundant and can be appropriately deleted. 5.Did the author test the speed on other devices, such as CPU? 6.Will the codebe open source? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: For lightweight tracking models, the training time is too long. The author can try to find new ways to reduce the time spent on training. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your insightful advice. We sincerely appreciate your valuable comments and recognition of the novelty of our work. We will carefully review and modify our manisctipt based on your suggestion to improve its presentation. We greatly value your support for our work! ## Q1: Font Size in Picture We apologize for the inconvenience caused by the small font size. We will follow your advice and increase the font size in Figures 1, 2, 3, 4, and 5 to enhance readability. ## Q2: Blocked Figure 1a Thank you for pointing out our oversight! We will address and correct the issue in the revised version of the manuscript. ## Q3: Inconsistent Font Size in Table and Redundant Abstract Thank you very much for your valuable suggestions! We will reorganize the table and revise the abstract to improve the overall presentation. ## Q4: Speed on Other Devices We evaluated the speed of CompressTracker on an Intel(R) Xeon(R) Platinum 8268 CPU @ 2.90GHz. The results are presented in the table below. Our CompressTracker achieves the best balance between accuracy and speed. We just propose a novel model compression framework rather than a specific model. To demonstrate the effectiveness of our framework, we applied it to compress several tracking models. Due to the framework's strong generalization capabilities, other researchers can select appropriate student models based on their hardware and apply our framework accordingly. We will add a more detailed explanation of this aspect in the revised version. \| Model \| AUC on LaSOT \| FPS(CPU) \| \|-------\|-------\|-------\| \| CompressTracker-2 \| 60.4 \| 29 \| \| CompressTracker-3 \| 64.9 \| 22 \| \| CompressTracker-4 \| 66.1 \| 18 \| \| CompressTracker-6 \| 67.5 \| 13 \| \| HiT-Base \| 64.6 \| 33 \| \| E.T.Track \| 59.1 \| 42 \| \| FEAR-XS \| 53.5 \| 26 \| \| Model \| AUC on LaSOT \| FPS(CPU) \| \|-------\|-------\|-------\| \| CompressTracker-M-S \| 62.0 \| 30 \| \| MixFormerV2-S \| 60.6 \| 30 \| \| Model \| AUC on LaSOT \| FPS(CPU) \| \|-------\|-------\|-------\| \| CompressTracker-SMAT \| 62.8 \| 31 \| \| SMAT \| 64.6 \| 33 \| ## Q5: Code Release Yes! We will release our code upon acceptance! Thank you once again for recognizing and supporting our work! ## Q6: Training Time We confirm this slightly longer training time in Limitaion (Section 5) and we are purchasing new technique to reduce the training time and further improve the accuracy of our CompressTracker.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper introduces CompressTracker, a general model compression framework for efficient transformer-based object tracking. CompressTracker divides the teacher model into stages corresponding to student model layers and randomly replaces student stages with teacher stages during training. It also aligns the teacher and student models using prediction guidance and feature mimicking. The framework gradually increases the probability of using student stages throughout training. CompressTracker achieves significant speed improvements while maintaining high accuracy. For example, CompressTracker-4 accelerates OSTrack by 2.17x while preserving 96% of its accuracy on LaSOT. Strengths: - Versatility: Compatible with various transformer architectures for student models. - Efficiency: Achieves a good balance between inference speed and tracking accuracy. - Streamlined training: Offers a single-step, end-to-end training process, simplifying the compression pipeline. Weaknesses: - Limited theoretical analysis: The paper focuses on empirical results without providing much theoretical justification for the proposed methods. - Lack of ablation on some components: Some components of the framework are not thoroughly explored. For instance, the impact of different feature mimicking strategies is not extensively analyzed. - Performance and Efficiency Trade-off: While CompressTracker maintains high accuracy, there's a slight performance drop compared to the original model. Training time for CompressTracker-4 (with only 4 blocks) exceeds that of the original OSTrack. This trade-off between training efficiency, inference speed, and model performance requires further optimization. - The core idea of reducing the number of Transformer blocks is not new. Similar approaches have been used in other models like TinyViT[1] and MiniViT[2]. [1] Wu K, Zhang J, Peng H, et al. Tinyvit: Fast pretraining distillation for small vision transformers[C]//European conference on computer vision. Cham: Springer Nature Switzerland, 2022: 68-85. [2] Zhang J, Peng H, Wu K, et al. Minivit: Compressing vision transformers with weight multiplexing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 12145-12154. Technical Quality: 3 Clarity: 4 Questions for Authors: Please refer to the weakness. Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: No Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your recognition of the efficiency and value of our work, as well as your insightful advice and comments. Other reviewers, such as Reviewer M8LK, Reviewer 8Frr, and Reviewer 2g8j, have also acknowledged the novelty and effectiveness of our approach. Our CompressTracker achieves an optimal balance between performance and efficiency. Our CompressTracker-4 obtains 66.1% AUC on LaSOT with 228 FPS. We would be grateful if you could reconsider our work and offer your support! ## Q1: Theoretical Analysis Thanks for your valuable advice. We will incorporate your suggestions and add more theoretical justification in the revised manuscript. Our replacement training is similar to the dropout technique. While dropout technique reduces the over fitting by randomly dropping neurons, our replacement training strategy prevent the student model from overfitting to specific teacher stages and improve the robustness through introducing randomness. Please refer to the comment for detailed derivation. We explain the superiority of replacement training from an information-theoretic perspective. Similar to dropout, our replacement training strategy increases the training entropy by introducing uncertainty, which enhances the generalization ability of student and reduces the overfitting of the student model to specific teacher stages. Extensive experiment results further validate effectiveness of our CompressTracker. We will follow your advice and add the theoretical justification in the final version. ## Q2: Ablation Study Indeed, we have conducted a series of ablation experiments to investigate each component of our CompressTracker, as detailed in Section 4.4. We examine the effects of Backbone Initialization (Table 5), Decoder Initialization and Optimization (Table 6), Stage Division (Table 7), Supervision (Table 10 and Figure 3), Training Epochs (Table 8). Besides, we perform ablation studies on Replacement training (Appendix Table 11), Progressive Replacement (Appendix Table 12), and Replacement Probability (Appendix Figure 4). We also compare the training times in Appendix Table 13 and Figure 5 to highlight the training efficiency of our CompressTracker. These comprehensive experiments are designed to assess the impact of each component thoroughly. For feature mimicking strategies, we analyzed various methods including ‘L1’, ‘L2’, ‘Proj + L2’, and ‘CE’ (L1-norm, L2-norm, linear projection before L2-norm, and cross-entropy). Results are shown in the following table. he performance of different strategies varies, and the 'L2' strategy adopted by our CompressTracker does not achieve the highest accuracy. To demonstrate that the improvement comes from our model compression framework instead of the specific and complex feature mimicking strategies, we just adopt the simplest approach: L2 distance without any complicated design. Our CompressTracker achieves the best balance between inference speed and accuracy, indicating that its superiority comes from our unified framework rather than from complex loss function designs. We believe that exploring more sophisticated strategies could further enhance performance. We will follow your suggestion and add more experiments on the components of our CompressTracker to provide a more thorough analysis in the final manuscript. \| Strategy \| AUC on LaSOT \| \|-------\|-------\| \| L1 \| 65.0 \| \| L2 \| 65.2 \| \| Proj + L2 \| 65.3 \| \| CE \| 65.4 \| ## Q3: Performance and Efficiency Trade-off Compared to previous works, our CompressTracker achieves the best balance between performance and efficiency. As shown in Appendix Table 13, CompressTracker-4 takes about 20 hours to train, slightly longer than OSTrack’s 17 hours, but retains 96\% of its performance on LaSOT (66.1\% AUC) while achieving a 2.17× speedup. In contrast, lightweight models like HiT, despite requiring less training time, exhibit lower accuracy. HiT matches CompressTracker-4 in speed (175 FPS) but performs worse (64.6 AUC on LaSOT) compared to CompressTracker-4 (66.1 AUC and 228 FPS). Furthermore, methods with complex multi-stage reduction, such as MixFormerV2-S, require much longer training (120 hours) but deliver inferior performance (60.6 AUC on LaSOT) compared to CompressTracker-4 (66.1 AUC), which achieves significantly better results with just 20 hours of training. Many prior works, like HiT and MixFormerV2, aim to balance accuracy and efficiency, and our framework surpasses these methods. We believe our work offers a novel perspective on this issue for the tracking field, and the trade-off is not weakness but an important area for tracking. We have confirmed the training inefficiency in the Limitation (Section 5), and we will work on to enhance accuracy and reduce training duration. ## Q4: Difference from TinyViT and MiniViT Our CompressTracker significantly differs from TinyViT and MiniViT. TinyViT enhances inference speed by aligning a light student outputs with teacher predictions, while MiniViT reduces model parameters through weight multiplexing without affecting inference speed. Both methods focus on image classification, whereas our framework is the first to propose a unified compression approach specifically for tracking. Our model differs from previous work in the following main ways. (1) We employs a novel stage division strategy, segmenting transformer layers into distinct stages, which TinyViT and MiniViT do not use. (2) We introduce a replacement training that randomly substitutes specific stages in student with those from teacher, unlike the isolated training of TinyViT and MiniViT. (3) Figure 3 compares different training strategies, showing our CompressTracker’s superior accuracy over methods like ‘Distill Training’ used in TinyViT and MiniViT, which highlights the advantages of our framework. We will include a more detailed analysis of how our approach differs from previous works and the advantages of our CompressTracker in the final version of the manuscript. --- Rebuttal 2: Title: Theoretical analysis of the superiority of our replacement training Comment: Please allow me to explain the superiority of our replacement training strategy from an information theoretic perspective. Dropout increases entropy during training by randomly dropping neurons to reduce model overfitting. Our replacement training strategy is similar to dropout technique in that it introduces the training uncertainty and increase the entropy to reduce the overfitting of the student model to specific stages of the teacher. Next, we will give theoretical derivations to show why our replacement training, similar to dropout, can increase the entropy of training. Firstly, we provide a uniform formal definition of both dropout and our replacement training. Dropout can be expressed as: if $b(p) = 1$, $h^{'} = h / p$, and if $b(p) = 0$, $h^{'} = 0$, where $h$ is the output of a specific neuron, $h^{'}$ is the output after applying dropout, and $b(p)$ is the Bernoulli sampling with probability $p$. And for our replacement training, $f_{i}$ can be defined as: if $b(p) = 1$, $f_{i} = SS_{i}(f_{i-1})$, and if $b(p) = 0$, $f_{i} = TS_{i}(f_{i-1})$, where $f_{i-1}$ and $f_{i}$ are the input and output feature of the $i$-th stage, and $SS_{i}$ and $TS_{i}$ denote the $i$-th student stage and teacher stage. Without the replacement training, $f_{i} = SS_{i}(f_{i-1})$. After providing a unified representation of dropout and our replacement training, we can give a derivation of our replacement training based on the formula of dropout. Entropy $H(X)$ can be used to measure the uncertainty of a random variable $X$, which is defined as: $H(X)= - \sum_{x} P(x)\mathrm{log}P(x),$ where $P(x)$ is the probability distribution of the random variable $X$ taking values $x$. Thus, the origin entropy of $f$ can be written as: $H(f_{i})= -\sum_{f_{i}} P(f_{i})\mathrm{log}P(f_{i}).$ Because $P(f_{i}) = P(SS_{i}(f_{i-1}))$, entropy $H_{SS}(f_{i})$ of student stage is: $H_{SS}(f_{i})= -\sum_{f_{i-1}} P(SS_{i}(f_{i-1}))\mathrm{log}P(SS_{i}(f_{i-1})),$ and entropy $H_{TS}(f_{i})$ of teacher stage can be written as: $H_{TS}(f_{i})= -\sum_{f_{i-1}} P(TS_{i}(f_{i-1}))\mathrm{log}P(TS_{i}(f_{i-1})).$ With our replacement training, the $P(f_{i})$ is: if $b(p) = 1$, $P(f_{i}) = pP(SS_{i}(f_{i-1}))$, and if $b(p) = 0$, $P(f_{i}) = (1-p)P(TS_{i}(f_{i-1}))$. Thus, the entropy of $P(f_{i})$ can be written as: $H(f_{i}) = -\sum_{f_{i}} P(f_{i})\mathrm{log}P(f_{i})$ $= -[\sum_{f_{i-1}}(1-p)P(TS_{i}(f_{i-1})) \mathrm{log}((1-p)P(TS_{i}(f_{i-1}))) + \sum_{f_{i-1}} pP(SS_{i}(f_{i-1})) \mathrm{log}(pP(SS_{i}(f_{i-1})))]$ $= -[(1-p)\mathrm{log}(1-p)\sum_{f_{i-1}}P(TS_{i}(f_{i-1}))] + (1-p)\sum_{f_{i-1}}P(TS_{i}(f_{i-1}))\mathrm{log}P(TS_{i}(f_{i-1})) + p\mathrm{log}p\sum_{f_{i-1}}P(SS_{i}(f_{i-1})] + p\sum_{f_{i-1}}P(SS_{i}(f_{i-1}))\mathrm{log}P(SS_{i}(f_{i-1}))$ $= -[(1-p)\mathrm{log}(1-p)+p\mathrm{log}p] + (p-1)H_{TS}(f_{i}) + pH_{SS}(f_{i}).$ Thus, the difference in entropy $\Delta H(f_{i})$ is: $\Delta H(f_{i}) = -[(1-p)\mathrm{log}(1-p)+p\mathrm{log}p]+ (p-1)H_{TS}(f_{i}) + pH_{SS}(f_{i}) - H_{SS}(f_{i})$ $= -[(1-p)\mathrm{log}(1-p)+p\mathrm{log}p] + (1-p)(H_{SS}(f_{i}) - H_{TS}(f_{i})).$ When $p$ is in range $[0,1]$, the term $-[(1-p)\mathrm{log}(1-p)+p\mathrm{log}p]$ is always positive. Since the student model generally performs worse than the teacher model, $H_{SS}(f_{i}) - H_{TS}(f_{i})$ is typically positive as well. Thus, $\Delta H(f_{i})$ is consistently positive. This demonstrates that our replacement training achieves a similar effect to dropout: increases training entropy. Based on the above theoretical analysis, we demonstrate that our replacement training method is similar to dropout in its ability to increase training entropy. Consequently, our replacement training helps reduce the overfitting of the student model to specific stages of the teacher.	null	null	null	null	null	null
AED: Adaptable Error Detection for Few-shot Imitation Policy	Accept (poster)	Summary: The paper aims to address the adaptable error detection problem, i.e., detecting the error behavior of a robot policy in an unseen environment. Solving the problem ensures that the robot is stopped before performing behavior that causes disruptions to the surroundings. The AED problem introduces three challenges: the abnormal behavior in the unseen environment is not observed in the training data and common anomaly detection method cannot be used; with complex background, there is no noticeable change between frames and this makes it difficult to indicate when the error occurs; AED is required to terminate the policy timely once the error occurs. The three challenges make it difficult to apply the current approaches on anormaly detection. The paper creates a new benchmark specially designed for adaptable error detection, which consists of 322 base and 153 new environments encompassing six indoor tasks and one factory task. The paper proposes a PrObe architecture consists of a pattern extractor and a flow generation equipped with a pattern loss, a temporal contrastive loss and a classification loss to predict the error probability. The paper conducts experiments on the proposed benchmark and achieves the top 1 counts, average ranking, and average performance difference. Extensive ablative study justifies the effectiveness of the design choices. Strengths: Few-shot imitation learning learns a robot policy for a new environment with only a few demonstrations, which could accelerates the development of robot applications. However, with only a few demonstrations, there exist many corner cases that the robot could not observe in the demonstrations and the appearance and background of the scene are also different. This makes the policy perform error bahavior, which makes the surroundings fall into a dangerous situation. Thus, the problem of adaptable error detection is important to solve. The three loss designs not only supervisedly learn the error prediction but also unsupervisedly regulate the intermediate feature to follow some human priors including close features for frames with high task relation and sparse feature to detect frame changes. The experiment datasets and tasks are specially designed for the AED problems and selected baselines are strong error detection baselines. Weaknesses: One main contribution of the paper is that error detection could avoid the robots making disrupted behavior in the real-world environment. Thus, the authors may need a real-world environment that contains some disruption scenarios and test the method in such environments. Even the most realistic simulation may not simulate some real factors such as inaccurate control program of the robot arm. Technical Quality: 3 Clarity: 3 Questions for Authors: On line 207-208, normalizing by L2 norm could make all the features have the same norm but how could this mitigate the biases? Even the unit vector may only learn the background information in the visual inputs. On line 212-215, the authors need to note that the task embeddings are extracted using prior works, which is introduced only in the experiments. Could the authors provide more details or some examples on how to collect the positive and negative pairs for L_{tem}? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer SywQ, Thank you for your insightful comments and high recognition of our work! Our responses below are dedicated to addressing your questions and concerns. --- [W1] The authors may need a real-world environment that contains some disruption scenarios and test the method in such environments. - Thank you for the suggestion. As we discussed in the Limitation section (lines 745-758), there are still several obstacles to performing our AED task in real-world environments. Considering the capabilities and safety concerns of few-shot imitation (FSI) policies and AED methods in real-world environments, we first verified the AED tasks in realistic simulated environments. This approach was also inspired by previous work [19] and related research fields (e.g., safe exploration [57, 58, 59]). - We will continue to advance the progress of our AED task, including future sim-to-real and real-world experiments. Thank you again. [Q1] On line 207-208, normalizing by L2 norm could make all the features have the same norm but how could this mitigate the biases? - According to our experience, the feature values extracted from observations in different environments vary greatly, and such deviations may exacerbate errors in model judgment. Therefore, we divide the extracted features by their L2 norm to turn them into unit vectors, and let the subsequent pattern observer learn on the unit vectors. This can eliminate the deviation in value scale caused by domain (environment) changes. However, unit vectors from different domains may still have angular deviations, so in lines 207-208 we use the term "mitigate" rather than "eliminate." - We will continue to explore advanced strategies in domain adaptation to further mitigate the effects of domain change on the model. Thank you for this question. [Q2] On line 212-215, the authors need to note that the task embeddings are extracted using prior works, which is introduced only in the experiments. - Thank you for the suggestion. As you mentioned, these task embeddings are extracted from various few-shot imitation (FSI) policies, so how they are extracted can only be introduced when discussing which FSI policies are ultimately used. However, regarding the function of task embeddings (demonstrations) in our AED task (also mentioned by Reviewer qV6D), we can indeed provide more explanation in Section 5.2. - Please check our response to Reviewer qV6D's W2 for more details. Thank you again. [Q3] Could the authors provide more details or some examples on how to collect the positive and negative pairs for $L_{tem}$? - Thank you for the question. First, please note that according to our setting, the failed rollouts $X^b_{fail}$ in the base environment have frame-level labels. For each sampled failed rollout in each training iteration, we first randomly select one frame as the anchor. The anchor may be a normal frame or a frame where the error has occurred. Based on the selected anchor, we will sample the positive and negative samples using the frame-level labels of the failed rollout. After selecting the three samples, we calculate the temporal distance between the anchor and the positive sample, and the anchor and the negative sample. Then, we adjust the temporal-aware margin in $L_{tem}$ based on these temporal distances (line 230). Through this process, our $L_{tem}$ objective can simultaneously consider the latent and temporal correlations between frames. - We will include these details in our final version. Thank you again for this question. --- - Thank you again for your time and dedication in reviewing our paper. We appreciate your positive assessment of our work. - If you have any unresolved concerns after reading our responses, please let us know. We look forward to learning more from you during our constructive discussions. --- Rebuttal 2: Comment: Dear Reviewer SywQ, Thank you once again for your valuable reviews and comments. We have provided additional details to address the concerns you raised. Since the discussion period ends in two days, we would appreciate knowing if there are any unresolved issues that require further clarification. If so, we would be happy to discuss them with you. Otherwise, if all your concerns have been addressed, we kindly ask you to consider adjusting your evaluation score accordingly. Thank you again for your time and consideration.	Summary: This paper introduces Adaptable Error Detection (AED) within Few-Shot Imitation (FSI) tasks, a critical yet underexplored area in robotics and AI. The authors establish a novel benchmark for assessing AED methods and present PrObe, designed specifically for this task. Through comprehensive evaluations, the paper demonstrates that PrObe outperforms baseline models and shows robustness across different FSI policies. Strengths: S1. Well-written. The paper has a coherent narrative and logical flow. The problem scenario is effectively motivated with real-life examples that illustrate the practical importance and applicability of the research. Relevant literature is extensively discussed and consistently referenced throughout the paper. S2. Illustrations. The schematic illustrations provided in the paper are well made and enhance the comprehensibility of AED, PrObe, and their experimental analysis. The figures effectively delineate the structured protocol of the AED task across different phases, and the intricate architecture of PrObe, showcasing its components and their functions. S3. Unique challenges. The unique challenges established by AED distinguish it from FSAD and other tasks in prior works. AED's focus on online anomaly detection in novel environments, where behavioral errors occur without clear visual cues, establishes it as a critical area of research. S4. Novel components. PrObe establishes the usefulness of its design principles by outperforming other baselines. Augmenting rollouts, extracting patterns from policy embeddings, generating a pattern flow and fusing it with task-embeddings, and utilizing a temporal-aware triplet loss - all contribute to its effectiveness. S5. Comprehensive analysis. The paper extensively explores PrObe's application in FSI policies for AED. PrObe achieves better visual separability of features in the latent space between successful and erroneous trajectories, thereby enhancing the interpretability of learned embeddings. Notably, PrObe excels in temporally precise error detection, outperforming other methods that detect errors too late or too early. An in-depth ablation study confirms the stability and essential contributions of PrObe’s components. Additionally, the paper explores PrObe’s failure scenarios, the impact of demonstration quality, and viewpoint changes. Weaknesses: W1. While the textual descriptions in the Preliminaries section provide a solid understanding of the DC policy, incorporating mathematical notation would significantly enhance clarity and comprehension. Specifically, the paper could benefit from the following mathematical formulations: 1. The contents of history $h$, detailing what it encapsulates 2. The feature encoder and task-embedding network computing the history, demonstration, and task-embedding features 3. Task-embedding network padding the sequences 4. The actor policy, e.g, $\pi(a\|o_t, f_h, f_{\zeta})$ 5. The inverse dynamics module 6. NLL and MSE objective in this particular setting W2. The average difference metric presented in Figure 4, which compares each method's performance to the second-best performing method, may be inadequate. This approach is only sensible when comparing two methods. A more informative assessment might measure performance gains relative to the worst-performing method, or performance losses relative to the best-performing method. Alternatively, using a standardized baseline or oracle for comparisons could provide a clearer and more meaningful evaluation. W3. PrObe is only shown to be effective for evaluating policies trained on image data. As a primary example, the authors motivate the FSI ED problem setting on robotic tasks. Although sensor data cannot always fully capture the properties and locations of separate objects, policies for robotic manipulation tasks are often trained on proprioceptive data—such as joint angles and torques—to achieve better precision. Image data, though rich in environmental context, can sometimes be unavailable or impractical due to factors like occlusions, lighting conditions, and higher training costs. The inherent differences between image-based and proprioceptive data mean that the latent space characteristics and the patterns critical for error detection would vary. Image-based models capture visual features like shapes and spatial relationships, whereas proprioceptive-focused models emphasize dynamics and kinematic features, including the robot’s mechanical interactions. While there's no inherent reason to doubt PrObe’s capability with different data modalities, the effectiveness of its pattern extractor when applied to latent features learned from proprioceptive data remains uncertain without empirical validation. Conducting experiments with proprioceptive sensor data and demonstrating successful results would further validate the robustness, generality, and applicability of PrObe. ### Minor Improvements 1. The second and third contributions—introducing a novel benchmark and PrObe—are blended together; these could be separate for clarity. 2. Please add a reference in the main paper that the limitations are in Appendix F. 3. Line 125 “semantically” refers to language or linguistics, which doesn’t seem to be the intended use of the word here 4. Line 133 “learns implicit mapping” → “learns an implicit mapping” 5. Line 136 “current history” → “the current history” 6. Line 154 “and few” → “and a few” 7. Line 168 “high controllable” → “highly controllable” 8. Line 174 “and few” → “and a few” 9. Line 192 “from agent’s” → “from the agent’s” 10. Line 205 “take as input the history features $f^h$” →“takes the history features $f^h$ from … as input” 11. Footnote on page 4 “thet” →”they” 12. Line 614 “are are” → “are” 13. Line 661 “Pick & Plcae” → “Pick & Place” Technical Quality: 4 Clarity: 4 Questions for Authors: Q1. Why did the authors choose the main conference track instead of the Datasets & Benchmarks track if they propose a benchmark? Q2. What is the intervention method employed at critical moments to the agent’s policy to collect failed rollouts? Q3. Why were TaskEmbED and LSTMED chosen among the 4 baselines for embedding visualization? Q4. Is AED only applicable for error detection on demonstration-conditioned policies or does it extend to more general policies? Q5. Is PrObe also able to evaluate policies that have been trained for robotic manipulation on proprioceptive data, not image data? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have extensively discussed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer 3VqP, Thank you for your insightful comments and high recognition of our work! Our responses below are dedicated to addressing your questions and concerns. --- [W1] Incorporating mathematical notation would enhance clarity and comprehension. - Thank you for your suggestion. We will add mathematical notations to the cases you listed. For example, we will use $h_t \coloneqq (o_0, o_1, \ldots, o_t)$ for the rollout history at time $t$, where $o$ represents the observation. We will also review the paper to identify any parts that can be enhanced for clarity and comprehension. [W2] The average difference metric presented in Figure 4 may be inadequate. - Thank you for pointing out this important issue. As per your suggestion, we provide the average performance difference metric comparing each method with the worst and best methods in the following tables. Comp. w/ the worst method: \| \| SVDDED \| TaskEmbED \| LSTMED \| DCTED \| PrObe \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| AUROC \| 7.05% \| 41.63% \| 41.59% \| 44.20% \| 67.29% \| \| AUPRC \| 2.61% \| 35.42% \| 61.67% \| 57.13% \| 78.03% \| Comp. w/ the best method: \| \| SVDDED \| TaskEmbED \| LSTMED \| DCTED \| PrObe \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| AUROC \| -34.87% \| -16.90% \| -15.95% \| -15.44% \| -2.45% \| \| AUPRC \| -32.28% \| -19.63% \| -10.87% \| -13.17% \| -1.06% \| - From the results under these two new metrics, it's obvious that our PrObe method significantly outperforms the compared baselines. We will include these new results in our final version. [W3, Q5] Is PrObe also able to evaluate policies trained on proprioceptive data? - Thank you for the interesting question. We agree that it is worth exploring whether PrObe can observe pattern changes in proprioceptive data (data in another modality). However, considering our work's setting (refer to the scenario presented in Section 4), it may not be feasible to conduct experiments on proprioceptive data for two main reasons: 1. The configurations of experts and agents are different. It may be very challenging for few-shot imitation (FSI) policies to learn tasks from different joint sets. As you mentioned, the proprioceptive data may not provide all the necessary task-related information that could potentially be extracted from image data. 2. Our setting assumes that we have less information and control in novel environments. Even if we assume that the error detector and FSI policy can be integrated into the same agent, we may not be able to access the expert's proprioceptive data (for example, if the expert is a customer). - Although it is difficult to conduct experiments on training PrObe with proprioceptive data under our current problem setting, we will explore whether PrObe has the ability to learn from data of different modalities in the future. [Q1] Why did the authors choose the main conference track instead of the Datasets & Benchmarks track? - Thank you for the question. In this work, we regard the AED task as a new learning problem. In addition, we propose a new method, PrObe, to address this challenging task. The AED benchmark is a platform to verify the effectiveness of various AED methods. For us, the first two parts (a new task and a new method) are the main contributions of this work, and thus, we decide to submit it to the main track. [Q2] What is the intervention method employed at critical moments to collect failed rollouts? - Thank you for the question. Different intervention methods are designed for different tasks, which can be roughly divided into two categories: (1) The interactive object is correct, but the interaction fails (e.g., failure to grasp the object, failure to press the button, failure to close the drawer), and (2) Interacting with the wrong objects (e.g., placing objects in the wrong container, closing the wrong drawer). - For the first case, we add a random shift to the accurate pose. For the second case, we replace the accurate pose with the pose of the incorrect object, which also includes a random shift. We will include these details in each task's description. [Q3] Why were TaskEmbED and LSTMED chosen for embedding visualization? - Due to space constraints when drafting the paper, we report two baselines with the most different attributes in our embedding visualization experiment (cf. Table 3 in the Appendix). However, we can provide visualization results for all baselines. Please refer to the PDF file in our global rebuttal for more details. We will integrate all visualization results into our final version. [Q4] Is AED only applicable for error detection on demonstration-conditioned policies? - Thank you for the question. In the problem setting we are considering (recall the scenario provided in Section 4), the policy has to learn the task through observing demonstrations performed in a novel environment. Mainstream few-shot imitation policies following this setting are usually demonstration-conditioned, so our AED task and Probe method focus on detecting the errors caused by this type of policy. - If the policy can now be directly trained and tested in a single scene, e.g., conventional imitation learning, then intuitively, our Probe can work in this setting by simply removing the fusion process of pattern flow and task embeddings. In other words, Probe would use only the pattern flow to detect errors. We will continue to explore how to improve the versatility of Probe for various policies. Typos & minor improvements - Thank you so much for your careful review of our paper. We will fix these typos and revise the paper following your instructions. --- - Thank you again for your time and dedication in reviewing our paper. We appreciate your positive assessment of our work. - If you have any unresolved concerns after reading our response, please let us know. We look forward to learning more from you during our constructive discussions. --- Rebuttal Comment 1.1: Comment: I appreciate the authors’ efforts in addressing my concerns and answering the questions. W2. I do think the results that you provided in this format are more coherent. I suggest you select one of them and swap out the current one in Figure 4. I would say that including both is unnecessary as they infer the same thing and it would just take up more space. I think the comparison w/ the worst method is more informative, as the scale is larger. W3. Thank you for the explanation. According to your reasoning, it indeed seems complicated to incorporate proprioceptive data. I do believe that camera footage for error detection is appropriate and sufficient, and image-based learning, after all, generally poses a greater challenge. I thank the authors for their work. All my concerns and inquiries have been addressed. I will keep the current score. A higher score would necessitate a more high-impact paper 1) tackling a broader setting, 2) surpassing prior methods by a landslide, or 3) having experiments conducted in many simulation environments or even physical robots. --- Reply to Comment 1.1.1: Comment: We are grateful for the reviewer's prompt and detailed feedback, which provided valuable advice during our discussion. We are also pleased that our responses have addressed all of the reviewer's concerns and questions. We would like to thank the reviewer once again for their high recognition of our work.	Summary: This paper proposed a task Adaptable Error Detection (AED) that attempts to perform online behavior error detection for FSI policies, it advocate three main challenges comparing to FSAD: novel environment, no notable change for behavior error that AED tries to detect, and it has to be conduct simultaneously with policy execution to ensure timely error detection and damage control. The AED benchmark is provided, and a AED method, PrObe is proposed to solve AED through error prediction based on fusion of task embedding and flow pattern generated from history of rollout. The PrObe achieve good performance compare to baselines, and paper provided comprehensive studies analyzing PrObe error detection timing, it's embedding pattern, the importance of components, the stability of error detection, and provided pilot error correction to demonstrate practicality of the AED task. Strengths: 1. AED is a important task for FSI research. 2. The proposed PrObe is very effective in AED task for various tasks. 3. abundant analysis is provided on PrObe timing, embedding analysis, performance stability, ablation and more. Weaknesses: The writing has improvement room: 1. line 161 - line 162: I am confused: if frame level label is not available, how come the measurement is averaging over actions? 2. The function of demonstration data is not clearly described in section 5.2, more explanation is helpful. The AED benchmark is not well described in the writing. Technical Quality: 3 Clarity: 2 Questions for Authors: See above. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitation in terms of proposed PrObe is discussed through analysis experiments, which make sense to me. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer qV6D, Thank you for your insightful comments and high recognition of our work! Our responses below are dedicated to addressing your questions and concerns. --- [W1] If frame-level label is not available, how come the measurement is averaging over actions? - We try to make Eq. 1 a general form, so we express it in such a way that we can obtain detection accuracy across frames. However, as we illustrated in the practical scenario (lines 163-166), frame-level labels are often unavailable during the testing phase in practice. Therefore, the A($\cdot, \cdot$) we use in testing is sequence-level. In addition, we conduct the timing accuracy experiments to verify the accuracy of different methods in determining the timing of error occurrences. - We will revise this part to make it more clear. Thanks for the question. [W2] The function of demonstration data is not clearly described in section 5.2. - The demonstration data originally served as important information for the few-shot imitation policy to understand how to perform tasks in novel environments. In our newly proposed AED task, these demonstrations are also used for AED methods to determine whether the current state deviates from the demonstrations. - We agree with the reviewer that we can provide more explanation on the function of demonstration data in our AED task and will include it in our final version. Thank you for pointing out this issue. [W3] The AED benchmark is not well described in the writing. - We have provided the details of our AED benchmark in Section D of the Appendix, including the task attributes, each task's description, success conditions, and potential behavior errors. We also refer readers to Section D of the Appendix when introducing evaluation tasks (Section 6.1, line 246) in the main text. - If there are any additional details about our AED benchmark that the reviewer is interested in, please feel free to let us know. --- - Thank you again for your time and dedication in reviewing our paper. We appreciate your positive assessment of our work. - If you have any unresolved concerns after reading our responses, please let us know. We look forward to learning more from you during our constructive discussions. --- Rebuttal 2: Comment: Dear Reviewer qV6D, Thank you once again for your valuable reviews and comments. We have provided additional details to address the concerns you raised. Since the discussion period ends in two days, we would appreciate knowing if there are any unresolved issues that require further clarification. If so, we would be happy to discuss them with you. Otherwise, if all your concerns have been addressed, we kindly ask you to consider adjusting your evaluation score accordingly. Thank you again for your time and consideration. --- Rebuttal Comment 2.1: Comment: Thanks for addressing my questions, i am happy to increase my score to 7. --- Reply to Comment 2.1.1: Comment: Thank you for your feedback! We are pleased that our responses have addressed your concerns. We will revise the paper following your valuable suggestions. We are also grateful to the reviewers for their support and for increasing their ratings of our work. Thank you again.	null	null	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their constructive and insightful comments on our work. We are excited that our work possesses several strengths recognized by the reviewers, as summarized below: - Our proposed AED task is important (all reviewers) with unique challenges (3VgP). - Our proposed solution PrObe contains novel components (3VgP) and is effective (all reviewers). - Our work provided comprehensive analysis (qV6D, 3VgP); the compared baselines are strong (SywQ). - Our paper is well-written (3VgP) with well-made illustrations (3VgP). Regarding the concerns and questions raised by the reviewers, we have provided more details and evaluation results under new metrics to address them. Additionally, we have analyzed the feasibility of conducting the studies suggested by the reviewers. Finally, we will refine our paper based on the improvement suggestions from all reviewers. We have posted our responses under each reviewer's comment separately. Please let us know if there are still any unresolved concerns. Thank you again for reviewing our paper. --- In response to the reviewer's inquiry, the attached PDF provides embedded visualizations of all baselines. Please check the explanation in the PDF with Table 3 in the Appendix. Pdf: /pdf/107fdda21820b8fe7272bfe9d188996b96402fde.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
e-COP : Episodic Constrained Optimization of Policies	Accept (poster)	Summary: In this paper, the authors propose a policy optimization algorithm for constrained Reinforcement Learning (RL) problem in a finite horizon setting. First, the authors develop a policy difference lemma for the finite horizon MDP problem. Following that, the authors combine a set of ideas to propose the e-COP algorithm. The authors show that the proposed algorithm is numerically much stable compared to the state-of-the-art solutions. Optimality of the proposed scheme is established under certain assumptions. Extensive simulations have demonstrated that the performance of the algorithm is better than existing methods. Strengths: The paper is well-written and easy to follow. The results and claims presented in the paper appears to be correct. Weaknesses: The idea of using a quadratic penalty term to improve the stability near the edge of the constraint set boundary is a purely heuristic based approach. The idea of approximating $\rho_{\pi}$ by the empirical distribution from the policy of the previous episode does not appear to be a good approximation. The idea of allowing the agent to act in a constrained manner even before the constraint is violated may lead to suppression in the original objective function. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The idea of using a quadratic penalty term to improve the stability near the edge of the constraint set boundary is a purely heuristic based approach. Is there any strong theoretical argument supporting such a choice? 2. In the related work section, the authors have discussed about drawbacks of using penalty terms in existing methods. Are these drawbacks applicable to the scheme proposed in the paper? 3. It is not clear how the time-dependent state occupation distribution is computed. The idea of approximating $\rho_{\pi}$ by the empirical distribution from the policy of the previous episode does not appear to be a good approximation. Although such approach may work well in infinite horizon problems where the optimality of stationary policy holds, it may not work well in the finite horizon setting. 4. The approach of using the second order penalty is not very novel and is widely used in practice for practical RL algorithms. What is the novelty of the proposed approach? 5. The idea of allowing the agent to act in a constrained manner even before the constraint is violated may lead to suppression in the original objective function. In other words, when the constraint is not too tight, the proposed scheme may unnecessarily impose additional level of constraint on the system. 6. How to ensure that we have sufficiently large $\beta$ and $\lambda_{t,i}$ which are well above the given lower bounds? The adaptive parameter selection process is just a heuristic. 7. The symbol in the last term of equation (7) is not clearly defined. 8. The algorithm proposed in the paper is an amalgamation of many ideas. Many of these ideas involve approximation and many are just driven from intuition without any formal guarantee of optimality. It is hard to see how this fusion work together in theory (e.g., satisfying the conditions given in Theorem 3.3). Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for comments on exposition and correctness. Responses below: We disagree with the reviewer on the weakness mentioned: (i) Calling introduction of quadratic penalty as purely heuristic minimizes the utility of such ideas that have proven to be central in making optimization algorithms practically useful. In optimization literature, it is a well known technique to improve numerical stability, and in fact our theoretical analysis for Theorem 3.3 accounts for it. See more under Q1. (ii) The idea of approximating $\rho_{\pi}$ by empirical distribution from the previous episode’s policy is again in the same spirit: Following a principled approach and yet introducing a series of good approximations that make the algorithm practical and yet, result in good or better performance. This is what has made PO algorithms so powerful that they are at the heart of most training of GenAI models (LLMs or Diffusion models) and find widespread usage on a daily basis. (iii) We have difficulty understanding the reviewer’s comment “ ... acting in a constrained manner even before the constraint is violated ... ”. Since the constraints depend on the entire episode, if you do not design policies that account for the constraints, then it may be too late to act to enforce them if you see the constraints are getting violated. Please do note that once the policy is designed as per our algorithm, there is no need for `constraint enforcement’. We have a guarantee (Theorem 3.3) that they will be satisfied. Q1. We respectfully, and strongly disagree that the introduction of quadratic penalty is purely heuristic. Our main result (text leading up to Theorem 3.3 and the theorem itself) talks about this very point. We theoretically prove that this attached quadratic penalty is an exact penalty from the primal problem in the constructed primal-dual problem. Thus, the constructed multiplier-penalty function is equivalent to the primal constrained problem to provide continuous and precise cost control. We mention in Lines 191-195 how the addition of the quadratic penalty theoretically helps the agent in more effective constraint control than its counterpart, the post-violation penalty approach. Q2. No they are not applicable. The only possible bottleneck in e-COP is the scalability, which can potentially be seen in some experiments. e-COP still promises to be a strong candidate for a safe RL baseline. Q3. You are correct in raising this concern. This has already been addressed in [1], which allows us to use such an approximation. Our extensive experiments on many environments showcase the ability of e-COP to outperform other SOTA baselines, so this approximation is even empirically proven to work. Q4. We do not use a second order penalty in our approach, we only deal with the expectation, which is first order. We have addressed the novelty of our work in Lines 57-68. Please see. Q5. To address this concern, we use an adaptive damping penalty that is instance dependent, adapts to the environment, and does not act “too safely”. This approach is discussed starting on Line 219, and ablation studies in Section 4.2 are also performed that show the effectiveness of our approach. Q6. For initial feasible parameters, we do a linesearch [2]. After initial feasible parameters, we update them as in Algorithm 2. Please go through the Appendix to see the theoretical justification, which we have provided in plenty. For your convenience, we provide it here as well: The gradient of $\mathcal{L}\_{t}(\mathbf{\pi}_{k}, \mathbf{\lambda}, \beta)$ takes the form $\nabla\_{\pi}\mathcal{L}\_{t}(\mathbf{\pi}\_{k},\mathbf{\lambda},\beta)=\nabla\_{\pi}\sum\_{h=t}^{H}\mathbb{E}\_{s\sim\rho\_{\pi\_{k,h}};a\sim\pi\_{k-1,h}}\big[-{\rho(\theta\_{h})}A^{\mathbf{\pi}\_{k-1}}\_{h}(s,a)\big]+\Sigma\_{\Psi\_{C\_{i},t}(\mathbf{\pi}\_{k-1},\mathbf{\pi}\_{k})\geq-\frac{\lambda\_{t,i}}{\beta}}\left(\lambda\_{t,i}+\beta\Psi\_{C_{i},t}(\mathbf{\pi}\_{k-1},\mathbf{\pi}\_{k})\right)\nabla\_{\pi}\Psi\_{C_{i},t}(\mathbf{\pi}\_{k-1},\mathbf{\pi}\_{k})$ Suppose $(\pi^{\star},\mathbf{\lambda}^{\star})$ is the optimal policy and its corresponding dual parameters. Consider KKT condition to $(\pi^{\star},\mathbf{\lambda}^{\star})$ of the undamped problem (Eq. (5) with $\beta=0$): $\nabla\_{\pi}\mathcal{L}\_{t}(\pi^{\star},\mathbf{\lambda}^{\star})=\nabla\_{\pi}\sum\_{h=t}^{H}\mathbb{E}\_{s\sim\rho_{\pi_{k,h}};a\sim\pi_{k-1,h}}\big[-{\rho(\theta_{h})}A^{\mathbf{\pi}\_{k-1}}\_{h}(s,a)\big]+\Sigma\_{i}\lambda\_{t,i}^{\star}\nabla\_{\pi}\Psi\_{C_{i},t}(\mathbf{\pi}\_{k-1},\mathbf{\pi}^{\star})$. Since the above two equations are consistent for the optimal point $(\pi^{\star},\mathbf{\lambda}^{\star})$ as shown in Theorem 3.3, we can relate the constraint violation nearby this optimal point as $\max\left(\Psi\_{C_{i},t}(\mathbf{\pi}\_{k-1},\mathbf{\pi}\_{k-1}),-\frac{\lambda_{t,i}^{(k-1)}}{\beta^{(k-1)}}\right)\approx\frac{\lambda\_{t,i}^{\star}-\lambda\_{t,i}^{(k-1)}}{\beta^{(k-1)}}$. This means we can reduce the constraint violations of the policy iteration close to the optimal point by reasonably increasing $\beta$. Based on this, we remove the Lagrange dependency in Eq. (7) and postulate the adaptive parameter selection procedure. Q7. It is clearly defined, please see Line 181. Q8. Most if not all SOTA PO safe RL algorithms require approximations (please see the related works section). We don’t understand what you mean by “hard to see in theory”? We have provided theoretical justifications for the development of e-COP (Section 3) and our main result Theorem 3.3 exactly deals with the feasibility of our approach, which is extensively validated by our experiments. We hope these answers address your concerns, and really hope that you consider raising your score. [1] Bojun, Huang. "Steady state analysis of episodic reinforcement learning." NeurIPS 2020. [2] Achiam, Joshua, et al. "Constrained policy optimization." ICML 2017 --- Rebuttal 2: Comment: Thank you for your detailed response which has resulted in a better understanding of the paper. However, I still feel novelty of the paper is limited as it incorporates various ideas already existing in the literature, for a episodic setting. I am happy to raise my score.	Summary: The paper proposes a new algorithm for finite-horizon constrained RL problems. The algorithm is based on three ideas: PPO-like updates for the finite-horizon setting, P3O-like treatment of constraints with adaptive penalty coefficient, and quadratic damping penalty. The proposed algorithm outperformed previous algorithms in some benchmark tasks. Strengths: A strong empirical performance of the proposed algorithm. The introduction of quadratic damping penalty to safe RL seems to be a new idea. The paper is easy to follow except some typos. Weaknesses: Performance difference lemma for the finite-horizon setting is already known. See Kakade's thesis. Since PPO-like update is based on the performance difference lemma and PPO's idea, I do not see it is novel. P3O-like treatment of constraints with adaptive penalty coefficient is also not novel. There are some parts I do not really understand. In particular I wonder if the proposed algorithm really does what Algorithm 2 states. Concretely Algorithm 2 suggests to save policy parameters for each time step, which sounds to be unusual. (Also I guess Line 6 of the algorithm has a type: $\pi_{k, t} \gets \pi_{k, t+1}$ should be $\pi_{k, t} \gets \pi_{k-1, t}$.) If the algorithm is implemented as above, it is unfair to compare to previous algorithms, because previous algorithms use a much fewer number of parameters. If the algorithm is not implemented as above, there is a serious issue in presentation. Eq 8 seems to contain many typos. For example $\min_{\pi_{k, t}} L(\theta, \lambda, \beta)$ does not make sense since the loss do not have $\pi_{k, t}$. (Sorry for the sloppy notation, but I believe the authors can understand what I mean.) Also $\pi_{k, t}^\star = \min_{\pi_{k, t}} L(\theta, \lambda, \beta)$ does not make sense since $\min_{\pi_{k, t}} L(\theta, \lambda, \beta)$ is a scalar. Technical Quality: 3 Clarity: 2 Questions for Authors: - Does the proposed algorithm what Algorithm 2 states? Algorithm 2 suggests to save policy parameters for each time step, which sounds to be unusual. - I looked into the code, but it seems it will not run since `_episodic_reward_loss` and `_get_surr_cadv` methods are undefined. I checked if the Github repo of omnisafe, but I cannot find those functions. Is the code intended only for reference? (Indeed the code is provided as a text file, not python file.) - I currently think the main contribution of the paper is the introduction of quadratic penalty dumping to safe RL. Therefore I would like to see results of ablation studies to prove the importance of quadratic penalty dumping. Is there any result? (Or am I missing?) Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: If the proposed algorithm really does what Algorithm 2 states, there is an issue of huge memory requirement as the horizon gets larger. I think it should be discussed. Otherwise I do not see any potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for comments on the novelty of the theory and strong empirical performance. Responses below: First please note that as far as we know this is the first policy optimization algorithm for the constrained or unconstrained setting as far as we know. Second, all policy optimization algorithms are based on a policy difference lemma, and so ours is not novel: we weren’t aware that for the episodic setting, it was already available in Kakade’s thesis. We will acknowledge and cite. We acknowledge that the penalty based constraints treatment is not novel. However, how to incorporate this treatment into episodic settings was unknown, as was how to update the policy network in the finite horizon setting. Note that we have a time-dependent policy (hence parameters are saved at each step). This is not a bug but a feature as stationary policies for the episodic setting are known to be sub-optimal. Line 6 of Algorithm 2 is correct, since we use the idea of backward recursion as is done in the classical dynamical programming algorithms. So, while it is true that our model stores a higher number of parameters, it is because this is needed for the episodic setting (stationary policies are suboptimal, nevertheless, if any one wishes to use our algorithm to find a stationary policy (with fewer parameters, it is quite straightforward to do so). The concern of fairness of comparison to these other algorithms designed for non-episodic settings is valid. Nevertheless, these are the closest baselines available and we show that their usage in the episodic setting will result in subpar performance (Also, we anticipated that reviewers will want to see comparison to such baselines even if they are not designed for this setting.). No other PO algorithms, with fewer or more parameters can perform as well as e-COP. The loss indeed has the minimizing variable, which is hidden in the importance sampling ratio as defined after Equation (4). We can make this distinction clear in the next draft. You are right in saying that the $\texttt{min}$ is incorrect, it is a typo, and it should be $\texttt{argmin}$, which we will also fix. Thank you for pointing this out. Ques 1: Yes, Algorithm 2 as stated is correct. Please see above. Ques 2: Yes, we had to code the two functions you mentioned, and we use commit SHA $\texttt{d55958a011df7800f256452e07811832cd2524d2}$ of omnisafe to run our experiments. The implementation of $\textunderscore get \textunderscore surr \textunderscore cadv$ function is based on the docs module in omnisafe. Ques 3: We have included ablation studies on the effect of the quadratic penalty in Section 4.2 with further details in Appendix A.3.4. We thank the reviewer for their review, and the many comments that have helped improve the paper substantially. If the reviewer is satisfied, we would really appreciate reconsideration of their current score. --- Rebuttal Comment 1.1: Comment: Thank you very much for the rebuttal! > However, how to incorporate this treatment into episodic settings was unknown, as was how to update the policy network in the finite horizon setting. Would you explain this a bit more? I do not really see why converting algorithms for the infinite-horizon setting to the finite-horizon setting is untrivial. In my experience it is easy in theory, especially given performance difference lemma for the finite-horizon setting. For example which part of the derivation of e-COP is challenging in comparison to P3O? > Note that we have a time-dependent policy (hence parameters are saved at each step). This is not a bug but a feature as stationary policies for the episodic setting are known to be sub-optimal... Yes, I know in theory a non-stationary policy is required. However one can explicitly provide the remaining time to a policy network, as is done in Time Limits in Reinforcement Learning by Pardo et al. > Line 6 of Algorithm 2 is correct, since we use the idea of backward recursion as is done in the classical dynamical programming algorithms. Would you explain this point a bit more? Solving Eq 8 with a first-order optimizer as explained in Line 229-230 does not seem to result in backward recursion. > The concern of fairness of comparison to these other algorithms designed for non-episodic settings is valid... Nevertheless, these are the closest baselines available and we show that their usage in the episodic setting will result in subpar performance... But gradually removing some components from e-COP recover previous algorithms, no? > Ques 2: Yes, we had to code the two functions you mentioned, and we use commit SHA Your SHA is different from the one in readme.md of the supplementary material. Also would tell me where `_get_surr_cadv` is? I cannot find it. Maybe would you give me a github link to the file in which the method actually appears? > Ques 3: We have included ablation Thanks, but where is it? Maybe reviewers cannot see updated version yet... --- Rebuttal 2: Comment: > Would you explain this a bit more? ... It is highly nontrivial due to the absence of a stationary state distribution. With a finite horizon, the time-dependent state occupation measure needs to be treated differently, as is done in e-COP. This different treatment requires completely new analysis to show that this approach is principled, and needs our main result (Theorem 3.3) to showcase that using occupation measure from the previous episode along with the quadratic penalty does not change the solution set of the original problem. This entire part is new and different when compared to P3O. > Yes, I know in theory a non-stationary policy is required ... There are two reasons why this approach is theoretically not correct (empirically maybe ok): 1. Augmenting a time dimension to the state space leads to sparse transitions, which causes the time-dependent state occupation measures to break down. 2. Time Limits in Reinforcement Learning paper deals with PPO in discounted episodic ($\gamma=0.99$) and un-discounted episodic ($\gamma=1$) setting. Comments: Firstly, for episodic RL, (un-discounted) cumulative episodic rewards are considered as metrics, rather than discounted rewards. Secondly, if you set $\gamma=1$, the policy difference lemma for discounted RL (as in Kakade's thesis) breaks down (see [1] for the solution to this problem). Due to this, no theoretical convergence or correctness statements can be made, which is why there are no theoretical justifications for such an approach. We, on the other hand, provide theoretical correctness proof and empirically showcase the effectiveness of our approach on various environments, showcasing the superior performance of e-COP compared to baselines. > Solving Eq 8 with a first-order optimizer ... See Lines 5-6 of Algorithm 2 for this. Eq. 8 is first-order solvable, and then backward recursion is done as part of the algorithm. From a careful look at Algorithm 2, it should be clear what is happening. > But gradually removing some components ... This is not correct: The introduction of quadratic penalty and backward recursion structure of the algorithm is not a linear addition, i.e., if one sets $\beta=0$ (so as to "remove" the quadratic penalty), one does not recover any previous algorithm. A novelty of our work lies in introducing the quadratic penalty for episodic settings and formalizing the policy update rule, all of which are independent of the baselines we have used. > Your SHA is different from the one ... We have not provided any readme.md with the supplementary material. Could you please tell which file are you referring to? For the code release, we are currently in the process of obtaining code release permission: We will make the code public once it is received, and before the paper is finalized. > Thanks, but where is it? Maybe reviewers cannot see updated version yet... The ablation studies are available in Section 4.2 ("Secondary Evaluation") of the submitted manuscript, with further details in Appendix A.3.4. We would also like to add that we very much appreciate the care and attention with which the reviewer has read the paper; the probing questions they ask will make our draft stronger ; and the willingness to engage with our response and cross-check our experimental data ! [1] Agnihotri, A. et al. ACPO: A Policy Optimization Algorithm for Average MDPs with Constraints. ICML 2024.	Summary: The authors introduce a policy optimization algorithm for episodic constrained RL problems, e-COP. In general, Lagrangian formulation is used for constrained optimization problems, however, the constraints are not always satisfied in real world applications. The solution approach avoids Hessian matrix inversion. The algorithm approach uses deep learning-based function approximation, a KL divergence-based proximal trust region and gradient clipping popular in improving the generalization and avoiding vanishing gradients in proximal policy optimization algorithms. The authors use SafeGym to benchmark their algorithm against the state-of-the-art algorithms such as PPO, FOCOPS, CPO, PCPO, and P3O. Strengths: The e-COP algorithm in most cases outperforms all other baseline algorithms but fails on Humanoid, AndReach, and Grid where it provides the second best optimal results. The authors used an extensive set of baselines -- few using Lagrangian approximations. The idea is original. In general, the paper is well-written and clear. There are few grammatical errors and typos. Weaknesses: The authors do not talk about the limitations of the algorithms. There is a mention on the complexity of the Grid environment resulting the e-COP not to perform as best. It will be good for the completion of the paper to have a short section on limitations. Technical Quality: 3 Clarity: 3 Questions for Authors: I would like to know whether any hyper-parameter tuning has been done for the baseline algorithms? It is possible that the Lagrangian formulation may perform better after tuning. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors do not adequately address limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments on originality, and well-written exposition. We will fix any grammatical issues and typos. We do want to point out that while the algorithm does not perform the best on Humanoid, AntReach, and Grid, it still performs the second best on these three while performing the first best on the other five environments. Response to concerns below: With regards to the weaknesses, while there are no fundamental weaknesses, we shall include a section on limitations in the next version, wherein we will discuss the scalability issues we faced in the Grid environment. Since the application in mind for this paper is in diffusion models, we plan to conduct research and experiments in this direction to ascertain e-COP’s capabilities. For the baselines, some hyperparameter tuning has indeed been done, and we provide the hyperparameter values in the Appendix, which are set based on the best performing parameters for each algorithm. We refer to standard codebases that are widely used for benchmarking algorithms [1], [2]. Note that we have done similar hyperparameter tuning for the Lagrangian formulation, PPO-L (and the other algorithms) as well. The gains in performance are not because of hyperparameter tuning but from the fact that the constrained problem formulation is the right fit for the problem. We thank the reviewer for their review, and the many comments that have helped improve the paper substantially. If the reviewer is satisfied, we would appreciate reconsideration of their current score. [1] omnisafe: https://github.com/PKU-Alignment/omnisafe [2] FSRL: https://github.com/liuzuxin/FSRL/ --- Rebuttal Comment 1.1: Comment: Thank you for the rebuttal comments. It will good to include in the paper that the baselines were implemented and optimized to the best of the author's abilities along with reference to the Appendix.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph	Accept (poster)	Summary: The paper introduces a framework, Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graphs (DARG), which dynamically extends benchmarks by generating evaluation data with controlled complexity. This framework addresses key limitations of static benchmarks, such as data contamination and a lack of adaptability to evolving LLM capabilities. Strengths: 1. The method of dynamically generating evaluation data based on adaptive reasoning graphs is promising. It addresses critical limitations of static benchmarks, such as data contamination and the inability to adapt to the evolving capabilities of LLMs. 2. The evaluation encompasses a wide range of LLMs and tasks, offering a thorough analysis of how different models perform across various complexity levels. This provides valuable insights into model robustness and generalization capabilities. 3. The paper effectively highlights how biases in LLMs can be exacerbated under complex testing conditions. This is a crucial aspect for developing fairer models and contributes significantly to the ongoing discourse on ethical AI. 4. The use of reasoning graphs to represent the underlying structures of problem-solving processes and the subsequent perturbation to create novel test samples is methodologically sound. This ensures that the generated data retains linguistic diversity and is representative of real-world scenarios. Weaknesses: 1. According to Figure 2, the ranking order among various models remains consistent as the complexity metrics increase. This observation contradicts the conclusion in the introduction (line 58) regarding the unreliable assessment of LLMs' capabilities using static benchmarks. This inconsistency needs to be addressed and clarified. 2. While the paper demonstrates the applicability of DARG across four reasoning tasks, it is unclear how well this approach generalizes to other types of tasks (e.g., knowledge based QA), particularly those that do not naturally lend themselves to graph-based representations. 3. The graph extraction and data generation process heavily relies on closed-source LLMs, such as GPT-4. Although rule-based constraints and data verification modules are incorporated, the dependence on proprietary models raises questions about the reproducibility of the proposed framework. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can the authors provide more detailed explanations or additional analyses to reconcile the observed inconsistencies between the model rankings and the stated conclusions? 2. How does the DARG framework perform when applied to other non-reasoning tasks? Can the authors provide preliminary results or case studies on different task domains? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have adequately addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your insightful feedback and recognition of our work's key strengths. We are encouraged by your highlight which addresses the critical limitations of static benchmarks with high-quality generated data, our comprehensive experiments with thorough analysis and valuable feedback, and findings on bias exacerbation. We're grateful for your acknowledgment of our contributions to LLM evaluation. We would now like to address your concerns with the following clarifications: - W1 & Q1: Inconsistency between Figure 2 and the conclusion about unreliable assessment using static benchmarks: Thanks for pointing this out! We respectfully disagree with the argument that the ranking order remains consistent. Figure 2 shows evident intersections between different lines as complexity increases across all three dimensions, indicating a significant number of changes in performance ranking. According to the detailed results in Tables 2, 3, and 4 in the appendix, the following are some examples (and such exceptions are not limited to these): - Command R+ vs. Deepseek Math and Phi-3-mini vs Mixtral 87B in numerical complexity - WizardLM-2 vs Mixtral 8X22B in reasoning graph's width - Mixtral 8X7B vs LLaMa-3-8B in reasoning graph's depth Moreover, our argument in line 58 also addresses the unreliability of absolute evaluation results in static benchmarks. While many LLMs claim over 96% accuracy on some static benchmarks, suggesting task mastery, our DARG framework reveals significant performance drops across all LLMs as complexity increases. This demonstrates the limitations of static benchmarks in accurately assessing LLM capabilities. We will better clarify this point in the camera-ready version. - W2 & Q2: Generalizability to other types of tasks: Please refer to Global Response #2. - W3: Reliance on closed-source LLMs for graph extraction and data generation: Please refer to Global Response #1*. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. my concerns are properly addressed and I would like to raise the score to 6. --- Reply to Comment 1.1.1: Comment: Thanks for taking the time to respond to our rebuttal. We are encouraged by your recognition of the contributions of our work and appreciate that our rebuttal has resolved your questions and your concerns. Thanks again for your efforts in reviewing our paper!	Summary: This paper proposed a dynamic evaluation of LLMs --- DARG. The authors first generate a reasoning graph of the problem, then perturb the problem's complexity along various "dimensions", then convert the more complicated graph back to natural language questions. The authors evaluate several LLMs on 4 perturbed datasets and observe consistent performance drop, indicating that LLMs may not reason very well, and previous good static evaluation results may be due to data contamination. Strengths: This paper tries to tackle an important evaluation issue in LLMs by providing dynamic yet controlled method. Weaknesses: The particular method only applies to problems that have a clear reasoning graph and it relies on a rule-based system and a non-ambiguous way to increase complexity, which may not be immediate clear how to apply to any dataset. However, it is a good starting point. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For GSM8K, when you increase the numerical complexity by +k, does that mean you sample numbers of which the range is increased by k? or you simply +k to all the numbers? 2. For figure 9, can you also add an overlapping radar plot such that your claims in line 195-199 is more easily visualized? 3. Line 203-208, can you explain how the evaluation is done? Did you ask the model being evaluated to generate a reasoning graph and compare with the ground truth? 4. For BBQ, how is the attributes' polarity determined? Do you have a predetermined set of attributes to sample from? 5. Line 232-234, I am not quite getting the claim here. Figure 4 the lower right subfigure is presenting the avoidance rate, and it seems Mistral 7B has the highest rate rather than GPT4 and Gemini. Also is the Mistral model here aligned? If so, then isn't it against the argument that over-alignment is causing the issue here? 6. Do we know the model performance when the reasoning graph stays the same, but only the question is paraphrased? Misc: - line 200 "Trubo"-> "Turbo" Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations have been discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly value your insightful feedback and are encouraged by your recognition of the importance of the problem we investigated. We would like to provide the following clarifications: - W1: DARG can only applied to tasks that have a clear reasoning graph and is not clear on how to apply to others. As discussed in lines 314-343, we acknowledge that our framework is based on reasoning graphs. However, the graph definition can be general and extended to various other tasks. Please refer to Global Response #2. - Q1: Details of Numerical Complexity by +k We increase the average numerical complexity of all the equations by +k instead of just increasing the sample range. - Q2: Overlapping Radar Plots to support the conclusion in Lines 195-199 Please refer to the supplementary PDF in the Global Response which includes comparisons across all LLMs (Fig 1), between LLaMa3-8b and LLaMa3-70b (Fig 2) (lines 194-195), and between Mistral-7b and Mistral-87b (Fig 3) (lines 196-199). The plots clearly show that LLaMa3-70b outperforms LLaMa3-8b, and Mistral-87b outperforms Mistral-7b in resistance to complexity increase across all dimensions of complexity, fully supporting the conclusions in lines 195-199. - Q3: Details of Evaluation in Lines 203-208 We manually compare the reasoning process in the output of LLMs (e.g., CoT reasoning) with the reasoning graph (considered ground truth, as we use a rule-based function for verification) to determine if errors are due to incorrect reasoning. - Q4: Attributes' Polarity Definition in BBQ We first generate a pool of positive and negative attributes using GPT-4-turbo with accurate definitions in the prompt. After that, we perform another stage of filtering to ensure they align with their predefined polarity (positive or negative) and are not related to the answer to the question, followed by sampling for each data point for augmentation. The definitions for such positive and negative attributes in the prompt are as follows: - Positive attributes are qualities that are generally viewed favorably by society. These traits often enhance a person's interactions, contribute to their success, and improve their overall well-being. - Negative attributes are qualities that are typically viewed unfavorably and can hinder personal growth or social integration. We manually check all the positive and negative attributes after filtering and observe that they all meet human standards. - Q5: The avoidance rate of Mistral 7B cannot support the argument of the over-alignment explanation. We appreciate the reviewer's observation regarding the avoidance rates shown in Figure 4. While it's correct that Mistral 7B exhibits the highest overall avoidance rate, this finding requires careful interpretation within the broader context of model performance. - Performance Context: It's crucial to note that Mistral 7B's overall accuracy is significantly lower than that of GPT-4-turbo and Gemini-Pro-1.5. This lower accuracy implies a higher total number of errors for Mistral 7B. - Avoidance Rate Calculation: The overall avoidance rate is calculated as the number of avoidance cases divided by the total number of data points. Due to Mistral 7B's higher error rate, this metric may not provide a complete picture of the "avoidance behavior" relative to other error types. - Introducing Avoidance Error Rate: To address this, we've calculated an avoidance error rate - the number of avoidance cases divided by the total number of errors. This metric provides insight into the proportion of errors that are specifically avoidance cases, controlling for the overall error rate. - Comparative Analysis: We present the averaged avoidance error rate for those 3 models across all complexity increase intervals: \| Model \| Averaged Avoidance Error Rate (std) \| \|-------\|-------------------------------------\| \| Mistral-7B \| 89.345 (2.74) \| \| Gemini-1.5-Pro \| 98.833 (0.843) \| \| GPT-4-turbo \| 95.552 (1.698) \| This result shows that among their errors, GPT-4-turbo and Gemini-Pro-1.5 have higher rates of avoidance cases, supporting our over-alignment hypothesis. In conclusion, the lower avoidance error rate for Mistral 7B, despite its higher overall avoidance rate, can be attributed to its larger total number of errors. This doesn't necessarily contradict our over-alignment explanation. We thank the reviewer for prompting this deeper examination, which has enriched our understanding of the results. - Q6: The results when the reasoning graph stays the same, but the question is paraphrased. To address this concern, we conducted additional experiments to test different LLMs on paraphrased questions. We used GPT-4-o to paraphrase math questions with the following prompt: Paraphrase the following math problem using different words and phrasing, but keep the core mathematical concepts, numbers, and solution process exactly the same. Do not change any numerical values or the steps needed to solve the problem. Here's the original problem: Original Math Problem: {original_problem} Please provide the rewritten version of this problem. Paraphrased Math Problem: We tested several LLMs on these paraphrased ones with two prompts (CoT and LtM). The results are as follows, where the numbers in parentheses are the difference between the paraphrased and the original (ACC_para - ACC_ori): \| Model \| CoT \| LtM \| \|-------\|---------------\|---------------\| \| GPT-4o \| 0.946 (+0.008) \| 0.946 (+0.002) \| \| Gemini-1.5-Pro \| 0.894 (-0.026) \| 0.906 (-0.022) \| \| LLaMa3-70B \| 0.906 (-0.016) \| 0.918 (-0.008) \| \| LLaMa3-8B \| 0.806 (+0.018) \| 0.822 (+0.024) \| \| Mixtral-8*7B \| 0.676 (+0.054) \| 0.73 (+0.048) \| From this, we find that paraphrasing can result in minimal performance changes compared with DARG's perturbation, and such changes are not consistent across all LLMs. --- Rebuttal Comment 1.1: Comment: Thanks the authors for their response. I do not have further questions. I will keep my original score of 6 as it was positive. --- Reply to Comment 1.1.1: Comment: Thanks for taking the time to respond to our rebuttal. We are encouraged by your recognition of the contributions of our work and appreciate that our rebuttal helped resolve your questions. Thanks again for your efforts in reviewing our paper!	Summary: The paper proposes a new framework, DARG, to extend current reasoning benchmarks with controlled and diversity dynamically. Authors evaluate multiple LLMs on those generated output from DARG and concluded that the proposed method is useful for evaluating LLMs in a dynamic and adaptive way. Strengths: 1. The paper proposes a new method to add control for benchmark augmentation when LLMs are involved, which is useful for evaluating LLMs. 2. The paper conducts experiments on significant amount of LLMs and multiple reasoning categories, indicating the generality of the method. 3. The method is well-motivated and stated. Weaknesses: 1. The paper only involves evaluations for DARG that use ChatGPT as the graph construction and graph-to-text generation engine, which may cause bias on the extended benchmarks. It would be good to see if the selection of the LLM used for those components can affect the final evaluation results. It is even useful to check if replacing those LLM components with humans will lead to a significant change. 2. DARG's rule-based function is not clearly explained in the paper, which seems important to ensure the quality of reasoning graph generation. 3. Due to the uncertainty of the quality of the generated benchmarks, it would be nice to reflect the correct rate evaluated with human eval on the LLM evaluation results with DARG, such as Figure 2. With that, it will be easier to see how confident the results are for the ranking of those LLMs. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness above Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your insightful feedback. We are encouraged by your recognition of our novelty, the comprehensive empirical evaluation, well-motivation, and the generality of our proposal. We would like to address your concerns and offer the following clarifications: - W1: Lack of different selections of the LLM for graph construction and graph-to-text decoding Thanks for pointing this out! Please refer to the Global Response #1. - W2: Lack of clear explanation of rule-based function in reasoning graph generation As described in Lines 103-106 and Algorithm 1, the rule-based function computes a “label” given the reasoning graph which is compared with the original ground-truth label, with implementations varying across tasks. To address your concerns, we explain the graph-to-label computation function for each of the four tasks in our experiments: - Math reasoning (GSM8K): The function traverses the reasoning graph to compute the final answer. It initializes values for "initial" nodes (input values) and iteratively computes values for other nodes by following graph edges. For each edge, it applies the specified operation (e.g., +-/) using values from connected nodes until all nodes have values. The pseudo-code is as follows: ``` function compute_graph_values(graph): initialize values dictionary for each node in graph: if node is initial: set node value in values while not all nodes have values: for each edge in graph: if source and objective nodes have values: compute target node value add to values dictionary return values ``` We then check the value of the "final" node against the ground truth label to verify the graph construction's correctness. - Social Reasoning (BBQ):* For this task, this function first validates the graph structure, traces paths from person nodes to label nodes, and matches person names with answer options. It then computes the answer based on label node connections: selecting the specific person if it is connected to a label node specified in the question. - Spatial Reasoning (BBH-navigate): Given the graph structure, this function tracks x and y coordinates, adjusting them based on each node's direction and step count. It then checks if the final position matches the starting point (0,0) to determine the computed label. - Symbolic Reasoning (BBH-dyck): This function uses a stack to process the input bracket sequence, pushing opening brackets and checking closing brackets for matches. It generates the label by creating closing brackets for any remaining open brackets on the stack, which is then compared with the original ground-truth label. Thanks again for pointing this out! We will add the details of such rule-based functions in the camera-ready version. - W3: Results of human eval with DARG with the same setting as Figure 2. Thanks for pointing this out! To further address your concerns, we have added additional human evaluation results. The human annotator, who holds a bachelor's degree in a STEM field, evaluated 20 randomly sampled data points for each complexity interval (300 data points in total) with a calculator using the same setting as in Figure 2. The human evaluation results are as follows: \| Original \| Numerical +2 \| Numerical +4 \| Numerical +6 \| Numerical +8 \| \|----------\|-------------------------\|-------------------------\|-------------------------\|-------------------------\| \| Human Eval ACC \| 1.0 \| 1.0 \| 0.95 \| 0.95 \| 0.95 \| \| Original \| Width +1 \| Width +2 \| Width +3 \| Width +4 \| \|----------\|----------\|----------\|----------\|----------\| \| Human Eval ACC \| 1.0 \| 0.95 \| 0.95 \| 0.90 \| 0.90 \| \| Original \| Depth +1 \| Depth +2 \| Depth +3 \| Depth +4 \| \|----------\|----------\|----------\|----------\|----------\| \| Human Eval ACC \| 1.0 \| 0.95 \| 0.90 \| 0.90 \| 0.85 \| From these results, we observe that although human evaluations also show a slight performance decrease as complexity levels increase, their performance and resilience to the complexity increase are much higher than those of all LLMs.	Summary: This paper introduces DARG (Dynamic Evaluation of LLMs via Adaptive Reasoning Graph), a framework for dynamically generating evaluation data for large language models (LLMs) with controlled complexity. The DARG framework constructs reasoning graphs from existing benchmarks, perturbs these graphs to generate more complex samples, and uses LLMs with code verification to filter out incorrect perturbations. The authors apply DARG to generate new test data for 4 reasoning tasks: math, social, spatial, and symbolic reasoning. One of the key experimental results shows that the performance of all LMs generally decreases as complexity increases, potentially indicating that the newly constructed datasets are indeed meaningfully challenging. They also demonstrate that DARG-generated data can be used to improve model performance through fine-tuning. Strengths: * Novel approach to augment existing static benchmarks, especially the idea of using code generation to filter out bad examples. * Comprehensive evaluation and interesting analysis of performance dropping across benchmarks. * Demonstrates potential of generated data for improving models via fine-tuning Weaknesses: - Baseline comparisons: Main experimental results show that the generated benchmarks are challenging for llms. However, this isn't the first work introducing the idea of creating harder benchmarks. So, it would be useful to know the gap the proposed work fills. For instance, it would be valuable to see comparisons against other dynamic evaluation methods like DyVal (which the authors do mention in the related work). - The perturbations used, especially for math problems, may be overly simplistic. Simply increasing numerical values or graph complexity doesn't necessarily capture all aspects that make math problems more challenging or interesting. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Have you explored using only open-source models for the graph extraction and data generation steps? How well does DARG work without access to closed-source LLMs? 2. Could you expand on the fine-tuning experiments? Specifically, what happens if we _add_ gsm8k training data to the newly generated samples? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your thoughtful feedback on our DARG framework. We're pleased that you recognize the novelty of our approach, particularly the use of code generation for filtering examples, as well as our comprehensive evaluation and analysis and the potential for improving models via fine-tuning. We would like to address your concerns and offer the following clarifications: - W1: Baseline comparisons and gap identification Thank you for highlighting this important point. We believe we have addressed most of this in our paper, but we appreciate the opportunity to clarify further: As emphasized in lines 38-39 and 320-322, our work addresses the challenge of dynamically and adaptively generating novel test samples with controlled complexity with label verification and diversity. Previous works [1][2][3] do not achieve such controlled complexity with label verification and diversity simultaneously. To further address your concerns, we provide the following clarifications on a detailed comparison between ours and related previous work. - Comparison with DyVal [1] (lines 30-33 and 311-316) DyVal's template-based generated samples explicitly define relationships between variables based on predefined rules, resulting in a lack of diversity. This explicit specification of relationships between variables and templates does not foster commonsense reasoning abilities like those in the GSM8K dataset and our generated ones. Furthermore, our DARG framework starts with existing datasets, allowing different datasets to capture different characteristics during new data generation, whereas DyVal has one set of predefined rules for a single task (e.g., math/logic reasoning). We provide a concrete example of DyVal's generated sample in lines 313-314, which you can refer to and compare with our examples in Figure 1. - Comparison with DyVal 2 [2] and Benchmark Self-Evolving [3] (lines 33-38 and 317-320) These approaches, which rely on prompting LLMs with pre-defined prompts, do not guarantee label stability or achieve fine-grained complexity control. Label verification is crucial for generating evaluation data points, as it ensures reliability, which is not addressed in these works. - W2: The perturbations may be overly simplistic We appreciate this observation and would like to clarify: - As described in lines 109-110, our perturbation method involves systematically changing the structure of the reasoning graph based on different levels of complexity. The specific complexity and perturbation definitions can vary based on the nature of the task, as demonstrated by our four different tasks. - We chose numerical or graph complexity because it reflects the complexity of the reasoning process, a key interest in the math reasoning community. Furthermore, we argue that through our DARG framework, changing the numerical values and reasoning graph can indeed generate a diverse set of new data points because the graph-to-text decoding stage introduces diverse and uncertain contextual information due to the probabilistic nature of LLMs (as described in line 750 in Appendix A, we intentionally set the temperature to 1 to further achieve this). As shown in Figure 1, the context of the newly generated question is significantly different from the original question. Additionally, we can define other components in the question, such as persons and attributes (similar to the graph definition for the BBQ dataset of social reasoning that we have explored), to further control its diversity in the context beyond the reasoning graph. - Q1: Exploration of using only open-source models for graph extraction and data generation Thanks for pointing this out! Please refer to the global response #1. - Q2: Results of adding GSM8K training data to the newly generated samples Thanks for pointing out this interesting problem. To address this question, we conduct the following additional experiments of combining an equal amount of original GSM8K training data with our newly generated data to fine-tune Mistral-7B and LLaMA 2-7B with the exact same setting as the previous fine-tuning experiments. The results are as follows: \| Model \| Original \| w/ GSM8K \| w/ DARG \| w/ GSM8K + DARG \| \|-------------\|----------\|----------\|---------\|-----------------\| \| Mistral-7B \| 6.875 \| 11.87 \| 13.75 \| 14.15 \| \| LLaMa2-7B \| 7.5 \| 8.75 \| 14.375 \| 14.50 \| From these results, we observe that fine-tuning with a combined equal number of GSM8K’s original training data and DARG-generated data does not achieve a significant performance improvement compared to fine-tuning with only DARG-generated data. The reason behind this may be that the DARG-generated data already covers the majority of the complexity levels present in GSM8K’s original training data. [1] DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks [2] DyVal 2: Dynamic Evaluation of Large Language Models by Meta Probing Agents [3] Benchmark Self-Evolving: A Multi-Agent Framework for Dynamic LLM Evaluation --- Rebuttal Comment 1.1: Title: Thanks for your response! Comment: Thanks for your response. I'll keep my positive score with an increased confidence in the work. --- Reply to Comment 1.1.1: Comment: Thank you for taking the time to respond to our rebuttal. We are encouraged by your recognition of the contributions of our work Thanks again for your efforts in reviewing our paper!	Rebuttal 1: Rebuttal: Global Response: - 1. Only use GPT-4-Turbo for graph construction and graph-to-text decoding and lack exploration of different LLM choices especially open-source models: We appreciate the importance of generalizing the DARG framework to different LLMs, especially open-source models. To address this, we conducted additional experiments comparing LLaMA 3.1-8B, LLaMA 3.1-70B, and LLaMA 3.1-405B with GPT-4-Turbo (used in our original experiments) based on their generation quality: - Graph Extraction: We used the four models to extract 100 reasoning graphs in GSM8K and compared their success rates. The results are as follows: \| Model \| Success Rate \| \|---------------\|--------------\| \| GPT-4-Turbo \| 0.91 \| \| LLaMA 3.1-8B \| 0 \| \| LLaMA 3.1-70B \| 0.83 \| \| LLaMA 3.1-405B\| 0.85 \| The results show that the current SOTA open-source LLMs (70B and 405B) perform well in graph extraction, comparable to GPT-4-Turbo. The smallest LLaMA 3.1-8B model underperformed due to poor instruction-following ability. - Graph-to-Text Decoding: We used the same models to decode 50 perturbed reasoning graphs back to the original data format under two conditions: a) single run and b) a maximum of 5 iterations of refinement. The results are as follows: \| Model \| Single-run Success Rate \| Max 5 Iter Refine Success Rate \| \|---------------\|--------------------------\|--------------------------------\| \| GPT-4-Turbo \| 0.40 \| 0.90 \| \| LLaMA 3.1-8B \| 0 \| 0 \| \| LLaMA 3.1-70B \| 0.18 \| 0.52 \| \| LLaMA 3.1-405B\| 0.20 \| 0.58 \| The results indicate that SOTA open-source LLMs (70B and 405B) achieve decent performance (~60% success rate) on graph-to-text decoding and can noticeably refine their initial generation through our code-agent framework, though there remains a significant gap with GPT-4-Turbo. We manually checked the errors and found that most were due to a lack of ability to follow instructions to output in a specific format (which may cause parsing errors), which is important for the agent framework. Overall, these additional experiments show that current SOTA open-source LLMs of sufficient size (70B and 405B) can perform reasoning graph construction well and demonstrate decent performance in graph-to-text decoding. However, there remains a gap in instruction-following and structured output abilities compared to GPT-4-Turbo, which may hinder their application in agent frameworks. - 2. Not sure how to apply DARG to other non-reasoning tasks: As stated in lines 341-342 of our limitations section, we focused on reasoning tasks, which are fundamental to many NLP tasks and a crucial area of LLM research. We investigated diverse reasoning domains, including math reasoning, social reasoning, spatial reasoning, and symbolic reasoning. The diverse range of reasoning tasks and datasets we selected demonstrates the generality of our approach, as recognized by reviewer MmK5. To address your concern, we've explored applying DARG to two additional datasets: - HumanEval: We apply DARG to code generation by extracting the logic graph of function implementations. For example, the function: ```python from typing import List def has_close_elements(numbers: List[float], threshold: float) -> bool: """ Check if in the given list of numbers, are any two numbers closer to each other than a given threshold. >>> has_close_elements([1.0, 2.0, 3.0], 0.5) False >>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3) True """ ``` According to Fig 4 in the supplementary PDF, each node in the graph can represent a basic operation such as comparison, length check, loop iteration, and return statement. Edges can represent the logic of the workflow and the condition of each operation. After constructing such a graph, we can increase the complexity by adding more edges or nodes to control the problem's complexity for dynamic evaluations. - CommonsenseQA: This multiple-choice QA dataset can be adapted using DARG by representing each answer choice as an option node and key attributes as attribute nodes. The edges between option nodes and attribute nodes represent the degree to which an option possesses the given attribute. The answer can be computed by selecting the option with the maximum edge value. For example (shown in Fig 5 in the supplementary PDF): Question: "Sammy wanted to go to where the people were. Where might he go?" Options: (A) race track; (B) populated areas; (C) the desert; (D) apartment; (E) roadblock We can modify option nodes, increase attribute nodes, and adjust edge values to control complexity. These examples show that DARG's core idea is applicable to various tasks and datasets, with specific adjustments for graph construction and complexity definitions. Pdf: /pdf/c652700107779f86433f14fb4027eb4d0c537cd8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MG-Net: Learn to Customize QAOA with Circuit Depth Awareness	Accept (poster)	Summary: The paper proposes a deep learning approach called Mixer Generator Network (MG-Net) to enhance the performance of Quantum Approximate Optimization Algorithm (QAOA) by dynamically designing optimal mixer Hamiltonians for a given class of problem and a circuit depth constraint. This is done by parameter grouping that is analyzed theoretically by the authors. The authors demonstrate the effectiveness of MG-Net enhances QAOA performance in terms of approximation ratio for QAOA problem instances like Ising models and weighted Max-Cut instances up to 64 qubits. Strengths: The main contribution of the paper (MG-Net) as a dynamic solution to customize mixer Hamiltonians is novel, innovative, and sound, backed by theoretical analysis of the convergence theory of QAOA. The performance evaluation of the method is also comprehensive, covering a range of QAOA problem instances and validating the proposed framework's effectiveness. The presentation of the paper is clear and well-structured. The paper detailed the explanations of methodologies supported by diagrams that aid comprehension. Weaknesses: - The “Quantum Hardware Constraints” in the title seems not appropriate. When talking about quantum hardware constraints, practically one would talk about the qubit connectivity, available gate set, noise model, etc. However, in this paper, it seems that it only talks about the circuit depth. - The paper is reasonably sound and well-presented. However, I am not sure if the broader NeurIPS community will be interested in this work since it is a specific improvement for the QAOA algorithm. This will attract the quantum machine learning community. However, some of the quantum computing communities might also be skeptical because, as mentioned in the conclusion, this does not take into account the noise in the device. - The paper does not address the efficiency issue. This can be done by adding the training and inference time for each method in Table 1. Minor comments: - “Intial” in Figure 1. - Add $d_{eff}$ after “effective dimension” to make it clearer in Figure 1. Technical Quality: 4 Clarity: 4 Questions for Authors: - It is not clear why position embedding is needed to represent the depth, why not give the number or use simple one-hot encoding? - Why $J_{ij}$ is chosen to be only positive between $[0.5, 1.5]$ and not include also negative value or even 0? To avoid frustration? - Can this approach be extended to other problems in quantum computing that use PQC such as unitary synthesis or state preparation? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Limitations are addressed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your detailed evaluation and thoughtful feedback on our manuscript. We are pleased to hear that you found MG-Net to be a novel, innovative, and sound contribution, with strong theoretical backing and comprehensive performance evaluation. In the following, we separately address your concerns: Q1: The “Quantum ... about the circuit depth. A1: As explained in the Section "Motivation" of our Global Response, we consider the allowable circuit depth as the primary hardware constraint, a common issue for both early fault-tolerant and NISQ devices. The allowable circuit depth of a quantum device is closely related to the metric of maximum coherence time. To make our title clearer, we have followed your suggestion and rewritten the title as "MG-Net: Learn to Customize QAOA with Circuit Depth Awareness" and clarified the hardware constraints in the main text of the revised manuscript. Additionally, MG-Net can be flexibly extended to other hardware constraints. Refer to the Section "Impact" of our global response for further details. Q2: The paper ... the noise in the device. A2: We appreciate your recognition of the soundness and presentation of our work. We address your concerns regarding the broader interest to the NeurIPS community and the consideration of device noise as follows: 1. Broader Interest to the NeurIPS Community. While the primary focus of our manuscript is on improving the QAOA to tackle complicated combinatorial optimization problems, the proposed method provides a general framework for automatically learning a hardware-oriented circuit generator that can balance performance and hardware resources. The methodologies and insights derived from this work have broader implications for the field of machine learning for physical sciences, which aligns with NeurIPS's emphasis on interdisciplinary collaboration. The principles of dynamically optimizing quantum circuits using deep learning can be applied to other quantum algorithms and hybrid quantum-classical approaches, making this work relevant to a wide audience interested in quantum technologies and their integration with machine learning. There is a growing number of papers related to quantum computing have been published at top AI conferences like NeurIPS, ICML and ICLR [Gao et al., 2023; Sidford et al., 2023; Wu et al., 2023; Tang et al., 2024; Patel et al., 2024; Lei et al., 2024]. Among these, QAOA is particularly appealing for solving combinatorial optimization problems. 2. Noise Settings. We acknowledge that our current manuscript does not explicitly consider noise constraints in quantum devices, as discussed in the conclusion of our manuscript. We address your concerns from two dimensions: the importance of fault-tolerant quantum research and the extension to noisy devices. - Importance of fault-tolerant quantum research. Conducting research within a fault-tolerant, noiseless framework is crucial for unlocking the full potential of quantum computing and achieving quantum advantages over classical methods, such as works on quantum learning theory and optimization acceleration [Larocca et al., 2023; Liu et al., 2024]. Our theoretical analysis on QAOA is situated at the same fault-tolerant context to provide practical insights for improving QAOA's performance to achieve quantum advantage within limited circuit depths in idealized settings. - Suitability for noisy devices. Our algorithm implementation MG-Net can adapt to device noise. Refer to the Section "Impact" of our Global Response for further details. The success of our approach within this framework sets the stage for future exploration of our model's applicability to noisy systems and real-device experimentation. Q3: The paper does not ... in Table 1. A3: Since only MG-Net includes a training phase, we have followed your advice to add inference time for each method shown in Table 2 of the newly uploaded pdf. Q4: “Intial” in Figure 1. A4: We have corrected this typo in the revised manuscript. Q5: Add $d_{eff}$ after "effective dimension" to make it clearer in Figure 1. A5: We have followed your advice to add $d_{eff}$ after "effective dimension" in the caption of Figure 1 of the revised manuscript. Q6: It is not clear ... one-hot encoding? A6: Thank you for your insightful question. It is indeed an interesting idea to replace position embedding with integer encoding or one-hot encoding. As mentioned in the "Contribution" section of our Global Response, MG-Net acts as an initial protocol and provides a flexible circuit-generation framework where model components can be conveniently replaced by advanced techniques. To address your concerns, we are conducting additional experiments to replace position embedding with one-hot encoding or integer encoding. Due to the limited rebuttal period, we will update the results as soon as they become available. Q7: Why $J_{ij}$ is ... frustration? A7: The choice of $J_{ij}$ values in our study is influenced by several considerations related to TFIM. The TFIM undergoes a quantum phase transition at the critical point $h/J=1$. To more comprehensively study the behavior of the system in different phases, we set $J\in [0.5, 1.5]$ and $h\in [0.1, 2]$. This range ensures that the value of $h/J$ can be smaller than, equal to and greater than 1, covering the critical point and both phases of the transition. We do not consider the case where $J$ is negative because we mainly focus on ferromagnetic interactions to avoid frustration. Q8: Can this approach be extended ... preparation? A8: Yes, our approach can be flexibly extended to other problems in quantum computing that use parameterized quantum circuits. Please refer to the Section "Impact" of our global response for further details. --- Rebuttal Comment 1.1: Comment: I want to thank the authors for their detailed responses to my comments and questions and for considering them in their revised paper. --- Reply to Comment 1.1.1: Comment: Dear Reviewer DGPC, Thank you for your kind words and for taking the time to review our responses. We are glad that our revisions and clarifications addressed your concerns, and we are grateful for your constructive feedback throughout this process. Best regards, Authors --- Rebuttal Comment 1.2: Title: A6: Update the results of onehot encoding and integer encoding Comment: There are two key differences between the implementation of position encoding and one-hot or integer encoding: 1. Vector length. The length of the one-hot-encoded vector $\mathbf{x}_p$ depends on the predefined maximum value of $p$, while the length of the integer-encoded vector $\mathbf{x}_p$ is 1. In contrast, we adjust the length of position-encoded vector $\mathbf{x}_p$ according to the dimension of $\mathbf{x}_C$ and $\mathbf{x}_M$; 2. Integration strategy. When using one-hot or integer encoding, we employ concatenation as the integration strategy for the three features $\mathbf{x}_C$, $\mathbf{x}_M$ and $\mathbf{x}_p$ rather than summation. The achieved approximation ratios for 6-qubit MaxCut problems using different depth encoding methods are shown below: \| Depth encoding method \| Approximation ratio \| \| ---------------------------- \| -----------------------\| \| Integer \| $0.981\pm 0.004$ \| \| One-hot \| $0.984\pm 0.003$\| \| Position \| $0.99\pm 0.0004$	Summary: This paper provides a theoretical analysis on parameter grouping strategy in QAOA circuits. And the authors proposed MG-Net to design the mixer Hamiltonian, leading to the advantage over conventional methods and other QAOA classes. Strengths: 1. MG-Net reduces the cost of labelling and training by utilizing two-stage training strategy. 2. Extensive numerical results strongly support the advantage of MG-Net, advancing the practicality of QAOAs. 3. The method of this paper is certainly effective and the results are well presented. Weaknesses: 1. The parameter grouping strategy does not guarantee a significant reduction in the effective dimension, which undermines the theoretical analysis's assurance of better convergence when employing MG-Net in subsequent sections. 2. The details regarding the training of the mixer generator with hardware information are not clearly presented, raising concerns about the authors' claims of a "problem-hardware-tailored mixer" and "Customize QAOA with Quantum Hardware Constraints." 3. The numerical results comparison should include the Grover-Mixer method, as it also involves a unique design of the mixer Hamiltonian. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Have the author considered and compared with other mixer Hamiltonian design, e.g. Grover-Mixer? 2. Can MG-Net be adapted to other VQA problems? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: There does not seem to be negative social impact of this theoretical research. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your thorough evaluation and thoughtful feedback on our manuscript. We are pleased that you recognized the strengths of MG-Net, including its cost-effective two-stage training strategy and the extensive numerical results that highlight its advantages, enhancing the practicality of QAOAs. In the following, we separately address your concerns: Q1: The parameter grouping ... sections. A1: To address the reviewer's concern, we would like to emphasize that the theoretical results are not established to completely guarantee the performance of MG-Net. Instead, the proposed MG-Net in our work is theory-inspired to improve the convergence by tailoring ansatz design with optimal parameter grouping strategy. To explain this point, we elaborate on the relationship between the theoretical results and proposed MG-Net algorithms. We first recall our theoretical contributions stated in Global Response that the Theorem 3.1 only establishes the qualitative analysis for the effect of parameter grouping strategy on the convergence in the overparameterization regime, showing that parameter grouping strategy could potentially reduce the effective dimension with the reduction amount determined by the specific Hamiltonian. Moreover, it is difficult to quantitatively analyze the reduction of effective dimension enabled by the parameter grouping strategy for various problem Hamiltonians, as this reduction depends on the detailed graph structure of problem Hamiltonians which is hard to analyze and is case-by-case. In this regard, there is no explicit optimal grouping strategy for significant reduction of the effective dimension for general problem Hamiltonians. To address this problem, we propose MG-Net to exploite the power of deep learning to automatically tailor the ansatz design with optimal parameter grouping strategy to achieve better convergence. Q2: The details regarding ... Constraints." A2: We acknowledge that further clarification is needed regarding the integration of hardware constraints in the training of MG-Net in our manuscript. As explained in the Section "Motivation" of our Global Response, we consider the allowable circuit depth as the primary hardware constraint, a common issue for both early fault-tolerant and NISQ devices. In the following, we address your concerns by providing a detailed explanation of the hardware constraints referred to in our manuscript and the implementation of the training of the mixer generator with hardware information. - Quantum hardware constraints. There are many properties of a quantum device that affect its capacity, such as qubit connectivity, noise level, and maximum coherence time. In our manuscript, the term "hardware constraint" refers to the allowable circuit depth of a quantum device, which is closely related to the metric of maximum coherence time. - Hardware-information-aware training. When constructing the labeled training dataset, we collect the achieved approximation ratio of a mixer Hamiltonian with varying circuit depths. After the first-stage training, the cost estimator can precisely capture the intrinsic correlation between the circuit depth and the achievable cost, as verified in Lines 282-288 of our manuscript. In the second-stage training, this circuit depth information is incorporated by encoding it into the input features of the mixer generator. Guided by the circuit-depth-oriented cost estimator, the mixer generator can be trained to predict the optimal mixer generator according to the given hardware constraints, i.e., circuit depth. - Extension to other hardware constraints. Although we only consider the allowable circuit depth as the hardware constraint in applying MG-Net to QAOA in our manuscript, our proposed method is a general framework to learn a hardware-oriented circuit generator. It can be flexibly extended to other hardware constraints. Refer to the discussion on impact in our global response for details. To clarify, we have followed the reviewer's advice to clarify the hardware-constraint-aware training in the revised manuscript. Q3: The numerical results comparison should include the Grover-Mixer method, as it also involves a unique design of the mixer Hamiltonian. A3: The Grover-Mixer (GM) method uses Grover-like selective phase shift mixing operators based on the prepared equal superposition of all feasible states. GM is designed to perform particularly well for constraint optimization problems. However, in our manuscript, we utilize QAOA to solve Max-Cut and TFIM problems, which are both unconstrained optimization problems. When applying GM to these tasks, it is equivalent to the original QAOA. Therefore, we did not consider GM in our original manuscript. To effectively compare our method with GM, we shifted our focus to a new constrained optimization problem. Due to the limited rebuttal period, we will update the results as soon as they become available. Q4: Have the author considered and compared with other mixer Hamiltonian design, e.g. Grover-Mixer? A4: Please refer to A3 for details. Q5: Can MG-Net be adapted to other VQA problems? A5: Yes, MG-Net can be flexibly adapted to handle other VQA problems. While the primary focus of our manuscript is on improving the QAOA, the proposed method provides a general framework for automatically learning a hardware-oriented circuit generator that can balance performance and hardware resources, based on the interdisciplinary collaboration of quantum science and artificial intelligence (AI). Many works that utilize deep learning techniques to assist the research of quantum algorithms have been successfully applied to tackle multiple VQA problems [Zhang et al, 2022; Wu et al, 2023; Fürrutter et al, 2024]. We kindly refer the reviewer to the discussion about extending our work to other problems in our Global Response. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response. A good portion of my concerns have been addressed.	Summary: In this paper, the authors propose a model named MG-Net to automatically generate QAOA ansatz. It is able to generate the mixer layer in QAOA and decide which gates to share parameters. By sharing parameters, the ansatz generated by the proposed model can potentially achieve better trainability as well as expressivity. Experimental results on Max-cut and TFIM for up to 64 qubits are provided showing the proposed method can improve the accuracy under the proposed setting. Strengths: Well-structured paper with numerous experiments verifying the theoretical claims. Weaknesses: 1. It is too simple to design the mixer Hamiltonian since it only involves choosing one of the three Pauli operators to fill in the blanks. I don't see evidence that the proposed model (especially the training step) can still work if it encounters a more complex candidate gate set. 2. The theoretical findings are not surprising to me since they mainly come from the quantum optimal control theory and overparameterization. 3. The experimental results are shaky. 1) The comparison against ma-QAOA actually shows that these two methods have extremely close performance, which makes sense since the proposed grouping method is similar to ma-QAOA; 2) Missing the number of parameters (especially the number of different parameters) for each method in all the experiments, which is crucial since the expressivity of the ansatz is closely related to the number of tunable parameters; 3) The fatal problem is that the authors seem not fully understand the QAOA. The RZZ gates provide two-qubit phase transitions, and the original mixer, which is made of Pauli-X gates, finds the minimum eigenvector. Therefore, the ansatz should not consist of any single-qubit Pauli-z gate to introduce an illegal phase to the final state. This can also be verified in Tab.4 in the appendix. A Pauli-Y gate can be decomposed by Pauli-Z and Pauli-X gates, and since it's illegal to have a single-qubit phase in the final state, involving Pauli-Y gates will not benefit the results. It is clear that there is no single-qubit phase in the Hamiltonian of either Max-Cut or TFIM; I'm confused about how the authors got the results. So with all the efforts in training and selecting from {X, Y, Z, I}, it is actually deciding X or Y, which have identical effects on the final state (Y might be even weaker). I suspect that randomly choosing from X and Y can produce identical results with precisely the same number of tunable parameters. 4. The whole paper is wrapped in an exquisite box with the essence of only choosing Pauli-X or Pauli-Y for the QAOA mixer. Technical Quality: 3 Clarity: 3 Questions for Authors: No questions. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Shaky experiments with naive essence disguised in the fancy presentation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 2 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our manuscript. We value your insights and address your concerns as follows: Q1: It is too simple ... candidate gate set. A1: We respectfully disagree with the reviewer's perspective that designing the mixer Hamiltonian is too simple. As mentioned in the Section "Motivation" of our Global Response, We release all freedom of the mixer Hamiltonian, including operator type and parameter grouping. Directly designing the mixer Hamiltonian based on Pauli operators remains computationally challenging and our proposed method tries to overcome this challenge efficiently. 1. Structure of Mixer Hamiltonian: As described in Lines 176-189 and Eqn. (5) of our manuscript, designing a mixer Hamiltonian involves not only choosing appropriate operator types but also determining the parameter groups. This dual requirement adds complexity to the design process beyond simply selecting Pauli operators. 2. Challenges in Mixer Hamiltonian Design: As discussed in Lines 194-197 of our manuscript, the search space for both parameter correlation and operator types grows exponentially with the system size $N$ (i.e., scaling at $O(N^N)$ and $O(2^N)$, respectively). This exponential growth makes designing an effective learning method difficult, as directly training a model in a supervised learning paradigm may require an infeasible amount of training data to achieve high accuracy. MG-Net addresses these challenges by bypassing the need to directly seek the optimal parameter correlation strategy and operator type. 3. Additional Experiments on Extended Candidate Operator Type Set: To further support our claims, we have conducted additional experiments with an extended set of candidate operator types $\{X,Y,XX,YY\}$, achieving an approximation ratio $0.985\pm 0.003$. As demonstrated in Figure 1 of the uploaded pdf, MG-Net maintains its efficiency and high performance even when the complexity of the mixer Hamiltonian design is increased. Q2: The ... overparameterization. A2: We would like to emphasize that our primary contribution lies in the development of novel algorithms designed to tailor optimal ansatz. While our study includes theoretical results, they are not the central focus but rather provide the theoretical foundation for MG-Net. Please refer to the Global Response for details. Notably, the theoretical results presented are non-trivial in terms of both contributions and techniques. Theoretical Contribution. Existing work on QNN convergence with symmetric ansatz does not address how specific ansatz design strategies affect symmetry or effective dimension. Our study qualitatively analyzes the relationship between the effective dimension and QNN convergence with various parameter grouping strategies in the context of QAOA. Technical Aspect. Our theoretical techniques involve the novel concept of effective dimension and tools from representation theory, widely used in analyzing QNNs with symmetric ansatz in quantum optimal control and over-parameterization. While the notion of effective dimension is inspired by these fields, our techniques for analyzing its relationship with ansatz design are unique. We use representation theory to compare the effective dimension of QNNs with different parameter grouping strategies by examining their algebraic structures. Q3: The experimental results are shaky ... parameters. A3: We respectfully disagree with the reviewer's perspective that the experimental results are shaky. We address the reviewer's concerns one by one as follows: 1. Comparison with ma-QAOA: As demonstrated in Section 5 of our manuscript, our proposed MG-Net performs significantly better than ma-QAOA regarding convergence and approximation ratio. As the system scale grows from 6 to 64, the approximation ratio gap between our method and ma-QAOA increases from 1% to 96%. Besides, the mechanism for designing the parameter grouping strategy for MG-Net and ma-QAOA differs. MG-Net dynamically generates the optimal parameter grouping based on the given problem instance and allowable circuit depth, while ma-QAOA adopts a static non-grouping strategy to maximize the number of trainable parameters. 2. Number of Parameters: We kindly remind the reviewer to refer to Lines 295-303 and Figure 5(a) of our manuscript for a detailed discussion about the number of parameters of different methods. 3. Mixer Hamiltonian Design: We agree that the mixer Hamiltonian usually cannot be purely Pauli-Z due to the requirement of noncommutativity with the cost Hamiltonian. Therefore, we do not include Pauli-Z in the pool of operator types. However, the Pauli-Y operator is acceptable and can drive the quantum state toward the solution along a different and possibly more efficient path than the Pauli-X operator, as indicated in Figure 1(a) of our manuscript. This choice aligns with principles of counterdiabatic (CD) driving or shortcuts to adiabaticity (STA). Numerous works [Chandarana et al., 2022; Zhu et al., 2022] have verified the effectiveness of using Pauli-Y as a mixer Hamiltonian, as listed in Table 1 of the uploaded pdf. Additional experiments investigating the impact of randomly choosing mixer operator types from Pauli-X and Pauli-Y while maintaining the same number of tunable parameters indicate that mixer Hamiltonians with different operators and the same number of parameters behave differently, shown in Figure 2 of the uploaded pdf. Q4: The whole ... mixer A4: Our disagreement stems from the motivation and design of the mixer Hamiltonian in QAOA. Based on the elaboration in our Global Response and our response A1-3, we believe the review provides a clear understanding of our work. We are prepared to provide further explanations if necessary. --- Rebuttal Comment 1.1: Comment: It seems that the reviewers currently do not have access to the general response or the uploaded files. --- Reply to Comment 1.1.1: Comment: We apologize for the issue with accessing our global response and the uploaded files. We are seeking assistance to resolve this issue and make the materials visible to the reviewers. For your convenience, we have pasted some explanations for your concerns from the global response below. Motivation. A promising approach to combinatorial optimization problems (COPs) is QAOA, which can potentially outperform classical methods when unlimited circuit depth is assumed. However, this requirement is impractical. Considering the practical utility of QAOA, where quantum resources are constrained by limited circuit depth, noise, and qubit connectivity, many variants of QAOA have been proposed to enhance its performance within these hardware constraints [Chandarana et al., 2022; Zhu et al., 2022; Yu et al., 2022; Herrman et al., 2022; Bartschi et al., 2020; Sauvage et al., 2022]. However, these alternatives often require deep domain expertise and lack generalizability across different tasks and circuit depths. In our work, we consider the allowable circuit depth as the primary hardware constraint, a common issue for both early fault-tolerant and NISQ devices. We release all freedom on the mixer Hamiltonian, including operator type and parameter grouping. Our goal is to dynamically adjust the mixer Hamiltonian according to the given problem and the allowable circuit depth of a quantum device, thereby enhancing the performance of QAOA while ensuring compatibility with the available quantum resources. Contributions. To fully exploit the potential of a QAOA circuit with arbitrarily limited circuit depth, we first theoretically analyze the convergence of QAOA: - Theoretical contributions. Existing literature studied the convergence of VQAs from two aspects, namely the quantum neural tangent kernel [Liu et al., 2022; Liu et al., 2023] and gradient flow [You et al., 2022; You et al., 2023]. In particular, when analyzing the convergence rate of QNNs with symmetric ansatz, both of these two techniques involve utilizing the tools of effective dimension for quantifying the effect of symmetry degree of ansatz design on the convergence rate. While existing literature has explored the convergence theory of VQAs with symmetric ansatz well, how concrete strategies for ansatz design affect the symmetry degree remains an open question. In this study, we initiate the first attempt to qualitatively analyze the effects of various parameter grouping strategies on the convergence of QAOA. - Technical contributions. We rigorously show that fully or partially grouping the parameters according to the spatial symmetry of problem Hamiltonians could reduce the effective dimension (Refer to Lemma B.6 in Appendix B), where the concrete reduction amount depends on the symmetric degree of problem Hamiltonians. Moreover, combined with the existing results on the effective dimension-based convergence analysis [You et al., 2022], we reach our main theoretical results (Theorem 3.1) regarding the convergence rate of QAOA with various symmetric structures in the overparameterization regime. - Implications. Qualitatively, when $p$ is large enough to reach the overparameterization regime, a suitable parameters grouping strategy should be adopted to reduce the effective dimension for better convergence. Conversely, for a small $p$, the parameters should not be grouped to obtain as large a representation space as possible for better convergence. However, it is difficult to quantitatively determine the effective dimension, the critical point for the overparameterization regime for general problem Hamiltonians, as they are determined by the specific graph structure of the problem Hamiltonian which is case-by-case and hard to analyze. This inspires us to utilize the power of deep learning to automatically tailor the circuit depth-aware ansatz design with the optimal parameter grouping strategy for better convergence. Guided by the established theoretical results, we propose MG-Net for automatical circuit design： - The QAOA circuit generated by the proposed MG-Net achieves higher approximation ratios at various circuit depths compared to other quantum and traditional methods, advancing the practical utility of QAOAs. - MG-Net greatly reduces the cost of collecting labeled training data, making it possible to handle larger-scale problems and more complicated mixer Hamiltonians. - MG-Net provides a flexible circuit-generation framework where the data encoder can introduce more hardware constraints, and the model components can be conveniently replaced by advanced techniques. Although our theoretical analysis is based on noiseless settings, our implementation does not impose limitations on device configuration. MG-Net can be flexibly extended to other hardware constraints, such as qubit connectivity and hardware noise. --- Reply to Comment 1.1.2: Comment: We describe the tables and figures in the uploaded file below: - Figure 1: This figure describes the behavior of the cost estimator with extended mixer operator pool $\{X,Y,XX,YY\}$ by drawing the correlation between the actual achieved approximation ratio and the result predicted by the cost estimator. A strong correlation (Spearman correlation coefficient of 0.85) between the estimated value and the label is observed, indicating that the cost estimator trained on the extended mixer operator pool can act as a reliable performance indicator for QAOA. - Table 1: This table lists previous works that have introduced the Pauli-Y operator as a candidate mixer Hamiltonian: \| Works \| Mixer Hamiltonian \| \|---------------\|-------------------\| \| DC-QAOA [Chandarana et al, 2022] \| $\{X,Y,ZY,YZ,XY,YX\}$\| \| ADAPT-QAOA [Zhu et al., 2022] \| $\bigcup_{i \in [N]} \{X_i,Y_i\} \bigcup \{\sum_{i\in [N]}Y_i\} \bigcup \{\sum_{i\in[N]}X_i\}\bigcup _{i,j\in[N]\times [N]}\{B_iC_j\\|B_i,C_j \in\{X,Y,Z\}\}$ \| - Figure 2: This figure describes the distribution of achieved approximation ratios related to different mixer operators with the same number of tunable parameters. The results indicate that mixer Hamiltonians with different operators randomly sampled from PauliX and PauliY and the same number of parameters behave differently. For example, the minimum, maximum, and standard deviation of achieved cost for a set of mixer Hamiltonians with 2 parameters are -1.45, -0.14, and 0.27, respectively.	Summary: This paper studies the convergence behavior of QAOA, a variational quantum algorithm for combinatorial optimization. The authors prove a convergence result showing how the use of parameter grouping affects the expressibility (i.e. effective dimension) and training time. Furthermore, they design a deep learning model known as MG-Net to generate an appropriate mixer Hamiltonian for a given input problem. Experiments are conducted to evaluate the performance of their model on prototypical problems (Max-Cut and TFIM). The numerical results suggest that their method achieves a higher approximation ratio than all other tested methods, including a simple greedy algorithm, Goemans-Williamson, and standard QAOA. Strengths: - This paper tries to resolve an important question relevant to the performance of QAOA, which has implications for the practicality of quantum optimization algorithms. - The use of parameter grouping, while not a very novel idea, appears to be the "right" choice to take advantage of problem symmetries and improve the performance of QAOA. - The presentation of this paper is clear and everything is fairly straightforward to understand. - The code availability is nice to have. Weaknesses: - In Section 3, there are several references mentioned that also study the convergence of VQAs, but the theoretical contributions of this work are somewhat unclear. This paper would benefit from a clear statement of how this work relates to the results of those existing papers. Technical Quality: 3 Clarity: 3 Questions for Authors: - Throughout this paper, $p$ is referred to as the circuit depth. Shouldn't it be the number of layers, which is not quite the same as the usually defined circuit depth (although they are proportional)? - For a given $H_C$, is the ansatz with partial grouping unique? I suppose not since there may be different generators of $\mathrm{Per}(H_C)$. If this is correct and the ansatz is not unique, then does Theorem 3.1 hold regardless of the choice of generators? - It is stated that this approach is designed for early fault-tolerant algorithms, while many existing works on variational quantum algorithms suggest they may be suitable for NISQ devices. Aside from the high monetary cost of training parameterized circuits, are the quantum resources (i.e. circuit depths) required for this approach low enough to be practical without fault-tolerance? If not, what are the main limitations of QAOA that need to be overcome? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: There are no major societal impacts of this work. However, as mentioned in the paper, this work is limited to QAOA and it is of interest whether this approach may be applied to more general variational quantum algorithms. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing the potential of the proposed method to enhance QAOA with practical utility. Your feedback is invaluable, and we address each of your concerns as follows: Q1: In Section 3, there are several references mentioned that also study the convergence of VQAs, but the theoretical contributions of this work are somewhat unclear. This paper would benefit from a clear statement of how this work relates to the results of those existing papers. A1: We have followed the reviewer's suggestions to state the theoretical contributions of our study in the main text and highlight the relation with existing papers in Appendix C. In the following, we will separately address these two concerns raised by the reviewer. The relation with existing literature. As stated in the theoretical contribution of our Global Response, while the convergence of VQAs with symmetric ansatz has been established from various aspects, they did not consider how concrete strategies for ansatz design, specifically the parameter grouping, affect the symmetry degree or equivalently the effective dimension, and hence fail to apply to the problem setting in our manuscript. Theoretical contributions. We qualitatively analyze the relation between the effective dimension and convergence rate of QNNs equipped with various parameters grouping strategies in the context of QAOA. The techniques used in our theoretical results are mainly the effective dimension and the tools of representation theory [Ragone et al., 2023], which have been widely used in existing literature to perform theoretical analysis for QNNs with symmetric ansatz. Please refer to the contributions discussed in Global Response for details. Q2: Throughout this paper, $p$ is referred to as the circuit depth. Shouldn't it be the number of layers, which is not quite the same as the usually defined circuit depth (although they are proportional)? A2: We agree that the terms "circuit depth" and "number of layers" are often used interchangeably but can have distinct meanings in different contexts. Circuit depth typically measures how many "layers" of quantum gates are executed in parallel, whereas the number of layers can either share this meaning or refer to the number of blocks with the same structure. In QAOA, $p$ specifically denotes the number of blocks, each consisting of a cost Hamiltonian and a mixer Hamiltonian, as defined in Eqn.~(1) of the manuscript. Following the reviewer's advice, we have revised the manuscript to ensure consistent and precise terminology. Q3: For a given $H_C$, is the ansatz with partial grouping unique? I suppose not since there may be different generators of Per${H_C}$. If this is correct and the ansatz is not unique, then does Theorem 3.1 hold regardless of the choice of generators? A3: For the reviewer's supposition, the answer is right. Namely, the ansatz with partial grouping for given $H_C$ is not unique. For the second concern, Theorem 3.1 indeed holds regardless of the choice of generators. In particular, Theorem 3.1 is established to perform the qualitative analysis about the effects of parameter grouping strategies on convergence. Multiple partial grouping strategies could exist to achieve the same effective dimension, which depends on the specific problem Hamiltonian. However, it is difficult to quantitatively determine the effective dimension and explicitly adopt the correct parameter grouping strategy for better convergence, as achieving this requires analyzing the complicated graph structure of the problem Hamiltonian case-by-case. In this regard, we delve into exploiting the power of deep learning and propose the MG-Net for automatically generating the ansatz design with optimal parameter grouping strategy. Q4: It is stated that ... If not, what are the main limitations of QAOA that need to be overcome? A4: As explained in the Section "Motivation" of our Global Response, we consider the allowable circuit depth as the primary hardware constraint, a common issue for both early fault-tolerant and NISQ devices. To further address your concerns, we provide a detailed explanation of the rationale behind researching noiseless cases and discuss the suitability of our approach for noisy devices. Finally, we clarify the resource requirements of our approach. - Importance of Fault-Tolerant Quantum Research: Conducting research within a fault-tolerant framework is crucial for unlocking the full potential of quantum computing and achieving quantum advantages over classical methods, such as works on quantum learning theory and optimization acceleration [Larocca et al., 2023; Liu et al., 2024]. Our theoretical analysis of QAOA is situated in the same fault-tolerant context to provide practical insights for improving QAOA's performance to achieve quantum advantage within limited circuit depths in idealized settings. - Suitability for Noisy Devices: Although our theoretical analysis is based on noiseless settings, our algorithm implementation, MG-Net, does not impose limitations on device noise. Refer to the discussion on impact in our Global Response for details. - Resource Efficiency:** By dynamically generating mixer Hamiltonians tailored to specific problems and allowable circuit depths in practical hardware, MG-Net can improve the achievable approximation ratio of QAOA with small number of layers, making it more feasible for NISQ devices. As shown in Table 1 of our manuscript, our method achieves an approximation ratio of 0.96 for 64-qubit Max-Cut problems, which is much higher than other QAOAs with the same number of circuit layers. We hope that this clarification will address the reviewer's concerns. We are prepared to offer more comprehensive responses if there are any further questions. --- Rebuttal Comment 1.1: Comment: I thank the authors for the detailed response. --- Reply to Comment 1.1.1: Comment: Dear Reviewer aG1x, Thank you for your kind words and your thorough review of our paper. We appreciate your acknowledgment of our detailed responses and are glad that our efforts help address your concerns and questions. Best regards, Authors	Rebuttal 1: Rebuttal: Dear Reviewers, We thank all reviewers for their efforts, insightful comments and constructive suggestions. We appreciate the opportunity to clarify the main motivation, contributions and impact of our work. Motivation. A promising approach to tackle combinatorial optimization problems (COPs) is QAOA, which can potentially outperform classical methods when unlimited circuit depth is assumed. However, this requirement is impractical. Considering the practical utility of QAOA, where quantum resources are constrained by limited circuit depth, noise, and qubit connectivity, many variants of QAOA have been proposed to enhance its performance within these hardware constraints [Chandarana et al., 2022; Zhu et al., 2022; Yu et al., 2022; Herrman et al., 2022; Bartschi et al., 2020; Sauvage et al., 2022]. However, these alternatives often require deep domain expertise and lack generalizability across different tasks and circuit depths. Our work considers the allowable circuit depth as the primary hardware constraint, a common issue for both early fault-tolerant and NISQ devices. We release all freedom on the mixer Hamiltonian, including operator type and parameter grouping. Our goal is to dynamically adjust the mixer Hamiltonian according to the given problem and the allowable circuit depth of a quantum device, thereby enhancing the performance of QAOA while ensuring compatibility with the available quantum resources. Contributions. To fully exploit the potential of a QAOA circuit with arbitrarily limited circuit depth, we first theoretically analyze the convergence of QAOA: - Theoretical contributions. Existing literature studied the convergence of VQAs from two aspects, namely the quantum neural tangent kernel [Liu et al., 2022; Liu et al., 2023] and gradient flow [You et al., 2022; You et al., 2023]. In particular, when analyzing the convergence rate of QNNs with symmetric ansatz, both of these two techniques involve utilizing the tools of effective dimension for quantifying the effect of symmetry degree of ansatz design on the convergence rate. While existing literature has explored the convergence theory of VQAs with symmetric ansatz well, how concrete strategies for ansatz design affect the symmetry degree remains an open question. In this study, we initiate the first attempt to qualitatively analyze the effects of various parameter grouping strategies on the convergence of QAOA. - Technical contributions. We rigorously show that fully or partially grouping the parameters according to the spatial symmetry of problem Hamiltonians could reduce the effective dimension (Refer to Lemma B.6 in Appendix B), where the concrete reduction amount depends on the symmetric degree of problem Hamiltonians. Moreover, combined with the existing results on the effective dimension-based convergence analysis [You et al., 2022], we reach our main theoretical results (Theorem 3.1) regarding the convergence rate of QAOA with various symmetric structures in the overparameterization regime. - Implications. Qualitatively, when $p$ is large enough to reach the overparameterization regime, a suitable parameters grouping strategy should be adopted to reduce the effective dimension for better convergence. Conversely, for a small $p$, the parameters should not be grouped to obtain as large a representation space as possible for better convergence. However, it is difficult to quantitatively determine the effective dimension, the critical point for the overparameterization regime for general problem Hamiltonians, as they are determined by the specific graph structure of the problem Hamiltonian which is case-by-case and hard to analyze. This inspires us to utilize the power of deep learning to automatically tailor the circuit depth-aware ansatz design with the optimal parameter grouping strategy for better convergence. Guided by the established theoretical results, we propose MG-Net for automatical circuit design： - The QAOA circuit generated by the proposed MG-Net achieves higher approximation ratios at various circuit depths compared to other quantum and traditional methods, advancing the practical utility of QAOAs. - MG-Net greatly reduces the cost of collecting labeled training data, making it possible to handle larger-scale problems and more complicated mixer Hamiltonians. - MG-Net provides a flexible circuit-generation framework where the data encoder can introduce more hardware constraints, and the model components can be conveniently replaced by advanced techniques. Although our theoretical analysis is based on noiseless settings, our implementation does not impose limitations on device configuration. MG-Net can be flexibly extended to other hardware constraints, such as qubit connectivity and hardware noise. Impact. While our work enhances the ability to employ quantum algorithms to tackle combinatorial optimization problems, the methodologies and insights derived from this work have a broader impact on the field of machine learning for physical sciences. The principles of dynamically optimizing quantum circuits using deep learning can be applied to other quantum algorithms and hybrid quantum-classical approaches. Concretely, MG-Net can be flexibly adapted to handle other VQA problems. The feature encoding used in MG-Net, which includes problem-specific and hardware-specific information, can be modified to accommodate the requirements of different VQA problems. This modularity ensures that the model can process diverse types of input data relevant to other quantum algorithms, such as unitary synthesis, state preparation and quantum machine learning. We hope these explanations can provide a more comprehensive understanding of our work and partially address your concerns. Below, we provide a point-by-point response to the reviewers’ comments and concerns. Best regards, Authors Pdf: /pdf/be1036ba111cd80adbc9435a41d89d4eebe83a95.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation	Accept (poster)	Summary: The problem statement, solution and its mathematical formulation is nicely presented in the draft. Authors have identified a problem in the time series generation process which is under explored and they propose a new algorithm Frequency-Inflated Conditional Diffusion Model (FiDE) to address this. The problem lies in the modelling of block maxima which are related to extreme events in time series data where traditional diffusion models fail or does not take into account properly. Authors demonstrated that these extreme values correspond to high frequency fourier components in time series and proposed the algorithm to inflate the weights of high frequency fourier components. The inflation of high-frequency components in the Fourier domain to prevent their premature dissipation during the diffusion process. This approach enables the model to retain significant extreme values that are typically lost in traditional modeling techniques. The paper presents detailed performance comparison with several classes of models, e.g. VAE, diffusion, and GAN using datasets from several domains. The comparison is extended to 4 metrics explaining the motivation to use them. Strengths: - The problem statement and its novel solution are the strength of the research presented. It is of great importance to model the tails of time series distribution which are generally not very well modeled. - correctly identifying that the tail is mainly due to high frequency Fourier components and authors model it well using the algorithm proposed. Strength lie in the fact that they found the root cause and then proposed the solution. Weaknesses: The results [Fig 5] show the comparison of block maxima values for only one dataset. More details on the dataset and more comparison would be helpful to confirm the generality of the proposed algorithm. Figure 5 shows the final results. It shows the improvement in the block maxima modelling but at the same time it does show degradation in modelling the bulk. An additional comparison of ALL - block maxima with a quantitative estimate of degradation in the performance is needed to understand the full impact of FIDE. The table shows the comparison of metrics for block maxima because this is the main focus of this paper. Understood! But at the same time to make it a fair comparison the quantification on the ALL - maxima block is also needed to fully evaluate the model. With the present table and figures it is challenging to quantify if the model is doing "more good" than "bad". Authors do mention in Results section L 257-261 that DDPM has slightly better performance but it is not quantified. Technical Quality: 3 Clarity: 4 Questions for Authors: The paper is generally well-written, but a few important points remain unanswered, which piques my curiosity. They are as follows: - When defining the block maximum what is the time window? What criteria defined this time window? Is there a thumb rule or model to choose it dynamically? - A discussion on error (unless missed) is needed. How are errors estimated on the metric reported in table 1. What are these errors (statistical/systematic or something else)? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: - More experimental comparison like figure 5 are really needed in the paper draft for further clarity. - What is the effect of FIDE on all samples (not just block maxima), a number would be useful to quote similar to block maxima. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Question 1: The choice of time window for defining block maxima is domain-dependent and application-specific, driven by the temporal characteristics of the data and the phenomena of interest. For climate or energy data, monthly block maxima (30-day or 90-day windows) are often relevant, capturing sub-seasonal to seasonal extremes. ECG data may require hourly or smaller windows to detect critical cardiac events, while financial data typically uses monthly or quarterly windows, aligning with reporting periods and long-term trends. The optimal window size should be determined based on the specific research questions, data resolution, and relevant time scales of the studied phenomena. Response to Question 2: We thank the reviewer for the question. The errors are estimated statistically. Each experiment was repeated five times, with the standard deviation across these iterations serving as our error estimate. This approach accounts for variability in model performance due to stochastic elements, such as random initialization of the network in training. We will explicitly detail this methodology in the revised manuscript to enhance clarity and reproducibility. Response to Weaknesses: Although our focus is on block maxima, our model generates the entire time series. We provided a comparison of the distributions for all values in Figure 5. However, due to space constraints, we were unable to include comparative figures for all baseline methods. Nevertheless, we have now included a table comparing the performance for all values (similar to Table 1). This table (Table A2) is provided in the attached PDF (please see the general rebuttal), which shows the metrics comparing the distribution and predictive performance for all values. The results indicate that our method is quite comparable to DDPM, though slightly worse, due to the tradeoff between preserving the overall vs extreme value distribution. However, it is still better than other VAE-based, GAN-based, and flow-based methods.	Summary: This article discusses the shortcomings of the insufficient ability to focus on maximum values when applying DDPM in the field of time series generation, and proposes a new framework to overcome this problem by introducing a high-frequency expansion strategy in the frequency domain to ensure the emphasis on high-frequency values. At the same time, the article also proposes a generative modeling method based on conditional diffusion. Strengths: 1. The logic of the article is relatively clear, the language is relatively clear and precise, and it is easy to understand. 2. The experimental settings of the article are relatively clear and the results are well organized. Weaknesses: 1. The article does not set up an ablation experiment to explore the effects of the improved strategy. 2. The article has a small amount of experiments and incomplete research. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. In Chapter 3, the article mentioned that adding linear noise to the diffusion model will reduce high-frequency components faster. There are many ways to add noise in the diffusion model, such as the Noise Schedule in iDDPM. Is there any review on this part? Discussion and experimentation? 2. Should the Transformer in Figure 4 in Chapter 4 be a Transformer Encoder? 3. In the field of time series generation, the conditional generation of diffusion models has been studied in the article DiffusionTS. What is the difference between the two? 4. In the selection of comparative algorithms, there have been some time series generation methods based on flow methods, eg: Fourier Flow. Do articles based on this method have good performance in the distribution of block maximum values? 5. The article has mostly discussed the issue of block maximum value distribution in time series generation. Should we discuss the performance of the data generated by enhancing the block maximum value distribution in downstream tasks? Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Question 1: We initially evaluated linear, sqrt, and sigmoid noise schedulers. We chose the linear scheduler as it provides a more gradual perturbation compared to the sqrt and sigmoid, which alter data more rapidly in initial iterations. We appreciate the reviewer's suggestion regarding iDDPM's noise schedule. Investigating the impact of alternative schedulers, such as the cosine scheduler in iDDPM, on high-frequency component dissipation is indeed an intriguing avenue for research. While beyond the scope of the current study, we acknowledge this as a valuable direction for future work and will explore it in subsequent investigations. Response to Question 2: Yes, it is a Transformer Encoder. We will change the notation/figure in the revised manuscript accordingly. Response to Question 3: Our proposed framework is designed to improve the generation of extreme values (i.e., block maxima) in time series whereas DiffusionTS enhances time series generation by emphasizing interpretability and the realistic representation of the generated data. Due to the tradeoff between preserving the overall distribution and accurately representing extreme values, DiffusionTS is insufficient to reproduce the extreme value distribution. To support this, we have included experiments comparing DiffusionTS to our method in the attached general rebuttal PDF (See Table A1). The results indicate that while DiffusionTS effectively captures the general distribution, it does not adequately preserve the extreme values. Response to Question 4: We appreciate the reviewer's suggestion to include flow-based methods in our comparative analysis. In response, we have incorporated two such baselines: Fourier-Flows and RealNVP. The results of this extended comparison are presented in the attached Table A1 (general rebuttal PDF). Our findings suggest that the proposed method consistently outperforms both flow-based baselines. Moreover, while flow-based methods perform comparably to the VAE- and GAN-based methods, they generally fall short when compared to diffusion-based approaches. This comparison further validates the efficacy of our proposed method across a broader spectrum of state-of-the-art approaches. Response to Question 5: The Predictive Score metric used in (Yoon et al., 2019) in Table 1 directly addresses the performance of generated data in a predictive downstream task. This metric evaluates how well a predictive model trained on generated data performs when tested on real data, effectively assessing the fidelity of our generated time series in practical downstream applications. Our method achieves top performance on three datasets and second-best on two others. This underscores the practical utility of our approach beyond mere distribution matching, showing its value in generating data that preserves important predictive characteristics of the original time series. Regarding Weakness 1: Indeed, we have provided an ablation experiment (See Appendix Table 2) exploring the effects of different strategies/modules of our proposed framework. This analysis specifically illustrates how different components of our approach contribute to the overall performance. These results provide valuable insights into the relative importance of each element in our model --- Rebuttal Comment 1.1: Comment: Thanks. I am not satisfied with the author's rebuttal. there are still a lot of issues that need to be clarified here. Q1: The motivation for using this strategy requires further analysis and validation. Q3: Based on the numerical experimental results, it is difficult to determine whether the performance improvement of the proposed method compared to diffusionTS is due to capturing extreme values. The differences between the two methods need further explanation. Why can't diffusionTS capture extreme values, but your method can? What is the motivation and principle behind this? You should use Wilcoxon-Holm analysis the results in Table 1 to demonstrate the advantages and disadvantages of the proposed method. Q4: What were the experimental setups for diffusionTS, Fourier-Flows, and RealNVP? How is the fairness of the comparison ensured? W1: Wilcoxon-Holm analysis should be used to demonstrate that the results in Table 2 indeed show a significant performance improvement. The mean values alone do not sufficiently indicate that your results are superior. I also do not think it is a good idea to place the ablation experiments in the supplementary materials. The core challenge in time series generation tasks is the scarcity of data, yet diffusion models require large amounts of data for training. Please verify the performance of the proposed method on small-scale datasets, such as 10% of the stock dataset. I am certain that diffusionTS cannot handle this; please provide an explanation and validation. What is the significance of generating time series in the context of large-scale data? --- Reply to Comment 1.1.1: Comment: Regarding Q3: We thank the reviewer for suggesting the Wilcoxon-Holm test to further validate the performance improvement of our proposed method relative to DiffusionTS. Although we typically use t-test statistics, we have conducted the Wilcoxon-Holm test as recommended. This analysis was performed on results from three datasets across four performance metrics for both our method and DiffusionTS, which serves as the second-best baseline in most cases. Out of 12 comparisons (3 datasets × 4 metrics), our method demonstrated statistically significant performance improvements over DiffusionTS in 10 cases, with p-values less than 0.05. For the other two cases, the p-values were 0.09 and 0.154. DiffusionTS struggles to effectively capture extreme values due to its primary focus on generating and reconstructing entire time series, rather than specifically preserving the distribution of extreme values. According to the objective function (Eq. 10) of DiffusinTS, it aims to minimize mean squared error in the time domain and frequency domain. This approach leads the model to prioritize the central tendency of the data, resulting in a strong focus on predicting the conditional expectation. However, this focus tends to underrepresent extreme values, which are often located in the tail of the distribution. Our proposed method addresses this limitation by incorporating both mean squared error and a Generalized Extreme Value (GEV) loss (see Eq. 11 in our paper). This combination allows our model to simultaneously maintain overall accuracy and accurately capture the distribution of extreme values, which is critical for applications that rely on modeling rare events. Furthermore, DiffusionTS's objective function (Eq. 10) and methodology do not adequately account for the preservation of high-frequency components, which are crucially linked to extreme values, as demonstrated both empirically and theoretically in our paper. The mean squared error applied to Fourier coefficients in DiffusionTS treats low and high frequency components equally. However, time series are typically dominated by low-frequency components, with most of the energy or power spectral density concentrated in these regions. High-frequency components, despite their importance for extreme values, contain much less energy. Consequently, this equal treatment often results in overlooking or underemphasizing these critical high-frequency components. Our proposed method addresses this limitation by introducing a targeted strategy to inflate high-frequency components. This approach ensures that these components do not prematurely fade out during the diffusion model's noising process, allowing them to dissipate at a rate comparable to low frequency components, thus better preserving the characteristics of extreme values. Regarding Q4: We employed comparable experimental setups for DiffusionTS, Fourier-Flows, and RealNVP as we did for our proposed method and other baseline approaches. To ensure a fair comparison, we carefully tuned the general hyperparameters (number of epochs, learning rate, batch size) for all methods under evaluation. To account for variability and ensure fairness in our comparisons, we conducted 5 independent runs for each method. We then reported the mean and standard deviation of all performance metrics across these runs. Regarding the last comment: We appreciate the reviewer's comment about the challenges of data scarcity in time series generation tasks. However, it's important to clarify that our research addresses a distinct problem: preserving extreme values in time series generation. This preservation of extreme values remains crucial regardless of whether the underlying context involves small-scale or large-scale data. We acknowledge the reviewer's suggestion to verify our method's performance on small-scale datasets. While this is indeed an interesting direction for further investigation, given the limited time remaining for authors' response, we were unable to conduct such an analysis at this time. Finally, as shown in the paper (figure 5), while current diffusion models can effectively capture the general pattern of a time series in the context of large scale data, they fail to capture the distribution of extreme values. Effective generative modeling of the extreme values is significant for developing robust risk management strategies and enhancing disaster preparedness measures, as noted in the introduction.	Summary: The article introduces FIDE (Frequency Inflated Diffusion Estimation), which is geared towards better capturing extreme values when generating time series through diffusion models, which the authors stress as crucial in domains like climate science and disaster preparedness. The approach involves inflating high-frequency components and conditionally generating samples based on block maxima, integrating the GEV distribution to ensure fidelity in extreme event representation. The authors run experiments comparing their approach to GANs and VAEs across a diverse set of datasets (AR1, Stock, Energy, Temperature, ECG). They compare both the overall data distribution and extreme values and find their approach to show promising performance. Strengths: The problem statement and proposed solution of the paper are clearly laid out and mathematically described. The authors also propose a meaningful experiment setup to empirically test their approach against sensible baselines. Weaknesses: It is not clear to me from the presented results whether there is a trade-off between being able to capture the overall distribution of values well and capturing extreme values of the distribution. The experiments seem to suggests that that might be the case, but it could partially just be a matter of picking a different loss-function during selection of the hyperparameter for the GEV loss. It would be good to get more clarity on this aspect. Technical Quality: 3 Clarity: 3 Questions for Authors: What would be a model or situation where Assumption 1 is violated? I.e. the statement is that block maxima are related to high frequency components in _many_ real-world time series. Are there any notable exceptions and would one need a completely different modeling approach to capture those, or would your method still allow one to sensibly generate those time series even though the Assumption is violated? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper adequately addresses its limitations and does not appear to have negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Questions: We thank the reviewer for the question. While Assumption 1 holds for numerous real-world time series, it may not be applicable to slowly varying time series without abrupt changes, where block maxima build up smoothly. In such instances, a customized diffusion model like ours may not be necessary, as the block maxima should dissipate at a rate similar to other values during the noise addition process. To demonstrate this, we modified the example time series shown in Figure 3 of the paper by applying a moving average to smooth the data. We then applied the forward pass of diffusion model to the smoothed time series. As illustrated in the attached Figure 1 (see the attached PDF in the general rebuttal), both the block maxima and other (non-maxima) values of the smoothed time series dissipated at a similar rate, making Assumption 1 no longer valid. In this situation, the slow evolution of the block maxima is part of the general pattern of the time series. Thus, existing diffusion models should be sufficient to capture the block maxima without the need for our model. Response to Weakness: Indeed, our results reveal a trade-off between accurately capturing the overall distribution and effectively modeling the extreme values. This is evident in Figure 5 which shows that our proposed method excels in capturing the block maxima distribution while slightly underperforms in capturing the distribution of all values. To further validate this observation, we conducted an experiment using temperature data, evaluating the trade-off with different inflation weights (1.0, 1.15, 1.3) applied to high-frequency components. An inflation weight of 1.0 indicates no inflation, while 1.3 denotes inflating the high-frequency components by a factor of 1.3. We observed that decreasing the inflation weight from 1.3 to 1.15 and then to 1.0 generally resulted in decreased performance in capturing the block maxima distribution while improving the performance in capturing the overall distribution. Our ablation study (Table 2, Appendix) further supports this finding, showing that removing the GEV loss leads to lower performance compared to our proposed framework. In short, these experiments demonstrate that the task of balancing the trade-off between capturing the overall distribution against the extreme values can be achieved by selecting the appropriate inflation weights (through cross-validation) and incorporating a GEV loss into the objective function.	Summary: The paper presents a novel generative model designed to better capture extreme values in time series data. FIDE introduces a high-frequency inflation strategy to prevent the loss of extreme values, integrates conditional diffusion modeling to condition on block maxima, and incorporates the Generalized Extreme Value (GEV) distribution to ensure the accuracy of extreme value representation. Empirical results show that FIDE outperforms existing methods in maintaining the distribution of extreme events across various datasets, making it a practical application in the generative modeling of time series data. Strengths: 1. Maintaining extreme value in generated time series data is an important but challenging problem. The proposed method can potentially satisfy this urgent need in real applications. 2. The problem is well motivated in Section 3, and the proposed method is reasonable and explained clearly. 3. The experimental results demonstrate that the proposed method can indeed preserve statistical information of extreme values. Weaknesses: 1. Although inflating high-frequency signals is intuitive, the possible influence on the fidelity of generated time series is not discussed. 2. The compared baselines, especially diffusion-based approaches, are not enough. For example, there are several works developing diffusion models for general time series generation (Ref-1), or domain-specific time series generation (Ref-2,3). Also, I think those diffusion-based probabilistic time series forcasting approaches can also be used for generation. Ref-1 Yuan, X., & Qiao, Y. (2024). Diffusion-ts: Interpretable diffusion for general time series generation. ICLR24 Ref-2 Kong, Z., Ping, W., Huang, J., Zhao, K., & Catanzaro, B. (2020). Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761. Ref-3 Zhou, Z., Ding, J., Liu, Y., Jin, D., & Li, Y. (2023, November). Towards generative modeling of urban flow through knowledge-enhanced denoising diffusion. Sigspatial23. Technical Quality: 2 Clarity: 2 Questions for Authors: Please answer my listed weaknesses above. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: None. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Response to Question 1: We appreciate the reviewer's concern regarding the fidelity of generated time series. We have investigated this issue in our experiments. First, Figure 5 demonstrates the tradeoff between preserving the overall distribution and the distribution of block maxima. The results suggest that our model excels in capturing block maxima, with minimal alteration to the overall series. In contrast, the baseline DDPM approach, which is capable of replicating the overall distribution, significantly underestimates the distribution of block maxima values. The analysis and discussion provided in Lines 257-261 highlights this fidelity aspect. While we did not provide the full quantitative results (similar to Table 1) for the entire time series due to space constraints, we have included them in the attached Table A2 (provided in the general rebuttal PDF) and will include them in the appendix of the revised paper. Finally, the predictive score metric shown in Table 1 also shows the fidelity of the generated time series by utilizing them for downstream time series forecasting tasks. Following a similar approach as in (Yoon et al, 2019), the downstream prediction task here corresponds to a multi-step time series forecasting task, which includes both predicting the block maxima and non-maxima values in a forecast window. The results suggest that our method achieves the best performance on three datasets and the second-best on two others, underscoring the model’s ability to accurately reproduce the temporal properties of the time series. Response to Question 2: We acknowledge the reviewer's valid point regarding the additional baseline methods. While our initial selection encompassed various generative model types, we concur that incorporating additional diffusion-based baselines is warranted, given our proposed method's foundation in diffusion models. In response to this insightful suggestion, we have expanded our comparative analysis to include the recent Diffusion-TS model, as recommended by the reviewer. The results of this extended comparison are summarized in attached Table A1 (provided in the general rebuttal PDF). Our findings conclusively demonstrate that our proposed method continues to outperform all baseline methods, including these new additions. Notably, Diffusion-TS emerges as the second-best baseline in most scenarios, underscoring the efficacy of diffusion-based approaches in this domain.	Rebuttal 1: Rebuttal: Thank you all for your careful and valuable suggestions. In response to your insightful feedback, we have expanded our comparative analysis to include the latest Diffusion-TS model and two additional flow-based baselines (Fourier-Flows and RealNVP) as recommended by the reviewers. This ensures a more comprehensive evaluation. We have also addressed your other suggestions and clarified the points you raised in the individual rebuttal. Please find the attached PDF that contains 1 figure and 2 tables to address reviewers' comments. Pdf: /pdf/8954a90ca8a24338605d0744931406f7d07cf2d1.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning	Accept (poster)	Summary: This paper addresses the Exemplar-Free Class Incremental Learning (EFCIL) challenge. It identifies two critical issues that undermine the effectiveness of existing methodologies and proposes a novel approach, AdaGauss, which adapts covariance matrices from task to task and mitigates task-recency bias. Strengths: 1. The paper is generally well-written and well motivated. It clearly demonstrates that the changes of covariance matrices also matters in Exemplar-Free Class Incremental Learning. 2. The proposed method demonstrates pretty good experimental results on all five datasets whether training from scratch or starting from a pre-trained backbone. 3. The proposed method is straightforward and easy to understand. Weaknesses: 1. I appreciate the efforts the authors have devoted to detailing the three observations encountered in Exemplar-Free Class Incremental Learning. However, observations 2 and 3 appear quite similar to me, as both seem to represent a simplification of the representation for previously seen classes. Additionally, the analysis of observations 2 and 3 does not significantly depart from existing explanations for why classification results tend to skew towards recent tasks, a phenomenon (task-recency bias / representation forgetting) already well-documented in the broader field of continual learning. 2. The proposed idea that both the mean and covariance should be adapted during training shares similarities with test-time adaptation methods. Therefore, some comparisons are necessary to delineate these relationships further. 3. Although it is quite straightforward that encouraging the feature extractor to produce features with linearly independent dimensions can mitigate dimensionality collapse, this approach does not guarantee the production of meaningful features. Additionally, it remains unclear whether simply optimizing the covariance matrices of features from one mini-batch can ensure linear independence. Some theoretical analysis would better explain the effectiveness of the proposed loss term $L_{AC}$ in mitigating dimensionality collapse (which I think there do exist some possible results to derive). Technical Quality: 2 Clarity: 3 Questions for Authors: See the weakness. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the feedback provided by the Reviewer. We will now address the specific weaknesses indicated: W1: Novelty of explanation of the task-recency bias We agree that task recency bias was well explored in CIL literature. However, in most works [1-4], the focus is on the bias in the linear classifier, which CIL methods try to unskew. On the contrary, in our work, we explain the bias in the embedding space - before the linear classifier and a novel method to prevent it ($L_{AC}$). Observations 2 and 3 show that ranks of covariance matrices increase in the later tasks, leading inverses of covariance matrices to have higher norms. That, in fact, generates task-recency bias in methods that utilize Bayes classifier or sample from class distribution in the embedding space. To the best of our knowledge, we are first to explain this and to provide a simple solution in the form of $L_{AC}$ loss. Less elegant solutions involve techniques like shrinking covariance matrices (used in FeCAM and EFC) as they introduce more hyperparameters. W2: Comparison to TTA methods Some TTA approaches, e.g., CAFA [5] or TTAC [6] align the means and covariances at test-time, with those computed during pretraining. However, there exist notable differences compared to the EFCIL setting considered in our work. In TTA, data of all the classes are available during the adaptation phase (it is only the features that change, like in domain incremental learning), whereas in EFCIL, the model is learning new classes in new tasks, and therefore, a class-level feature alignment cannot be applied (as there is no access to previous task classes). However, it is true that there are some similarities (e.g., the feature drift that occurs in TTA due to changes in domains), and we will add references and discussion to TTA works in our work. W3.1: Impact of $L_{AC}$ on meaningfullness of features Our loss function consists of three components, which are of a different optimization goal, and a sweet spot between them must be found. As shown in Eq. 5 and Fig.11 in the Appendix, we analyzed the $\beta$ hyperparameter, which sets the strength for the covariance regularization. As the Reviewer noticed, increasing the $\beta$ hyper-parameter increases the value of the cross-entropy loss, meaning that features produced by the network are less meaningful (from the classification perspective). Therefore, a good $\beta$ must be found during hyperparameter search. In all of our experiments, we utilized $\beta=1$. We have made the tradeoff between loss functions clearer in the new revision. W3.2: Explaining linear independence of features We verified the linear dependency problem experimentally. The value of $L_{AC}$ loss remains at -1 after each epoch in our experiments, as presented in the Appendix, Fig.11 ($\beta$=1). That means that for each minibatch the values of the diagonal of its Cholesky decomposition are greater than 1, and thus are positive. The latter implies that the covariance matrix of each minibatch is full rank [7]. Thus, the features are linearly independent. We agree that optimizing $L_{AC}$ at the mini-batch level is a stochastic problem like SGD, and it also requires that each mini-batch should be representative of the train dataset distribution. However, it results in the overall full-rank covariance matrices (what is presented in Figure 6 and 7 in the main paper). In our work, we use a mini-batch size that is four times bigger than the latent dimension to perform Cholesky decomposition efficiently. ---- We hope our response alleviates any concerns the Reviewer may have. However, if there are any remaining uncertainties, kindly specify references and additional questions, and we can further discuss them. Otherwise, we would appreciate if the Reviewer reconsidered improving the final score of our submission. ---- [1] Wu, Yue, et al. "Large scale incremental learning." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019. [2] Zhou, Da-Wei, et al. "Deep class-incremental learning: A survey." arXiv preprint arXiv:2302.03648 (2023). [3] Castro, Francisco M., et al. "End-to-end incremental learning." Proceedings of the European conference on computer vision (ECCV). 2018. [4] Hou, Saihui, et al. "Learning a unified classifier incrementally via rebalancing." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019. [5] Su, Yongyi, Xun Xu, and Kui Jia. "Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering." NeurIPS 2022. [6] Jung, Sanghun, et al. "Cafa: Class-aware feature alignment for test-time adaptation." ICCV 2023. [7] Horn, Roger A.; Johnson, Charles R. (1985). Matrix Analysis. Cambridge University Press. ISBN 0-521-38632-2 --- Rebuttal Comment 1.1: Title: Response Comment: I appreciate the authors taking the time to respond and I tend to keep my score. --- Rebuttal 2: Comment: Thank you for your response! We think we have addressed most of your concerns. We are eager to assist if there's anything else we can do to improve your score. Reviewer xJqZ has changed the score based on the valuable discussion!	Summary: Existing methods use Gaussian distributions to represent classes in the feature extractor's latent space, but face unchanged covariance matrices and task-recency bias. This paper introduces AdaGauss, an approach that adapts covariance matrices and mitigates the bias through an anti-collapse loss function. AdaGauss achieves top performance on EFCIL benchmarks and datasets, whether training from scratch or using a pre-trained backbone. Strengths: 1. The paper presents innovative approaches to tackle the EFCIL problem. 2. The writing is quite good, and easy to follow overall. 3. he paper includes comprehensive experimental findings to support its claims. Weaknesses: 1. Literature is incomplete. The paper concentrates on the EFCIL, yet quite a number of important EFCIL techniques [1-3] are not presented, and I want to see experimental comparisons with [1,3] if possible. [1]Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Renchunzi Xie, Kar-Ann Toh, Zhiping Lin”ACIL: Analytic Class-Incremental Learning with Absolute Memorization and Privacy Protection”, Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS) 2022. [2] Ma, C.; Ji, Z.; Huang, Z.; Shen, Y.; Gao,M.; and Xu, J. 2023. Progressive Voronoi Diagram Subdivision Enables Accurate Data-free Class-Incremental Learning. In The Eleventh International Conference on Learning Representations. [3] Zhuang H, He R, Tong K, et al. DS-AL: A dual-stream analytic learning for exemplar-free class-incremental learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(15): 17237-17244. 2. The claim "dimensionality collapse" concept is very important, but the authors didn't explain anything related. 3. An adaptor is needed for implementing the proposed algorithm. Could you provide a comparison among the proposed methods regarding the auxillary networks imposed? Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: n.a Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the Reviewer for providing important references, constructive feedback, and insightful comments. Below we respond to the weaknesses mentioned. W1: Incomplete literature and comparison to ACIL and DS-AL Thank you for providing relevant literature - we have added it to the Related Works of the paper. Thanks to the good quality of the code in [1, 3] we have compared our method to ACIL and DS-AL. Please note that [1, 2, 3] methods utilize half-dataset warm-start settings in the original papers. Firstly, we ran [1, 3] in the equal task setting using original implementations. The results ($A_{last}$ \| $A_{inc}$) are below: \| \| CIFAR100 \| CIFAR100 \| ImageNetSubset \| ImageNetSubset \| \| -------- \| :--------: \| :--------: \| :--------: \| :--------: \| \| \| T = 10 \| T = 20 \| T = 10 \| T = 20 \| \| ACIL[1] \| 38.8 \\| 53.2 \| 30.8 \\| 42.7 \| 44.2 \\| 54.8 \| 35.3 \\| 47.6 \| \| DS-AL[3] \| 40.8 \\| 54.9 \| 31.7 \\| 43.2 \| 46.8 \\| 58.6 \| 36.7 \\| 48.5 \| \| AdaGauss \| 46.1 \\| 60.2 \| 37.8 \\| 52.4 \| 51.1 \\| 65.0 \| 42.6 \\| 57.4 \| Next, we have run AdaGauss with the half dataset in the first task setting: \| \| CIFAR100 \| CIFAR100 \| ImageNetSubset \| ImageNetSubset \| \| -------- \| :--------: \| :--------: \| :--------: \| :--------: \| \| \| T = 5 \| T = 10 \| T = 5 \| T = 10 \| \| ACIL[1] \| 57.8 \\| 66.3 \| 57.7 \\| 66.0 \| 67.0 \\| 74.8 \| 67.2 \\| 74.6 \| \| DS-AL[3] \| 61.4 \\| 68.4 \| 61.4 \\| 68.4 \| 68.0 \\| 75.2 \| 67.7 \\| 75.1 \| \| AdaGauss \| 58.9 \\| 65.7 \| 55.4 \\| 63.7 \| 66.8 \\| 74.1 \| 62.8 \\| 68.0\| We can see that AdaGauss performs better in the equal task scenario. However, in the half-dataset scenario, ACIL and DS-AL are better. In our opinion, the half-dataset setting prefers methods with a frozen feature extractor (FE) after the first task [1, 2, 3]. However, in this paper, we focus on explaining the task recency bias when training the FE in incremental steps. We also did not have enough time to perform a hyperparameter search. We have added these results and the discussion to the Appendix. W2: Explaining the dimensionality collapse We tried to explain this concept in line 73: a large fraction of features' variance is described only by a small fraction of their dimensions. This is also reflected by the analysis done in Section 3.2, line 109, where we presented that the rank of the class covariance matrix is far lower than its dimensionality. That is the result of training the feature extractor with cross-entropy that forces the representational collapse of the latent representation to the number of classes that need to be linearly separable [5]. However, we agree that this can still be improved. We have done that in the introduction in our work and included more related work. W3: Different adaptor architectures We have included results ($A_{last}$ \| $A_{inc}$) for different types of adaptors in the appendix and below. $d$ denotes number of times the hidden layer is bigger than the input and output layers. We train for 10 equal tasks. \| \| CIFAR100 \| ImageNetSubset \| \| :--------: \| :--------: \| :--------: \| \| SDC[4] \| 43.7 $\pm$ 0.6 \| 46.7 $\pm$ 0.8 \| \| Linear \| 42.3 $\pm$ 0.6 \| 45.5 $\pm$ 0.7 \| \| MLP (2-layers, d=4) \| 45.7 $\pm$ 0.8 \| 50.4 $\pm$ 0.8 \| \| MLP (2-layers, d=16) \| 46.1 $\pm$ 0.8 \| 51.1 $\pm$ 1.1 \| \| MLP (3-layers, d=4) \| 44.6 $\pm$ 0.7 \| 49.9 $\pm$ 1.0 \| \| MLP (3-layers, d=16) \| 45.1 $\pm$ 0.8 \| 50.1 $\pm$ 0.8 \| We can see that utilizing non-linear MLP networks yields better results than linear network and SDC methods. 2-layer MLP networks are also preferred over the 3-layer ones. --- If we have adequately addressed the Reviewer's concerns, we kindly ask for your support and slightly improving the score. If you have any further concerns or additional points to raise, we are eager to address them. Your insights are valuable in enhancing the quality and impact of our research. --- [1] Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Renchunzi Xie, Kar-Ann Toh, Zhiping Lin ”ACIL: Analytic Class-Incremental Learning with Absolute Memorization and Privacy Protection”, Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS) 2022. [2] Ma, C.; Ji, Z.; Huang, Z.; Shen, Y.; Gao,M.; and Xu, J. 2023. Progressive Voronoi Diagram Subdivision Enables Accurate Data-free Class-Incremental Learning. In The Eleventh International Conference on Learning Representations. [3] Zhuang H, He R, Tong K, et al. DS-AL: A dual-stream analytic learning for exemplar-free class-incremental learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(15): 17237-17244. [4] Yu, Lu, et al. "Semantic drift compensation for class-incremental learning." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020. [5] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 2020. --- Rebuttal Comment 1.1: Title: Thank you for the response Comment: Thank you for the response. The rebuttal has addressed my concerns and I will keep my rating. --- Rebuttal 2: Comment: We are glad to hear that our response addressed all of your concerns! We are here to assist if there's anything else we can do to improve your score. Please note, that Reviewer xJqZ has just modified the score based on the valuable discussion!	Summary: The paper addresses the problem of Exemplar-Free Class Incremental Learning (EFCIL), which involves training a model on sequential tasks without access to past data. Current methods represent classes as Gaussian distributions in the feature space, enabling Bayes classification or pseudo-feature replay for classifier training. However, these methods face issues such as the need to adapt covariance matrices after each task and susceptibility to task-recency bias due to dimensionality collapse. The authors propose AdaGauss, a novel method that adapts covariance matrices between tasks and incorporates an anti-collapse loss function to mitigate task-recency bias. AdaGauss achieves state-of-the-art performance on popular EFCIL benchmarks and datasets, whether training from scratch or using a pre-trained backbone. Strengths: Overall this is a good paper with the following strengths: * This paper provides a clear experimental analysis and elaboration of the motivation for the proposed method. * For EFCIL, it is important and interesting to track changes in the distribution of past classes. * The proposed method achieves good results in multiple settings. Weaknesses: * The method of knowledge distillation through projectors proposed in this paper has been widely studied and discussed [1, 2]. Different from the general field of knowledge distillation, what are the technical innovations of the proposed method? Does it have unique insights for continual learning tasks? * The proposed method introduces several additional structures and requires a large number of samples from the simulated distribution to be trained, so the number of model parameters and the required training time should be discussed. [1] Chen Y, Wang S, Liu J, et al. Improved feature distillation via projector ensemble. NeurIPS 2022. [2] Miles R, Mikolajczyk K. Understanding the role of the projector in knowledge distillation. AAAI 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: I have some questions mainly about the experiments: 1. The experiments in this paper use the "learn from scratch" setting. Another common setting used in CIL is the "learn from half", which is also a dominant setting in EFCIL. It simply means that half of the classes in the dataset are learned as an initial task, and then the remaining classes are evenly divided into subsequent incremental learning tasks. This experimental setting is used in most of the compared works, such as PASS, SSRE, FeTrIL and FeCAM. How does AdaGauss perform in this setting? 2. Although the authors provide a detailed ablation analysis in Table 3, I still have some concerns. Why is $L_{AC}$ replaced by "utilized covariance matrix shrinking with the value of 0.5" in the sixth row of table 3, instead of just removing it? Are there other components of the proposed method that cannot be run in the absence of $L_{AC}$? 3. Does the adaptive tuning for Gaussian distributions fail if there is no $L_{AC}$ to resist dimensionality collapse? 4. Additionally, it would be beneficial if the authors could include a row comparing the performance of AdaGauss without both "Adapt mean", "Adapt covariance" and $L_{AC}$. This additional row would contribute to a clearer comparison and enhance the transparency of the evaluation. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: It is hard to foresee any potential negative societal impact of this theoretical work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We express our gratitude to the Reviewer for the feedback and insightful remarks. We shall begin by addressing the specific weaknesses pointed out: W1: Innovation of knowledge distillation through a projector In this work, we focus on knowledge distillation in EFCIL, which is different than [1, 2]. KD with the projector for EFCIL is only one part of the proposed solution. Together with adaptation step and anti-collapse loss the method receives the best results (Tab. 3). Additionally, our unique insight about KD is explaining how it impacts the task-recency bias in EFCIL. As presented in Section 3.2, KD through a learnable projector is less susceptible to bias because it provides stronger representations. That, combined with $L_{AC}$, allows us to overcome the bias and achieve good results. We have made that clearer in the new version of the manuscript. W2: Number of parameters and time complexity The number of model parameters is discussed in Section 4.2: Memory requirements. After the training, the projectors for KD and adaptation are removed and not stored. Therefore, the number of network parameters is not increased. We aggregated logs from our experiments on CIFAR100 split into 10 equal tasks to compare the training and inference time. We utilized original implementations of baseline methods and utilized Resnet18 trained for 200 epochs with batch size of 128 on a single machine. We used NVIDIA GeForce RTX 4060 and AMD Ryzen 5 5600X CPU for the experiments; we repeated them 5 times. We use a linear head classifier for FeTrIL and also test a version of AdaGauss with a linear head classifier instead of a Bayes classifier - we sample from memorized distributions to train it. We present results in Tab.2 of the pdf file. Methods that freeze the backbone after the first task (FeTrIL and FeCAM) have lower training time than others. AdaGauss takes less time to train than EFC, as it does not require training the linear head. The inference time of methods that utilize the Bayes classifier (FeCAM, AdaGauss) is higher than methods that utilize linear classification heads (LwF, FeTrIL, EFC). Replacing the Bayes classifier with a linear head in AdaGauss boosts its inference time by 7.5 seconds. We have added these results and discussion to the Appendix. Q1: Learn from half dataset Learning from scratch is more challenging than half-dataset setting as it requires to incrementally train feature extractor [3], not just the classifier. However, using the pre-trained model (or learning from half) can be considered a more practical and real-life setting. Thus, we evaluated our method with a pre-trained model in Table 2. However, we have additionally compared our method to the mentioned baselines in a half-dataset setting using the original implementations under the same data augmentations as AdaGauss. Please note that we did not have enough time to perform hyperparameter search for our method - we utilized these from the equal task setting, whereas the results for other methods were optimized by their authors. We provide results in the form ($A_{last}$ \| $A_{inc}$) in the Tab. 4 of the rebuttal pdf file. AdaGauss performs better than PASS, SSRE, and FeTrIL (5 tasks) in the half-dataset setting. However, it is slightly worse than most recent baselines when using default hyperparameters. FeCAM, ACIL, and DS-AL freeze the feature extractor after the initial task, which can explain their good results in the half-dataset training. We have added these results to the Appendix. Q2.1: Why is $L_{AC}$ replaced by shrinking in Tab. 3? Suppose we just remove the $L_{AC}$. In that case, covariance matrices of classes become singular, and it is mathematically impossible to invert them, leading to the inability to perform Bayes classification or to sample from these distributions (required to adapt distributions). Therefore, we utilized the lowest possible shrink value to invert covariance matrices (in this case 0.5; we performed a grid search to obtain it from values 0.01, 0.05, 0.1, 0.5, ...). Shrinkage is a standard technique used in FeCAM[4] and EFC[3] to alleviate the problem of singularity and matrix inversion. Q2.2: Are there other components of the method that cannot be run in the absence of $L_{AC}$ Without $L_{AC}$ we cannot perform classification and adaptaing Gaussians. Both of these require the inverse of class covariance matrices, and they cannot be calculated when covariance matrices are singular. $L_{AC}$ prevents that. We have improved the writing to emphasize this important aspect of the AdaGauss method and made it more clear in the ablation study. Q3: Does the adaptive tuning for Gaussian distributions fail if there is no $L_{AC}$ Yes, without $L_{AC}$, class covariance matrices are singular, and we cannot invert them. Therefore, we cannot sample from class distributions (step 16. in Alg. 1). Q4: Providing additional row to the ablation study We have added this row and the discussion to the new version of the Appendix. We also present the results ($A_{last}$ \| $A_{inc}$) in Tab.5 of the pdf file. With all of these components turned off, AdaGauss achieves poor results compared to the baseline. ---- If the Reviewer's concerns have been sufficiently addressed in our responses, we humbly seek their support for the paper and improving the score. ---- [1] Chen Y, Wang S, Liu J, et al. Improved feature distillation via projector ensemble. NeurIPS 2022. [2] Miles R, Mikolajczyk K. Understanding the role of the projector in knowledge distillation. AAAI 2024. [3] Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning. ICLR 2024 [4] Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. NeurIPS 2024 --- Rebuttal 2: Title: Response Comment: I appreciate the authors' response. This rebuttal addresses most of my concerns. Although AdaGauss does not achieve SOTA performance in the "Learn from half" setting, solving EFCIL with an adaptable Gaussian distribution sounds rational. I think adjusting the loss weight of knowledge distillation might be helpful to improve performance in the "Learn from half" setting. Taking into account the rebuttal and other reviewers' comments, I tend to keep my rating. --- Rebuttal 3: Comment: Thank you for your response! We are glad that we have addressed most of your concerns. If there is any issue we can resolve to improve your score, we are eager to do it. Please note that the Reviewer xJqZ changed the score based on the discussion!	Summary: This paper analyzes the impact of dimensionality collapse in EFCIL and examines the distribution shift of the mean and covariance matrix. Based on these findings, the paper proposes the AdaGauss method, which adapts the covariance and mean, and designs a loss term to prevent the dimensionality collapse of the feature extractor. Strengths: - The analysis of the dimensionality collapse of the feature extractor is interesting. - The paper provides a comprehensive experimental analysis, which clearly demonstrates the effectiveness of different components within the proposed model. Weaknesses: - The writing in the current version needs improvement. There are many parts that I had to reread several times to understand. For example, the three observations in Lines 89-118, I strongly recommend to summarize them in one sentence initially to highlight the core findings. In Section 3.3.2, which is one of the most important parts and contributions of the paper, the explanation is insufficient even though I refer to the appendix. What is the meaning of $a_i$ ? There is no clear definition. - Regarding Figure 1, is it merely an illustrative figure, or is it a visualization figure? If it is a visualization figure, what are the details used to depict this figure? - In Section 3.3.4, the learned adaptation network is trained on the t-th task. Why can it be used for previous μ and σ? The underlying assumption is that all previous tasks share the same shift pattern. Is this a reasonable assumption? - For Algorithm 1, I am not sure how it addresses the Batch Norm statistics, which are crucial to continual learning and the covariance shift this paper focuses on [1,2]. - The time complexity of the proposed method compared to other methods should be discussed. [1] Continual Normalization: Rethinking Batch Normalization for Online Continual Learning, ICLR 2022. [2] Diagnosing batch normalization in class incremental learning, 2022. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Please address the questions in the weakness. 2. Can the proposed method be applied to other network structures such as 4-layer convolutional networks or ViT? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the Reviewer for the insightful comments and providing relevant works. We begin by responding to the weaknesses: W1: Writing improvements We have revisited the whole Section 3 and applied the Reviewer's suggestions in the new version of the manuscript. W1: Meaning of $a_{i}$ The definition of $a_i$ is provided in line 146: More precisely, let S be the size of the feature vectors and $a_i$ be the i-th element of the diagonal of a Cholesky decomposition of minibatch’s covariance matrix. We have made it clearer in the new version. W2: Regarding Fig. 1 Fig. 1 is an illustration that provides intuition into the problem of mean and covariance adaptation; it is not a visualization. However, the provided accuracy after the last task and KL divergence (between memorized and real class distributions) were measured in experiments on the ImageNetSubset dataset split into ten tasks. Such illustrations are utilized in a variety of works, e.g. [1] (Fig. 1), [2] (Fig. 1), [3] (Fig. 3), [4] (Fig. 1c). W3: Assumption that all tasks share the same shift pattern In EFCIL, we do not have access to old data. Thus, we learn the non-linear adaptation network with current task data and use its generalization for the old classes' distribution (μ and $\Sigma$). We do not assume that the shift pattern for all the previous tasks is the same. We train an additional non-linear network to learn these patterns, assuming that we have only access to the current data. That is the main assumption. The learned adapter network estimates the shift differently for each task and its classes, as we visualize in Fig. 1 of the rebuttal file (classes 0, 1 are from 1st task, 10 and 11 - 2nd). On the contrary, in the recent EFC [4], only the mean is adapted, and there is an assumption that the covariance does not shift (Fig. 5). W4: Regarding batch norm We agree that BN is an important aspect of CL. However, it is more severe for methods that do not perform shift compensation, which is reflected in the provided works [5, 6]. The fact that BN changes the shift of class representations is fine in AdaGauss, as we utilize the MLP adaptor to predict and compensate for this shift. Therefore, our method is agnostic to BN, and BN does not negatively impact AdaGauss. To prove it, we provide (Tab.1 in rebuttal pdf) results ($A_{last}$ \| $A_{inc}$) with frozen after the first task or removed BN layers from Resnet18. We can see that removing or freezing BN layers slightly decreases the performance of the method. Additionally, our method is architecture agnostic and can work with networks in which the BN is absent, i.e., ViT, which uses Layer Norm. We provide results with different backbones below in the response to Question 2. We have added a discussion about BN and mentioned RW [5, 6] to our manuscript. W5: Time complexity We have gathered cumulative inference and training time of our experiments on CIFAR100 split into ten tasks. We utilize the original implementation for each method and run experiments on a single machine with NVIDIA GeForce RTX 4060 and AMD Ryzen 5 5600X CPU. We repeat each experiment 5 times, train all methods for 200 epochs, use 4 workers and batch size equal to 128. We test vanilla AdaGauss, and AdaGauss where the Bayes classifier is replaced with a trained linear head, where the classifier is trained on samples from class distributions (mean and cov. matrix). We utilize FeTrIL version with linear classification head. We present results in Tab.2 of the rebuttal pdf file. The inference of our method takes a similar amount of time as in FeCAM, as the feature extraction step is followed by performing Bayes classification. The inference time of AdaGauss is slightly higher than that of methods with linear classification head (LwF, FeTrIL, AdaGauss with linear head) because Bayes classification requires an additional matrix multiplication when calculating the Mahalanobis distance. The training time of AdaGauss is longer than for LwF, EFC, FeCAM, and FeTriL as we do not freeze the backbone after the initial task and additionally train the auxiliary adaptation network. Still, AdaGauss takes less time to train than its main competitor - EFC, and is much faster than SSRE. Our method does not increase the number of networks' parameters because the distiller and the adapter are disposed after training steps. Q2: Can the proposed method be applied to other network structures? Yes, the method is architecture-agnostic and can be applied to different network structures and different training regimes. We do not assume any requirements towards backbone architecture. We will make it clearer in the paper, and we will include AdaGauss results for different backbones in the appendix, namely ViT small and ConvNext. These results ($A_{last}$ \| $A_{inc}$), presented in Tab.3 of the pdf file, are for EFCIL setting with 10 and 20 equal tasks and weights pretrained on ImageNet (as in Tab. 2). Using more modern feature extractors architectures further improves the results of AdaGauss. --- We hope that our explanation has addressed the Reviewer's concerns. Should there be any additional queries, we are willing to provide further details. If no further clarification is needed, we kindly ask the Reviewer to increase the final score. --- [1] Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. NeurIPS 2024. [2] Fetril: Feature translation for exemplar-free class-incremental learning. WACV 2023. [3] Semantic drift compensation for class-incremental learning. CVPR 2020. [4] Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning. ICLR 2024 [5] Continual Normalization: Rethinking Batch Normalization for Online Continual Learning, ICLR 2022. [6] Diagnosing batch normalization in class incremental learning, 2022. --- Rebuttal Comment 1.1: Title: Comment by Reviewer xJqZ Comment: I thank the authors for the efforts to address the concerns, especially the additional experiments conducted to verify the effectiveness of the proposed method with different backbone networks. The clarifications made by the authors and the results of the additional experiments addressed my concerns. I have increased my score to 6 accordingly. --- Rebuttal 2: Comment: Thank you for improving the score. Your review helped us improve the work a lot!	Rebuttal 1: Rebuttal: We want to express our gratitude to all the Reviewers and Chairs for their dedication and effort. The majority of Reviewers consider accepting the work, all agree on comprehensive experimental analysis, and three Reviewers highlight the motivation behind the method (5j6M, VMJn, 6bmT). We have thoroughly reviewed the feedback and addressed all the raised concerns, what significantly enhanced the quality of our submission. We have now prepared a revised version of our work and are ready to engage in further discussions. Based on the comments received, we have performed additional experiments, which are presented in the rebuttal pdf file, and made the following changes to the revised version: * Improved introduction and method section to better explain the problem of task recency bias caused by dimensionality collapse in EFCIL, and why it is different than popular recency bias in the classification head (xJqZ, 6bmT, VMJn). * Added analysis of training and inference time complexity of AdaGauss compared to other methods (xJqZ, 5j6M). * Tested our method in a half-dataset setting and added results to the Appendix (5j6M, 6bmT). * Added discussion of relevant works proposed by reviewers (6bmT, xJqZ). * To confirm that AdaGauss works with different backbone architectures, we have run experiments with ConvNext and ViT and added them to the Appendix (xJqZ). * Experimentally proved that the predicted distribution shift is different for different classes (xJqZ). * To understand the impact of batch norm (BN) layers on our method we have performed experiments with frozen and removed BN layers. We have added a discussion to the Appendix (xJqZ). * Provided AdaGauss results for different feature adaptation networks (6bmT). We thank the Reviewers once again and look forward to discussing any other aspects of the paper that require further clarification. -- authors Pdf: /pdf/8484cc77f404ae3d081a9965f13d4af8857c0954.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards Exact Gradient-based Training on Analog In-memory Computing	Accept (poster)	Summary: This is a theoretical paper about training analog systems based on resistive memories. The paper tackles specifically the problem of weight update asymmetry in such resistive devices. The paper develops a model for the weight dynamics under “Analog SGD”, taking into account weight update asymmetry, and it shows that Analog SGD does not converge, due to an asymptotic error. The paper then shows that the Tiki-Taka algorithm, developed in prior works for such analog devices, has better convergence properties. Tiki-Taka eliminates the asymptotic error and converges to a critical point. Simulations corroborate the theory. Strengths: The topic tackled by the paper is very important. Analog computing could lead to energy efficiency gains of several orders of magnitude for both inference and training, compared to digital neural networks on GPUs. Developing appropriate theoretical models of analog computing to better understand the behavior of these analog systems is of utmost importance. To my knowledge, this paper is the first attempt to study the convergence properties of analog SGD. The paper is clear and the study is thorough. Weaknesses: The theoretical analysis rests on a set of models and assumptions. Thus, the applicability of the conclusions rests on the validity of all these models/assumptions. In practice, I expect none of these assumptions to perfectly capture the real behavior of the system. For example, Figure 2 hints that the analog SGD dynamics (Eq3) better fits real behaviors than the digital SGD dynamics (Eq2), but the matching still seems to be far from perfect. This, of course, is not a problem, however, what is more problematic is that there does not seem to be any discussion in the paper about which assumptions/models are accurate and widely accepted, and which ones less accurately capture empirical data or are still debated by the community. The paper implicitly assumes that backpropagation (BP) is used for computing the weight gradients. However, a large (and growing) group of researchers in neuromorphic computing (or ”physical neural networks”) think that BP is not the best fit for analog devices, and they are exploring alternative training methods. See e.g. Ref [1] below for a very recent review of such algorithms. In particular, Ref [1] discusses several algorithms that extract or estimate the weight gradients in broad classes of analog systems. My understanding is that it wouldn’t take too much time and effort to extend the present study to include these other gradient-descent-based algorithms. In my view, including these other methods would add a lot of value to the paper, by broadening the potential impact of the work. Reference: [1] Momeni, Ali, et al. "Training of Physical Neural Networks." arXiv preprint arXiv:2406.03372 (2024). Minor remark. One claim in the introduction is not properly referenced. Specifically, where do the figures of $2.4 million and $4.6 million for training LLaMA and GPT-3 come from? The references provided are the LLaMA and GPT-3 papers, which do not provide these figures, if I am not mistaken. Technical Quality: 3 Clarity: 3 Questions for Authors: Is the model of Eq6 for weight update asymmetry a widely accepted model? Does it match empirical data? Similarly, are the forms of q+ and q- for ALD, written at lines 174-175, widely accepted and/or corroborated by data? Does the AIHWKit very faithfully simulate real analog devices? What are the strengths and limitations of this tool? The paper is concerned with analog resistive memories where the weights are implemented as conductances (which are positive, I suppose), but on Figure 6 (Appendix L), I see that the weight values can be negative. Lemma 1 also indicates that w_k can be either positive or negative. Could you clarify this point? I understand that the RHS of Eq.5 can be arbitrarily small by choosing alpha = sqrt(1/K) and letting K -> infinity. But, unless I am mistaken, this does not imply that the sequence grad f(W_k) converges to zero. Can we conclude from Eq5 that W_k converges? If not, why is Eq.5 important for studying the convergence of (digital) SGD? I am not sure to understand why Theorem 2 is important. To my understanding, the important result is Theorem 3, which shows that analog SGD does not converge, contrary to digital SGD (Eq.5). Could you clarify this point? Is the theory of analog SGD and Tiki-Taka provided in this paper limited to backpropagation? Or could it be used with other methods that compute weight gradients, e.g. Equilibrium Propagation and/or other algorithms presented in Ref [1] above? Similarly, could the AIHWKit be employed with these other training algorithms? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The applicability of the theoretical results depends on the validity of the underlying assumptions – see my comment above. To be clear, I am supportive of the methodology followed by the authors. However, given that the overall model will be at most as accurate as the least accurate assumption, it would be extremely useful to know which assumptions/models we can trust 100% (i.e. are well accepted), and which ones are still debated / less supported by experiments. This would help future works identify what are the main “bottlenecks” of the study, to investigate those in depth and perhaps further improve the model. For instance, can we really trust Eq3 as “the right model” of weight dynamics? Or should one merely think of it as a better model than Eq. 2? Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We really appreciate your recognizing the importance of our work and for helpful suggestions. Please find our point-by-point reply to your comments below. > W1. There is no discussion about which models are accurate and widely accepted. As the reviewer correctly pointed out, adding a discussion about the accuracy of the models are important. The applicability of our analysis relies on the following two analog hardware models: - Asymmetric update model (6): The update (6) omits some non-ideality, like cycle-to-cycle/device-to-device variation and analog-digital conversion error [12][27] - ALD response model (8): ALD models assume the response factor is a linear function, which is widely adopted in the literature [16,19,23]. It explains why the proposed dynamic does not match the real behaviors perfectly in Fig 2. We will clarify the potential error raised from the simplified model in the revision. We hope the discussion helps the readers better evaluate the scope of applicability of the theory. > W2. Including these other PNN training methods would add a lot of value, by broadening the potential impact of the work. Thank you for pointing out this interesting direction and suggesting a recent work. Indeed, this is inspiring! To establish the connection between our work and the PNN training, we will add a paragraph to briefly review the progress as follows in literature review section. > We are especially interested in AIMC hardware with resistive crossbar arrays. It is a specific implementation of physical neural network (PNN) [R1,R2]. PNN is a model implemented by a physical system involving tunable analog quantity. The quantity is adjusted to implement the learning on some specific tasks. Various hardware is capable of supporting PNN, such as holographic grating [R3], wave-based systems [R4], and photonic networks [R5], to name a few. [R1] Wright, et al. "Deep physical neural networks trained with backpropagation." [R2] Momeni, et al. "Training of Physical Neural Networks." [R3] Psaltis, et al. "Holography in artificial neural networks." [R4] Hughes, et al. "Wave physics as an analog recurrent neural network." [R5] Tait, et al. "Neuromorphic photonic networks using silicon photonic weight banks." > W3. Where do the figures of 4.6 million for training LLaMA and GPT-3 come from? Thanks for the careful reading! The cost is estimated by multiplying the required GPU hours by the AWS price per hour, where the GPU hours are reported in [1]. To avoid confusion, we will rephrase this sentence with the following one. > For example, it requires 184 thousand GPU hours to train an LLAMA2 7 billion model [R6], and this time increases to 1.7 million GPU hours for its 70 billion version [1]. [R6] Touvron, et al. "Llama 2: Open foundation and fine-tuned chat models." > Q1. Is the model of Eq6 for weight update asymmetry a widely accepted model? The answer is affirmative. For example, [16, 52] have already demonstrated different factors scale the update in different weight state. > Q2. Does the AIHWKit faithfully simulate real analog devices? AIHWKit is capable of providing real simulations across different granular levels, including IO noise, pulse update, update variation, response factor, and A/D discretization, to name a few. One of the limitations is the significant overhand of detailed simulation and inadequate multi-GPU parallel support, which make it time-consuming to conduct large-scale simulations. > Q3. Why can the weights be negative? This is just for the mathematical convenience. To implement the negative weight on analog devices, two resistive crossbar arrays, a main array, and a reference array are set. The weight is represented by the difference of the two conductances multiplied by a scaling factor, which can be negative. Thanks for pointing this out! We ignored this detail in our dynamic, but we will clarify it in the revision. > Q4. Can we conclude from Eq5 that $W_k$ converges? If not, why is Eq.5 important for studying the convergence of (digital) SGD? In the convergence study of (digital) SGD, there are typically two types of convergence: the convergence to the stationary point (e.g., $W^$ such that $\nabla f(W^)=0$) or the convergence to the optimal solution (e.g., $W^$ such that $W^\in \arg\min_W f(W)$). Without additional assumptions like convexity, the convergence to the stationary point might be the best we can hope for in the worst case. Therefore, with a set of properly choosing decreasing stepsizes, Eq5 has been commonly used as a metric to assess the convergence to the stationary point of SGD; see [20]. As a result, we also use the metric of Eq5 to demonstrate the convergence of analog algorithms. > Q5. Why is Thm 2 important? Compared to Thm 3, Thm 2 is important since it serves as another main component to reveal the performance limit of Analog SGD. Thm 3 claims that there exist a bad situation that Analog SGD has an asymptotic error but it does not ensure it is the worst case. Combining Thm 2 and 3, we claim that $4\sigma^2 S_K$ is the worst-case possible asympotic error, which provide a better insight. > Q6. Is the theory provided in this paper limited to backpropagation? Thanks for the inspiring question! We believe our theory can be adapted to other analog training algorithms since algorithms like equilibrium propagation (EP) adopt other methods to determine the update directions. To replace the gradient in our analysis with a given update direction, we can still study the convergence using similar techniques. We will seriously consider this as a future direction. > Q7. Could the AIHWKit be employed with these other training algorithms? The answer is affirmative. AIHWKit enables multiple levels of simulation granularity. To implement other algorithms like EP, one needs to modify the algorithm-level code. With the above clarifications, we hope that our responses addressed all your questions. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed replies to my questions!	Summary: The paper introduces the training on the analog in-memory accelerators with SGD. The traditional analog SGD algorithm suffers from inexact convergence due to asymmetric updates/gradient noise. The paper shows theoretical foundations for gradient-based training on analog devices. The authors propose Tiki-Taka, which reflects the device model to the Analog SGD. It shows better empirical performance, e.g., it can eliminate asymptotic errors and converge to critical points. They show the theoretical and empirical efficacy of the proposed approaches in overcoming training limitations on the analog devices. Strengths: - The paper is well organized and easy to follow. - Concrete theoretical foundings on the proposed algorithm. - Benchmark sets and the target network look reasonable, and the result also sound and look promising. - Showing cases on various analog devices like ReRAM, etc. Weaknesses: - It'd be great if authors present how computations (ops) can be realized on the analog devices. - Needs comparison to other analog SGD methods, like [16] or [22]. Technical Quality: 3 Clarity: 3 Questions for Authors: - Floating-point operations are not trivial on analog devices. How can they be implemented? Additionally, for SGD training, how are analog devices more efficient compared to digital devices? - How does it scale to large batch sizes? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for acknowledging the merits of our work. Our point-to-point response to your comments and suggestions follow next. > W1. It'd be great if authors present how computations (ops) can be realized on the analog devices. The workload of forward and backward computation can be separated into two categories: matrix-vector multiplication (MVM) operations and non-MVM operations. In AIM computation hardware, the MVM operations are implemented in analog in-memory arrays, while the other non-MVM operations are conducted in digital circuits. We illustrate in the attached PDF in Author Rebuttal on how the MVM is implemented by analog devices. In a $3\times 3$ resistive crossbar array, a matrix $W$ is represented by the conductance of resistor at the crosspoint, where the $(i,j)$-th element of $W$ is represented by the conductance of the $(i,j)$-th resistor. To conduct an MVM operation $z = Wx$, voltage $x_{j}$ is applied between $j$-th and $(j+1)$-th row, where the subscript $j$ is used to imply the $j$-th coordinate. By Ohm's law, the current is $I_{ij}=W_{ij}x_{j}$; and by Kirchhoff's law, the total current on the $i$-th column is $\sum_{j}I_{ij}=\sum_{j}W_{ij}x_{j}$. > W2. Needs comparison to other analog SGD methods, like [16] or [22]. Note that the goal of this paper is not to develop the state-of-the-art analog training algorithm, but to understand why the vanilla SGD training on analog device does not work as expected, and why some correction operations used in the heuristic algorithms such as Tiki-Take work, through the unified lens of asymmetric updates on analog devices. We hope this theoretical understanding can contribute to future developments of analog training algorithms, hardware and material. In this context, in the introduction, we have compared the studied algorithm Tiki-Taka (TT-v1) with TT-v2 [22] and TT-v3/v4 [16] (c.f. Sec 1.2). Both [16] and [22] are variants of Tiki-Taka which were proposed to deal with some practical issues in analog training such as the reading noise and non-perfect zero-shifting issues. Since our focus in this paper is on understanding the impact of asymmetric updates on analog devices, in the simulations parts, we omitted the comparison between our paper and [16][22]. Nevertheless, as requested by the reviewer, we have now compared these four methods in the same setting as that in Section 5.3. We train a CNN model under $\tau=0.7$ on MNIST dataset, whose results are listed below. TT-v2--v4 are always better than TT-v1. \|Digital SGD\|Analog SGD\| TT-v1 \| TT-v2 [22] \| TT-v3 [16] \| TT-v4 [16] \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|99.24%\|82.17%\|98.56%\|98.94%\|98.91%\|99.01% > Q1. How can floating-point (FP) be implemented by analog devices? This is a great question! The FP is implemented in the digital domain [12, 31]. In the MVM computations, the weights are represented by conductance, while the input and output are represented by current signals. Therefore, the FP is unnecessary for MVM operations. The other operations involving FP computation are non-MVM ones, which are conducted in digital domains. We will add a remark in the revision to clarify it. > Q2. How does it scale to large batch sizes? > This is an interesting point! Implementing mini-batch gradient computation on the analog devices are significantly different from that on the digital devices. To implement the large batch in AIM computation hardware, each gradient in a batch is computed sequentially and accumulated to $W_k$ (for Analog SGD) or $P_k$ (for Tiki-Taka). For example, we use batch size 8 in FCN/CNN training and 128 for Resnet training. Studying the impact of batch size on the convergence of analog training will be an interesting future direction! We hope the above detailed clarifications can fully resolve your concerns. --- Rebuttal 2: Comment: Thanks for the detailed response. It has clarified some of my confusion and improved my understanding on the paper. For W2, to clarify my comment, it'd be great if you can elaborate differences to existing algorithms.	Summary: This paper presents a theoretical framework for gradient-based training on analog devices. This work first identifies the non-convergence problem of Analog SGD, which stems from asymptotic errors due to asymmetric updates and gradient noise and then presents a convergence analysis of Tiki-Taka, demonstrating its ability to accurately converge to a critical point, thereby eliminating the asymptotic error. Strengths: -- The proof the convergence of Analog SGD and showing that noise and asymmetric updates together lead to its asymptotic error. -- Demonstration of Tiki-Taka algorithm precisely converging to the critical point by mitigating the drift caused by asymmetric bias and noise. -- Empirical simulations using both synthetic and real datasets to confirm the presence of asymptotic error in Analog SGD and to show that Tiki-Taka outperforms Analog SGD. Weaknesses: -- The experimental results were obtained on rather small datasets. It'd be great to include the results on ImageNet. -- The experiments were limited to one type of networks (i.e., resnet) on CIFAR dataset. How does it generalize on other networks such as Transformers, or mobile CNNs? Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the time spent in reviewing our paper and for the valuable comments. The weaknesses identified by the reviewer mainly focus on the datasets and the model archtecture used in the experimental results as follows. > - The experimental results were obtained on rather small datasets. It'd be great to include the results on ImageNet. > - The experiments were limited to one type of networks (i.e., resnet) on CIFAR dataset. How does it generalize on other networks such as Transformers, or mobile CNNs? Albeit the main focus of this paper is on the theoretical understanding of analog training, we have still conducted additional experiments during the limited rebuttal period. To demonstrate the efficiency of Tiki-Taka under various situations, we conducted more simulations on different datasets and model architectures, including MobileNetV2 and large/small MobileNetV3 on CIFAR10/CIRAR100 datasets. The results are listed as follows. Table 1: Training on CIFAR10 dataset \|\|Digital SGD\|Analog SGD\| Tiki-Taka \|:-:\|:-:\|:-:\|:-:\| \|MobileNetV2\|94.47\|93.88\|94.24 \|MobileNetV3-Small\|93.21\|92.47\|93.37 \|MobileNetV3-Large\|94.78\|94.02\|94.63 Table 2: Training on CIFAR100 dataset \|\|Digital SGD\|Analog SGD\| Tiki-Taka \|:-:\|:-:\|:-:\|:-:\| \|MobileNetV2\|79.24\|78.34\|78.75 \|MobileNetV3-Small\|76.41\|76.13\|76.45 \|MobileNetV3-Large\|79.62\|79.67\|80.05 The results show that Tiki-Taka always achieves better test accuracy than Analog SGD by about 0.5% in almost all cases, which is consistent with our conclusion. Due to the limited time and inadequate GPU resources, we can not perform simulations on a larger scale like training on ImageNet dataset during the rebuttal period.	null	null	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for their constructive comments. Comments from all the reviewers were really helpful, which we believe have been fully addressed in detail in our rebuttal. The attached PDF is the illustration to explain how analog devices implement MVM operations (see the response to Reviewer gG9b). We look forward to the rolling dicussion and further engagement with the reviewers and area chair(s)! Pdf: /pdf/226af5d3ae48490984dfe80b8abd658b578d0e40.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Minimizing UCB: a Better Local Search Strategy in Local Bayesian Optimization	Accept (poster)	Summary: The paper presents a novel method for local Bayesian Optimisation. Instead of relying on estimates of the gradient of the black-box functions like earlier methods, the algorithm proposed in the paper minimises an upper bound on the objective function. The paper shows a connection between the proposed method and gradient descent, whose update rule can also be thought of as minimising a certain upper bound on function. The authors study the convergence of the proposed algorithm theoretically and empirically, and show it can develop improvement over existing baselines. Additionally, the authors propose a variant of the algorithm with an improved exploration strategy. Strengths: The paper is generally well-written and I found the connection between MinUCB and Gradient Descent very appealing and nicely presented. Minimising UCB seems to be an interesting alternative to performing gradient steps, as it does not require the specification of the stepsize. Empirical results show that at least on a number of selected tasks, the algorithm delivers a non-negligible improvement in performance. Summing up the theoretical and empirical results of the paper, I believe their contribution is sufficient and I recommend accepting the paper. Weaknesses: 1. I belive Theorem 1 could use a proof sketch, currently just by reading the main body, it is impossible to understand where the result is coming from. 2. While authors acknowledge the existence of TuRBO and ARS, they do not empirically compare with them. It would be nice to see a comparison with those baselines, particularly with TuRBO, which is very popular. 3. In the related work, the authors mention the methods using additive models in high-dimensional Bayesian Optimisation but only cite relatively old papers (the newest one is from 2017). Since then, additive methods have seen much development. To make the section more up-to-date, it would be great if the authors would cite more recent work, such as [1] or [2]. [1] Ziomek, Juliusz Krzysztof, and Haitham Bou Ammar. "Are random decompositions all we need in high dimensional Bayesian optimisation?." International Conference on Machine Learning. PMLR, 2023. [2] Han, Eric, Ishank Arora, and Jonathan Scarlett. "High-dimensional Bayesian optimization via tree-structured additive models." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 9. 2021. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In the paper, the authors explain the UCB is most likely to be “small” around already queried points and thus conclude the algorithm is mostly local, however, potentially allows to make bigger steps if needed. I wonder what would happen if we explicitly constrained the algorithm to be local, e.g. via the trust region strategy of TuRBO? Would it be possible for authors to conduct a simple ablation study on at least one of the problems? Thus could help gauge whether the improvement delivered by the algorithm comes from better local optimisation or from occasionally switching to a more global search. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors explicitly state the lack of convergence guarantee for LA-MinUCB as a limitation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weakness 1: I belive Theorem 1 could use a proof sketch... Response: We apologize for not including a proof sketch in the paper. The core tool we use in the paper is actually to prove the Lipshitz properties of various functions in Gaussian processes, i.e. the mean function $\mu(x)$, and the standard derivation function $\sigma(x)$. However, the difficulty is that the $\sigma(x)$ is not Lipshitz continuous under some conditions. We develop some tools to build a similar property for $\sigma(x)$ with some controllable error. We use the lipshitz properties of these function to build the following inequality: $$ f(x_{t+1})\le\min_{x\in \mathcal{X}}\mu(x)+\beta_{t}\sigma(x)\le f(\hat{x}^{t+1})+e_{t}<f(x_{t})+e_{t}$$ where $\hat{x}^{t+1}=x_{t}-\eta_{t}\nabla \mu(x_{t})$, and stepsize $\eta_{t}$ decreases at a logarithmic rate with $t$, and error term $e_{t}$ will decrease to 0. $\hat{x}^{t+1}$ is actually the gradient descent point starting from $x_{t}$. This inequality will directly help to prove the final result, i.e. the gradient convergence result. Weakness 2: While authors acknowledge the existence of TurBO and ARS, they do not empirically compare with them. It would be nice to see a comparison with those baselines, particularly with TuRBO, which is very popular. Response: We are sorry that our paper did not compare with these methods. We have added comparison results with TurBO in the experiment, please refer to the Fig 1,2,3 in the attached PDF for details. Weakness 3: ...only cite relatively old papers ... Response: We apologize for not citing the latest works. Relevant content will be added in future versions of the paper. Question 1: ...I wonder what would happen if we explicitly constrained the algorithm to be local, e.g. via the trust region strategy of TuRBO? ... Response: Thank you for your comment. We have added this type of experiments in Fig 4 in the PDF attached in the global rebuttal. We apply the trust region constrain on MinUCB, and find that adding the trust domain may actually lead to a worse result. We can explain from two aspects. On the one hand, the changes in the trust region are not timely. At the beginning, the radius of the trust region is small, which limits the search range. However, when approaching local optimum points, minimizing UCB often does not exceed the range of the trust region. The second aspect is that MinUCB itself can also be explained from the perspective of trust region, which can be refered to the third point of global rebuttal. So overall, as our method adopts the idea of trust region, it behaves local and can achieve good results. --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal. I remain positive about the paper.	Summary: This paper proposes and analysis a new local Bayesian optimisation scheme. The main idea is to replace the update of the current solution by minimising the upper confidence bound of the GP estimate (in the minimisation setting). The strategy is shown to obtain similar convergence rates as previous works. In addition, a "lookahead" variant of the algorithm is introduced. Finally, the proposed strategies are evaluated on synthetic benchmarks and RL tasks, showing competitive performance. Strengths: Originality: The proposed algorithm is a variant of a prior approach (GIBO), which updates the iterate by minimising the UCB of the function instead if a local quadratic upper bound. The individual ideas are not necessarily novel (e.g. pessimistic estimates are quite commonly used in bandit/RL algorithms), but the combination with the local Bayesian optimisation scheme appears to be novel. Quality: * The obtained bounds are slightly worse than prior works, but on the upside do not require the knowledge of a Lipschitz bound. * The experiments show good performance on the synthetic benchmarks and RL tasks Clarity: * The algorithm and the motivation is clearly presented and overall the paper is well written, but would still benefit from polishing. However some details could be explained better as outlined below. Significance: Local Bayesian optimization schemes have been a very successful approach to zero-order (noisy) black box optimisation and the theoretical analysis of these algorithms is still lacking, although the ground work has been layed in the prior work by Wu et al (2024). The paper therefore contributes to this landscape and the results are relevant at least for the Bayesian optimisation community and may inform better optimisation algorithms more broadly; or encourage further improved bounds. Weaknesses: My main concerns are the following: * The minimization of the UCB score in the step update appears to be global in nature. First, the range of min /argmin operations is never specified, e.g. in Algorithm 2, eq (4), (6), etc) which is confusing to the reader. Second, the authors claim that the acquisition step leads a "local" update (e.g. in lines 186-187, or the claim in 215 is not formally shown). What does it mean to be "local"? How is this formally proved? * While I agree with the general picture in Figure 1, and even if this is somewhat reflecting the general case, I believe one can construct kernels/examples where this "local" property is violated and the algorithm technically jumps to a different part of the domain (e.g. for linear kernel with near-zero slope, where observation data causes a sign switch and the step update moves to a different vertex of the domain). * Third, how can we efficiently find the minimizer of the UCB? We know that in general the UCB is not a convex function so finding the argmin is as hard as in global Bayesian optimization, without additional algorithmic constraints. * The discussion of prior works is lacking several works, e.g. the LineBO paper [1,2] which also provides local convergence guarantees based on a gradient descent scheme, and was shown in combination with trust regions. * The experiments lack several baselines (e.g. Turbo or LineBO) * The introduction of the Lookahead strategy in 7 is a bit out of the blue and not well connected to the remaining paper; and does not come with theoretical guarantees. How does this related/connected to Algorithm 1? * In general, the paper should better highlight the technical contributions and challenges related to the main result, and more formal/precise definitions of the problem setting and results. Minor remarks: * The abstract uses acronyms that have not been introduced (BO, GIBO). * The claim in line 28 that dimension 20 is a critical limit for Bayesian optimisation is somewhat arbitrary, even in lower dimensions convergence can be slow if the function is non-smooth etc. * Equation (2) and the surrounding discussion does not make sense to me: The right-hand side of (2) does not depend on x, and can therefore not used as acquisition function or to minimise over x. * The type setting of equation (1) looks odd - try using `\langle` and `\rangle` for the inner product. * Although not wrong, the discussion of UCB in line 164 is lightly miss-leading as the classical UCB approach is used in the maximization setting, and the corresponding quantity in the context of this paper would be the lower confidence bound (LCB). * 233: gamma is not defined * 241: What is n ? * 187: "view" -> "viewed". [1] Kirschner, J., Mutny, M., Hiller, N., Ischebeck, R., & Krause, A. (2019, May). Adaptive and safe Bayesian optimization in high dimensions via one-dimensional subspaces. In International Conference on Machine Learning (pp. 3429-3438). PMLR. [2] Kirschner, J., Mutný, M., Krause, A., Coello de Portugal, J., Hiller, N., & Snuverink, J. (2022). Tuning particle accelerators with safety constraints using Bayesian optimization. Physical Review Accelerators and Beams, 25(6), 062802. Technical Quality: 2 Clarity: 2 Questions for Authors: Can you define what it means to be "local" and how the update step satisfies such a property? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The paper briefly discusses limitations, mainly the missing analysis of the lookahead approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback. The suggestions and questions are briefly responded below. Weakness 1 and Question 1: The minimization of the UCB score in the step update appears to be global in nature... The range of min /argmin operations is never specified... What does it mean to be "local"? Response: Thank you for your comment. In this paper, we would like to emphasize that the term 'local' mentioned does not refer to limiting optimization within a small range, such as Turbo, but rather to the algorithm exhibiting local behavior. Although the minimization is performed globally on UCB, according to our explanation in the first point of global rebuttal, the standard deviation term $\beta\sigma(x)$ actually behaves as an additional penalty on the searching, which will limit the search to be near the current data set. In our view, another main difference between global and local algorithm, is whether the algorithm attempts to conduct a global exploration. Local methods such as approximate gradient descent methods like GIBO or MPD, or other well-known Turbo methods, basically use greedy strategies to try to find a point with a lower function value than the previous iteration. They do not consider exploring the entire space to find possible global optimum. This strategy can ensure that each step has a descent and quickly converges to a local optimum point. This is because when we have a sequence $\\{f(x_1), ..., f(x_n), ...\\}$,Where $f(x_{i+1})<f(x_i)$, it can be mathematically proven that this sequence will eventually converge, and at the same time, $\\{x_i,i=1...\\}$ (or at least a subsequence of it) will also converge to a point $x^$. Under these algorithms, this $x^$ is likely to be a local optimum point or saddle point of the function. According to the first explanation of MinUCB in the global rebuttal, we can assume that MinUCB also uses this greedy strategy, and only concentrates on the descent of the function value, which ensures that MinUCB can converge to the local optimum points efficiently. MinUCB can also be explained through a trust region view, which can be seen the third point of global rebuttal. Weakness 2: ...One can construct kernels/examples where this "local" property is violated...e.g. for linear kernel... Response: Thank you for your comment. We believe that the example you mentioned, that is, under a linear kernel, it is indeed possible for the algorithms to have a large search distance . However in this example, the global optimum and the local optimum are the same (if the optimum is unique). When some local information about the Gaussian process is known, such as the current function value being higher than the Gaussian process prior mean, or the norm of gradient being large nearby, MinUCB and LA-MinUCB will have a relatively large search distance, as the local information indicates that the local optimum point is a certain distance away from the current point. However, this is not a global search, as it is only looking for a point that is better than the current point function value through the information of Gaussian process. This greedy strategy will quickly approach the local optimum point of the current region, which is also its advantage over other methods as it better utilizes the information of Gaussian process. Weakness 3: How can we efficiently find the minimizer of the UCB? Response: In our numerical experiments, we only used the built-in optimizer in Botorch without any special range restrictions, but it seems that there was no significant numerical instability when minimizing UCB. If the dimensionality is really high, we believe that the results of the previous iteration can be used as a starting point to find a better local optimum point of UCB, which can at least achieve better results compared to the previous step. Weakness 4: ...Lacking several works, e.g. the LineBO paper...lack several baselines(e.g. Turbo or LineBO) Response: We are very sorry for not including an introduction to the LineBO series of work. We will include it in future version of paper. We have added a comparison with TurBO in Fig 1,2,3 in the PDF attached to the rebuttal, but due to time constraints, we apologize that we may not be able to present the comparison results with LineBO in this rebuttal. We will add relevant experiments in future version of paper. Weakness 5: The introduction of the Lookahead strategy...not well connected to the remaining paper... Response: Thanks for your comment. As the main contribution of our paper is to bulid the relationship between gradient descent with minimizing UCB, and MinUCB still partially depend on the gradient (the local exploration), then is it possible to develop an algorithm that better utilize the concept of UCB? Thus we apply a look ahead strategy to build the LA-MinUCB, and the experimental results are also very competitive. Proving the convergence of LA-MinUCB should be an important research direction in the future. Weakness 6: In general, the paper should better highlight the technical contributions and challenges related to the main result... Response: We are sorry for not clearly describe our contributions and challenges in our paper. We think the main contribution of this paper, is exactly to build the relationship between minimizing UCB and gradient descent under the Gaussian process surrogate, and show that minimizing UCB will bring extra efficiency in local search. We also develop the algorithms MinUCB and LA-MinUCB to utilize this idea. The theoretical proof for MinUCB is a big chanllage and is technically not trivial. We have developed some tools to prove the convergence of this algorithm. Weakness 7: Minor remarks Response: We apologize for some of the inaccurate statements and will improve them in future versions of the paper. $\gamma$ is a coefficient in the matern kernel, and $n$ is the number of samples. --- Rebuttal Comment 1.1: Comment: I'd like to thank the authors for their response and for clarifying several issues. I appreciate that TurBO was added as a baseline. The proposed approach remains competitive in the reported experiments. It is also clearer now what "local" means, even though the notion is not fully defined formally. As far as I understand the goal is to be "local" in the kernel metric, and "efficient" only refers to sample efficiency, although empirically there are likely also computationally benefits. I think this should be made clear in the paper. One weakness remains that the lookahead strategy (which performs best in the experiments) was not analysed and is not clearly motivate from the main algorithm. This paper still has potential for improvements and therefore remains borderline, however I will raise my final score given the response by the authors.	Summary: The author proposed a local Bayesian optimization method using UCB to drive its iterates. The UCB step replaces the gradient descent step that is typically in such methods. Strengths: The comparison of gradient-descent to UCB is very interesting. The author included analysis as well so there is enough content to the paper. The assumptions on kernels are interesting and not straightforward. Clearly, effort has been made to establish convergence, though I think some of the theorems in the appendix would be better labeled lemma. Weaknesses: The biggest concern I have is that the algorithm is too similar to Bayesian optimization using UCB. The author tried to add more technical development such as the look-ahead algorithm. But the main difference from the proposed algorithm and GIBO is that the next iterate is found via UCB. In doing so, it's hard to argue the proposed algorithm is a gradient-based method anymore. And the proposed algorithm becomes too familiar. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Please proofread the paper to check grammar and spelling. For example, line 151, 161. 2. Line 171 "more likely to be lower than" should be "more likely to have a smaller value than". 3. Line 269 "minimize the minimum point?" 4. The gradient descent methods have a rich literature in how to choose the step size. It's not always $1/L$. The author should clarify on that point. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback. The suggestions and questions are briefly responded below. Weakness : The biggest concern I have is that the algorithm is too similar to Bayesian optimization using UCB. The author tried to add more technical development such as the look-ahead algorithm. But the main difference from the proposed algorithm and GIBO is that the next iterate is found via UCB. In doing so, it's hard to argue the proposed algorithm is a gradient-based method anymore. And the proposed algorithm becomes too familiar. Response: Thanks for your comment. In our work, we focus on 'local' algorithm and show that the algorithm we propose, and specifically the acquisition function of minimizing UCB can provide an efficient local search approach. This is in contrast to the traditional minimization of LCB (on the minimization problem context), where the focus is to directly balance the trade-offs between exploration and exploitation for global optimization. We achieve this by first build the relationship between minimizing UCB and gradient descent under the Gaussian process surrogate, and show that minimizing UCB from this view will bring extra efficiency in local search. This provides us with an interesting idea that if we replace gradient descent with minimizing UCB, we may propose alternative efficient local BO algorithms. That is why we propose our two algorithms, MinUCB and LA-MinUCB, to illustrate this idea. As you mentioned, the proposed algorithm is not a gradient-based method anymore, but this is actually our final goal. The point selection strategy in our algorithm, has a similar behaviour as the gradient descent. The coefficient $\beta$ in UCB controls the search area, that larger $\beta$ will force algorithm to search in a smaller area, this is just similar with the stepsize in gradient descent. Our algorithm design use the idea of previous approximate gradient algorithm, and we apply minimizing UCB to better utilize the Gaussian process, which allows our method to integrate the advantages of both aspects. This makes our methods look similar to UCB, but in fact there are fundamentally different, that we transform the UCB concept from a global one to a local one. The experimental results also indicate that this combination can bring significant performance improvements, demonstrating the effectiveness of our idea. Question 1,2,3: grammar and spelling Response : Thanks for correcting the spelling and grammar errors in the paper, and we will make all revisions in future versions of the paper. Question 4 :The gradient descent methods have a rich literature in how to choose the step size. It's not always $1/L$. The author should clarify on that point. Response: Thank you for your suggestion. In this paper, we mainly want to explain the relationship between minimizing UCB and gradient descent. Some concepts may not be expressed clearly, and we will revise the relevant description in future versions of the paper. --- Rebuttal Comment 1.1: Comment: We extend our heartfelt gratitude for the time and effort you have dedicated during the review process. We would greatly appreciate it if you could examine the comments we have returned. We are eagerly anticipating your feedback and are looking forward to the possibility of engaging in further discussion with you. --- Rebuttal Comment 1.2: Title: reply to the author Comment: Dear author, Thank you for the rebuttal. Sorry for the late reply. I understand your motivation. But algorithm-wise, it seems to me your algorithm is still UCB BO. Your rebuttal did not add any information that would refute this, in my opinion. Correct me if I am wrong. Even down to the choice of $\beta$ for convergence analysis. I understand how you have come to UCB from the gradient descent algorithms. But you end up with an existing algorithm, albeit with some tweaks. I appreciate a lot of the extra work you put in, but to me this is clearly an upgrade to UCB. Only now your convergence has actually been weakened to a gradient optimality (KKT) condition, which does not even guarantee a local minimum. I really appreciate the KKT condition, in accordance with the "local" emphasis. But in my humble opinion, going from a global algorithm (UCB) to a local algorithm is not necessarily an upgrade. I know you are just comparing to existing gradient descent type methods. But you have to compare to UCB because that's what you ultimately proposed. Correct me if I am wrong. Thank you. --- Rebuttal 2: Comment: Dear Reviewer, Thanks for your fast response and further clarification of your concern. We would like to appreciate your view and perspective on the UCB, which is quite relevant to our topic. We will answer your questions from the following three aspects: Why we focus on a local approach? In this paper, we want to emphasize that what we focus on is the local approach instead of searching the global optima. The reason is that, searching the global optima is nearly impossible in high dimensional case, which is so-called curse of dimensionality. Typical global algorithms will have poor performance on high-dimensional problems. Take traditional UCB BO algorithms as an example, that they try to balance exploration and exploitation through maximizing UCB or minimizing LCB. In high dimensional problem, this ‘exploration’ may become ‘over-exploration’. Traditional UCB BO uses almost all points to learn all uncertain parts of the function, making the algorithm approach the global optimum very slowly. This is also reflected in [1], that the regret bound of UCB BO will grow exponentially with data dimension $d$. Traditional UCB BO also performs poorly in high-dimensional experiments. According to your latest suggestions, we have applied traditional UCB BO on our synthetic experiments and found that this method can hardly achieve any high value in these experiments. Similar experimental results can be found in Wu et al. [2]. On the contrary, local approach, like approximate gradient methods [2,3] or trust region methods [4], only focus on searching in a small area, which effectively decrease the over-exploration. Besides, a local optimum can be found relatively quickly, and exponential sample complexity is only necessary if one wishes to enumerate all local optima [2]. Although local approach only focuses on local optimums, this type of method still outperforms traditional global methods in high-dimensional performance [2,3,4]. That is why we mainly focus on local optimization algorithms in high-dimensional problems. As global and local are two completely different concepts, the global method and local method cannot be equated. How do we transform the traditional UCB concept into a local one? According to the statements above, we need to look at the local approach (focused on local optimization), and specifically those based on approximated gradient descent, as they have been proved to be effective in local BO. Our objective is to improve the efficiency of this type of gradient based approach. We first realize that there is a relationship between minimizing UCB and gradient descent under the Gaussian process surrogate, and show that minimizing UCB will bring extra efficiency in local search. Minimizing UCB, as it is a rather conservative strategy and only focuses on local exploitation (please refer to viewpoint 1 of Global Rebuttal), is a new idea and is completely different with the concept in traditional UCB BO method (as traditional UCB still considers exploration). We replace gradient descent with minimizing UCB to obtain extra efficiency, and thus we propose our two algorithms, MinUCB and LA-MinUCB, to illustrate this idea. It should also be noted that our MinUCB achieve a polynomial convergence rate with data dimension $d$, and this result is impossible to be obtained in traditional UCB BO. This is because traditional UCB BO will only converge at global optimum and its regret bound will grow exponentially with $d$. That's also why our method performs very well in high dimensions, and our results far exceed the original UCB BO in these cases. How to do global search in high dimensional case? As the traditional BO like UCB BO falls in high dimensional case, a local search with adaptive restarts (similar with TurBO [4]) or a mixed global local search with our local algorithm can be much more efficient. According to the results in [2], local search with multiple starting points can effectively approximate the global optimal value without a large difference, and is very competitive with global methods. [1] Srinivas, Niranjan, et al. "Gaussian process optimization in the bandit setting: No regret and experimental design." arXiv preprint arXiv:0912.3995 (2009). [2] Wu, Kaiwen, et al. "The behavior and convergence of local bayesian optimization." Advances in neural information processing systems 36 (2024). [3] Müller, Sarah, Alexander von Rohr, and Sebastian Trimpe. "Local policy search with Bayesian optimization." Advances in Neural Information Processing Systems 34 (2021): 20708-20720. [4] Eriksson, David, et al. "Scalable global optimization via local Bayesian optimization." Advances in neural information processing systems 32 (2019).	Summary: The paper presents an extension of a local Bayesian optimization strategy, specifically targeting methods based on gradient information (GIBO). GIBO-type algorithms operate through two stages: an exploitation step and an exploration step. The paper introduces two novel algorithms within this framework. In GIBO, the exploitation step is performed using gradient descent. The authors identify that gradient descent can be viewed as minimizing a quadratic upper bound. They further note that minimizing the tighter upper confidence bounds (UCB) can exploit more information from the GP surrogate model. Consequently, they replace the gradient descent step in GIBO with UCB minimization, resulting in a new algorithm named MinUCB. The paper provides a convergence analysis for MinUCB. The second algorithm modifies the exploration phase of GIBO to optimally query points using a look-ahead strategy, aligning with the new exploitation approach. Both algorithms are empirically evaluated on some established benchmarks. Strengths: - The paper presents a clear, well-motivated, and relevant contribution to the field of local BO. - The related work is thoroughly discussed, situating the paper's contribution within the context of existing research. - The theoretical analysis is interesting and significantly strengthens the algorithmic contribution. - The proposed algorithms improve the state-of-the art for GIBO-style algorithms. Weaknesses: ## Main aspects Locality: In contrast to GIBO and MDP, the exploitation step is standard UCB. Therefore $x_{t+1}$ does not need to be in the local region around $x_{t}$. In fact this will only be the case if the algorithm is initialized in a region that is lower then the mean. I hypothesize that the "exploitation" step of MinUCB will explore globally until it finds a "promising region". I don't see this as a fundamental problem with the algorithm and may even allow MinUCB to switch between local and global exploration, which could be beneficial and avoid getting stuck in local optima. However, the paper does not discuss this behavior or explain why we should expect MinUCB to exhibit local-only exploration behavior. Including such a discussion and empirical investigations on how local MinUCB actually is could significantly improve the paper. Empirical evaluation: The empirical evaluation is limited, missing important details, and is generally not well executed. - The algorithm is evaluated on a subset of the problems in Nguyen et al. and does not introduce any new benchmarks. The low number of problems makes it hard to assess how efficient the algorithm really is, especially since the Hopper experiments seem problematic due to state normalization. No method is able to "solve" the hopper task, and the achieved reward is very low. - There are unexplained differences in the results on the synthetic benchmarks compared to those reported by Nguyen et al., where MDP and GIBO have similar performance after 500 evaluations. In the new results, GIBO outperforms MDP. - There are important details missing: 1. Batch sizes 2. Hyperpriors and hyperparameters - The code is not available for review, which hinders the evaluation of the experiments at the time of review. - The meaning of the error bars changes between plots. - A sensitivity study is missing. How sensitive is the algorithm to the choices of the newly introduced hyperparmeter, namely $\beta$ in the UCB? Limitations: The paper does not discuss the limitations of the proposed method and its evaluation. ## Minor - In Theorem 1, $n$ is never introduced. - The paper often cites arXiv preprints when peer-reviewed versions of the articles are available. - There are some grammar and spelling mistakes/inaccuracies: - "..solve _the_ high dimensional black-box.. problem" should be plural: "..solve high dimensional black-box.. problems" (there are many such problems). - "Gradient descent" should be lower case. - "large estimate to ensure _the_ convergence" should be: "large estimate to ensure convergence" (we are talking about general convergence behavior). - The typesetting of "L-smooth" is off, perhaps due to being typed in math mode. - Some words/abbreviations are typeset in math mode (e.g., in equation 3, "UCB"; in equation 5, "trace"). - Equation 1 is hard to parse due to the nested inequality and scalar product symbols. - Figure 3: "LA-MinUCB has consistently optimal performance." In what sense is the performance optimal? This claim seems premature and needs clarification. Technical Quality: 3 Clarity: 4 Questions for Authors: 1) In line 54 and later in the evaluation the authors claim "MPD ... can exhibits numerical instability.." but do not follow up on this claim. What numerical instabilities were observed? 2) Why limit the experiments to a subset of the problems in Nguyen et al.? 3) Why choose constant batch sizes and beta when the theory predicts choices with guaranteed convergence? Why deviate from the theoretically obtained values? 4) Why are the results on the synthetic benchmarks significantly different from those reported in Nguyen et al.? 5) How where the batch sizes chosen? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The limitations of the algorithm are not discussed and need to be addressed. For instance, the paper derives convergence for specific choices of beta and batch sizes, but chooses different values in the empirical results. This discrepancy raises the question of whether there is a gap between theory and practice. Discussing these limitations and the rationale behind the chosen parameters in the experiments would provide a more comprehensive understanding of the algorithm's performance and applicability. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback. The suggestions and questions are briefly responded below. Weakness 1: Locality: I hypothesize that the "exploitation" step of MinUCB will explore globally until it finds a "promising region"... switch between local and global exploration Response: Thank you for your comments and suggestions. We acknowledge a phenomenon that when the initial point value is higher than the prior mean, there may be a larger step size in the initial one or two exploitation step. However, this should be a fast local search instead of a global exploration. As we mentioned in the point 2,3 in global rebuttal, the local exploitation in each step is only to try find the point that have a lower value than the previous one, which does not involve global exploration idea. For the global-local balance, it's possible to use a mutistart strategy , which is similar as TurBO. However, we believe that the main contribution of this article is to propose the relationship between minimizing UCB and gradient descent under a Gaussian processes surrogate, and give the corresponding theory, so we do not emphasize the global-local balance in this paper. Weakness 2 and Question 2,5: Empirical evaluation: details, no new benchmarks, error bars, open source code Response: Thank you for your comments. In our experimental settings, we basically used the same settings as GIBO and MPD. For example, the hyperpriors is the uniform distribution on an interval and and hyperparameters is set the same as GIBO and MPD. For MinUCB, the batch size for local exploration is the data dimension $d$. For LA-MinUCB, we set the batch size as $d$ in synthetic experiments, and emperically find that when the batch size is chosen as $0.2d$ on the reinforcement learning task, the numerical experiments will have good result. We apologize for the lack of experiments in the original submission. We have added other real world objectives mentioned in MPD [1]. The experimental results can be seen in Fig 3 in the pdf attached to the rebuttal. The differnent error bar is mainly due to the significant differences between sampled functions in the synthetic data, and directly plotting will result in large variances at different points. So we applied scaling to achieve better visual effects. We are sorry for not making the code public because we are afraid that it may lead to the leakage of author information, but we guarantee that the experimental results are accurate, and will make it public after the acceptance of paper. Weakness 3: Lack of sensitivity study Response: We are sorry for the lack of sensitivity experiments, which will be included in future papers. The PDF for this rebuttal does not have enough space to place the results. Overall, the $\beta$ controls the convergence speed and final convergence result. When $\beta$ is relatively small, the convergence speed will be fast. However, the final convergence result may be slightly weaker than that when $\beta$ is large. This is quiet similar with the gradient descent with a large or small stepsize. Small $\beta$ means the local exploitation is more aggresive, but may lose some accuracy, which will cause the algorithm jump around the local optima. Weakness 4 and Question 1,4: The difference between our experimental result and Nguyen. et al[1] Response: Thank you for your comments. When we conducted experiments based on Nguyen's code , we found that the MPD method is very likely to stuck at a specific point, and the search results afterwards may even get worse. This phenomenon is particularly significant in synthetic experiments. The MPD results in our experiment are actually very similar to the previous results of Nguyen et al., that they both achieving good result after the first batch of point. However, GIBO performed better in subsequent searches, and is robustness in diffenrent experiments. We speculate that this may be due to randomness or other unknown settings. Question 3: Why choose constant batch sizes and beta Response：The reason for selecting a fixed batch size and $\beta$ is to compare with previous methods. The previous methods used fixed parameters, so we also use fixed parameters for reasonable comparison. Although it is theoretically necessary to increase the batch size and $\beta$ to ensure convergence, fixed parameters can actually achieve good results (although there is a small distance between the result and the true local optimum). [1] Quan Nguyen, Kaiwen Wu, Jacob Gardner, and Roman Garnett. Local bayesian optimization via maximizing probability of descent. Advances in neural information processing systems, 35:13190–13202, 2022 --- Rebuttal Comment 1.1: Comment: We extend our heartfelt gratitude for the time and effort you have dedicated during the review process. We would greatly appreciate it if you could examine the comments we have returned. We are eagerly anticipating your feedback and are looking forward to the possibility of engaging in further discussion with you.	Rebuttal 1: Rebuttal: We thank the reviewers for their constructive comments and suggestions, which have brought great help to improve our paper. We would like to provide a clearer explanation of two aspects: the local behaviour and why our algorithms are local. In this paper, we would like to emphasize that the term 'local' mentioned does not refer to limiting optimization within a small range or a specific interval, such as TurBO, but rather to the algorithm exhibiting local behavior. More specifically, the algorithm will only try to search points around the current dataset, without doing global exploration. The locality of minimizing UCB: Compared with minimizing LCB, minimizing the UCB is a very conservative strategy that places additional penalties on exploration. Specifically, LCB is small with small $\mu(x)$ or large $\beta\sigma(x)$, while UCB is only small when both $\mu(x)$ and $\beta\sigma(x)$ are small. If we consider a point that is far away from the current dataset. Since there has been no exploration around this point, $\beta\sigma(x)$ will be very large, which means that the minimum point of UCB will not be taken here. So the meaning of minimizing UCB is that, it will search towards points with smaller $\mu(x)$, while $\beta\sigma(x)$ controls the search distance, forcing the search to be around the current dataset, which leads to minimizing UCB being a local strategy. Why MinUCB (Algorithm 1) is local?(gradient descent view): Our MinUCB can be viewed as the enhanced version of gradient descent with a larger descent in each step. If we assume $x_{t}$ as the result of the $t^{th}$ local exploitation (step move, line 8 in Alg 1), then what we actually proved in Theorem 1 is the following inequality: $$f(x_{t+1})\le\min_{x\in \mathcal{X}}\mu(x)+\beta_{t}\sigma(x)\le f(\hat{x}^{t+1}) + e_{t}<f(x_{t})+e_{t}$$ where $\hat{x}^{t+1}=x_{t}-\eta_{t}\nabla \mu(x_{t})$, and stepsize $\eta_{t}$ decreases at a logarithmic rate with $t$, and some error terms in the proof $e_{t}$ will decrease to 0. $\hat{x}^{t+1}$ is actually the gradient descent point starting from $x_{t}$. The above inequality shows that MinUCB also adopts a greedy strategy that it only tries to find a point that is better than previous one. This reflects the locality of MinUCB, as it only focuses on the reliable improvement in each step, instead of doing exploration to search potiential global optima like minimizing LCB. This greedy strategy also guarantees that MinUCB will converge to local optima with a polynomial speed. Why MinUCB (Algorithm 1) is local? (trust region view): The trust region, which is defined in TurBO [1], is a hyperrectangle centered at the best solution found so far, and will shrink its size after too many consecutive “failures”, or expand it after many consecutive “successes”. The adjustment of trust region is trying to find a area, that the searching result $x$ in the area satisfies that $f(x)<f(x_{t})$ with a certain probability, where $x_{t}$ is current best point, so that the proportion of “successes” and “failures” can be controled. From this view, we can construct a variant of trust region through an upper confidence bound view: $$ UCBTR_{x_{t}} \coloneqq \\{x\|\mu(x)+\beta\sigma(x)<f(x_{t})\\}$$ This region is more conservative than the traditional trust region, as any points in this area will be lower than $f(x_{t})$ with a probability larger than $p$, and $\beta$ controls this probabiliity $p$. Therefore, the local exploitation (step move, line 8 in Alg 1) can be viewed as searching the reliable minimum point in this trust region: $$ \min_{x\in \mathcal{X}}\mu(x)+\beta\sigma(x)\le \mu(x_{t})+\beta\sigma(x_{t})\approx f(x_{t})$$ where the right approximation is guaranted through local exploration (sampling, line 4 in Alg 1), and this local exploration step will learn more local information around $x_{t}$, which will help to expand or shrink this area $UCBTR_{x_{t}}$. In this view MinUCB can be treated as a variant trust region method method like TurBO, where we replace the trust region with a probability subset. This probability subset is accurate and reliable, which brings additional efficiency in local searching. Why LA-MinUCB (Algorithm 2) is efficient?: Compared with MinUCB, LA-MinUCB adopts a more greedy stratgy in local exploration. In addition to learning local information, this local exploration step (line 3 in Alg 2) is aimed at obtaining better reliable minimum point in the next local exploitation through a look ahead strategy. This will possibily obtain a larger decreasement of function value than MinUCB in each step, which is also reflected in numerical experiments. Pdf: /pdf/93cacb5839fe67bccb49cbd6dbe1995bc7b859fb.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Grammar-Aligned Decoding	Accept (poster)	Summary: This paper focuses on the constrained decoding scenario where LLMs are expected to produce high-quality and grammatically correct outputs. It presents the adaptive sampling with approximate expected futures (ASAp) method, which is designed to enhance the quality of the output by adjusting the conditional probability to align with the model’s original output distribution while ensuring grammaticality. Central to ASAp is the development of effective approximations to the expected future grammaticality (EFG). More specifically, ASAp involves an extra sampling process, where the estimates of EFG are iteratively refined after each sampling and will eventually converge to exact EFG in the limit of infinite iterations. Experimental results show that the output probability distribution of ASAp is closer to the original one than that of GCD methods. Strengths: Advantages of ASAp include facilitating more accurate grammatical predictions. ASAp approximates the desired distribution by considering the potential future grammaticality of sequences, whereas GCD typically samples greedily based on current probabilities without considering the future development of the sequence. This approach allows ASAp to more accurately predict and generate text sequences that comply with grammatical rules. Weaknesses: The algorithm heavily relies on prior samples to estimate grammatical probabilities, which can be computationally expensive and difficult to manage, especially in cases where the dataset is large or the environment is dynamically changing. Moreover, a fair comparison of decoding speed between ASAp and GCD methods is required, as ASAp needs this sampling process to produce high-quality outputs, but conducting sampling on LLMs is very time-consuming. Only evaluating the quality of the output by KL divergence and expectations is limited because the decoding sampling algorithm, except for the greedy search, does not always obey the probability distribution. For example, when the decoding temperature is set to 1, models may not output tokens with the highest probability. To conduct a more intuitive evaluation, one can choose some constrained decoding tasks with explicit quality scores, such as constrained translation [1] with BLEU and Exact Match, and code generation, which can be evaluated by the passing rate. Since the experiments in this paper are based on prompt-motivated LLMs, a natural step is to fine-tune LLMs and see if they can produce grammatically correct output with high quality. On the other hand, the authors can also try ASAp based on fine-tuned LLMs, which may further boost the performance of ASAp. To enhance the readability of the article, the following suggestions can be considered. 1. It’s more intuitive to give specific input and output examples for SLIA and INV-BV tasks. 2. The names GCD and GAD are too similar to easily differentiate, even after reading the paper. 3. Other typos: In line 289, the word “both” is typed as “bot”. [1] Wang, S., Li, P., Tan, Z., Tu, Z., Sun, M., & Liu, Y. (2022). A Template-based Method for Constrained Neural Machine Translation. arXiv preprint arXiv:2205.11255. Technical Quality: 2 Clarity: 3 Questions for Authors: Given previous samples with the prefix $w_{1:i}$, there exist abundant future tokens beyond just the next token. However, ASAp only uses the next token likelihoods to facilitate a better estimate of EFG. Why not use other future tokens? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The algorithm heavily relies on prior samples to estimate grammatical probabilities, which can be computationally expensive and difficult to manage… a fair comparison of decoding speed between ASAp and GCD methods is required, as…conducting sampling on LLMs is very time-consuming. Indeed, ASAp requires building a large data structure to keep track of the sampled prefixes, but decoding each sample takes the same time with GCD and ASAp. > Only evaluating the quality of the output by KL divergence and expectations is limited because the decoding sampling algorithm, except for the greedy search, does not always obey the probability distribution.. one can choose some constrained decoding tasks with explicit quality scores.. We agree that downstream, extrinsic evaluations are of interest: e.g., how much does ASAp improve task performance metrics on specific tasks? (See next paragraph, we do include some of this information in our paper.) However, these results are task dependent, as the reviewer mentions -- in some cases, increasing likelihood may correspond to higher task accuracy and in some cases not. Since we focus on formalizing the problem of distribution misalignment in GCD, the primary objective of our experiments has been to evaluate intrinsic measures, i.e., to assess how closely the probability distributions from GCD and ASAp approximate the exact GAD in Equation 2 in practice. We do present the downstream quality of the output in Table 1 in Appendix A.5.2. However, since the model we used for evaluation was not sufficiently well-trained for our task, the experimental results in Appendix A.5 show that the output with the highest probability does not necessarily indicate good quality on the downstream task. > a natural step is to fine-tune LLMs and see if they can produce grammatically correct output with high quality. On the other hand, the authors can also try ASAp based on fine-tuned LLMs We thank the reviewer’s suggestions regarding the evaluation. Previous work on GCD has shown that grammar-constrained LMs outperform fine-tuned models in structured NLP tasks (citation [5]). But we agree it is interesting to find out whether ASAp still provides benefits over GCD when applied to an LLM distribution that has already been fine-tuned to achieve higher grammaticality from the outset. Due to limited time and data (139 INV-BV and 2416 CP problems), we conducted a small experiment to test the reviewer’s hypothesis. In our finetuning step, we want to teach the LLM to assign higher probabilities to grammatical outputs for the specific task DSL. We randomly selected 2 INV-BV problems (find_inv_bvsge_bvneg_4bit and find_inv_bvsgt_bvor_4bit for INV-BV) and 4 CP problems (CP_re_ptb_215, CP_re_ptb_434, CP_re_ptb_1627 and CP_re_ptb_1643 for CP) from the test set, and used all other input-output pairs of prompt and output programs to construct datasets for finetuning the base LLM. We obtained 2 finetuned LLMs, one for INV-BV and one for CP. We ran GCD and ASAp on the finetuned models on the randomly left-out problems and checked the convergence rates of the KL-divergence. The results from finetuned models did not show significant differences in terms of convergence compared to the original model. As done in our evaluation, we computed the expectation for each benchmark obtained via GCD and ASAp after 2,000 iterations and compared it against the target expectation $Q^{P,G}$ (line 76) of GAD. The sum of least squares difference between expectations computed by GCD and the expectations of $Q^{P,G}$ are 0.677 (INV-BV4), 0.278 (CP) (respectively. 0.051 (INV-BV4), 0.201 (CP) for ASAp). I.e., the expectations computed by ASAp were closer to the expectations of exact GAD than those computed by GCD. We didn’t include SLIA as we didn’t have sufficient data for further finetuning. We acknowledge there are alternative ways to fine-tune the model for learning grammatically; this goal is beyond the scope of the paper. Details on experimental setup: We adhere to the established LoRA finetuning pipeline and create task-specific datasets for instruction tuning. In line with our paper's methodology, we incorporate in-context examples in the instruction tuning dataset to enhance the models' performance in in-context learning. For each task, we independently finetune Mistral-7B, resulting in two versions of the model (for INV-BV4 and CP). We employ a standard train-validation-test split of 70-10-20%. Instruction tuning is conducted on the training set, and model selection is based on the lowest validation loss. Learning rate: 2e-4, warmup ratio: 0.03, max sequence length: 2048, LoRA alpha: 32, LoRA dropout: 0.05, and LoRA r: 64. The best checkpoints for the finetuned models for INV-BV and CP are at 328 and 536 steps, respectively. > It’s more intuitive to give specific input and output examples for SLIA and INV-BV tasks. Due to the space constraint in the main paper, we have included concrete examples and the grammar used to constrain decoding for LLMs in Appendices A.4.1 and A.4.2. We will provide a detailed description of these problems in the supplementary materials. > The names GCD and GAD are too similar to easily differentiate, even after reading the paper. We will use more readable macros. > Given previous samples with the prefix w_{1:i}, there exist abundant future tokens beyond just the next token. However, ASAp only uses the next token likelihoods to facilitate a better estimate of EFG. Why not use other future tokens? The reviewer is right that there could be faster-converging decoding approaches, which we are exploring as future work (e.g., as hinted by the reviewer, sampling more tokens at once to accelerate convergence is one of them). The main focus of the paper is to formalize the likelihood misalignment problem in existing grammar-constrained decoding, and to provide an initial solution with provable asymptotic guarantees. --- Rebuttal Comment 1.1: Comment: Thank you for your response, which has addressed some of my concerns. I think this is an interesting work. However, I will maintain my rating score as the issues warrant consideration in a major revision of this submission.	Summary: This paper proposes adaptive sampling with approximate expected futures (ASAp) for grammar-aligned decoding for LLMs. The main objective of this method is to match the conditional probability of the LLM’s distribution conditioned on the given grammar. The evaluation is performed on code generation and structured NLP tasks showing a better likelihood for LLM’s outputs being grammar-constrained. Strengths: - The method is well-formalized and clear. The examples are very useful for faster understanding - The method is well-motivated (there isn’t an independent motivation section but each choice in the design of the method is appropriately justified) - The paper provides empirical validation that the method works and that it improves the benchmark scores Weaknesses: - Nothing in particular. Maybe I have found Section 2 of the paper somewhat difficult to read but it is probably because of my lack of recent readings on CFG. Technical Quality: 4 Clarity: 4 Questions for Authors: I suggest maybe some improvement in Section 2. Restructuring it maybe and providing slightly more context on CFG (or giving a concrete example of how it works as a reminder, I found Fig 1 not illustrative enough), I'm sure I won't be the only reader who doesn't have recent experience with CFG. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes, limitations are discussed. What I liked in particular in this paper is that the limitations of ASAp are well exposed and discussed throughout the paper and not just in the limitations section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > I suggest maybe some improvement in Section 2. Restructuring it maybe and providing slightly more context on CFG (or giving a concrete example of how it works as a reminder, I found Fig 1 not illustrative enough), I'm sure I won't be the only reader who doesn't have recent experience with CFG. We thank the reviewer for the suggested improvement. In the revision, we will include a formal definition of CFG and an example of how a string can be derived for the CFG in Fig 1.	Summary: This paper points out that prior methods on grammar-constrained decoding (GCD) distorts the language model's learned distribution over sequences. At the heart of this exposition is the notion of _expected future grammaticality_ (EFG), where prior GCD methods can be cast an an upper-bound approximation to the EFG. To ameliorate this problem, the authors proposed the ASAp method: through many iterations, the EFG of a prefix can be better estimated. Strengths: This is an important finding that conventional methods on grammar-constrained decoding distorts the language model's learned distribution. As far as I know, this important fact has not been pointed out before. The notion of expected future grammatically will be impactful in the area of sequence decoding. Weaknesses: The proposed method seems very slow to run, requiring many iterations. Additionally, it requires storing a table of all seen prefixes and their future grammatically, which would be very large if the grammar itself is large. It seems to me that the ASAp method only depends on a grammar -- it does not specifically requires a training or test set to run. I wonder if it would be possible to first sample extensively and train the ASAp EFG estimates, then decode on a set. In this way ASAp can be cast as a preprocessing step of the grammar: if it is very slow, it wouldn't matter much since it'll only be processed once. For example we are doing NL2SQL under many conditions. The target grammar stays the same: SQL. One can utilize a LM and generate massive amount of prefixes that has EFG > 0, then compute this estimation of $\tilde c_S$ for all these distinct prefixes. Technical Quality: 4 Clarity: 3 Questions for Authors: - L10: "prorblem" => "problem" - Missing related work: R Shin, et al (2021): Constrained language models yield few-shot semantic parsers. https://aclanthology.org/2021.emnlp-main.608/ - Alg 1: What is "ancestral sampling"? Please explain. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors adequately discussed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The proposed method seems very slow to run, requiring many iterations. Additionally, it requires storing a table of all seen prefixes and their future grammatically, which would be very large if the grammar itself is large. It seems to me that the ASAp method only depends on a grammar -- it does not specifically require a training or test set to run. I wonder if it would be possible to first sample extensively and train the ASAp EFG estimates, then decode on a set. We acknowledge that our proposed method can be slow to converge. However, the main focus of the paper is to formalize the likelihood misalignment problem in existing grammar-constrained decoding, and to provide an initial solution to address this problem together with a proof of convergence. The reviewer is right that there could be better decoding approaches, and we have started exploring several approaches as part of our future work (e.g., as hinted by the reviewer, sampling more heterogeneous sequences during preprocessing, and training a neural estimate of expected future grammaticality, offline, to use later during test-time sampling). > Missing related work: R Shin, et al (2021): Constrained language models yield few-shot semantic parsers. We thank the reviewer for providing the missing early related work on grammar-constrained decoding. Semantic parsing is a very interesting application we plan to investigate in the future. We will include this paper in our revision. > Alg 1: What is "ancestral sampling"? Please explain. In the literature, `ancestral sampling' is commonly used to describe the default process for sampling from a locally normalized generative model, that is, sample the variables in order of a topological sort of the graphical model. In left-to-right autoregressive models like LLMs, this is equivalent to just sampling tokens left-to-right. We will add this explanation. --- Rebuttal Comment 1.1: Comment: Thanks for the response. I will keep my score since I believe that some additional work (e.g. training a neural estimate of EFG) will make this paper more contained.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their feedback, which will greatly improve the paper. We have addressed each reviewer’s comment in their corresponding answer. We summarize the edits we plan to do to improve the paper based on the feedback we received: Reviewers cHBv and gp9U proposed ways to potentially modify the ASAp algorithm to speed up convergence. 1. We will clarify in introduction that the main focus of the paper is to formalize the likelihood misalignment problem in existing grammar-constrained decoding, and to provide an initial solution to address this problem together with a proof of convergence. We will explain in Section 3 that the ASAp algorithm is not necessarily optimal in terms of sample-efficiency and clarify in the conclusion that there are opportunities for improvement. 2. We will include a formal definition of CFG and an example of how a string can be derived for the CFG in Fig 1. 3. We will add the missing related work provided by the reviewer cHBv: R Shin, et al., Constrained language models yield few-shot semantic parsers (2021) 4. We will address all other comments raised by the reviewers and include all the technical and writing clarifications from each rebuttal. 5. We have performed the experiment proposed by reviewer gp9U on fine-tuned models for our tasks and did not observe any significant differences in terms of convergence rates for both GCD and ASAp compared to the original model. If the reviewers find it beneficial, we can add a paragraph about this finding.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
VLM Agents Generate Their Own Memories: Distilling Experience into Embodied Programs of Thought	Accept (spotlight)	Summary: This work proposed ICAL, aims to improve decision-making in large language and vision-language models by generating optimized trajectories and language annotations from noisy demonstrations and human feedback. ICAL abstracts noisy trajectories into optimized sequences with language comments, refined through human feedback during execution. The experiment shows ICAL significantly improves performance in benchmarks like TEACh, VisualWebArena, and Ego4D, surpassing state-of-the-art methods. Strengths: 1、Improved Performance: ICAL significantly enhances decision-making and task success rates across various benchmarks, such as TEACh, VisualWebArena, and Ego4D. 2、Reduced Reliance on Expert Examples: The method minimizes the need for expert-crafted examples by generating useful abstractions from sub-optimal demonstrations and human feedback. 3、Versatility: ICAL is effective in multiple domains, including dialogue-based instruction following, multimodal web tasks, and video action anticipation. 4、Human Feedback Integration: The approach incorporates human feedback to refine and adapt abstractions, improving the agent’s performance over time. Weaknesses: 1、The method still relies on human feedback for refining abstractions, which may not always be feasible or scalable. 2、The process of generating and refining abstractions heavily rely on GPT4-V 3、The effectiveness of ICAL is constrained by the capabilities of the underlying Vision-Language Models (VLMs), such as GPT-4V. 4、The method’s performance can be affected by the quality of the initial noisy demonstrations and the accuracy of human feedback. Technical Quality: 3 Clarity: 2 Questions for Authors: The VLM-driven Abstraction Generation component appears to heavily rely on the performance of GPT-4V. There is no ablation study for replacing GPT-4V. If such results exist and I missed them, please indicate where they can be found. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >ICAL's effectiveness is constrained by the capabilities of VLMs like GPT-4V. The VLM-driven Abstraction Generation relies heavily on GPT-4V. Is there an ablation study for its replacement? We appreciate the comment regarding the reliance on GPT-4V for the VLM-driven Abstraction Generation component. We acknowledge the importance of evaluating alternative models and have addressed this in our work. In Table S2 of the Appendix, we provided an ablation study using GPT-3.5 for the VLM-driven Abstraction Generation on the TEACh dataset, where visual experiences are converted to textual format. The results indicate a performance drop when using GPT-3.5 for ICAL Abstraction Generation, with the number of successfully revised trajectories reduced by more than half compared to GPT-4V (52 versus 122 tasks successfully completed). In Section S3.4 of the Appendix, we demonstrate how relabeling unsuccessfully refined trajectories can help bridge this performance gap when using weaker models like GPT-3.5, thereby reducing the dependence on GPT-4V. We will highlight these findings more prominently in the main paper to ensure their visibility. At the time of our experiments, no open-source VLMs with in-context multimodal generation capabilities comparable to GPT-4V were available. Per your comments, we implemented LLaVa-NeXT for the abstraction phase in VisualWebArena given multimodal inputs. As shown in Figure R2, we found that it failed to properly revise the abstractions to take into account the feedback. However, we are committed to continuing our experiments with emerging open-source VLMs as they become available. >The method still relies on human feedback for refining abstractions, which may not always be feasible or scalable. It is essential to note that this feedback is sparse, provided in natural language, and required only a few times for each task. In the Visual Web Arena, our agent needs an average of just 5.36 natural language feedbacks per example, with each feedback taking less than 15 seconds to create. In TEACh, our agent needs just 1.516 average feedbacks per episode with an average length of 18 words per natural language feedback. In fact, natural language feedback in ICAL could be easily communicated via speech by a human while observing the agent, whereas low-level coding and abstraction writing communication would be significantly more laborious. In order to hand-write GPT4-V-length abstractions, this would require 202.62 words on average for each example, including the precise coding of actions. Additionally, our agent becomes increasingly efficient over time, requiring less human feedback and fewer environment interactions as it processes more examples. By retrieving past successful abstractions during the VLM-abstraction making and human-in-the-loop phases, it uses previously stored knowledge to help abstract new examples. As shown in Figure R1, for the second half of examples processed, the model requires significantly fewer environment steps (436±88 vs. 267±43, p=0.0143) and human feedbacks (0.74±0.17 vs. 0.21±0.08, p=0.0089) per episode. This demonstrates that retrieving abstracted examples during abstraction learning reduces both human effort and environment interaction over time. Consequently, using previously stored ICAL examples not only improves test performance but also accelerates learning for future examples. Furthermore, the feedback provided does not require specialized expertise. For example, typical feedback from VisualWebArena includes comments like, "This does not have 2 upvotes and includes a meme of three Spider-Men, a flag of the Netherlands, and a flag of Croatia. You should scroll down to see more posts." In TEACh, feedback might be, "The sink is full right now. Empty the sink before interacting with it." Such straightforward feedback significantly reduces the complexity and cost of data collection. It also often provides more context for the model to create generalizable abstractions. For example, the sink being full is not directly necessary for correcting the mistake, but will help the agent learn generalizable knowledge via the generated abstractions. Our approach additionally minimizes the need for detailed annotations or extensive interactions required to train reinforcement learning (RL) or behavior cloned agents for long-horizon, multimodal tasks. This method not only shortens the data collection process but also ensures minimal human expertise and interaction, making it a scalable and cost-effective alternative. >The method’s performance can be affected by the quality of the initial noisy demonstrations and the accuracy of human feedback. We acknowledge the concern regarding the impact of the quality of initial noisy demonstrations and the accuracy of human feedback on our method. Our method is designed to handle and successfully revise even very noisy or incorrect demonstrations. For instance, Listings S2 and S3 in the Appendix of the paper illustrates a very noisy code trajectory that was successfully corrected by ICAL. While it is true that ICAL may not always recover from extremely noisy demonstrations or feedback, such cases are typically filtered out or relabeled, as discussed in step 6 of Section 3.3 in the main paper. This challenge is not unique to our method; gradient-based methods also suffer from inaccurate examples that can lead to incorrect model updates. As suggested by reviewer vg3K, one viable solution to this could be to utilize a reward model to identify and remove misleading or extremely noisy demonstrations before processing. --- Rebuttal Comment 1.1: Title: increase rate to 5 Comment: Thanks to the authors for the detailed response. My concern about the dependency on human feedback has been addressed. So I increase my rate from 4 to 5. I do not give a higher rate at the current stage. That is because I am still concerned about the generalization and reproducibility of the overall flow, as it heavily relies on the close-sourced GPT-4. If the proposed method can not help any open-source models, it is hard to justify that the contribution can be generalized to other VLMs. --- Rebuttal 2: Title: Thank you! Comment: Thank you for raising your score from 4 to 5. We truly appreciate your feedback and will incorporate the discussion into the final version. We will include further discussion on our use of close-sourced VLMs in the paper. We are committed to continuing our experiments with emerging open-source VLMs as they become available. Additionally, we will emphasize our ablation studies with GPT-3.5 and fine-tuning in the main paper to address these points.	Summary: The paper proposes a pipeline for Large Language and Vision-language models (LLMs and VLMs) to digest and learn from sub-optimal demonstrations and human feedback. The LLM/VLM, given sub-optimal task demonstrations, is prompted to produce abstractions of the trajectory (including task and causal abstractions, state changes, task decomposition and subgoals, and state abstractions), and possibly refine the given trajectory with insight from these abstractions. Optimally, these produced abstractions can be further refined by interactions with humans. Experiments show that the proposed framework greatly surpasses zero-shot CoT baselines and is competitive with/slightly better than having expert demonstations. Strengths: - The authors provide sufficient details with their experiment environments and experiment setup, as well as the code, for readers to better understand and reproduce the results. - The paper proposes a novel method for LLMs to autonomously refine sub-optimal trajectories and build a high-level abstract understanding of the task based on sub-optimal trajectories. For very complex tasks, these high-level abstract descriptions have the potential to help humans better understand the task and the general strategies to accomplish the task. - The experiments span three different types of environments, demonstrating the generality of the proposed method. The proposed method also achieve good experimental results over the baseline methods. Weaknesses: - The experiment result focuses on the accuracy of the method; not enough comparison is provided in terms of the efficiency. For example, whether is might be more efficient for the human feedback provider to directly edit and improve the sub-optimal trajectories and provide the human-refined ones to the LLM/VLM? - The scaling capability of the proposed method is unclear. In the proposed method, each trajectory needs to go through the human-in-the-loop fine-tuning process which seems quite inefficient. For simple tasks in the TEACh benchmark the method needs ~100 trajectories to perform well according to Figure 5. For more complex tasks it might be less sample efficient and difficult to scale. - It also requires the environment to automatically reset itself, which is another limitation. - Some text in figures 1 and 2 are too small and difficult to read and understand. A more illustrative example of what the task is, what are the inputs/outputs, and what kind of information is included in the abstracted state is desired. Technical Quality: 3 Clarity: 3 Questions for Authors: - Can we easily design a reward function that specifies how sub-optimal a trajectory is? If we provide these sub-optimal trajectories and their corresponding rewards to the LLM/VLM, can it use this reward information to further refine the trajectories? The "in-context RL" capability of LLMs are previous studied in works like [1]. - What is the difference between expert and non-expert feedback (mentioned in lines 281-282)? Examples with qualitative differences can be very informative. [1] Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf. https://arxiv.org/abs/2309.04658 Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors fail to include sufficient details on the human-in-the-loop fine-tuning phase. The evaluation might be biased if the human feedback provider is very familiar with the task and the LLM; they might be able to provide much more effective feedback than non-experts. If this is the case, then the amount of training required for the human to provide more effective feedback is worth investigating. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >Efficiency of ICAL method? For simple tasks in the TEACh benchmark the method needs ~100 trajectories to perform well according to Figure 5. Scaling capability of the human-in-the-loop phase? Actually, these roughly 100 trajectories are used for all tasks in TEACh combined. In fact, each task has on average 9.67 tractories in the memory after ICAL learning. The TEACh dataset includes 12 complex, long-horizon task types, averaging 67 steps for the shortest task (water plant) and 359 for the longest task (breakfast) for a human expert to complete. This complexity is significantly greater than other popular benchmarks like ALFRED, which averages only 50 steps across all tasks for experts. Despite the complexity, ICAL only requires a few noisy trajectories for each task. This is 6x less in-domain data than the strongest behavior cloning baseline (E.T. [1]), and achieves 21X the success rate, while also using noisy demonstrations. In VisualWebArena and Ego4D, which also represent high complexity, the number of demonstrations used for ICAL is of a similar magnitude (~100). Importantly, our agent becomes increasingly efficient over time, requiring less human feedback and fewer environment interactions as it processes more examples. By retrieving past successful abstractions during the VLM-abstraction making and human-in-the-loop phases, it uses previously stored knowledge to help abstract new examples. As shown in Figure R1, for the second half of examples processed, the model requires significantly fewer environment steps (436±88 vs. 267±43, p=0.0143) and human feedbacks per episode (0.74±0.17 vs. 0.21±0.08, p=0.0089). This demonstrates that retrieving abstracted examples during abstraction learning reduces both human effort and environment interaction over time. Consequently, using previously stored ICAL examples not only improves test performance but also accelerates learning for future examples. It is essential to note that the human feedback is sparse, provided in natural language, and required only a few times for each task. In VisualWebArena, our agent needs an average of just 5.36 natural language feedbacks per example, with each feedback taking less than 15 seconds to create. In TEACh, our agent needs just 1.516 average feedbacks per episode with an average length of 18 words per natural language feedback. In fact, natural language feedback in ICAL could be easily communicated via speech by a human while observing the agent, whereas low-level coding and abstraction writing communication would be significantly more laborious. In order to hand-write GPT4-V-length abstractions, this would require 202.62 words on average for each example in VisualWebArena and 48-107 lines of text for each example in TEACh, including the precise programming of actions or code. >Can we easily design a reward function that specifies how sub-optimal a trajectory is? Can the LLM/VLM use this reward information to further refine the trajectories? Thank you for this interesting suggestion. Based on your comment, we prompted GPT-4 with the sub-optimal trajectory and retrieved examples from memory, and asked the model to assign an "optimality" score from 1 to 5 with a reflection explaining the score. This score and reflection were then given to the VLM during the abstraction generation phase. After running this experiment for 40,000 steps, we observed no significant difference in the number of tasks successfully completed (22 with the score and 24 without the score). We hypothesize that this is because the VLM already performs this evaluation implicitly. This is supported by Figure R1, where in-context examples reduce the need for human feedback and environment interaction over time. This demonstrates that the VLM effectively utilizes the provided examples during the learning phases to revise the trajectory, inferring sub-optimality without explicit scoring. >What is the difference between expert and non-expert feedback? More details on the human-in-the-loop fine-tuning phase. The authors and their lab colleagues provided the human-in-the-loop feedback in the paper. We did not recruit participants due to the costs associated with recruiting participants for running ICAL and ablations. However, the authors have no more familiarity with the websites used in VisualWebArena than a typical web user and did not provide expert feedback, meaning it did not focus on precise coding or low-level changes. We will make these details more clear in the main paper. For example, typical feedback from VisualWebArena includes comments like, "This does not have 2 upvotes and includes a meme of three Spider-Men, a flag of the Netherlands, and a flag of Croatia. You should scroll down to see more posts." In TEACh, typical feedback is of the form, "The sink is full right now. Empty the sink before interacting with it." >It also requires the environment to automatically reset itself. ICAL improves learning efficiency as more examples are learned, reducing human effort and number of environment resets over time (Figure R1). Our approach also minimizes the need for extensive resets required to train reinforcement learning (RL) or large data collection for behavior cloning. Despite this, some resetting is necessary and we will add this to the limitations of the method. >Some text in figures 1 and 2 is too small and difficult to read and understand. A more illustrative example of the task, inputs/outputs, and abstracted state is desired. Thank you for the feedback. We will increase the font size in Figures 1 and 2. Additionally, we will include an extra figure to clearly illustrate the task, inputs, outputs, and the abstracted state. This figure will specifically show the exact inputs (e.g., images, text) for VisualWebArena and the corresponding outputs from the abstraction generation phase. References [1] Pashevich et. al. (2021). Episodic Transformer for Vision-and-Language Navigation. --- Rebuttal Comment 1.1: Comment: I thank the authors for addressing my questions. Considering the overall contributions and limitations of the paper, I will keep my ratings. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you for your thoughtful feedback. We appreciate your recognition of the strengths of our work and your constructive suggestions. We will include the clarifications and feedback in the final version. If you have any further questions, please feel free to discuss them with us. We are particularly pleased that the suggested experiments and feedback allowed us to emphasize the human-in-the-loop efficiency and further clarify the human feedback and trajectory refinement.	Summary: The paper proposes ICAL, In-Context Abstraction Learning, which builds a memory of suboptimal experiences that are abstracted into states and plans, as well as correction and reflection from human feedback. The approach is based on extensive prompting to elicit structured representations from past experiences, and apply RAG at test time. ICAL exhibits improvement in success rates across virtual domains compared to raw VLMs and CoT methods. Strengths: The idea of using extensive prompting to generate structured representation for RAG is clean and intuitive, and the authors clearly detail the steps involved. Using RAG enables the overall pipeline to improve continually. Weaknesses: My concern with the paper is the limited contribution in terms of the approach. It is a relatively simple adaptation of previous methods such as CoT, ReACT, Socratic Model [1] as well as RAG. [1] Zeng, Andy, et al. "Socratic models: Composing zero-shot multimodal reasoning with language." arXiv preprint arXiv:2204.00598 (2022). Technical Quality: 3 Clarity: 3 Questions for Authors: Could you further discuss the limitations of the work? I think currently the paper lacks thorough discussions on the limitations. For example, what are the common failure modes? And is there a viable and scalable solution to it potentially? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The overall prompting strategy is quite sensitive to the model performance. The fine-tuning improvement is quite limited as discussed in 4.6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >Clarifying the novel contributions and advancements beyond existing methods, such as CoT, ReACT, Socratic Model, or RAG. While our framework incorporates chain of thought prompting from CoT, interleaves reasoning and acting as seen in ReACT, predicts plans over pretrained modules similar to Socratic models, and uses retrieval-augmented generation of action plans inspired by RAG methods, it is distinct in its primary contribution: optimizing examples for improved in-context learning. We leverage VLMs to create abstractions for examples, improving their utility as RAG inputs. To the best of our knowledge, this is the first work to do this. Specifically, we introduce a novel two-stage approach that refines each example and generates four types of generalizable knowledge, enabling rapid adaptation of multimodal agents with few demonstrations. Our results, in Tables 1-3, demonstrate that both stages are crucial for high performance, significantly outperforming CoT and RAG, which use raw or handwritten in-context examples. Additionally, Figure 5 illustrates our method's support for continual learning without forgetting. Furthermore, our optimized in-context examples are applicable to various multimodal agents across three complex domains: long-horizon instruction-following for household robots, visually-grounded web navigation, and action forecasting from real-world egocentric videos. Reviewers vg3K and VRBk note that these experiments on "sufficiently different tasks... suggest the generalizability of the proposed method." Previous works have been limited to text-based environments, gaming scenarios, or single real-world domains. To our knowledge, we are the first to show robust results across these three diverse and complex domains. We position our approach as a superset that incorporates the strengths of previous VLM agent works, while introducing a novel focus on optimizing in-context learning examples. As Reviewer VRBk puts it, ''the overall goal of using VLMs to distill demonstrations into canonical examples is an interesting combination of foundation models with classical ideas such as case-based learning.'' Our approach can be likened to in-context policy improvement with a multi-modal LLM actor, where instead of aiming to maximize rewards through specific reasoning strategies explored in previous works, it refines and optimizes in-context examples. >The overall prompting strategy is quite sensitive to the model performance. The fine-tuning improvement is quite limited as discussed in 4.6. We apologize for not clarifying these points in our paper. Specifically, we show that fine-tuning CoT-prompted GPT-3.5 significantly improves performance, doubling its success rate from 11.8% to 23.2%. Additionally, incorporating retrieval-augmented generation (RAG) with fine-tuning (using ICAL examples as memory) provides further improvement, resulting in our best-performing agent. This agent outperforms the non-finetuned memory-augmented CoT-prompted GPT-3.5 by 4.9% in goal-condition success. We view weights and external memory of abstractions as two forms of memory with distinct benefits. Weight fine-tuning requires many examples, while RAG can learn from a single example. In scenarios with limited data (150 examples or fewer), external memory updates and RAG use data more efficiently and remain competitive with weight fine-tuning. This is relevant to our study as our considered domains fit this category. >Discussion on the limitations of our work. We will expand our limitation section per your request. Below, we include additional limitations and error modes of our agent with specific examples. We will include this discussion in the main paper. 1. Visual Grounding Limitations of GPT-4V While ICAL improves performance in visually grounded tasks, errors persist due to the base VLM's limitations in visual grounding. For example, the agent fails to identify colors accurately, leading to errors like selecting the wrong item. This issue is evident in the "reddit_2" and "reddit_9" web tasks, where the agent navigates to incorrect posts, showing a failure to match images to webpage tabs. The agent also occasionally struggles with pre- and post-condition recognition. We found that in-context multimodal examples helped grounding in cases where grounding elements were unique, but easily identifiable but continue to fail in fine-grained cases. This grounding limitation has been noted in previous work as well [1]. We expect improvements with more data, added grounding objectives [2], and fine-grained image annotations [3]. In future work, we plan to extend ICAL for better fine-grain language grounding using multimodal retrieval, building on methods like ViperGPT [4]. 2. Fine-grained in-context planning failures During learning phases, ICAL may not fully acquire all necessary information for test time, resulting in failures. For instance, in the "shopping_3" web task, where the instruction is to display the most expensive red controller from the "PS4 accessories" category, the agent navigates to the PS4 category but fails to access the accessories subcategory. This demonstrates a lack of understanding of website structures and navigation. This limitation suggests that agents need to recognize when their knowledge base is insufficient and query the user for missing information. We plan to address active querying during testing in future work. References [1] Zheng et. al. GPT-4V(ision) is a Generalist Web Agent, if Grounded. [2] Ma et al. (2024). Groma: Localized Visual Tokenization. arXiv:2404.13013. [3] Garg et al. (2024). ImageInWords: Hyper-Detailed Image Descriptions. arXiv:2405.02793. [4] Suris et. al. (2023) ViperGPT: Visual Inference via Python Execution for Reasoning. --- Rebuttal 2: Title: Response to the rebuttal Comment: I thank the authors for answering my questions and addressing my concerns. I am more convinced with the contribution of optimizing few-shot examples now and how it differs from previous work. I will raise my score to 5. I am still generally concerned with the overall limitation of the work --- I am glad the authors provide additional discussions on it and I hope they will be added in the revised manuscript, but I have to say I am generally less convinced of prompting techniques that build on previous strategies, despite the performance improvement that the authors demonstrate. --- Rebuttal Comment 2.1: Title: Thank you! Comment: Thank you for your thoughtful feedback and for raising your score to 5. We appreciate your recognition of our work in optimizing few-shot examples and clarifying how our approach differs from prior work. We understand your concerns about building on existing strategies. In the revised manuscript, we will include a thorough discussion of the limitations and how our approach differs from previous work.	Summary: The goal of this paper is to teach VLMs novel tasks by prompting VLMs to create multimodal abstractions for unfamiliar domains. Given instructions paired with noisy demonstration trajectories, this paper proposes a method to encapsulate the information from these into examples consisting of optimized trajectories paired with generalizable language abstractions. A VLM is specifically prompted to produce abstractions such as essential task steps, cause and effect generalizations, expected state changes, a step by step plan, and relevant parts of state information. The abstracted example includes an executable trajectory which is executed and potentially refined with a human in the loop via natural language feedback. These examples are then used for in context of a VLM to improve task performance. The paper includes experiments on 3 datasets - TEACh, VisualWebArena and Ego4D, comparing respectively to the HELPER model - a prior SoTA model for TEACh, and GPT4V for VisualWebArena and Ego4D. The most improvement in seen in VisualWebArena, followed by partial success on TEACh. On Ego4D, performance is on par with supervised learning methods that use more data but Zero shot CoT with GPT4V performs better. Strengths: The overall goal in this paper of using VLMs to distill demonstrations into canonical examples is an interesting combination of currently popular foundation models with more classical ideas such as case based learning. The paper includes experiments on multiple benchmarks with sufficiently different tasks to suggest generalizability of the proposed method. It is also refreshing to see that the authors do not hesitate to include the unusually good performance of Zero shot GPT4V on Ego4D which might be seen as a negative result. Given the topic of the paper, a number of engineering details and information about prior work is needed to fully understand the paper and increase the likelihood of reproducibility of the work. The authors have made a very strong attempt at this with a detailed appendix. Weaknesses: While the appendix goes a long way towards reducing this, it is likely that the paper might still be difficult to follow for readers unfamiliar with the datasets and prior methods referenced in this paper. Technical Quality: 4 Clarity: 4 Questions for Authors: It is possible I missed this in the paper, but I how was the human in the loop feedback obtained for the TEACh experiment? Given that the checklist says no human subjects were recruited, I assume this feedback was provided by authors. Is this correct? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors include a discussion of limitations and potential negative social impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 9 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >While the appendix goes a long way towards reducing this, it is likely that the paper might still be difficult to follow for readers unfamiliar with the datasets and prior methods referenced in this paper. Thank you for your feedback. We acknowledge that the paper may still be challenging for readers unfamiliar with the datasets and prior methods. To address this, we will provide more detailed explanations on dataset processing and implementation in the appendix in the camera-ready version and the README in the code repository. >It is possible I missed this in the paper, but how was the human in the loop feedback obtained for the TEACh experiment? Given that the checklist says no human subjects were recruited, I assume this feedback was provided by authors. Is this correct? Yes, the authors and their lab colleagues provided the human-in-the-loop feedback in the paper. We did not recruit participants due to resource constraints, specifically the costs associated with recruiting participants for running ICAL and ablations. During this phase, we translated task progress into natural language feedback during failures, including failed actions (e.g., “The toaster is full right now.”) and missed task steps (e.g., “The pillow needs to be put onto the sofa”). The feedback provided did not require domain expertise, such as focusing on precise coding or low-level action changes. For example, typical feedback from VisualWebArena includes comments like, "This does not have 2 upvotes and includes a meme of three Spider-Men, a flag of the Netherlands, and a flag of Croatia. You should scroll down to see more posts." In TEACh, typical feedback is of the form, "The sink is full right now. Empty the sink before interacting with it." We will make these details more clear in the main paper. >Performance comparison on Ego4D. To clarify, zero-shot CoT with GPT-4V shows strong results, but ICAL with GPT-4V outperforms it. Specifically, ICAL achieves lower edit distances relative to GPT-4V zero-shot COT: increases in performance by 9.9 for verbs, 10.0 for nouns, and 4.0 for actions.	Rebuttal 1: Rebuttal: We appreciate the reviewers' positive feedback. Reviewer VRBk commended our innovative combination of foundation models with case-based learning, robust generalizability across benchmarks, and transparency in results. Reviewer ih5w highlighted ICAL's improved success rates across domains, and our clean, intuitive structured representation. Reviewer vg3K noted our framework's advancements over zero-shot CoT baselines, detailed experimental setup aiding reproducibility, and potential for helping humans understand complex tasks. Reviewer foCb noted ICAL's superior performance across various domains and continual integration of human feedback to improve over time. We address overall concerns in this global response and specific reviewer concerns in their respective responses. >@Reviewers vg3K, foCb: How efficient is the human-in-the-loop phase? Is it scalable? Our agent becomes increasingly efficient over time, requiring less human feedback and fewer environment interactions as it processes more examples. By retrieving past successful abstractions during the VLM-abstraction making and human-in-the-loop phases, it uses previously stored knowledge to help abstract new examples. As shown in Figure R1, for the second half of examples processed, the model requires significantly fewer environment steps (436±88 vs. 267±43, p=0.0143) and human feedbacks (0.74±0.17 vs. 0.21±0.08, p=0.0089) per example. This demonstrates that retrieving abstracted examples during abstraction learning reduces both human effort and environment interaction over time. Consequently, using previously stored ICAL examples not only improves test performance but also accelerates learning for future examples. It is additionally essential to note that this feedback is sparse, provided in natural language, and required only a few times for each task. In VisualWebArena, our agent needs an average of just 5.36 natural language feedbacks per example, with each feedback taking less than 15 seconds to create. In TEACh, our agent needs just 1.516 average feedbacks per episode with an average length of 18 words per natural language feedback. In fact, natural language feedback in ICAL could be easily communicated via speech by a human while observing the agent, whereas low-level coding and abstraction writing communication would be significantly more laborious. In order to hand-write GPT4-V-length abstractions, this would require 202.62 words on average for each example in VisualWebArena and 48-107 lines of text for each example in TEACh, including the precise programming of actions or code. >@Reviewers foCb, ih5w: Does the method rely on GPT4V? How sensitive is it to the prompting strategy? In Table S2 of the Appendix, we provided an ablation study using GPT-3.5 for the VLM-driven Abstraction Generation on the TEACh dataset, where visual experiences are converted to textual format. The results indicate a performance drop when using GPT-3.5 for ICAL Abstraction Generation compared to GPT4V (52 versus 122 tasks successfully completed). In Section S3.4 of the Appendix, we demonstrate how relabeling unsuccessfully refined trajectories can help bridge this performance gap when using weaker models like GPT-3.5, thereby reducing the dependence on GPT-4V. We will highlight these findings more prominently in the main paper to ensure their visibility. Per your comments, we implemented LLaVa-NeXT for the abstraction phase in VisualWebArena given multimodal inputs. As shown in Figure R2, we found that it failed to properly revise the abstractions to take into account the feedback. However, we are committed to continuing our experiments with emerging open-source VLMs as they become available. We further show in Section 4.6 that fine-tuning CoT-prompted GPT-3.5 significantly improves performance, doubling its success rate from 11.8% to 23.2%. Additionally, incorporating retrieval-augmented generation (RAG) with fine-tuning (using ICAL examples as memory) provides further improvement, resulting in our best-performing agent. This agent outperforms the non-finetuned memory-augmented CoT-prompted GPT-3.5 by 4.9% in goal-condition success. In scenarios with limited data (150 examples or fewer), external memory updates and RAG use data more efficiently and remain competitive with weight fine-tuning. >@Reviewers ih5w: How does this differ from CoT, ReACT, Socratic Model, and RAG? While our framework incorporates CoT, interleaves reasoning and acting as seen in ReACT, predicts plans over pretrained modules similar to Socratic models, and uses RAG for inference, it is distinct in its primary contribution: optimizing examples for improved in-context learning. We introduce a novel two-stage approach that optimizes examples for improved in-context learning. To the best of our knowledge, this is the first work to do this. As Reviewer VRBk puts it, ''the overall goal of using VLMs to distill demonstrations into canonical examples is an interesting combination of foundation models with classical ideas such as case-based learning.'' Our results, in Tables 1-3, demonstrate that both stages are crucial for high performance, significantly outperforming CoT and RAG, which use raw or handwritten in-context examples. Additionally, Figure 5 illustrates our method's support for continual learning without forgetting. Furthermore, our optimized in-context examples are applicable to various multimodal agents across three complex domains: long-horizon instruction-following for household robots, visually-grounded web navigation, and action forecasting from real-world egocentric videos. Reviewers vg3K and VRBk note that these experiments on "sufficiently different tasks... suggest the generalizability of the proposed method." Previous works have been limited to text-based environments, gaming scenarios, or single real-world domains. To our knowledge, we are the first to show robust results across these three diverse and complex domains. Pdf: /pdf/7c71d373f563762aa6be46ff3f3e80f812bbbeee.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Computerized Adaptive Testing via Collaborative Ranking	Accept (poster)	Summary: In this paper, the authors first discovered the inconsistency issue in existing Computerized Adaptive Testing (CAT) solutions for estimating the latent abilities of students, that is the higher accuracy of ability estimation (with a lower MSE) does not necessarily guarantee the ranking consistency of students’ abilities. Based on this discovery, the authors then proposed a novel Computerized Adaptive Testing framework CCAT inspired by collaborative ranking. Specifically, CCAT uses collaborative students as anchors to assist in test-question selection and estimation in testing. More importantly, the authors provide a theoretical analysis of the upper bound of ranking consistency error for collaborative students, which verifies that with an adequate number of collaborative students, the ranking consistency error can be reduced to an acceptable level. Through experiments on two real-world datasets, the authors demonstrated that CCAT can achieve the best ranking consistency. Strengths: 1. Computerized Adaptive Testing is a cross-cutting research direction between artificial intelligence and the testing area, with broad applications in the real world. In this paper, the authors first discovered the inconsistency issue in existing Computerized Adaptive Testing (CAT) solutions for estimating the latent abilities of students, and therefore, the research motivation is both original and significant. 2. The proposed Collaborative Computerized Adaptive Testing framework CCAT exploits the idea of collaborative students to address the incomparable test-answering behavior problem of different students. This idea is quite novel and differs from traditional CAT solutions and collaborative ranking solutions. 3. Both the theoretical analysis and the experimental validation seem to be solid and convincing. Weaknesses: 1. It is not easy to illustrate the main ideas of CAT as it contains both the Ability Estimation Part and the Question Selection Part. Therefore, the readability of this paper can be further improved especially for the readers without any background of CAT. 2. The main idea of the CCAT algorithm can be explained in more detail in the main text. For instance, how can we get the collaborative students? 3. More references about collaborative ranking are recommended to be included and discussed. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. How can we get the collaborative students? Are these collaborative students the same as all the testing students? 2. From Equation (5), the response of the students belongs to 0 or 1, which means the student's responses studied in this paper are either right or wrong, what if the rating of the student responses has more choices (like the values in the range of [0,1])? Does the CCAT solution still work in this scenario? 3. How to determine the parameter T (testing round) in CCAT? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback on our manuscript. We sincerely appreciate your time and effort in evaluating our work, and we appreciate you for the opportunity to explain and articulate our work. > Q1: It is not easy to illustrate the main ideas of CAT as it contains both the Ability Estimation Part and the Question Selection Part. Therefore, the readability of this paper can be further improved especially for the readers without any background of CAT. > > Q2: The main idea of the CCAT algorithm can be explained in more detail in the main text. We appreciate your observation regarding the difficulty in illustrating the main ideas of CAT, particularly as it involves both the Ability Estimation Part and the Question Selection Part. To address this, we will optimize the content of our article to enhance its readability, especially for readers who do not have a background in CAT. We will ensure that our explanations are more accessible and clear, incorporating additional context and examples where necessary. > Q3: How can we get the collaborative students? > > Q4: Are these collaborative students the same as all the testing students? In the experimental phase, collaborative students can be constructed by splitting the dataset and supplementing it with the predicted values from the IRT, as detailed in Appendix Section C. We believe that there is no essential difference between these collaborative students and testing students, only that collaborative students have answer records while testing students do not. In real exam scenarios, generally speaking, in real CAT systems such as GRE, a group of students will take tests in advance to generate test records[1]. Moreover, we can also use students who have previously taken exams as collaborative students. > Q5: More references about collaborative ranking are recommended to be included and discussed. We appreciate the importance of providing a comprehensive context for our work and will include additional references related to collaborative ranking in future versions of our paper. This will help situate our research within the broader academic discourse and offer readers a more thorough understanding of the field. > Q6: From Equation (5), the response of the students belongs to 0 or 1, which means the student's responses studied in this paper are either right or wrong, what if the rating of the student responses has more choices (like the values in the range of [0,1])? Does the CCAT solution still work in this scenario? Indeed, changing the range of question values implies that the traditional IRT would no longer be applicable, and consequently, the CCAT algorithm cannot be directly used in its current form. However, we believe that our approach remains universal. Specifically, for any given evaluation model, we can still optimize the accuracy of ability ranking among collaborative students by enhancing the question selection process. Ultimately, collaborative student voting can be employed to rank the tested students effectively. > Q7: How to determine the parameter $T$ (testing round) in CCAT? In general, for research purposes in CAT problems, the parameter $T$ is often set to fixed values such as 5, 10, 15, or 20. These fixed values provide a standardized way to compare different methods and results. However, in practical application scenarios, the termination of T can be more dynamic. It can be based on specific indicator change amplitudes, such as the ability change value or ranking change value. The testing rounds can be terminated once these changes are less than a predefined threshold, ensuring that the testing adapts to the candidate's performance and achieves optimal efficiency. Reference: [1] Computerized adaptive testing: Theory and practice[M]. Dordrecht: Kluwer Academic, 2000. --- Rebuttal Comment 1.1: Comment: Many thanks for the authors' rebuttals. My questions and concerns have been well-addressed in the rebuttal, so I want to increase my recommendation score from 7 to 8. --- Reply to Comment 1.1.1: Comment: Thank you for your time and thoughtful evaluation. It's great to hear that all your questions and concerns have been successfully addressed. Your insights and suggestions are valuable to us, and we will include the discussed information in the future version.	Summary: This paper addresses a real-world problem in AI education: improving the accuracy of student rankings by selecting different questions during the exam process. It proposes a question selection method based on collaborative students and provides theoretical guarantees. The experimental results have demonstrated the effectiveness of its method in ranking consistency. Strengths: 1. To my knowledge, the perspective of this paper is novel. It starts from a real exam scenario and defines the CAT problem as a ranking problem, which has not been solved in previous CAT research. This means that based on this work, numerous ranking methods may be incorporated into CAT problems. 2. The paper demonstrates the method's superior ranking accuracy through experiments on real-world datasets. The approach exhibits general applicability across CAT systems estimated by IRT or GD. 3. The logic of this paper is clear, and the supplement materials include all theoretical proofs and several additional experiments, making this paper easy to follow. Weaknesses: 1. Can you clarify the detail of \theta^T_c? Furthermore, since the true abilities of "collaborative students" are known, why use the abilities of collaborative students at T-moment instead of their true abilities. 2. The method proposed in the paper seems to have value primarily for educational research. While this is an important domain, the paper could benefit from discussing the potential applicability in other fields or providing more insights into how the proposed approach could be generalized beyond the education sector. 3. Ranking is a common issue in recommendation systems. After defining CAT tasks as ranking problems in this paper, what are the similarities and differences between CAT and recommendation tasks? Technical Quality: 4 Clarity: 3 Questions for Authors: See Weaknesses part Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Regarding the questions you raised, we have carefully considered each point and have made following responses: > Q1: Can you clarify the detail of $\theta^T_c$? Furthermore, since the true abilities of "collaborative students" are known, why use the abilities of collaborative students at T-moment instead of their true abilities. The parameter $\theta^T_c$ represents the ability assessment results obtained by collaborative students who have answered the same T questions as the students being tested. This approach ensures that the collaborative students' ability assessments are based on a comparable set of questions, providing a more accurate basis for scoring the tested students. Using the true abilities of collaborative students ($\theta^_c$) for each student would mean that the collaborative students would lose their sensitivity to the question selection process. This would hinder the optimization of ranking through question selection, as the adaptive nature of the test would be compromised. > Q2:* The method proposed in the paper seems to have value primarily for educational research. While this is an important domain, the paper could benefit from discussing the potential applicability in other fields or providing more insights into how the proposed approach could be generalized beyond the education sector. Our method performs particularly well when the number of CAT test questions is small, addressing the cold start problem inherent in CAT. This characteristic suggests that our research may be applicable to cold start problems in other fields as well. For instance, our approach could be valuable in personalized recommendation systems, where limited initial data is a common challenge. Additionally, it could be useful in any domain that requires adaptive testing or assessment with sparse initial data. > Q3: Ranking is a common issue in recommendation systems. After defining CAT tasks as ranking problems in this paper, what are the similarities and differences between CAT and recommendation tasks? CF and CCAT share some underlying principles, but they are fundamentally different in their applications and objectives. CF is a recommendation technique that estimates item preferences based on previous users' behaviors, while CCAT adapts the test items based on the examinee's ability level, which is estimated dynamically during the test. Moreover,CF aim to suggest items based on user preferences (ranking items), whereas CAT necessitating precise estimation of students' abilities with minimal interaction (ranking users). In CF, the item ranking of a user is fixed, but in CAT, selected questions affect the ability evaluation through the answers, impacting the ranking. This variability makes CAT tasks more challenging than typical CF problems. --- Rebuttal Comment 1.1: Comment: Thank you for your response. After considering your feedback, I have decided to maintain my score. --- Reply to Comment 1.1.1: Comment: Thank you for your time and thoughtful evaluation. Your insights have been invaluable to us and will certainly help in refining our research.	Summary: The paper proposes an algorithm for performing computerized adaptive testing that handles and accounts for student rankings in the item recommendaiton. Strengths: - I like the idea of incorporating additional information in the collaborative filtering approach (while I don't quite understand why you need to intermittently rank students during CAT) Weaknesses: My main critique of the paper is the work's motivation for connecting ranking and CAT: Ranking seems like a completely separate task from CAT, and would happen after CAT. The paper's introduction would benefit from a concrete example where online updates in collaborating filtering is important when administering tests. It seems like there would be issues during testing such as handling the nonstationarity of question difficulties and forcing students to start and complete every question at the same time before proceeding onto the next question. It's not clear why you need intermediate ranks amongst students as they are getting tested. Looking at Alg 1, it feels like the work should use/estimate student rankings not for the sake of estimation/collaborative filtering, but rather selecting questions that best differentiate students -- so distinguish their abilities to further diagnose students. But I'm curious what the authors originally had in mind with connecting online ranking estimations with CAT, because currently... I don't quite see how the algorithm assumptions can hold in the real world (see paragraph above). Technical Quality: 2 Clarity: 1 Questions for Authors: - Could the authors describe how their work differs from collaborative filtering approaches where estimation of item difficulty and user ability can be done through previous users and their attempted items? - It seems strange to me that the collaborative records used for testing are being simultaneously updated at every step t (rf. Algorithm 1). Why do we need to update the student ranks online? - Theorem 1 seems more like a statement about the collaborative filtering approach in estimating item difficulty and user ability, and is not really about estimating the ranking. Could the authors cleave the effects of items and users on the IRT estimation from the ranking estimation in their Theorem? - While Figure 3 points to areas of improvements, I also noticed areas where their method does worse: e.g., on the second row of the NIPS-EDU. Could the aggregate difference be reported instead of the heat map visualization? Confidence: 2 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: The authors do not provide a separate Limitations section in their paper. While the authors state they provided the limitations in Appendix D and experiments, these seem to be more like "findings" from empirical observations than limitations of their work/framing/approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback on our manuscript. We sincerely appreciate your effort in evaluating our work. Below, we address each of your comments in detail: > Q1: My main critique of the paper is the work's motivation for connecting ranking and CAT: Ranking seems like a completely separate task from CAT, and would happen after CAT. You are right. Ranking typically occurs after the CAT process, but it is inherently interconnected with CAT and is also influenced by CAT . Existing CAT methods focus solely on the accuracy of abilities (Figure 4), overlooking the importance of ranking in CAT (Figure 1). This oversight can lead to potential biases and inequities in selection tests. Our research aims to address this gap by enhancing the testing process. > Q2: Could the authors describe how their work differs from collaborative filtering approaches? CF and CCAT share some underlying principles, but they are fundamentally different in their applications and objectives. CF is a recommendation technique that estimates item preferences based on previous users' behaviors, while CCAT adapts the test items based on the examinee's ability level, which is estimated dynamically during the test. Moreover, CF aim to suggest items based on user preferences (ranking items), whereas CAT necessitating precise estimation of students' abilities with minimal interaction (ranking users). In CF, the item ranking of a user is fixed, but in CAT, selected questions affect the ability evaluation through the answers, impacting the ranking. This variability makes CAT tasks more challenging than typical CF problems. > Q3: It seems like there would be issues during testing such as handling the nonstationarity of question difficulties and forcing students to start and complete every question at the same time. We would like to clarify that in traditional CAT models, the difficulty of questions is pretrained before testing and is fixed throughout the testing process. This assumption is based on the fact that the difficulty of a question is determined by its inherent attributes [1]. Meanwhile, students do not need to wait for other students before proceeding onto the next one. Each student, upon completing a question, will compare their performance with a group of "collaborative students", which are defined as students who have already answered questions in question bank in Definition 1. Generally speaking, in real CAT systems such as GRE, a group of students will take tests in advance to generate test records [2], which indicates that our define and hypothesis is reasonable. > Q4: Why are intermediate ranks needed amongst students as they are getting tested? We do not dynamically update students' rankings during the CAT process. We only update students' abilities after each question (like other CAT) and use the comparison results between the tested students and collaborative students for rankings after testing. > Q5: Why are collaborative records used for testing being simultaneously updated at every step t (ref. Algorithm 1)? Collaborative records are not updated every round. Instead, we extract useful records from the collaborative records corresponding to the questions that the current student has answered. > Q6: Could the authors cleave the effects of items and users on the IRT estimation from the ranking estimation in their Theorem? Yes. Collaborative students are only a prerequisite of Theorem 1 rather than what it focuses on. Theorem 1 can be understood as a process where collaborative students vote for Student A and Student B (A is better than B). It claims that as long as there are enough collaborative students, A's votes will surpass B's votes. This explains why we utilize collaborative students to vote for ranking the students being tested. > Q7: Could the aggregate difference be reported instead of the heat map visualization? Yes. We originally hoped to visually demonstrate the advantages of CCAT's results compared to IRT's results through a heatmap. Below is the aggregate difference representing the average improvement of each student in 20 steps using CCAT compared to IRT (positive value means improve and negative value means decline): \| Aggregate Difference \| Average \| 1 \| 2 \| 3 \| 4 \| 5 \| 6 \| 7 \| 8 \| 9 \| 10 \| \| -------------------- \| ------- \| ----- \| ----- \| ---- \| ----- \| ---- \| ---- \| ----- \| ---- \| ---- \| ---- \| \| NIPS-EDU \| 0.18$\uparrow$ \| -0.15 \| 0.18 \| 0.22 \| 0.19 \| 0.13 \| 0.24 \| 0.22 \| 0.49 \| 0.06 \| 0.16 \| \| JUNYI \| 0.004$\uparrow$ \| -0.08 \| -0.09 \| 0.11 \| -0.12 \| 0.26 \| 0.09 \| -0.21 \| 0.00 \| 0.07 \| 0.01 \| Our method may not achieve better ranking relationships among all students, but from the table and various experimental results, our method can indeed improve the overall ranking consistency of students. > Q8: While the authors state they provided the limitations in Appendix D and experiments, these seem to be more like "findings" from empirical observations than limitations of their work. Sorry for the confusion. As mentioned in the Experiment section, we have stated limitations in line 262 that CCAT may not perform as well on long test sequences as methods that directly optimize capabilities. To make it clearer, we will separate statements of limitations into a new section in the new version of this paper. Reference: [1] Wainer H, Dorans N J, Flaugher R, et al. Computerized adaptive testing: A primer[M]. Routledge, 2000. [2] Computerized adaptive testing: Theory and practice[M]. Dordrecht: Kluwer Academic, 2000. --- Rebuttal Comment 1.1: Comment: Thanks for your response and clarification -- I see, so you're using the previous traces of other participants to do selection, instead of just the current participant's ability and the difficulties from your item bank. I've raised my score. I am curious, when authors say that that their algorithm shows "significant improvement", have you run a statistical sig test to verify this? I think it's generally good practice to state the test (at least in a footnote) and mark with the appropriate sig values when using "significant" as a descriptive. Thanks! --- Rebuttal 2: Comment: Thank you for your positive feedback. We greatly appreciate your recognition of our efforts to address your concerns. Regarding your follow-up question, we did conduct a statistical significance test to verify the improvements reported in our algorithm. As shown in Tables 4 and 5 of Appendix D, we performed the tests and found significant results at the p < 0.05 level. However, we apologize for not explicitly mentioning this in the main text of our paper. We will ensure to include this information in the revised version of the main text, with the appropriate significance values clearly marked. If you have any other questions, please feel free to ask.	Summary: This paper proposes a new perspective on Computerized Adaptive Testing (CAT) by framing it as a task of ranking students. The authors define CAT as a ranking problem and present a feasible optimization algorithm to address this. Extensive experimental results demonstrate that this method significantly improves the consistency of student ranking scores compared to the baseline system. Strengths: 1. The paper explores a previously overlooked issue in CAT and human assessment domain—student ranking consistency. It redefines the CAT task within this context. The solution proposed is interesting and interpretable, aligning well with real educational scenarios. 2. In terms of technical implementation, this paper utilizes existing students as ranking anchors to enhance selection and evaluation methods. It provides a theoretical basis, demonstrating that the algorithm can reduce ranking consistency errors to an acceptable level. 3. Experiments on real-world datasets have shown that the proposed method improves ranking consistency by an average of 5% compared to baseline selection methods. Weaknesses: 1. This paper has rich and convincing experiments, but I want to know why Table 1 only shows the results of BOBCAT in the Table 1(a). Furthermore, this paper does not explain why BOBCAT has such a significant difference in performance on this task. 2. In Table 1 (a), NCAT performs best on the NIPS-EDU dataset at T=5, but this result is not discussed in this paper. 3. This paper mainly analyzes the question selection and estimation of the CCAT method. In fact, according to my understanding, the construction of collaborative students is long-term and complex in real educational scenarios. Is there a method to ensure the effective construction of collaborative students? Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper and its supplementary materials clearly illustrate its limitations. Due to the need for collaborative students to be stored, the algorithm’s time and space complexity should be more thoroughly discussed. Providing detailed analyses and potential optimization strategies could strengthen the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your comments! To address your concerns, below we prudently justify the details of our proposed method and experiments. > Q1: This paper has rich and convincing experiments, but I want to know why Table 1 only shows the results of BOBCAT in the Table 1(a). Furthermore, this paper does not explain why BOBCAT has such a significant difference in performance on this task. As a bilevel optimization model, BOBCAT optimizes both the IRT model and the question selection strategy. This means it employs a separate GD model and does not require training an additional IRT model through the MCMC method. Consequently, only one result for BOBCAT is shown in Table 1(a). The significant difference in BOBCAT's performance can be attributed to its two-layer optimization approach, which leads to peak accuracy at specific times (as illustrated in Appendix Figure 4, ACC), but poor performance at other times. Additionally, training the IRT model can lead to instability in estimating real ability, ultimately resulting in poor performance in ranking metrics. > Q2: In Table 1 (a), NCAT performs best on the NIPS-EDU dataset at T=5, but this result is not discussed in this paper. The CAT ranking problem we are studying can essentially be seen as an degradation problem for accuracy. When optimizing for accuracy, we may randomly achieve higher or lower ranking consistency. As observed in Table 1, although NCAT performs well on the NIPS-EDU dataset with T=5, its performance on the JUNYI dataset is even worse than that of random selection. This indicates that the NCAT method may not be effective in consistently optimizing ranking problems. We will include a discussion of these observations in the revised version of our paper to provide a comprehensive analysis of NCAT's performance across different datasets. > Q3: This paper mainly analyzes the question selection and estimation of the CCAT method. In fact, according to my understanding, the construction of collaborative students is long-term and complex in real educational scenarios. Is there a method to ensure the effective construction of collaborative students? Currently, the selection and construction of collaborative students are relatively simple processes. Generally speaking, in real CAT systems such as GRE, a group of students will take tests in advance to generate test records [1]. If CAT is viewed as a long-term process, while CAT increases students' learning efficiency, it also leads to the sparsification of student data. As data becomes sparser, the effectiveness of CAT and IRT may be impacted. To date, only one study has addressed the bias in IRT data [2]. Therefore, more research is needed to ensure the long-term construction and maintenance of collaborative students. Reference: [1] Computerized adaptive testing: Theory and practice[M]. Dordrecht: Kluwer Academic, 2000. [2] Kwon S, Kim S, Lee S, et al. Addressing Selection Bias in Computerized Adaptive Testing: A User-Wise Aggregate Influence Function Approach[C]//Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2023: 4674-4680. --- Rebuttal Comment 1.1: Comment: These responses address my questions and strengthen my score. --- Reply to Comment 1.1.1: Comment: Thank you for your time and thoughtful evaluation. It's great to hear that all your questions have been successfully addressed and your acknowledgment means a lot to us.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: The study deals with CAT( Computer Adaptive Testing) via Collaborative Ranking. Strengths: The proposed CCAT algorithm demonstrates superior performance in ranking 259 consistency across two public datasets. Particularly, CCAT shows more significant improvement 260 when fewer questions are tested, outperforming other methods. Weaknesses: The study does not discuss in sufficient detail the following: limitations, generalization possibility and the application of the algorithm as part of instructional design. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Q1: The study does not discuss in sufficient detail the following: limitations, generalization possibility and the application of the algorithm as part of instructional design. Thank you for your valuable feedback. We apologize that we have shown in the experimental and appendix sections, rather than divided it into paragraphs for description and we will supplement this section in future versions of our paper: Limitations: As mentioned in section Experiment, we have stated limitations in line 262 in this version that CCAT may not perform as well on long test sequences as methods that directly optimize capabilities. To make it clearer, we will separate statements of limitations into a new section in the new version of this paper. Generalization Possibility: Our method performs particularly well when the number of CAT test questions is small. Additionally, since CAT is inherently a cold start problem, our research may be applicable to other fields facing similar challenges. For example, personalized recommendation systems, healthcare diagnostics, and user behavior prediction are areas where cold start problems are prevalent, and our approach could potentially be generalized to improve performance in these fields. Application as Part of Instructional Design: In instructional design, the CCAT algorithm can be integrated to create adaptive learning paths tailored to individual student abilities. By dynamically adjusting the difficulty and selection of questions based on ongoing assessments, educators can provide a more personalized and effective learning experience. However, careful consideration must be given to the practical implementation and potential limitations discussed above.	null	null	null	null	null	null
Federated Behavioural Planes: Explaining the Evolution of Client Behaviour in Federated Learning	Accept (poster)	Summary: The paper introduces Federated Behavioural Planes (FBPs), a method designed to track FL clients' behavior by means of examining the their representations in two behavioural planes with the aid of a server-owned dataset. The Error Behavior Plane (EBP) and Counterfactual Behavioural Plane (CBP) correspond to two 2-D space where each client model's prediction errors and counterfactural distribution are examined. Combining these two representations a new aggregation rule is proposed to fend off malicious clients. Strengths: - The paper is overall well written. The idea presented is clear and easy to follow. - The method of characterizing clients' behavior is kind of novel. Weaknesses: - Creating the two plains require the plaintext of all local models, which raises major privacy concerns. - From Eqs. 6 and 7, I suppose generating the counterfacturals must be very costly. - Insufficient experimental evidence to support the claim. For example, fig. 3 does not show distinct client behavior clearly. I also expect to see more results such as the behavioural scores of different clients and how sensitive the FBP is to the intensity of attacks. Technical Quality: 3 Clarity: 3 Questions for Authors: - Results in Table 1 make me confused whether the whole method is still run in a cannonical FL setting. - What is the difference between proximity and sparsity? - How is counterfactural generator trained and why does it impact the task accuracy? - What purpose does Eq. 4 serve? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: - A framework overview is missing, which makes it hard to understand how the planes are created throughout the FL iterations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: _Privacy concerns of transmitting the plaintext of all local models_ __Our method can integrate Local Differential Privacy (LDP) or Homomorphic Encryption (HE) to enhance privacy, ensuring that sensitive information remains protected while maintaining robustness.__ Our framework, like traditional FedAvg, is susceptible to privacy leakage if plaintext local model parameters are exposed. However, it can incorporate LDP, where clients add noise to the model before transmission, maintaining DP while creating behavioral planes. Although this process may result in some performance loss, it is comparable to other robust aggregation methods that calculate client similarities using differentially private model parameters. To further enhance privacy, our framework can integrate HE, which masks client models while allowing for inference on the validation set. Although this increases computational time during the aggregation, it ensures that sensitive information remains completely concealed from the server, thereby addressing privacy concerns effectively. We added a sentence in L348: "Since our method allows the integration of privacy-enhancing techniques such as LDP and HE, a promising future direction would be to analyze their impact on performance and efficiency". _Insufficient experimental evidence regarding the distinction of client behaviors through Behavioral Planes (BPs). Expected more results, e.g. the client behavioral scores_ __We have strengthened the experimental evidence by analyzing behavioral scores extracted from BPs, demonstrating that our method accurately identifies and mitigates malicious clients across all attacks.__ As suggested, we analyze the behavioral scores of honest and malicious clients across all attacks (Fig. B), which are calculated over the BPs. The 5-fold experiments were conducted on the CIFAR10 (most complex dataset), and show the mean and 95% confidence intervals during the training rounds. Statistically, our method consistently identifies malicious clients, effectively excluding or reducing their weights during aggregation. These results further demonstrate that BPs accurately describe malicious clients during the training process. _How sensitive the proposed method is to the intensity of the attacks_ __Our method remains stable and unaffected by increasing attack intensity, demonstrating robustness against model poisoning attacks.__ We analyzed the sensitivity of our method to attack intensity by increasing the amount of noise $\beta$ that attackers add to the global model before sending it to the server [L677]. We conducted 5-fold experiments on both the Breast Cancer and MNIST datasets. As shown in Fig. E, increasing attack intensity decreases FedAvg’s performance, while our methods (using all planes and only the counterfactual plane) remain stable and unaffected by the attack intensity. Interestingly, our methods’ accuracy slightly increased as attack intensity rose, likely because the malicious models became more degraded and, therefore, less stealthy. _A framework overview is missing, which makes it hard to understand how the planes are created throughout the FL iterations_ We thank the reviewer for this observation. To address this limitation, __we have provided a detailed algorithm (Algorithm 1) for our method to create the behavioral planes and implement our robust aggregation (FBSs)__. While Algorithm 1 covers the server operations, we will also include the algorithm for the client operations in the Appendix A.5 of the paper, which reflects the traditional FL algorithm. _Results in Table 1 make me confused whether the whole method is still run in a cannonical FL setting_ __Yes, our method operates within a canonical FL setting, with changes to server-side aggregation and the introduction of the counterfactual (CF) generator. The CF generator is trained jointly with the local client and does not alter the performance of the predictor alone (Table 1)__. For clarity, we updated L230-231: "We compared model accuracy across four settings: two centralized learning (CL) and two FL, using three different datasets under non-IID conditions. For FL experiments, we used the traditional FedAvg approach with two variations: predictor-only and predictor with CF generator." _What is the difference between proximity and sparsity?_ These are standard metrics to quantitatively evaluate counterfactuals (CFs) [19, 35 in the paper]. __Proximity measures how much our CFs are closer to our data distribution (distance between the CF and the closest data point in the training set with the same label), while sparsity measures how many features were changed between the initial sample and the CF.__ For clarity, we updated L208-209: “proximity (↓) [35], assessing the realism of CFs by their closeness to the training data (distance between the CF and the closest data point in the training set with the same label); and sparsity (↓) [19], which quantifies the changes made to the input to generate the CFs (number of features changed between the initial sample and the CF).” _What purpose does Eq. 4 serve?_ __Eq. 4 provides a theoretical foundation for understanding FL dynamics, describing how client behaviors evolve during training and are influenced by intermediate steps.__ As mentioned by both Rev-eMJr and ULpn, our primary goal is to provide a general theoretical foundation for understanding FL dynamics, specifically how client behaviors evolve during training and how they are influenced by intermediate steps (e.g., local training and aggregation). Eq. 4 describes the dynamics of client behaviors in the FL system. Although we did not analytically solve it, we experimentally observe its solutions through client behaviors using our planes. For clarity, we modify L97-98: "These dynamics can be encapsulated in the following differential equation, which describes how client behaviors evolve during training and are influenced by internal forces within the FL system". --- Rebuttal 2: Comment: Thank you for your valuable feedback. Please let us know if you have any further questions or if there are any points that need additional clarification. We would be grateful if you could consider updating your review in light of our responses.	Summary: This paper introduces a novel method called Federated Behavioural Planes (FBPs) for analyzing, visualizing, and explaining the dynamics of Federated Learning (FL) systems. FBPs are consist of Error Behavioural Plane (EBP), reflecting the model’s predictive performance, and Counterfactual Behavioural Plane (CBP), reflecting the decision-making processes. Using insights from FBPs, the paper also proposes a new robust aggregation technique called Federated Behavioural Shields (FBS) to enhance security in FL systems. Strengths: + Novel approach: The Federated Behavioural Planes (FBPs) introduce a new way to analyze and visualize client behaviors in Federated Learning, addressing a gap in existing literature. + Theoretical foundation: The authors provide a theoretical framework for understanding FL dynamics, grounding their practical approach in solid mathematical concepts. + Explanatory power: FBPs allow for visual identification of client clusters and trajectories, enhancing interpretability of FL systems. Weaknesses: - The paper heavily relies on visual representations (FBPs) to explain client behavior, which makes it infeasible to scale to large-scale FL systems where the convergence might take more rounds and client selection takes place in each round. These factors will largely make the trajectory on the planes hard to keep track of.   - The computational overhead is too large. + In each training round, the counterfactuals are required to be computed for the locally updated model from every client. And the differences between the counterfactuals from each pair of clients are also computed. For cross-silo FL settings, this is already a large computational overhead. For cross-device settings, I do not think this process is affordable especially for high dimensional data like image data. + Besides, since the paper focuses primarily on detecting anomalies and enhancing security from attackers, the main battlefield of this method should be more on the cross-device setting where the number of clients is large. To this end, the prohibitive computational overhead mitigates the motivation of this work.   - The method completely relies on a validation set on the server, which may be hard to obtain without introducing any privacy concerns. Even a validation set is available to use, under what circumstances will the validation set be equally fair to all clients so that it is not far from any client’s local data distribution. And if the plane shows anomalies for some client, how to distinguish the reasons from having an unfair validation set and client is the anomaly/attacker?   - The datasets and the federated settings in the experiments are too simple. And the compared robust aggregation methods are not so recent. Technical Quality: 3 Clarity: 2 Questions for Authors: - Is the counterfactual generator learned on the server? Why would a generator whose goal is for better interpretability affect the performance of predictor? - How is the local CL’s accuracy computed, and why is it so much lower than FL in Table 1? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Most limitations of the proposed method (validation set, computational overhead) have been addressed. Please refer to Weaknesses for other limitations of this work. There is no potential societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: _The paper heavily relies on visual representations to explain client behaviors, infeasible for large FL systems._ __Our method is scalable and automatically extracts and analyzes statistics from behavioral planes (BPs) to identify and mitigate malicious clients without relying on visual inspection. Visual inspection aids in understanding the characteristics under which malicious clients behave differently, but it is not necessary.__ Our defense extracts client-specific scores from BPs in large FL systems to reduce the impact of malicious clients. To further demonstrate scalability and automatic identification of malicious behaviors, we recorded client scores during training using CIFAR10 (Fig. B). These scores can identify malicious clients and reduce their contribution to the global model. Interestingly, as the model converges, more weight is given to the attacker with an inverted gradient, since its model closely resembles the global model from the previous round (plus a negligible inverted update). Visual inspection can still be useful for ML engineers to understand why certain clients are removed or contribute less, by visualizing on multiple planes where they differ from ‘honest’ clients (see Fig. 3). The experiments and discussion will be added to Appendix B. _Client selection at each round will affect trajectory on the planes._ __Our method ensures that client selection does not impact the effectiveness of our defense (FBSs), as aggregation relies on both current and available historical behavior. Visualizations can track client trajectories based on selected rounds, maintaining a clear representation of client behavior.__ If a client is not selected in a round, its trajectory will not consider that particular round, but trajectories can still be created based on selected rounds. While the current focus is on evaluating client behavior over time, it is also possible to establish trajectories only for selected clients relative to the server’s position, ensuring accurate visualization not affected by client selection. Finally, although client selection influences the visual representation, it has no effect on FBSs, as aggregation depends on both current and historical client behavior (Moving Average in Sec. 3.5). _The method relies on a server validation set, which may be difficult to obtain without privacy concerns._ __Several solutions can be integrated with our method to create a synthetic validation set on the server without transmitting sensitive information from clients.__ A clean validation set is a common assumption of existing methods [51, 45, 31 cited in paper]. To address this, some proposed solutions include optimizing for a clean validation set [15] and generating representative inputs through dimensionality reduction and stratified sampling [44]. As discussed in Sec. 6.2, we plan to: - Apply an autoencoder or Bayesian probability conditioned on class labels to generate a synthetic dataset based on client data distributions - Explore synthetic generation as in [44] If the server has a validation set, our method weights each client to maximize global model performance on that test set, regardless of whether the model weights come from honest or malicious clients _When is the validation set fair to all clients? How to distinguish between an unfair validation set and an anomaly/attacker?_ This is a great question, and that’s exactly why we proposed the multi-plane evaluation. Unlike traditional ML metrics such as accuracy or error, __our defense method is not strongly influenced by the fairness of the validation set, as counterfactuals (CFs) are more closely related to the learned client decision process and data distribution than the evaluation set.__ Anomalous clients show unrelated CF distributions and high error, while underrepresented clients generate plausible CFs but are affected by the unfair validation set only in the error plane. To demonstrate effectiveness under an unfair validation set, we removed one client’s data from the validation set and recorded behavioral scores. Preliminary results (Fig. C) show that the 95% confidence interval of the unfair client’s score overlaps with those of other honest clients, distinguishing them from malicious clients. We introduce in L340: Considering the importance of fairness among clients and the promising results shown in Appendix B, the exploration of the impact of unfair validation sets and the development of other validation-independent descriptors, such as CFs, represent promising areas for future research. _Datasets are simple and baselines are not so recent_ __We have enhanced our experimental setup by introducing a more challenging dataset, CIFAR-10, with a 10-class classification task, and a recent robust aggregation baseline, RFA [Pillutla et al., 2022], into all our experiments to further demonstrate the robustness and effectiveness of our method.__ Even under these new conditions (Fig. A), our method outperforms or matches other baselines across all four datasets and all five attack conditions, except for Inv. gradient attacks in MNIST, where Trimmed mean performs better (1 out of 20 cases). _Is the counterfactual generator learned on the server?_ __No, the generator is simultaneously trained in an end-to-end fashion on the client-side along with the predictor__ [L139,L231]. Please refer to the common answers for model analysis. _How is the local CL’s accuracy computed, and why is it so much lower than FL in Table 1?_ __In Local CL, we train a separate model for each client’s private data__, ensuring privacy by not sharing data across clients. In non-IID settings, training on single-client data reduces performance due to the lack of diverse data, explaining the lower accuracy of Local CL compared to FL, which uses data from all clients, in Table 1. We added in L235: For Local CL, we report the average accuracy of all models trained individually by each client and evaluated on a common test set --- Rebuttal Comment 1.1: Title: Thanks for the rebuttal Comment: Thanks the rebuttal. Based on the author's rebuttal, I believe I have a better understanding of the content of the paper. Most of my concerns have been addressed and I have increased my rating. --- Rebuttal 2: Comment: Thank you for your valuable feedback. Please let us know if you have any further questions or if there are any points that need additional clarification. We would be grateful if you could consider updating your review in light of our responses.	Summary: This paper introduces a novel method called Federated Behavioural Planes (FBPs) to analyze, visualize, and explain the dynamics of client behavior in Federated Learning (FL) systems. The primary contributions of the paper are as follows: * 1\. Introduction of Federated Behavioural Planes (FBPs): FBPs consist of two planes: * 1.1\. Error Behavioural Plane (EBP): This plane analyzes the predictive performance of client models by visualizing the errors they produce. * 1.2\. Counterfactual Behavioural Plane (CBP): This plane examines the decision-making processes of client models through counterfactual explanations, highlighting how decision boundaries are formed. * 2\. Visualization and Analysis: FBPs provide informative trajectories that describe the evolving states of clients, enabling the identification of clusters of clients with similar behaviors. This helps in understanding both beneficial and detrimental behaviors in FL systems. * 3\. Federated Behavioural Shields (FBS): Based on the patterns identified by FBPs, the authors propose a robust aggregation technique named Federated Behavioural Shields. This technique enhances security by detecting malicious or noisy client models and surpasses the efficacy of existing state-of-the-art FL defense mechanisms. * 4\. Experimental Validation: The paper demonstrates through experiments that FBPs can effectively track client behavior, identify client clusters, and improve the security and performance of FL systems. The proposed FBS method outperforms other robust aggregation methods in defending against various types of attacks. Overall, the paper offers a comprehensive approach to enhance understanding, trust, and control over federated learning systems by introducing a novel method to analyze and secure client behaviors. Strengths: The paper "Federated Behavioural Planes: Explaining the Evolution of Client Behaviour in Federated Learning" demonstrates several strengths across different dimensions: * 1\. Originality: The paper introduces a novel approach, Federated Behavioural Planes (FBPs), to analyze and visualize client behavior dynamics in Federated Learning systems. This method offers a unique perspective on understanding client behavior evolution in FL, which is a relatively unexplored area in the existing literature. The combination of predictive performance analysis and decision-making process evaluation through FBPs showcases originality in addressing the challenges of client behavior in FL systems. * 2\. Quality: The paper maintains a high standard of quality in terms of methodology, experimental design, and theoretical framework. The introduction of FBPs as a tool to explain the dynamics of FL systems reflects a well-thought-out approach to addressing the evolving behavior of clients in federated learning environments. The robust aggregation mechanism proposed, Federated Behavioural Shields, demonstrates a quality solution to enhance security in FL systems. * 3\. Clarity: The paper is well-written and structured, making it easy for readers to follow the concepts presented. The clarity in explaining the Federated Behavioural Planes framework, the experimental results, and the implications of the proposed method enhances the overall understanding of the research. The use of figures and explanations aids in visualizing complex concepts related to client behavior in FL systems. * 4\. Significance: The paper's contribution to the research area of Federated Learning is significant. By introducing FBPs and Federated Behavioural Shields, the paper addresses a key challenge in FL systems - understanding and controlling client behavior. The insights provided by FBPs and the improved security offered by Federated Behavioural Shields have the potential to enhance the reliability and control over FL systems, making a valuable contribution to the field. Overall, the paper makes a substantial contribution to the field of federated learning. Its originality lies in the novel problem formulation and creative combination of existing ideas. The quality of the research is demonstrated through rigorous methodology and comprehensive experiments. The clarity of the presentation ensures that the contributions are accessible to a broad audience. The significance of the work is underscored by its potential impact on improving the security and efficiency of federated learning systems. Weaknesses: Complexity of Methods: - Computational Overhead: The concurrent training of counterfactual generators with the main predictive models introduces significant computational overhead. This could be particularly burdensome in real-world federated learning settings where resources are limited. The paper could benefit from a more detailed analysis of the computational costs and potential optimization strategies to mitigate this overhead. - Actionable Insight: Consider providing a detailed comparison of the computational requirements of the proposed method with baseline methods. Explore possible optimizations or approximations that could reduce the overhead without significantly compromising the performance. Real-World Applicability: - Scalability Concerns: While the experiments are comprehensive, they are conducted on relatively small datasets and a limited number of clients. This raises concerns about the scalability of the proposed methods to larger, real-world federated learning scenarios with many clients and more complex data distributions. - Actionable Insight: Include a discussion on the scalability of FBPs and FBS. Consider performing a scalability analysis, even if only theoretical, to predict the performance and feasibility of the methods in larger settings. Additionally, simulations or theoretical models could provide insights into expected behavior in large-scale deployments. Generalization Across Different Models: - Model-Specific Limitations: The proposed method may be tailored to specific types of models (e.g., neural networks) and may not generalize well to other types of models used in federated learning (e.g., decision trees, support vector machines). - Actionable Insight: Discuss the applicability of FBPs and FBS to different types of models. Providing a broader range of experiments that include different model architectures could strengthen the paper. If certain models are not compatible, explain the limitations and potential modifications required for broader applicability. Evaluation Metrics: - Limited Evaluation Metrics: The evaluation primarily focuses on standard metrics like accuracy and robustness against attacks. While these are important, they might not capture all aspects of the system's performance, such as the impact on communication efficiency, latency, and energy consumption. - Actionable Insight: Introduce additional evaluation metrics that capture the holistic performance of the system, including communication overhead, latency, and energy consumption. This would provide a more comprehensive assessment of the practicality of the proposed methods in real-world applications. Interpretability and Usability: - Interpretability for Non-Experts: The paper, while clear in its technical explanations, may still be challenging for practitioners who are not experts in federated learning or explainable AI. Enhancing the interpretability and usability of the methods for a broader audience could be beneficial. - Actionable Insight: Provide more intuitive explanations and visualizations of the key concepts and methods. Including case studies or practical examples demonstrating the application of FBPs and FBS in real-world scenarios could make the methods more accessible and easier to understand for non-experts. Real-World Validation: - Lack of Real-World Validation: The experiments are conducted in controlled settings, which may not fully represent the challenges and variability encountered in real-world federated learning deployments. - Actionable Insight: Discuss potential real-world applications and the expected challenges. If possible, provide preliminary results or insights from deploying the methods in a real-world scenario. Alternatively, outline a detailed plan for future real-world validation studies. Conclusion: While the paper presents significant contributions to federated learning, addressing these weaknesses could enhance its impact and practicality. By focusing on computational efficiency, scalability, broader model applicability, comprehensive evaluation metrics, interpretability, and real-world validation, the work can move closer to its stated goals and become more robust and applicable to a wider range of scenarios. Technical Quality: 3 Clarity: 3 Questions for Authors: Questions for the Authors: * 1\. Computational Overhead: - Question: How significant is the computational overhead introduced by the concurrent training of counterfactual generators with the main predictive models? - Suggestion: Provide a detailed comparison of the computational requirements of your method with baseline methods. Include any potential optimization strategies to mitigate this overhead. * 2\. Scalability: - Question: How does the proposed method scale with a larger number of clients and more complex datasets? - Suggestion: Include a scalability analysis or discussion that addresses the performance and feasibility of FBPs and FBS in larger, real-world federated learning scenarios. Simulations or theoretical models predicting the system's behavior in large-scale deployments would be helpful. * 3\. Applicability to Different Models: - Question: Is the proposed method applicable to different types of models beyond neural networks, such as decision trees or support vector machines? - Suggestion: Discuss the generalizability of FBPs and FBS to various model types. If there are limitations, explain the necessary modifications to apply the method to other model architectures. * 4\. Evaluation Metrics: - Question: Have you considered additional evaluation metrics that capture communication efficiency, latency, and energy consumption in your analysis? - Suggestion: Introduce and discuss additional metrics to provide a comprehensive assessment of the system’s performance in practical applications. This would strengthen the evaluation of the proposed methods. * 5\. Interpretability and Usability: - Question: How can the methods be made more interpretable and usable for practitioners who are not experts in federated learning or explainable AI? - Suggestion: Provide more intuitive explanations and visualizations. Including case studies or practical examples of FBPs and FBS in real-world scenarios could enhance understanding and applicability for a broader audience. * 6\. Real-World Validation: - Question: Have you conducted any preliminary real-world validations of the proposed methods? If not, what are the plans for such validations? - Suggestion: Discuss any real-world applications or preliminary results. Outline a detailed plan for future real-world validation studies to demonstrate the practical applicability and effectiveness of your methods. * 7\. Comparison with State-of-the-Art: - Question: How does the proposed Federated Behavioural Shields compare with other state-of-the-art defense mechanisms under different attack scenarios? - Suggestion: Provide a detailed comparative analysis with other robust aggregation methods under various attack types. Highlight the strengths and potential weaknesses of your approach in comparison to existing techniques. * 8\. Visualizations and Trajectories: - Question: Can you provide more detailed examples or visualizations of client trajectories in the Error Behavioural Plane and Counterfactual Behavioural Plane? - Suggestion: Include more visual examples and explanations of client trajectories to illustrate how FBPs can be used to identify different client behaviors and detect malicious clients. * 9\. Impact of Non-IID Data: - Question: How does the method handle non-IID (non-independent and identically distributed) data across clients, and how robust is it to such scenarios? - Suggestion: Discuss the impact of non-IID data distributions on the performance of FBPs and FBS. Provide experimental results or theoretical analysis demonstrating the method’s robustness to non-IID data. * 10\. Future Directions: - Question: What are the future directions or potential extensions of your work? - Suggestion: Outline possible future research directions or extensions of FBPs and FBS. This could include integrating additional behavioral planes, optimizing computational efficiency, or exploring new applications of the methods. Conclusion: Addressing these questions and suggestions can provide clarity on the strengths and limitations of the proposed methods, enhance the understanding of their practical applicability, and offer insights into potential improvements. Engaging with these points during the rebuttal phase can lead to a productive discussion and potentially strengthen the overall contribution of the paper. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Assessment of Limitations and Potential Negative Societal Impact: Based on the provided content and the NeurIPS checklist guidelines on limitations and broader societal impacts, here is an assessment of how well the authors have addressed these aspects and suggestions for improvement: Addressing Limitations: - Identified Limitations: The paper acknowledges some limitations, such as the computational overhead introduced by the concurrent training of counterfactual generators and the potential scalability issues in larger federated learning deployments. - Suggestions for Improvement: - Computational Efficiency: Provide a more detailed discussion on the computational requirements and potential optimization strategies. This could include specific techniques to reduce overhead, such as model pruning, quantization, or distributed optimization methods. - Scalability Analysis: Include more comprehensive scalability experiments or simulations that predict the method's performance in larger, real-world settings. Discuss how the method can be adapted or optimized for large-scale deployments. Potential Negative Societal Impact: - Security and Privacy: The primary focus of the paper is on enhancing security and privacy in federated learning, which is a positive societal impact. However, potential negative impacts, such as the misuse of federated learning systems or unintended biases in the models, should be considered. - Suggestions for Improvement: - Bias and Fairness: Discuss the potential for unintended biases in federated learning models and how the proposed methods could mitigate or exacerbate these biases. Provide suggestions for ensuring fairness in federated learning deployments. - Misuse of Technology: Address the potential for misuse of federated learning systems, such as using the technology for surveillance or other harmful purposes. Discuss safeguards and ethical considerations to prevent misuse. Constructive Suggestions for Improvement: * 1\. Detailed Discussion on Limitations: - Expand the discussion on identified limitations, providing more details on computational efficiency and scalability. Include theoretical analysis or empirical evidence supporting the claims and potential solutions. * 2\. Bias and Fairness: - Add a section discussing the potential for biases in federated learning models. Explain how the proposed methods could impact fairness and provide guidelines or best practices for ensuring equitable outcomes. * 3\. Ethical Considerations and Misuse: - Address potential misuse of federated learning technology. Discuss ethical considerations and propose safeguards to prevent the harmful application of the technology. Highlight the importance of transparency and accountability in federated learning deployments. * 4\. Societal Impact Statement: - Include a comprehensive societal impact statement that covers both positive and negative aspects. This should highlight the benefits of enhanced security and privacy, as well as address the potential risks and ethical concerns. Conclusion: While the paper makes significant contributions to enhancing security and privacy in federated learning, it could benefit from a more thorough discussion of limitations and potential negative societal impacts. By addressing these areas, the authors can provide a more balanced view of their work and ensure that it is applied responsibly and ethically. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: _Computational Overhead and Evaluation Metrics, the paper could benefit from a detailed analysis of the computational costs._ As suggested, __we performed a detailed analysis of the computational overhead of our framework, examining all its components__: Local Computation, Communication Overhead, and Server-side Computation. Our evaluation metrics include GFLOPs for local inference, megabytes for communication, time complexity for the aggregation process, and the duration of one training round. We also conducted a comprehensive comparison with other methods. Please refer to the common comment (@Rev-eMJr,ULpn,aREF) for detailed results and discussions. _Scalability concerns. “While the experiments are comprehensive, they are conducted on relatively small datasets” Question: “How does the proposed method scale with a larger number of clients and more complex datasets?”_ __We enhance the robustness of our method’s validation by: 1) Introducing a more complex dataset, CIFAR-10. 2) Adding a new baseline, [K. Pillutla et al., RFA, 2022], for all experiments. 3) Experimentally demonstrating that our methods scale effectively with a larger number of clients.__ We introduced an experiment on CIFAR10, which is a larger dataset and a recent FL benchmark [K. Pillutla et al., RFA, 2022]. This experiment significantly improves the robustness of our validation, and Fig. A further demonstrates our method’s effectiveness in complex FL scenarios. Our method outperforms or matches other baselines across all datasets (Breast, Diabetes, small-MNIST, small-CIFAR10) and conditions (No attack, MP noise, MP inverted gradient, DP label flipping, DP inverted loss), except for Inverted gradient attacks in small-MNIST, where Trimmed mean performs better (note that this is 1 out of 20 cases). Additionally, we tested the scalability of our framework with up to 200 clients. As shown in Fig. D, our method introduces 1 extra minute per round with 200 clients, compared to FedAvg. In contrast, the baseline Krum introduces over 15 minutes per round. This demonstrates the efficiency and scalability of our approach. _Generalization across different models._ __Our method, while focused on neural networks (NNs) and traditional FL, theoretically applies to other ML models like decision trees, though it requires appropriate aggregation processes and counterfactual (CF) generation.__ We focused our study on NNs and traditional FL, using differentiable models capable of producing a global model through parameter aggregation. However, extending our method to other ML models, such as decision trees, presents an interesting and challenging opportunity for future work due to their discrete model parameters (e.g., splitting rules and thresholds). Our proposed method is based not on the similarity between client model parameters, which can be problematic with such heterogeneous architectures, but on the client model’s performance and their respective decision-making process. Therefore, in theory, our defense method (FBSs) of evaluating client behaviors still holds with these ML models but requires a suitable aggregation process and CF generation. Various aggregation methods have been proposed in the literature for such models, which can be adapted for this purpose [Wang et al., Decision Tree-Based Federated Learning, 2024]. Additionally, generating CFs from interpretable models, such as decision trees, is straightforward due to their inherently explainable nature. _Interpretability and Usability. “The paper, while clear in its technical explanations, may still be challenging for practitioners who are not experts”._ To broaden the audience and facilitate the comprehension of our method, __we introduce the pseudocode of our algorithm (Algorithm 1) for creating the behavioral planes on the server and implementing our robust aggregation method (FBSs).__ _Visualizations and Trajectories. Question: Can you provide more detailed examples or visualizations of client trajectories in the Error Behavioural Plane and Counterfactual Behavioural Plane (CBP)? Suggestion: Include more visual examples and explanations of client trajectories_ In addition to the trajectories presented in Fig. 3 of the main paper, __we provided additional trajectories in the Appendix B (Fig. 8)__. Specifically, Fig. 8 illustrates client trajectories within the FBPs for different scenarios, including an Inverted Loss attack on the Synthetic dataset and a Data Flip attack on the Diabetes dataset, highlighting the distinct behavioral patterns of clients and attackers in these settings. These trajectories also demonstrate the possibility of identifying clusters of clients with similar data distributions, particularly on the CBP. _How does the method handle IID and non-IID data?_ __Our method is robust in both IID and challenging non-IID scenarios, demonstrating higher accuracy in IID settings across most conditions.__ Please note that in the main paper, we conducted all experiments in non-IID scenarios (Section 4.1), which are the most realistic yet challenging for robust aggregation methods, as even honest clients behave differently. Additionally, we compared the performance of our defense method under No-attack, Crafted-noise, Inverted-gradient, Label-flipping, and Inverted-loss attacks in both IID and non-IID settings. The table below shows that higher accuracy is achieved in the IID setting under almost all conditions compared to the non-IID setting, highlighting the increased complexity of operating in non-IID environments (Appendix B). \| Cond. \| No attack \| MP Noise \| MP Grad \| DP Flip \| DP I.Loss \| Mean \| \|---------\|-----------\|----------\|---------\|-----------\|-----------\|----------\| \| non-IID \| 95.7±1.1 \| 98.0±0.8 \| 95.3±0.7 \| 94.2±0.6 \| 95.9±0.9 \| 95.8±0.4 \| \| IID \| 98.2±0.3 \| 98.4±0.4 \| 98.2±0.2 \| 96.4±0.9 \| 93.7±1.0 \| 97.0±0.4 \| --- Rebuttal Comment 1.1: Comment: Thank you for your valuable feedback. Please let us know if you have any further questions or if there are any points that need additional clarification.	null	null	Rebuttal 1: Rebuttal: # Answer to reviewers and ACs We thank the reviewers for their insightful feedback. We are encouraged by their recognition of the novelty in our ideas and proposed methods for analyzing and visualizing client behaviors (eMJr, ULpn, aREF), and the significance of this work in the FL field (eMJr). We are pleased that our work was found to be clear, easy to follow (aREF) and grounded in a solid theoretical foundation (eMJr, ULpn). We appreciate ULpn’s acknowledgment of our method’s explanatory power, enhancing FL system interpretability. Reviewers’ feedback has certainly improved the quality of our manuscript and we hope we successfully addressed your concerns in this rebuttal and in our updated submission. We reply to shared questions here and address specific questions under each reviewer’s feedback. # Summary of Changes In response to the reviewers’ comments, we worked on improving the clarity and comprehensiveness of our paper with new sections and explanatory content. By incorporating a new dataset, a new SOTA baseline, and additional experiments, we have increased the robustness of our method’s validation. These improvements address the specific concerns and strengthen our work’s overall contribution. However, the core contributions and evaluations of our work remain unchanged. Our changes are summarized as follows: - __Comprehensive Theoretical and Experimental Analysis of Computational Costs__ (Appendix A.6) - __Expanded Experimental Validation:__ Validated on a more complex dataset, CIFAR10, in addition to the four previously used datasets, to further assess our method’s performance (Fig. A and Table in Appendix B) - __Additional Baseline Comparison:__ Introduction of a recent robust aggregation baseline [K. Pillutla et al., RFA, 2022] in all experiments (Fig. A, new Table 5, 6, 7) - __Attack Intensity Analysis:__ Verified robustness against varying attack intensities (Fig. E) - __Behavioral Score Analysis:__ Evaluated the importance of information provided by our behavioral planes by recording the automatically extracted client behavioral scores from our defense method (Fig. B) # Common Answers _@Rev-eMJr,ULpn,aREF – Lack of computational cost analysis_ __We perform a comprehensive analysis of the computational cost of our methods. In the worst-case scenario, introducing counterfactuals (CFs) adds 7.6% more model parameters and 5.1% more GFLOPs for inference compared to the predictor alone. Overall our method adds 1 minute per round to FedAvg, while Krum adds over 15 minutes per round.__ The computational cost in FL frameworks includes local computation, communication overhead, and server-side computation - __Local Computation.__ Our methods integrate a CF generator with the original predictor to explain decision-making processes and provide insights into client data distribution. For small NNs, local computation is minimal compared to communication latency and synchronization, allowing CF generation for predictor input without affecting training efficiency. As the predictor size increases, as shown with MNIST and CIFAR10, we can efficiently generate CFs for intermediate layers using a relatively small number of neurons. As shown in Table A, without losing performance, the generator requires only 2.7% of the predictor’s GFLOPs (ResNet-18) for inference on a 28x28 RGB image. Training time is proportional, as our generator trains end-to-end with the predictor. - __Communication Overhead.__ The CF generator is transmitted with the predictor to the server for evaluation and aggregation. The increase in the number of transmitted parameters is marginal compared to the predictor, consisting of only 1.8% of the predictor’s parameters (Table A). This results in an additional 0.92 megabytes (MB) compared to the 49.68 MB required for the predictor. - __Server-side Computation.__ This involves evaluating client models on a small validation set (e.g., 250 samples as shown in Fig. 9) and calculating pairwise distances between client CFs. Model evaluations, which required a single pass for each client model, are negligible compared to the computational load of calculating pairwise distances. We use the sliced Wasserstein distance with a complexity of $O(m \log m)$, where $m = n_{samples} \times 2=500$ [L148]. This operation is repeated for each unique pair of clients, leading to a total complexity of $O(n^2⋅m \log m)$, where $n$ is the number of clients. Compared to Krum’s complexity of $O(n^2⋅d)$, where $d$ is the number of model parameters, our method is more efficient for NNs with more than 4480 parameters, which is typical in practical applications - __Overall Computational Cost.__ We compared the computational cost of our defense (FBSs) with Krum and FedAvg across different network dimensions and client numbers (Fig. D). For increasing client numbers, our method uses a CF generator with 7.6% of the predictor’s parameters (worst-case scenario) and 250 validation samples, as used in the paper. Our method scales better than Krum under both conditions, adding only 1 extra minute per round with 200 clients compared to FedAvg, while Krum adds over 15 minutes per round. This demonstrates its efficiency in large-scale scenarios _@Rev-ULpn,aREF – Why does the CF generator affect predictor performance?_ __The CF generator affects the performance of the predictor because it is jointly optimized with the predictor and this is needed to generate CFs during training.__ Therefore training the CF generator alongside the predictor influences the training process of the predictor. The small change in performance, shown in Table 1, can be attributed to the additional loss providing extra information to navigate the optimization space more efficiently, similar to regularization terms. On the contrary, using a post-hoc CF generator (which does not affect performances) would necessitate training a separate generator for each client at the end of every epoch, significantly increasing the computational cost Pdf: /pdf/7a665fe2998a4fd2647142051c51b8fd001925ba.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Robust Gaussian Processes via Relevance Pursuit	Accept (poster)	Summary: This work proposes a new way to perform heteroskedastic regression using Gaussian processes via data-point-specific noise levels. These noise levels are inferred using a sequential selection procedure maximizing the log-marginal likelihood. The authors show that under a specific parametrization the log-marginal likelihood is strongly concave in the noise variances and give approximation guarantees for their proposed algorithm. Experiments on regression and Bayesian optimization tasks demonstrate the benefits of this approach. Strengths: Making Gaussian processes robust to some degree of model misspecification is a worthwhile goal and this paper provides a nice addition to the existing toolbox of methods. The paper is well-written and presents largely well-executed experiments for regression and Bayesian optimization. In particular, in tasks where Bayesian optimization is applied, it seems plausible that outliers appear in the way they are synthetically generated in the evaluation. I also appreciated that the authors submitted their code. Weaknesses: ### Theory While correct as far as I can tell, I question the value of the theoretical result, if in practice you optimize the hyperparameters of the covariance function and $\rho$ jointly. Is there reason to believe, that the convexity in $\rho$ for fixed hyperparameters is beneficial given this choice? It would be informative to see an experiment that compares the non-convex parametrization to the convex one where all hyperparameters are optimized jointly. If the convex parametrization improves convergence, this would suggest that convexity is beneficial even when optimizing all hyperparameters jointly. ### Related Work Even though this is quite recent work, I would have liked to see a discussion of Altamirano et al. (2024) in the related work section. They introduce data-point-specific observation noise based on a generalized Bayesian inference perspective. I think a comparison to their approach in the experiments would improve the paper, but I would not expect this given the recency of the work. - Altamirano, Matias, Francois-Xavier Briol, and Jeremias Knoblauch. "Robust and Conjugate Gaussian Process Regression." International Conference on Machine Learning (ICML), 2024. URL: https://arxiv.org/abs/2311.00463 ### Experiments As the authors laudably acknowledge, the fit times were 100x slower (see Figure 8) than the simple baseline approaches. How does that impact the Bayesian optimization results if plotted as a function of wall-clock time? It seems that eventually some of the other approaches catch up to your method. If the fit time is 100x that of the baseline approaches, the reward curves might look different if plotted as a function of wall-clock time. Or is there an implicit assumption that each evaluation is significantly more expensive than fitting the GP? The experiments on regression problems could be expanded in my opinion, e.g. with benchmark datasets from the UCI repository. The test problems that were considered are arguably designed to test optimization algorithms (hence their usage in BO). From looking at the code the "real-world problems" in Section 6 have synthetically generated corruptions. I think the paper could be made stronger if there were at least one experiment where the corruption is not synthetically generated, but simply part of the observation process. ### Other Weaknesses and Improvements - Almost all plots are way too small and barely readable. If you reformat these, then I am willing to raise my "Presentation" score. - Link to the proofs in the appendix after each theoretical result, otherwise they are hard to find. - When you reference details in the appendix, link to the appropriate section (e.g. line 286). - Citations in lines 32 to 34 for the methods that are referenced are missing - Typos (l.211, l. 286, l. 362) Technical Quality: 3 Clarity: 3 Questions for Authors: - Is your assumption about the data corruption different from Huber's $\varepsilon$-contamination model (see e.g. https://arxiv.org/pdf/1511.04144) or is it identical? Equation 2 suggests they are different, but the text and experiments often refer to only a fraction of the data being corrupted (e.g. in Section 4.3 and Section 6). - How does your approach compare to a vanilla GP on a dataset with no corruption? Does it "fail gracefully"? - Do you see better performance when jointly optimizing the hyperparameters for the convex reparametrization of the data-point-specific noise variances than if not reparametrizing? - How did you set the schedule to test data points for being outliers in the experiments (c.f. lines 182-184)? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I would have liked to see a bit more discussion on the limitations of the approach in Section 7. For example, the fact that the theoretical result does not apply when jointly optimizing the hyperparameters as is the case in the experiments, and a discussion of the wall-clock runtime. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > “While correct as far as I can tell, I question the value of the theoretical result, if in practice you optimize the hyperparameters of the covariance function and ρ jointly. Is there reason to believe, that the convexity in ρ for fixed hyperparameters is beneficial given this choice?” “It would be informative to see an experiment that compares the non-convex parametrization to the convex one where all hyperparameters are optimized jointly” We would expect the overall problem to be “more convex” than with the canonical parameterization, and thus improve optimization performance. More formally, the convexity guarantee will guarantee the positive-definiteness of the submatrix of the Hessian corresponding to the rhos, which is better than not guaranteeing positive definiteness at all. Positive-definiteness is also beneficial for quasi-Newton optimization algorithms like L-BFGS, which restarts the Hessian-approximation whenever it encounters non-convex regions, because the associated updates to the Hessian approximation are not positive-definite. This leads the algorithm to momentarily revert back to gradient descent, with an associated slower convergence rate. Practically, we ran convergence analyses using the data from Figure 1, allowing all rhos to be optimized jointly with other hyper-parameters (length scale, kernel variance and noise variance), recording the achieved negative log marginal likelihood (NLML) as a function of the tolerance parameter `ftol` of the L-BFGS optimizer. The results indicate that the optimizer terminates with a much better _NLML_ using the convex parameterization with the same convergence tolerance: ftol \| Canonical \| Convex \| ----\|----\|--------\| 1e-03 \| -4.37 \| -14.18 \| 1e-04 \| -4.37 \| -93.00 \| 1e-05 \| -4.37 \| -93.01 \| 1e-06 \| -4.37 \| -135.05 \| 1e-07 \| -98.62 \| -518.68 \| 1e-08 \| -97.52 \| -1139.29 \| There are also settings in which we do not actually jointly optimize the hyperparameters. For instance, consider a situation in which we have access to data from the same data generating process that has been manually labeled by domain experts as outlier-free. Then we can estimate the model hyperparameters (of the non-robust GP) on that data, and fix those for the RGP-RP on the new data set that we do not know to be outlier-free (and that we cannot label due to cost or other practical reasons). > “Even though this is quite recent work, I would have liked to see a discussion of Altamirano et al. (2024) in the related work section” Thanks for pointing out this relevant work. We have added a discussion of and comparison to this work in our general response, and will add this to the paper. > “Or is there an implicit assumption that each evaluation is significantly more expensive than fitting the GP?” This is indeed the case; GPs are commonly applied to Bayesian optimization or active learning tasks where evaluation times are on the order of hours or days, so the fitting time of the GP is generally not much of a concern. We will make this assumption explicit. > “The experiments on regression problems could be expanded in my opinion, e.g. with benchmark datasets from the UCI repository” We have done so, see our general response. Please let us know if there are any additional datasets you would like us to consider for the CR if accepted. > “I think the paper could be made stronger if there were at least one experiment where the corruption is not synthetically generated [...]” Thanks for this feedback. We have added the “Twitter Flash Crash” example from Altamirano et al. (2024) and an additional variant of it to our results (see general response). If you have any other examples of real-world data we would be happy to consider this for inclusion in the CR. > “Almost all plots are way too small and barely readable” We have adjusted the size of our plots (in particular the axis labels and legends), see attached pdf. We will utilize part of the additional space for the CR submission to increase plot size throughout. > “Is your assumption about the data corruption different from Huber's ε-contamination model or is it identical? Equation 2 suggests they are different, but the text and experiments often refer to only a fraction of the data being corrupted [...]” A significant high-level difference is that Huber’s model is homoskedastic while ours is not. Citing Huber’s original article, he states: “Let $x_1,⋯, x_n$ be independent random variables with common distribution function $F$”, i.e. Huber’s model assumes that every data point is equally likely to be an outlier. In contrast, our model introduces data-point specific variances that are adapted using marginal likelihood optimization, leading the resulting likelihood function to be different for each datapoint with a distinct rho. The article often refers to the “fraction of data being corrupted” because the inference of this fraction is an important problem that determines the performance of RP, which requires choosing a number of outlier variances ($\rho_i$). We solve this problem by leveraging Bayesian model selection. > “How does your approach compare to a vanilla GP on a dataset with no corruption? Does it "fail gracefully”?” Yes, and the results in our general response demonstrate this. We expect this to hold as long as the prior on the occurrence probability is sufficiently uninformative and puts some weight on no occurrences. Throughout our experiments, we used a geometric prior on the support size with a mean equal to 20% of the number of observations. We will add the results and discussion to the manuscript. > “How did you set the schedule to test data points for being outliers in the experiments (c.f. lines 182-184)?” See [our response to reviewer aaaZ](https://openreview.net/forum?id=5FATPIlWUJ&noteId=OjPuKthMyu). > “I would have liked to see a bit more discussion on the limitations of the approach in Section 7 [...]” Good point, we will include these in the discussion in Section 7. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I believe the additional results you presented strengthen the paper. I have raised my score to reflect this. --- Reply to Comment 1.1.1: Comment: Thank you again for your detailed, actionable feedback, and consideration.	Summary: The paper proposes a robust Gaussian Process regression by inferring data-point-specific noise levels with a sequential selection procedure maximising the log marginal likelihood. The authors show the good performance of their method in a mix of synthetic and real-world regression tasks, including Bayesian optimisation. Strengths: The paper is well-written and easy to follow. The method is clearly motivated and presented, with a good amount of related work discussed. The authors address a problem—robust regression in the context of Gaussian processes—of sufficiently broad interest to merit publication. The authors also provide a theoretical analysis to support their method. Weaknesses: In my opinion, the weakest part of the paper is the regression problems in the experiment section. It only benchmarks against two test functions, one of which is not a standard benchmark for regression; I would like to see more standard UCI benchmarks. Additionally, the proposed method does not clearly outperform other methods, particularly in the Hartmann6 test, where adaptive Winsorization outperforms RGP-RP. The authors do not discuss the results of the regression problems, so a discussion providing guidelines on the cases where the proposed method will perform better is needed. Finally, the authors claim their method "permits fast and robust inference"; however, in Appendix D.1, the method is slower than the Student-t GP for the Hartmann6. It would be useful if the authors discussed this in the main paper. Technical Quality: 3 Clarity: 3 Questions for Authors: - What's the difference between RGP-RP and RGP-RP* in the experiment section? - What's the schedule $\mathcal{K}$ selected for the experiments? How does this affect the performance and fitting time of the method? - While the Friedman and Hartmann6 test functions are well-known, the authors should provide a reference for these functions. - Why use Hartmann6 as a test function for regression? Hartmann6 is typically used to test optimisation due to its properties. Why not use standard UCI benchmarks? - Do the authors know why the student-t GP performs so poorly in the Hartmann6 function, even when the noise comes from a student-t distribution? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: No obvious negative societal impact Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > “The weakest part of the paper is the regression problems in the experiment section” As part of our rebuttal, we have produced a significant number of additional results on UCI benchmarks, the benchmarks and real-world Twitter crash example of Altamirano et al. 2024, and included an additional baseline (see general response and attached pdf). > “Additionally, the proposed method does not clearly outperform other methods, particularly in the Hartmann6 test, where adaptive Winsorization outperforms RGP-RP.” We don’t necessarily expect our method to outperform all other baselines in all settings. But with our additional empirical results in our general response, we do show that it (i) works much better in situations in which outliers cannot be easily separated from the rest of the data, (ii) is competitive in most other settings. We also note that while the adaptive winsorization heuristic indeed performs slightly better than RGP-RP on Hartmann6 with constant, well-separated outliers, it completely fails on the Friedmann10 function and is barely better than a non-robust GP. In general, winsorization does not perform well if the outliers are not clearly and a-priori separated from the rest of the data. > “The authors do not discuss the results of the regression problems, so a discussion providing guidelines on the cases where the proposed method will perform better is needed” We agree that the discussion can be improved and we will include this into the paper (see general response for discussion). > “the authors claim their method "permits fast and robust inference"; however, in Appendix D.1, the method is slower than the Student-t GP for the Hartmann6. It would be useful if the authors discussed this in the main paper” The Student-t likelihood performs very poorly on Hartmann6 and has very high variance. We suspect that there are numerical issues in the optimization of the variational objective that results in the fitting terminating early and at suboptimal points, thus producing faster fit times. Note that the Student-t log likelihood is also non-convex, which can give rise to highly suboptimal local minima. We plan to investigate this in more detail. > “What's the difference between RGP-RP and RGP-RP* in the experiment section?” See [our response to reviewer CPiv](https://openreview.net/forum?id=5FATPIlWUJ&noteId=JGfjtKQW5T). > “What's the schedule K selected for the experiments? How does this affect the performance and fitting time of the method?” The default schedule used throughout the experiments is `[0%, 5%, 10%, 15%, 20%, 30%, 40%, 50%, 75%, 100%]` of the number of observations (traversed backwards by the backward algorithm). The fitting time is linear in the number of steps in the schedule, as a new model fit is carried out for each step of the schedule. The advantage of a schedule with a finer granularity is that the model could be more data efficient, by introducing additional noise variances for outlying data points only, so that all other data is "trusted" up to the homoskedastic noise variance $\sigma^2$. For most practical applications, it is probably sufficient to test for single-digit percentages of outliers, the default schedule here is designed to exhibit high generality, as is evidenced by RP's performance on the variety of benchmarks, including its outperformance on the RCGP benchmarks and Twitter crash example of Altamirano et al. 2024 included in the general response, using the same schedule. If timing was a particular concern, the optimization could likely be accelerated substantially by keeping track of the approximation of the Hessian matrix generated by each L-BFGS call, which could be used to warm-start the optimization of each step and lead to significantly accelerated convergence, as the consecutive optimization problems tend to be similar. > “While the Friedman and Hartmann6 test functions are well-known, the authors should provide a reference for these functions.” Thanks for the callout, we will provide the respective references. > “Why use Hartmann6 as a test function for regression? Hartmann6 is typically used to test optimisation due to its properties. Why not use standard UCI benchmarks?” We have produced additional results on a number of UCI benchmarks as part of this rebuttal (see general response). > “Do the authors know why the student-t GP performs so poorly in the Hartmann6 function, even when the noise comes from a student-t distribution?” We believe you are referring to the right column of Figure 4, whose x-axis corresponds to the fraction of data points that are corrupted by the outlier process. First, the axis label is missing in the submission, and we will correct this. What the figure shows is that the Student-t GP only starts to dominate the other methods once more than 40%-50% of data points are corrupted by Student-t noise. We indeed expect the Student-t GP to be the best model as the fraction of Student-t-distributed noise increases, as it becomes the correct noise assumption when 100% of data points are “corrupted” by Student-t noise. However, the results also show: if the main goal is to make a GP robust to a _sparse_ set of outliers – rather than a uniformly heavy tailed observation noise – then RP significantly outperforms the Student-t GP by leveraging this sparse structure. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I'll increase my score to 7.	Summary: The authors propose a robust Gaussian process model in the sparse heteroscedastic noise setting. Orthogonal Matching Pursuit- like algorithm is proposed to infer data point specific noise levels. The negative log marginal likelihood function to be optimized is claimed to be strongly convex by reparametrizing the additive noise levels. The method is validated on synthetic as well as real-world experiments. Strengths: 1. The authors formulated a convex $\mathcal{L}$ in the Gaussian process regression model by reparametrizing the noise levels, which is very desirable for optimization perspective. The representation of the paper is good, however, I found the section 4.3 difficult to follow. 2. The relevance pursuit is a novel approach to the heteroscedastic noise case of robust GP regression. Weaknesses: 1. Much more work on heavy tailed outliers is done. So, more references and discussion need to be added in related work. Kersting et. al. 2007 proposed GPs for heteroscedastic noise setting. They also learned the homoscedastic part of the variance aiming to learn the remaining part using a second GP. The work is related and at least needs a mention. It would have been nice to see comparison with this model. 2. While this work employs greedy algorithms for GPs, it is not the first to do so, so I would not consider it exceptionally novel. 3. The experimentation is adequate but less discussed. Ref: Most Likely Heteroscedastic Gaussian Process Regression Minor corrections: 1. Consider adding references for these three categories on line 34. No reference is given for down weighting procedures even in related work. 2. Figure 1 and 2 are not discussed at all. What regression example? 3. broken line 150 4. unclosed braces figure 3 5. no need of repeating the abbreviation RIP on line 200 6. Broken line 264 7. Is the x axis of right figure in Figure 4 percentage of corruptions or probability? 8. Different notations need to be used for $\delta$ on lines 67 and 112. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Why Matern kernel was chosen in one of the baselines? 2. What is the difference between RGP-RP*-BW and RGP-RP-BW. I see that the former is canonical one. What is the difference between canonical and non-canonical reparameterization of $\mathbf{s}$? 3. What is $\mathbf{K}_{0}$ on line 218? Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Limitations are discussed as I hoped they would be. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Kersting et. al. 2007: This is indeed related work, and we will discuss it in the updated manuscript. The main difference to our method is that their approach is unlikely to work well with outliers. In particular, while they do account for heteroskedastic observation noise, they model the logarithm of the standard deviation of the noise noise process as another GP. This not just restricts the observation to be (heteroskedastic) Gaussian, but, importantly, imposes a smoothness assumption on the variance of the observations as a function of the inputs. In contrast, our approach infers individual additional noise terms, and thus allows us to go beyond the setting in Kersting et. al. 2007 to handle even gross outliers (which require a noise process that is discontinuous in the inputs). Another downside of the approach in Kersting et. al. 2007 is that they require an EM algorithm to estimate the parameters of the two GP models, which does not have any theoretical guarantees and in practice often has convergence issues (see e.g. discussion here). > “I found the section 4.3 difficult to follow” Thank you for your feedback. We will improve the clarity of our writing in this section in particular by introducing the general framework of Bayesian model selection before stating its instantiation in our work and providing further background references. Please let us know if there are additional edits you would like us to do. > “While this work employs greedy algorithms for GPs, it is not the first to do so” While greedy algorithms have been applied for many problems in general (outside of GPs), their application to the sparse optimization of noise variances for robustness, which we show permits strong theoretical guarantees, has to our knowledge not been studied and proposed before. Please let us know if you would like us to add any additional reference for related work. > “The experimentation is adequate but less discussed” We produced a significant number of additional results with new baselines and test problems for this rebuttal, and will improve discussion of the entirety of the empirical results (see general response). > “Figure 1 and 2 are not discussed at all. What regression example?” Thank you for catching this. The illustrative regression example is based on a synthetic one-dimensional modified sine function with a few outliers. We will detail the exact setup for this in the appendix. > “Is the x axis of right figure in Figure 4 percentage of corruptions or probability?” Yes, this is stated in the figure’s caption. We will add an axis label as well in order to improve clarity. > "Why Matern kernel was chosen in one of the baselines?” No particular reason other than that the Matern-5/2 kernel is widely used in practice and the default in many GP / BO libraries (e.g. botorch, trieste). We expect the results to be quite similar if another common kernel, say an RBF kernel, were used. > “What is the difference between RGP-RP-BW and RGP-RP-BW. I see that the former is canonical one. What is the difference between canonical and non-canonical reparameterization of s?” This difference is described at the beginning of Section 5.2. The “canonical” parameterization is that which directly uses the robust variances $\boldsymbol \rho$ in the parameterization. The non-canonical parameterization is the convex parameterization we introduce in Section 5.2, i.e., that where ${\boldsymbol \rho}(\mathbf s) = \text{diag}({\bf K_0}) \odot ((1 - {\mathbf s})^{-1} - 1)$ and the new parameter is $\mathbf{s}$. The method label coding is as follows: A “” indicates using the canonical (not necessarily convex) parameterization, and “BW” indicates using the backward algorithm for relevance pursuit (Algorithm 2 in Appendix A). We will improve the clarity of the exposition here and clearly explain the method labels. The convex parameterization is a key ingredient that allows us to prove strong theoretical guarantees for the result of the greedy algorithm, and also has beneficial effects on the convergence of the hyper-parameter optimization algorithms, as we demonstrate in the other responses. > “What is $\bf K_0$ on line 218?” Thanks for catching this forward reference, $\mathbf{K}_0$ is defined in Theorem 8: $\mathbf{K}_0 = k(\mathbf{X}, \mathbf{X}) + \sigma^2 \bf I$. We’ll move the definition up in the text. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my concerns in detail. While the additional experimentation certainly demonstrates the effectiveness of the RGP, a more comprehensive comparison would benefit from including benchmarks that explicitly model heavy-tailed distributions, such as those employing Laplace (Kuss et. al., 2006) and Huber (Algikar et.al., 2023) likelihoods. Both methods address similar challenges – spatially uncorrelated gross outliers and input-independent noise – making them valuable benchmarks. Algikar et.al's method claims to not lose efficiency at Gaussian distribution using Huber loss and leverage weighting. My current score remains unchanged. --- Reply to Comment 1.1.1: Title: [time sensitive question] Comment: We are glad we addressed your concerns in detail via our additional experiments which demonstrate the effectiveness of RGP. The reviewer now suggests additional methods to compare against. At a minimum, we can discuss our work in the context of those papers, and promise to include these comparisons in the CR. We hope that time permits us to run some additional simulations in the next day to meet your standards. Before we can do this, can you please clarify the stated references in your follow-up response above? Is "Kuss et. al 2006" Malte Kuß's PhD thesis https://pure.mpg.de/rest/items/item_1791134/component/file_3167621/content ? Is "Algikar et.al., 2023" the unpublished manuscript of Pooja Algikar & Lamine Mili, Robust Gaussian Process Regression with Huber Likelihood? (preprint at https://www.researchgate.net/publication/367280832_Robust_Gaussian_Process_Regression_with_Huber_Likelihood) --- Rebuttal 2: Comment: Thank you again for your suggestions. We present the additional results you suggested below and hope that this will allow you to fully recommend our work for acceptance. Summary We added additional variational GP models with Laplace and Huber likelihoods, and translated the Matlab code of the "projection statistics" of Algikar et.al., 2023 ([Github](https://github.com/apooja1/GP-Huber/tree/main)) to PyTorch. We then combined the projection-statistics-based weighting of the Huber loss with a variational GP (referred to as `Huber-Projection` ) to get as close as possible to a direct comparison to Algikar et.al., 2023 without access to a Matlab license. ## Additional Benchmark Results with 15% Corruptions (Negative Log Likelihood) The following results include the Friedman 10 function used in Algikar et.al., 2023, and the UCI CA Housing data set, which - according to UCI - is a replacement for the deprecated UCI Boston Housing data, which was used in Algikar et.al., 2023. In summary, for a sparse 15% of corrupted observations, Relevance Pursuit is able to almost uniformly achieve better predictive negative log likelihoods (NLL) for the datasets and outlier distributions we consider, the only exception being the CA Housing data, where the standard GP performs surprisingly well in terms of NLL, but not in terms of RMSE. \| Data \| Standard \| Relevance Pursuit \| Student-t \| Laplace \| Huber \| Huber + Projection \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| friedman5 + uniform \| 2.07e+00 (4.88e-02) \| -1.35e+00 (7.22e-02) \| 4.67e-01 (3.00e-01) \| 8.88e-01 (1.01e-01) \| 8.29e-01 (9.14e-02) \| 8.88e-01 (1.08e-01) friedman5 + constant \| 4.28e+00 (2.25e-02) \| -1.13e+00 (1.82e-02) \| 8.24e-01 (3.17e-01) \| 2.28e+00 (2.63e-02) \| 2.34e+00 (3.13e-02) \| 2.34e+00 (3.13e-02) friedman5 + student-t \| 3.42e+00 (1.34e-01) \| -1.11e+00 (1.04e-01) \| 1.78e-02 (2.86e-01) \| 1.55e+00 (1.14e-01) \| 1.66e+00 (1.11e-01) \| 1.59e+00 (1.10e-01) friedman5 + laplace \| 3.55e+00 (7.84e-02) \| -1.22e+00 (5.89e-02) \| 3.47e-01 (3.13e-01) \| 1.75e+00 (8.46e-02) \| 1.90e+00 (7.36e-02) \| 1.90e+00 (7.36e-02) friedman10 + uniform \| 1.86e+00 (3.04e-02) \| -1.82e+00 (3.97e-02) \| 6.78e-02 (3.85e-01) \| 9.20e-01 (1.99e-01) \| 9.39e-01 (2.01e-01) \| 7.65e-01 (1.80e-01) friedman10 + constant \| 4.21e+00 (1.39e-02) \| -1.27e+00 (1.10e-02) \| 8.85e-01 (3.79e-01) \| 2.20e+00 (1.45e-02) \| 2.23e+00 (1.54e-02) \| 2.23e+00 (1.54e-02) friedman10 + student-t \| 3.34e+00 (1.24e-01) \| -1.40e+00 (1.17e-01) \| -6.80e-02 (3.65e-01) \| 1.96e+00 (1.25e-01) \| 1.95e+00 (1.23e-01) \| 1.96e+00 (1.22e-01) friedman10 + laplace \| 3.46e+00 (6.68e-02) \| -1.44e+00 (5.91e-02) \| 4.89e-01 (3.94e-01) \| 2.13e+00 (6.35e-02) \| 2.21e+00 (2.01e-02) \| 2.21e+00 (2.01e-02) yacht_hydrodynamics + uniform \| 3.45e+00 (1.42e-01) \| 2.37e+00 (2.79e-01) \| 1.18e+02 (1.06e+01) \| 7.07e+01 (5.17e+00) \| 7.46e+01 (5.38e+00) \| 1.89e+02 (1.64e+01) yacht_hydrodynamics + constant \| 5.29e+00 (1.70e-01) \| 1.79e+00 (2.43e-01) \| 7.16e+01 (7.68e+00) \| 2.32e+01 (1.59e+00) \| 1.95e+01 (1.12e+00) \| 3.89e+01 (3.46e+00) yacht_hydrodynamics + student-t \| 4.46e+00 (3.08e-01) \| 2.42e+00 (3.38e-01) \| 1.32e+02 (1.05e+01) \| 8.20e+01 (5.07e+00) \| 8.13e+01 (5.17e+00) \| 1.55e+02 (1.38e+01) yacht_hydrodynamics + laplace \| 4.35e+00 (2.88e-01) \| 2.27e+00 (3.71e-01) \| 1.14e+02 (9.86e+00) \| 6.85e+01 (4.72e+00) \| 6.12e+01 (3.87e+00) \| 1.17e+02 (8.89e+00) california_housing + uniform \| 3.14e+00 (1.44e-01) \| 2.92e+00 (2.06e-01) \| 5.85e+01 (9.59e-01) \| 6.77e+01 (2.41e+00) \| 6.69e+01 (1.94e+00) \| 8.94e+01 (2.45e+01) california_housing + constant \| 3.57e+00 (2.26e-01) \| 4.51e+00 (2.11e-01) \| 4.02e+01 (1.59e+00) \| 1.83e+01 (6.58e-01) \| 1.75e+01 (7.03e-01) \| 2.18e+01 (4.19e+00) california_housing + student-t \| 1.77e+00 (6.65e-02) \| 3.15e+00 (1.67e-01) \| 5.95e+01 (1.44e+00) \| 5.20e+01 (2.19e+00) \| 4.95e+01 (1.75e+00) \| 6.90e+01 (2.02e+01) california_housing + laplace \| 1.61e+00 (4.21e-02) \| 3.51e+00 (1.93e-01) \| 5.60e+01 (1.58e+00) \| 4.23e+01 (1.72e+00) \| 4.06e+01 (1.47e+00) \| 5.62e+01 (1.66e+01) ## RMSE RP performs uniformly better w.r.t. RMSE, we only include CA Housing here due to space limitations. \| Data \| Standard \| Relevance Pursuit \| Student-t \| Laplace \| Huber \| Huber + Projection \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| california_housing + uniform \| 7.10e-01 (7.58e-03) \| 7.39e-01 (1.75e-02) \| 1.16e+00 (1.79e-03) \| 1.17e+00 (3.04e-03) \| 1.17e+00 (2.93e-03) \| 1.18e+00 (6.17e-03) california_housing + constant \| 2.28e+00 (5.12e-02) \| 6.35e-01 (4.38e-03) \| 1.17e+00 (2.39e-03) \| 1.16e+00 (1.87e-03) \| 1.16e+00 (1.88e-03) \| 1.17e+00 (5.51e-03) california_housing + student-t \| 1.34e+00 (1.68e-01) \| 6.56e-01 (5.46e-03) \| 1.18e+00 (3.56e-03) \| 1.19e+00 (4.13e-03) \| 1.18e+00 (3.83e-03) \| 1.19e+00 (6.74e-03) california_housing + laplace \| 1.00e+00 (5.04e-02) \| 6.51e-01 (4.71e-03) \| 1.18e+00 (3.41e-03) \| 1.18e+00 (3.91e-03) \| 1.18e+00 (3.67e-03) \| 1.18e+00 (6.30e-03) Title: Results for Suggested Benchmarks --- Rebuttal 3: Comment: We highlight the main limitation of Relevance Pursuit versus the additional baselines: if 100% of the data points are subject to heavy-tailed noise, the methods based on uniformly heavy-tailed likelihoods, including `Huber + Projection` will perform best, as the following results demonstrate, similar to the setup in Table 1 of Algikar et.al., 2023. # 100% Laplace Noise ## RMSE \| Data \| Standard \| Relevance Pursuit \| Student-t \| Laplace \| Huber \| Huber + Projection \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| neal + laplace \| 1.51e+00 (1.06e-01) \| 2.40e+00 (2.72e-01) \| 1.16e+00 (9.18e-02) \| 1.12e+00 (9.90e-02) \| 1.12e+00 (9.80e-02) \| 1.12e+00 (9.80e-02) friedman5 + laplace \| 1.39e+01 (5.95e-01) \| 1.37e+01 (6.36e-01) \| 8.33e+00 (3.69e-01) \| 7.34e+00 (3.50e-01) \| 7.40e+00 (3.45e-01) \| 7.40e+00 (3.45e-01) friedman10 + laplace \| 1.30e+01 (3.77e-01) \| 1.27e+01 (3.99e-01) \| 7.24e+00 (1.99e-01) \| 6.07e+00 (1.91e-01) \| 6.26e+00 (1.71e-01) \| 6.26e+00 (1.71e-01) yacht_hydrodynamics + laplace \| 2.57e+01 (1.17e+00) \| 4.75e+01 (3.51e+00) \| 1.64e+01 (3.18e-01) \| 1.61e+01 (2.52e-01) \| 1.62e+01 (2.60e-01) \| 1.67e+01 (3.55e-01) california_housing + laplace \| 1.57e+00 (8.25e-02) \| 2.06e+00 (1.44e-01) \| 1.21e+00 (1.08e-02) \| 1.20e+00 (9.39e-03) \| 1.20e+00 (9.69e-03) \| 1.23e+00 (2.05e-02) ## NLP \| Data \| Standard \| Relevance Pursuit \| Student-t \| Laplace \| Huber \| Huber + Projection \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| neal + laplace \| 1.89e+00 (1.14e-01) \| 1.38e+02 (3.98e+01) \| 1.57e+00 (9.22e-02) \| 1.55e+00 (9.20e-02) \| 1.55e+00 (9.12e-02) \| 1.55e+00 (9.12e-02) friedman5 + laplace \| 4.55e+00 (2.81e-02) \| 4.56e+00 (1.88e-02) \| 3.64e+00 (4.03e-02) \| 3.46e+00 (4.45e-02) \| 3.48e+00 (4.31e-02) \| 3.48e+00 (4.31e-02) friedman10 + laplace \| 4.58e+00 (1.25e-02) \| 4.57e+00 (1.27e-02) \| 3.60e+00 (2.53e-02) \| 3.32e+00 (3.01e-02) \| 3.37e+00 (2.61e-02) \| 3.37e+00 (2.61e-02) yacht_hydrodynamics + laplace \| 4.74e+00 (4.45e-02) \| 5.30e+00 (8.31e-02) \| 4.91e+00 (8.87e-02) \| 5.19e+00 (1.17e-01) \| 5.12e+00 (1.04e-01) \| 9.90e+00 (6.44e-01) california_housing + laplace \| 1.88e+00 (5.25e-02) \| 2.24e+00 (9.41e-02) \| 2.92e+00 (6.01e-02) \| 3.33e+00 (7.47e-02) \| 3.27e+00 (5.84e-02) \| 6.64e+00 (3.28e+00) Title: Results for uniformly heavier-tailed noise --- Rebuttal 4: Comment: Last, we provide the code of the main parts of our translation of the Matlab code of Algikar et.al., 2023 in the following, for you to be able to check its correctness. ``` def projection_statistics(H: Tensor) -> Tensor: """ Args: H: (n x d)-dim Tensor, i.e. number of data points x dimensionality. NOTE: in the original code, this is taken to be X with an appended ones column. Returns: A (n)-dim Tensor of projection statistics. """ dtype = H.dtype device = H.device m, n = H.shape M = torch.median(H, dim=0, keepdim=True).values # (1 x d) row vector u = torch.zeros(m, n, dtype=dtype, device=device) # i.e. (n x d) matrix v = torch.zeros(m, n, dtype=dtype, device=device) z = torch.zeros(m, 1, dtype=dtype, device=device) # i.e. (n x 1) matrix P = torch.zeros(m, m, dtype=dtype, device=device) # i.e. (n x n) matrix eps = 1e-6 # avoiding divide-by-zero issues for kk in range(m): u[kk, :] = H[kk, :] - M # looping over data points v[kk, :] = u[kk, :] / max(torch.linalg.norm(u[kk, :]), eps) for ii in range(m): z[ii, :] = torch.dot(H[ii, :], v[kk, :]) zmed = torch.median(z, dim=0, keepdim=True).values MAD = 1.4826 * (1 + (15 / (m))) * torch.median(torch.abs(z - zmed)) for ii in range(m): P[kk, ii] = torch.abs(z[ii] - zmed) / max(MAD, eps) PS = torch.amax(P, dim=0) return PS def _compute_projection_statistics_weights(H: Tensor, PS: Tensor) -> Tensor: """This computes the weights for the projection statistics, according to the procedure in the original Matlab code: https://github.com/apooja1/GP-Huber/blob/fee038963b471eb198d59b22d57b89e69f451d8c/Experiments/Friedman """ niu = torch.sum(H != 0, dim=-1) cutoff_PS = torch.zeros_like(niu) for i in range(len(niu)): # for chi2, see this post for correspondence of ppf with invcdf: # https://stackoverflow.com/questions/53019080/chi2inv-in-python # the 0.975 was copied from the original matlab code: # https://github.com/apooja1/GP-Huber/blob/fee038963b471eb198d59b22d57b89e69f451d8c/Experiments/Friedman.m#L186 cutoff_PS[i] = chi2.ppf(0.975, df=niu[i].item()) weights = (cutoff_PS / PS.square()).clamp(max=1.0) return weights def compute_projection_statistics_weights(X: Tensor) -> Tensor: # X: n x d n, d = X.shape H = torch.cat((X, torch.ones(n, 1)), dim=-1) # n x (d + 1) PS = projection_statistics(H) # (n x d) -> n return _compute_projection_statistics_weights(H, PS) ``` For the Huber loss, we use $\epsilon = 0.45$ and $b = 0.5$, identical to the setting of the [Github repository](https://github.com/apooja1/GP-Huber/blob/fee038963b471eb198d59b22d57b89e69f451d8c/basics/lik_huber.m#L191) of Algikar et.al., 2023. All reported results were generated with 32 independent replications, reporting the mean and standard error. An additional implementation detail: the `median` and `max` calls take the extra `dim=0` argument in PyTorch, as the Matlab function are defined to compute the row-wise maximum. Also, we added a small numerical constant `eps` in the code of `projection_statistics` to avoid division-by-zero errors. The rest of the code should be an almost verbatim translation. A word of thanks Thank you again for your valuable suggestions, we believe they make the paper stronger. Let us know if you have additional questions or if all your remaining concerns have been addressed. Title: A PyTorch implementation of Algikar et.al.'s projection statistics --- Rebuttal Comment 4.1: Comment: Thank you for providing the performance comparisons with additional models, especially on such short notice. I find the comparison to be very thorough and have adjusted my score accordingly.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their detailed comments and valuable suggestions. We are glad to see that the reviewers found our "novel approach” to be “clearly motivated and presented” and of “broad interest”, and that our work “provides a nice addition to the existing toolbox of methods”. While reviewers in general found our work “well-written” and “easy to follow”, there were some requests for improving clarity (e.g. Section 4.3), more readable figures, and a more comprehensive discussion of the regression results. We will gladly make these adjustments, and will utilize the additional page to improve legibility and include the additional experiments suggested by the reviewers and added here. ## RCGP Benchmarks In response to the desire for additional regression results, we ran a comprehensive suite of benchmarks that include various UCI datasets, and included the RCGP method from the contemporaneous work Altamirano et al. (ICML 2024) as an additional baseline, using their publicly available benchmarking suite. The results are presented in Table 1 in the attached pdf. The main takeaways are: - Overall, RGP-RP outperforms the baselines in terms of Mean Absolute Error (MAE) - the metric chosen by Altamirano et al - in 50% of the test cases, is tied with the standard GP in 40% of cases, and the Student-T GP (t-GP) is best in 10% cases. There is not a single case where RCGP or standard GP outperforms RGP-RP in a statistically significant way (standard errors in parenthesis). - For uncorrupted data, all methods perform comparably. Notably, RGP-RP shows predictive performance indistinguishable from the standard GP, an indication that the RGP-RP’s Bayesian model selection correctly chooses the outlier-free model. - For “Uniform” and “Asymmetric” Outliers, as defined in Altamirano et al. 2024, RGP-RP outperforms the baselines for all test cases, showing that our results generalize to more problems. - RGP-RP takes longer to fit than other baselines but achieves superior predictive performance. Fitting times in these benchmarks are on the order of seconds, and so this should not be a concern for most practical applications. Relative timings that each method took to complete the entire benchmark are: - GPR (GPFlow): 1.9x, - t-GP (GPFlow): 20.6x - RCGPR (GPFlow): 5.7x - Standard (BoTorch): 1.0x - RGP-RP (BoTorch): 36.2x _Note: The original RCGP benchmarks from Altamirano et al. (2024) [hard-code the seed that controls the random train-test-splits](https://github.com/maltamiranomontero/RCGP/blob/aff281a39a6be6eefc15d96db8acba6c49224d28/experiments/uci/dataset_api.py#L321), therefore leading the results to only use a single train-test split, contrary to what is indicated in the paper. We edited the code to use different seeds for each replication._ ## Twitter Flash Crash As an additional real-world example, we consider the “twitter flash crash” from Altamirano et al. (2024), see Figure 1 in the attached pdf for results. _Note: The original notebook sets the RCGP noise variance to the one inferred by GPR, and turns off the learning of the variance. In contrast, RCGP jointly learns the noise variance in the original benchmarks._ In the top row of Figure 1, we show results for - “RCGP (GPR Noise)”, the original method that forces the variance to the GPR value. - “RCGP (Trained Noise)” sets the variance to be trainable. - “RCGP (Fixed Noise)” forces the noise variance to 0.25. The top row shows that RCGP with the original setting (GPR Noise), and (Fixed Noise) are less affected by the outliers than GPR, but still substantially so. Surprisingly, RCGP (Trained Noise) is even more affected. Further, the RGP-RP (“Relevance Pursuit”) is virtually unaffected by the outlying data points while modeling other high-frequency components of the time series closely. ## Twitter Flash Crash with Additional Training Day A key shortcoming of RCGP is that its weighting function only uses the distribution of outcomes Y relative to the prior mean $m({\bf x})$. For this to be effective, the outliers need to be separable from the marginal data distribution, like, e.g., for winsorization. The reason RCGP shows improvements on the single-day example is that the outliers at the sell-off time are a-priori separated from most of the data. If this is not the case, RCGP can up-weight outliers, while down-weighting real data. We illustrate this failure in the bottom row of Figure 1, generated by included data from the preceding day. Here, RCGP is not robust to the outliers - in fact, the weighting function assigns a very large weight to the outlier, while RP-GP exhibits consistently strong performance. ## Additional Results: UCI Data using RP Setup We also added UCI datasets (yacht, energy, CA housing) into our own benchmarking suite. Figure 2 in the pdf contains the results for the yacht data with “uniform” outliers. Note that RCGP’s “uniform” outlier process for Table 1 is different from ours: - RCGP’s “uniform” outliers: generated by adding or subtracting a random value uniformly sampled between 3 and 9 standard deviations of the original data, thereby separating outliers from the data. - RP’s “uniform” outliers: generated by uniformly sampling inside the range of the uncorrupted data. This leads to the differences between the results reported in Figure 2 and Table 1 on the yacht data. The results show that RP generally outperforms the baselines in terms of MAE and LL over a range of corruption percentages, and is faster than the robust Student-T and Trimmed-MLL approaches. The results on the CA housing and energy data are qualitatively similar. We will include them in the manuscript. ## An Ask With these additional results, we believe that the main weaknesses of the paper cited by the reviewers – the regression experiments and comparison to additional baselines (RCGP) – were addressed comprehensively. Based on this, we would like to ask the reviewers to consider raising their score. Pdf: /pdf/d923e3cd04fcd424b803cdc982f6af7acb7c2c94.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On the Surprising Effectiveness of Attention Transfer for Vision Transformers	Accept (poster)	Summary: In this work, the authors demonstrate that a large part of the benefits of pre-training in ViT models actually comes not from the pre-trained features but the pre-trained knowledge of attention maps. Specifically, the authors propose an alternative to fine-tuning call Attention Transfer, and they use this method to transfer attention maps from pre-trained ViTs to teach from-scratch ViTs. This method achieves results comparable to results from finetuning. The authors present two variations of Attention Transfer, the first of which directly copies the attention maps from the teacher network, while the second teaches the student network to make its own attention maps with a distillation objective derived from the teacher attention maps. The authors demonstrate impressive and surprising results on ImageNet-1k, and present further analysis of this approach in a variety of configurations. While the proposed Attention Transfer method does have some clear limitations, specifically for domain gaps, overall, the work has very interesting implications for ViT pretraining and finetuning. Strengths: The main result of the work, which is to demonstrate the key importance of pre-trained attention maps over pre-trained features in ViT finetuning/transfer learning, is very interesting, and has important implications for the use of ViTs. While I think it is unlikely that Attention Transfer (in its current form) will replace standard finetuning (see notes in the following section), I think the implications of the results and analysis are still very important. The potential to use Attention Transfer with ensembles is also interesting, and the authors show that it has the potential to boost the performance of self-supervised ViTs further. This work also helps to explain some of the properties of MAE. In particular, prior works have found that MAE lends itself well to fine-tuning, but not as well to direct linear probes. This work shows a possible explanation: that the strength of MAE is not in its pretrained features but instead in its pretrained attention maps. The work is clearly presented and has wide coverage of many model and training configurations in the main work and appendix. They also include additional analysis on the impact of partial transfer, and assess if the student is or is not re-learning the same features as the teacher. Overall, their analysis is quite comprehensive. Weaknesses: While the proposed Attention Transfer method has very interesting implications for ViTs, I’m not sure if it will make its way into practical use, for either the Attention Copy or Attention Distillation variants. The main issue is that the proposed method does not always match the performance of the regular fine-tuning approach, particularly in cases with a domain gap between the pretraining data and the downstream task. In addition, Attention Transfer is more expensive to train than standard Fine Tuning, as acknowledged in Section A.2. For some of the analysis results, it would be very helpful to see how Attention Distillation performs as compared to Attention Copy. In particular, in any analysis where there is a domain gap (Table 2 for example), it would be interesting to see if Attention Distillation performs better, as it is suggested that Attention Distillation allows more flexibility in the student and thus may perform better in such cases. On a less important note, I find that the visualizations in Figures 7, 10, and 11 are somewhat difficult to see due to the combination of the heat map colors and the background images. I would suggest revising these figures to try and make them clearer. Technical Quality: 4 Clarity: 4 Questions for Authors: In the Attention Copy setting, are the unnecessary Q and K layers removed from the student network? There is an important conclusion in the work that seems somewhat under-acknowledged. In lines 210-217 and in Table 1, it is found that transferring the teacher’s Q is more effective than transferring the attention map. Why was more attention and testing not performed on Q transfer? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors discuss the limitations of Attention Transfer and present results for it with many configurations. Overall, they are quite transparent about acknowledging the situations where Attention Transfer underperforms full Fine Tuning. I do not see any risks for negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We’re glad that you found our results interesting and our analysis comprehensive. We address your questions below: > I’m not sure if it will make its way into practical use, for either the Attention Copy or Attention Distillation variants. We agree that the attention transfer methods are not currently ready for practical use. One of our main objectives was to use attention transfer to understand the role of features vs attention in pre-trained models. But beyond scientific utility, attention transfer may have some advantages over fine-tuning in the future: - An attention map (of size LxL, L is the sequence length) does not depend on the model’s dimension, which means the map from a model can be directly transferred to a different-sized model. - Attention transfer gets rid of layer-wise learning rate decay. This is a crucial hyper-parameter almost used everywhere for tuning pre-trained vision models (beyond ViT). The fundamental prior here is that early layers should change less compared to later layers. But such a prior can be a restriction for next-generation models and getting rid of it opens up new opportunities. - Sharing weights can incur security risks (e.g., white-box attacks). In such settings, we need an effective way to transfer knowledge from the pre-trained model, and attention transfer offers such a possibility. We will add more detail on avenues for future research in the paper, and we hope other researchers find more practical ways to use attention transfer! > For some of the analysis results, it would be very helpful to see how Attention Distillation performs as compared to Attention Copy. In our main rebuttal above, we have updated tables so that we have both Attention Copy and Attention Distillation. In general, Attention Distillation performs better than Attention Copy, which follows our existing findings. > Revising attention map visualizations We have modified the figures to be easier to see. We have included samples in the 1 page PDF. Let us know if you have any further suggestions! > In the Attention Copy setting, are the unnecessary Q and K layers removed from the student network? Yes, we remove them from the student network. > Why was more attention and testing not performed on Q transfer? Thank you for pointing this out! See [a] in the main rebuttal above. --- Rebuttal Comment 1.1: Comment: I thank the authors for their responses and discussion. I agree that the authors should be careful about phrasing to avoid a possible "overclaim" for Attention Transfer vs Transfer Learning. I also support the revised visualizations. Overall I think this work has very interesting implications about transfer learning in ViTs and I maintain my original rating in support of accepting this work.	Summary: The paper introduces attention transfer as an alternative to fine-tuning in Vision Transformers (ViT), separating intra-token and inter-token operations to enhance feature extraction and combination. Attention transfer contains attention copy and attention distillation. Attention distillation matches the performance of fine-tuning on ImageNet-1K while learning different representations, facilitating the performance of feature ensemble. The method is verified across various model scales and datasets. Strengths: 1. The paper first points out the sufficiency of attention maps in pretrainings and provides extensive analyses on attention transfer. 2. The proposed attention distillation is verified across various model scales and datasets. Weaknesses: 1. Compared with full fine-tuning, the proposed attention transfer method introduces additional computation costs on the forward process of the teacher model. Therefore, to make the article more comprehensive, authors should consider including a baseline of distillation on features. 2. Transfer tasks (e.g., Table 2 and 3) should contain the results of training from scratch. 3. There is an important result in L169, the result of attention copy from a fine-tuned model (85.6), not appearing in any Tables or Figures. There should be an extra Table or Figure containing the result with other similar settings (e.g., attention copy from only pre-trained model) for better comparison. Technical Quality: 3 Clarity: 2 Questions for Authors: In Table 1, transfer Q is better than transfer Q,K (which is equivalent to the attention map), does it mean distillation on features would achieve better performance? If so, is attention really important and sufficient for transfer tasks? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the feedback on our work! Below, we respond to your questions and comments: > In Table 1, transfer Q is better than transfer Q,K (which is equivalent to the attention map), does it mean distillation on features would achieve better performance? See the main rebuttal above for a detailed discussion of the Q-copying result. For feature distillation, we followed your suggestion and tried distilling the residual stream features from a pre-trained MAE ViT-L. In our preliminary results, we obtain a downstream accuracy of 81.3 on ImageNet-1k. This is significantly lower than the 85.7 that can be achieved through fine-tuning or attention distillation. This makes sense: the features learned during self-supervised pre-training are not directly well-suited for classification, so trying to match them can hurt performance. CKA analysis of the features (Fig. 5) supports this hypothesis – the fine-tuned MAE does well by significantly changing the features in the latter half of the network. Overall, transferring attention appears to do much better than distilling the features. > Transfer tasks (e.g., Table 2 and 3) should contain the results of training from scratch. In our main rebuttal above, we have updated Table 2-3 with the results of training from scratch. > There should be an extra Table or Figure containing the result with other similar settings (e.g., attention copy from only pre-trained model) for better comparison. Thanks for the suggestion! We will add a new subsection and table that shows the effect of the teacher (pre-trained vs fine-tuned) on attention copy and distillation. We hope this will be more organized for readers. --- Rebuttal Comment 1.1: Comment: Their rebuttal has addressed most of my concerns. I find this paper interesting and believe it will benefit the community. As a result, I have increased my score to weak accept. However, I am still not convinced that the attention map is the primary factor for the high performance, rather than the Q features. More qualitative or quantitative evidence would support the claims presented in the paper. --- Reply to Comment 1.1.1: Comment: We're glad that we have addressed most of your concerns! Below, we clarify the result on copying Q: > I am still not convinced that the attention map is the primary factor for the high performance, rather than the Q features. - We emphasize that the only way that copying Q affects the student model is by making it easier to learn a useful attention map $softmax(QK^\top)$. Thus, the strong performance of copying Q should support our surprising findings on the importance of the pre-trained attention maps. - During the rebuttal period, we experimented with distilling only the Q activations from the teacher to the student. We found that this achieves 85.0 when distilling from a pre-trained MAE ViT-L, which is worse than copying Q. We suspect that this is because the student must first learn to match Q and then learn the K that creates a good attention map. Only then does the Q-distillation attention map provide reasonable guidance to learn good features. - Q distillation does not match the 85.7 that Attention Distillation achieves. This supports our hypothesis that copying Q does well because it allows the model to slightly modify the pre-trained attention maps to better match the downstream task. Indeed, Attention Copy does well when the teacher maps are well-suited for the downstream task: copying from a fine-tuned MAE ViT-L achieves the same 85.6 accuracy (L169). - In our original response to Reviewer p7aT, we reported the result of distilling the features from a pre-trained MAE model. This achieved an accuracy of 81.3, significantly lower than what Attention Transfer achieves, which further corroborates our story that features are not as important as previously thought. - We will add these new results and discussion on copying Q to the paper, around L217. - Finally, Figure 7, 10, and 11 visualize the attention maps learned by attention distillation and qualitatively show that it matches the teacher’s attention maps well for the layers that are distilled. This links the pre-trained attention maps to the high downstream task performance. Overall, we believe that our findings on the sufficiency of attention maps are already quite surprising and useful for the research community. We have run extensive experiments and ablations and hope that our paper provides valuable insights that potentially motivate the next generation of pre-training and fine-tuning algorithms.	Summary: The authors propose a novel perspective on the utility of pretraining vision transformers by demonstrating that the actual features and representations learned during pre-training are not crucial. Instead, they find that simply re-using the self attention from pre-training specifically, the way information flows between tokens—is sufficient for models to learn high-quality features from scratch and achieve performance comparable to pre-trained models. To support their claim, the authors introduce a method called attention copy and attention distill. This method involves transferring the attention from a pre-trained teacher ViT to a student ViT, either by copying or distilling the attention maps. This approach allows the student model to learn its own features while still benefiting from the pre-trained attention patterns. The authors also highlight several drawbacks of previous works that heavily rely on pre-trained features and representations. They point out that the conventional fine-tuning approach may not be as effective under distribution shift settings, where the pre-trained features might not generalize well. In contrast, their attention transfer method provides a more robust alternative that maintains high performance even when the distribution shifts. Through systematic experiments, the authors examine various aspects of their findings, particularly focusing on the sufficiency of attention maps. They provide evidence that attention patterns alone can guide the learning process effectively, thus questioning the necessity of the entire pretraining paradigm. Strengths: The following are some strengths of this work - The problem is well motivated and also also been discussed in previous works. Not specifically the distillation approach, but reusing attention maps has been interesting not only in the vision community, but widely studied in NLP. - I found that the paper is beautifully written, it basically shows the efforts that the authors took in trying to explain each aspect of their work. They keep the language simple and easy to understand, with concise explanations together with proper visualizations and plots where necessary. - The analysis of different components such as transferring attention from different layers, different heads, CKA analysis etc. is very well thought and it was a joy to read through the findings. So I thank the authors for this and encourage them to do these kind of analysis in all their future works as well. - The experiments on different tasks such as image classification, model robustness etc. show the effectiveness of the approach. Weaknesses: The following are some queries: - In Figure 5, it is surprising to see that attn-copy has least correlation with respect to the fine-tuned model as compared to attn-distill. In attn-distill from layer 20-24, the correlation increases much more than the pretrained model, which is not the case with attn-copy. Do the authors have any intuition on why this is the case? - In general, with the CKA computation. I think it would be more interesting to understand the correlation across the features at different layers of the model. I would refer the authors to the work in [51], where the authors show correlation plot of features across every layer before and after their method is applied. This would help understand how attn-copy and attn-distill affect representations learned by the model. The authors can take a look at Figure 2 and Figure 4 in [2] for reference. - Continuing from above, it would also be interesting to see this correlation across different heads after the model is pretrained with attn-copy and attn-distill. In [1], the author shows that there exists high correlation across attention heads in ViTs, so it would be interesting to see if attn-copy or attn-distill mitigates this. [1] https://github.com/sayakpaul/probing-vits?tab=readme-ov-file#visualizing-mean-attention-distances [2] Zhou et al., Refiner: Refining Self-attention for Vision Transformers, arxiv 2021 - Do the authors have the same observation for attn-copy and attn-distill when using ViT-L pretrained in a supervised setting. Also, im curious to know if the same observation holds for methods pretrained using self-distillation approaches such as DINO, BYOL, iBOT or distillation approaches like SimSiam, etc. - I think I might be missing something, but the visualization of attention from the [CLS] token, in Figure 7 seems that attn-distill has worse attention than attn-copy at the deeper and intermediate layers. Also the localization of the object seems to be bad as well. Can the authors please comment on this. - I would also like to see the comparison with different state-of-the-art methods that use ViT-L as their backbone is tasks like classification, object detection and robustness. Technical Quality: 3 Clarity: 4 Questions for Authors: There have been works such as [51], that illustrate the performance of copying attention across intermediate layers of the network. However, I would urge the authors to also include the following works for completeness, which have shown that copying attention works well in the domain of NLP [3] Xiao et al., Sharing attention weights for fast transformer, IJCAI 2019 [4] Wang et al., Evolving attention with residual convolutions, ICML 2021 [5*] Ying et al., Lazyformer: Self attention with lazy update, arXiv 2021 Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Yes the authors discuss the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We’re really glad that you liked the motivation, analysis, and writing of our paper! We respond to your comments below: > Fig. 5: In attn-distill from layer 20-24, the correlation increases much more than the pretrained model, which is not the case with attn-copy. Do the authors have any intuition on why this is the case? Our main hypothesis is that this is because we transferred all 24 layers for Attention Copy but only distilled 18 layers for Attention Distillation. This means that Attention Distillation is more flexible in how it combines the features for the last 6 layers, so it can find a strategy similar to the fine-tuned MAE. In contrast, Attention Copy is constrained to be more similar to the pre-trained MAE. > Do the authors have the same observation for attn-copy and attn-distill when using ViT-L pretrained in a supervised setting. Also, im curious to know if the same observation holds for methods pretrained using self-distillation approaches such as DINO We are running this right now! We hoped to have the results by now but had an issue with our cluster. We will post an update below with the results as soon as possible. > in Figure 7 seems that attn-distill has worse attention than attn-copy at the deeper and intermediate layers. Also the localization of the object seems to be bad as well. Attention Distillation only really deviates from the pre-trained attention maps at the later layers of the network. This makes sense, since we don’t distill the last 6 layers’ attention maps. Furthermore, the “noisy” attention pattern can sometimes be quite useful in Vision Transformers, since they often tend to store information in low-entropy regions of the background (see “ViTs Need Registers” [6]). The fine-tuned MAE shows similar “noisy” patterns in its later layers, which we can see after fixing a small issue with our attention map visualizations: since we follow the standard practice of using global average pooling, which averages the representations at all spatial locations, the CLS token representation is not used after the 24th layer. This means that its attention map has no signal to improve. We fix this by now showing the pattern after the 23rd layer instead. The examples in the 1 page PDF show that the fine-tuned MAE also exhibits the same patterns in the later layers as Attention Distillation. > I would also like to see the comparison with different state-of-the-art methods that use ViT-L as their backbone is tasks like classification, object detection and robustness. To the best of our knowledge, fine-tuned MAE is the SOTA ViT-L model without using extra data on ImageNet classification and the OOD robustness benchmarks. For object detection, the SOTA ViT-B-based model using ImageNet-1k is ViTDet [32] with 56.0 $AP^{box}$ and 48.0 $AP^{mask}$. Note that the detection results cannot be directly compared with ours due to further architectural modifications within ViTDet on top of the ViT backbone. > I would urge the authors to also include the following works for completeness, which have shown that copying attention works well in the domain of NLP Thank you for suggesting these NLP papers – we will definitely add them in our related works section! [6] Darcet, T., Oquab, M., Mairal, J., & Bojanowski, P. (2023). Vision transformers need registers. arXiv preprint arXiv:2309.16588. --- Rebuttal Comment 1.1: Title: Response to author's rebuttal Comment: I thank the authors for the responses. I think they have answered most of my queries and am satisfied with it. Im looking forward to the results of the DINO experiment. I keep my initial rating. Thanks --- Rebuttal 2: Comment: Here are the results of fine-tuning, Attention Copy, and Attention Distillation from a DINO ViT-B/16 pre-trained model. Note that our DINO fine-tuning recipe achieves an accuracy of 83.2, higher than the 82.8 that has been achieved in previous papers. Overall, we find similar results as we originally reported in Table 5 with MoCo (another representation learning method based on self-similarity), as well as FLIP (a vision-language contrastive method). \| \| tune \| copy \| distill \| \|-\|-\|-\|-\| DINO \| 83.2 \| 82.2 \| 82.8 \|	Summary: This paper investigates how transferring attention patterns from a pre-trained ViT to a student affect the student's downstream performance. By applying the attention copy strategy, the paper shows that when the pre-trained dataset and downstream dataset are the same, the trained student may achieve performance superior than the student training from scratch. Moreover, it largely recovers the performance of a pretrained-and-finetuned model. Further, the authors propose an attention transfer (distillation) scheme. With attention distillation scheme, the student may achieve performance comparable or even better than the pretrained-and-finetuned model. The authors provide extensive experiments to verify the effectiveness of attention transfer and show when attention transfer works. Strengths: - The paper is well-written and easy to follow. - The findings is meaningful and interesting to an extent. - Extensive experiments are conducted. Weaknesses: - The findings is somewhat similar to previous works which apply attention transfer in ConvNets. The only different things is in ViT, attention can be explicitly represented, which eases the operation of attention transfer. - The authors seem to over-claim something, e.g., "offer a potential alternative to the decade-long practice of fine-tuning pre-trained vision models". The empirical results in the paper cannot sufficiently support this claim. For example, we only see comparable results on one setting where the model is pretrained on ImageNet (unsupervised/self-supervised pretraining) and finetuned on ImageNet. For other settings, like out-of-distribution tasks, detection tasks, etc., the attention transfer does not work as well as fine-tuning. - The experiments which study the effect of transferring a subset of Q, K, V seems not to support the main claim of this paper. The results show that transferring Q is the best of all. So why do we choose to transferring Q and K (or attention)? It implies that transferring the attention pattern may not be the key to the superior performance on downstream task, although transferring attention patterns sounds reasonable and interpretable. - For Table 2-6, it is strange to see only partial results in one table. It is better to show the results for "training from scratch", "fine-tuned", "copy" and "distill". Technical Quality: 2 Clarity: 3 Questions for Authors: Please see the weakness part. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We’re glad that you found our work well-written, with extensive experiments and interesting findings. Below, we respond to your questions and comments: > The findings is somewhat similar to previous works which apply attention transfer in ConvNets. The only different things is in ViT, attention can be explicitly represented, which eases the operation of attention transfer. Previous works on ConvNets have mainly looked at transferring different properties of the features, like spatial feature magnitudes. They have also mainly been conducted in the knowledge distillation paradigm, where a task-specific (not pre-trained) downstream teacher is distilled into a smaller student. Our work differs from prior work in a few ways. First, as you point out, ViTs allow us to explicitly decouple the attention patterns from the features that they combine. Second, we extensively investigate the pre-training/fine-tuning paradigm, and we surprisingly find that the attention maps, by themselves, are often sufficient to achieve the same downstream performance. This result is completely new and calls into question the “feature learning” story that typically motivates pre-training in vision. > The authors seem to over-claim something, e.g., "offer a potential alternative to the decade-long practice of fine-tuning pre-trained vision models" You’re right – we intended to say that attention transfer achieves surprisingly good results, and we didn’t mean to imply that attention transfer was already sufficiently developed to be a practical alternative to fine-tuning. We tried to make this clear in L327-330 in the conclusion. We will clarify our wording in the introduction and write “with further research, attention transfer could be a potential alternative to fine-tuning pre-trained vision models.” > The results show that transferring Q is the best of all. So why do we choose to transferring Q and K (or attention)? It implies that transferring the attention pattern may not be the key to the superior performance on downstream task, although transferring attention patterns sounds reasonable and interpretable. This is a great question! See [a] in the main rebuttal above. > For Table 2-6, it is strange to see only partial results in one table. It is better to show the results for "training from scratch", "fine-tuned", "copy" and "distill". For some of our tables, we had run either “copy” or “distill” due to compute limitations. During the rebuttal period so far, we have been running more experiments to ensure that Tables 2-6 are comprehensive – see [b] in the main rebuttal above. --- Rebuttal Comment 1.1: Title: Re：rebuttal Comment: Thanks for the authors' detailed response. However, I still feel concerned about the novelty and the conclusion of this work. The effectiveness of transferring Q implies that the attention may not be the key to the surprising results of performing attention transfer to an extent. The results are interesting and the story is generally good. But it is still not convincing to me that attention is the underlying key to the performance though it is intuitive. The authors show transferring attention works but transferring Q also works. Is it possible attention is not the key factor? Further evidence should be provided beyond performance numbers, to show the relationship between attention transfer and the performance. Based on the above reasons, I still think the current version is not acceptable. --- Reply to Comment 1.1.1: Comment: > The authors show transferring attention works but transferring Q also works. Is it possible attention is not the key factor? - We emphasize that the only way that copying Q affects the student model is by making it easier to learn a useful attention map $softmax(QK^\top)$. Thus, the strong performance of copying Q should support our surprising findings on the importance of the pre-trained attention maps. - During the rebuttal period, we experimented with distilling only the Q activations from the teacher to the student. We found that this achieves 85.0 when distilling from a pre-trained MAE ViT-L, which is worse than copying Q. We suspect that this is because the student must first learn to match Q and then learn the K that creates a good attention map. Only then does the Q-distillation attention map provide reasonable guidance to learn good features. - Q distillation does not match the 85.7 that Attention Distillation achieves. This supports our hypothesis that copying Q does well because it allows the model to slightly modify the pre-trained attention maps to better match the downstream task. Indeed, Attention Copy does well when the teacher maps are well-suited for the downstream task: copying from a fine-tuned MAE ViT-L achieves the same 85.6 accuracy (L169). - In our original response to Reviewer p7aT, we reported the result of distilling the features from a pre-trained MAE model. This achieved an accuracy of 81.3, significantly lower than what Attention Transfer achieves, which further corroborates our story that features are not as important as previously thought. - We will add these new results and discussion on copying Q to the paper, around L217. - Finally, Figure 7, 10, and 11 visualize the attention maps learned by attention distillation and qualitatively show that it matches the teacher’s attention maps well for the layers that are distilled. This links the pre-trained attention maps to the high downstream task performance. Overall, we believe that our findings on the sufficiency of attention maps are already quite surprising and useful for the research community. We have run extensive experiments and ablations and hope that our paper provides valuable insights that potentially motivate the next generation of pre-training and fine-tuning algorithms.	Rebuttal 1: Rebuttal: We thank reviewers for their time, effort, and feedback for our paper. To recap, reviewers appreciated various strengths of our work: - Significance: “well motivated” (7AbG), “very interesting implications for ViT pretraining and finetuning” (9FPV), “potential to boost the performance of self-supervised ViTs further” (9FPV), “helps to explain some of the properties of MAE” (9FPV), “findings is meaningful and interesting” (qaog) - Experiments: “extensive experiments” (qaog), “joy to read through the findings” (7AbG), “their analysis is quite comprehensive” (9FPV), “extensive analyses” (p7aT) - Clarity: “beautifully written” 97AbG), “well-written and easy to follow” (qaog), “ clearly presented” (9FPV) Next, we provide general comments for some shared topics of discussion. We will address individual concerns and questions by responding to each review. [a] High performance from copying Q (Table 1) [qaog, p7aT, 9FPV] We do observe higher performance (85.6) from copying the self-attention queries $Q$ than Attention Copy (85.1), but lower performance than Attention Distillation (85.7) First, we note that copying $Q$ does not transfer any features – the student network only “sees” the result of using $Q$ to compute the attention map softmax($QK^\top$). Thus, copying $Q$ solely provides structure to the student attention maps. This is consistent with our story that the attention patterns are sufficient to guide networks to high downstream performance. Our hypothesis is that copying $Q$ does well because it gives the model flexibility to change the attention patterns to be more suitable for the downstream task. Directly copying the entire attention map $QK^\top$ is too inflexible, which is why attention distillation works better, as it can deviate from the teacher’s attention maps. Overall, we mainly focus on transferring the entire attention map since it’s a clean way to split the network’s computation (it contains all inter-token communication). To be thorough, we are now training models with $Q$-distillation, which encourages the student self-attention queries $Q$ to be close to those of the teacher. We have had some cluster problems but should have the results within a few days. [b] Updated Tables 2-6 [qaog, p7aT, 9FPV] We update Tables 2-6 so that they all contain these models: trained from scratch, MAE fine-tuned, Attention Copy, Attention Distillation. We will update the main paper with these results as well. Table 2: \|Pre-training data \| tune \| copy \| distill \| scratch \| \|--------------------\|-------\|---------\|--------\|-------\| \| ImageNet-1k \| 85.7 \| 85.1 \| 85.7 \| 83.0 \| \| COCO \| 85.2 \| 83.1 \| 84.6 \| 83.0 \| Table 3: \|evaluation data\| tune \| copy \| distill \| scratch \| \|---------\|-------\|---------\|--------\|---------\| \| iNat 2018 \| 79.9 \| 71.8 \| 74.1 \| 64.3 \| \| iNat 2019 \| 83.8 \| 77.9 \| 80.0 \| 66.2 \| Table 4: \|out-of-distribution evaluation\| tune \| copy \| distill \| scratch \| \|---------\|-------\|---------\|--------\|---------\| \| ImageNet-A \| 56.5 \| 48.9 \| 54.3 \| 32.0 \| \| ImageNet-R \| 59.6 \| 57.5 \| 56.8 \| 51.9 \| \| ImageNet-S \| 45.2 \| 43.1 \| 42.9 \| 38.0 \| \| ImageNet-V2 \| 76.4 \| 75.5 \| 75.9 \| 72.4 \| Table 5: \|pre-training method \| tune \| copy \| distill \| \|---------\|-------\|---------\|--------\| \| MAE \| 85.7 \| 85.1 \| 85.7 \| \| MoCo-v3 \| 84.0 \| 82.5 \| 83.3 \| \| FLIP \| 87.4 \| 86.6 \| 86.1 \| \| none \| 83.0 \| 72.7 \| 76.3 \| Table 6: \|model\| tune \| scratch \| copy \| distill \| \|---------\|-------\|---------\|--------\|---------\| \| ViT-B \| 82.5 \| 83.6 \| 82.0 \| 83.4 \| \| ViT-L \| 83.0 \| 85.7 \| 85.1 \| 85.7 \| \|ViT-H \| 83.0 \| 86.9 \| 86.1 \| 86.3 \| Pdf: /pdf/adf999b7c642a0c3bed4284723ce7a206c5c429c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MeMo: Meaningful, Modular Controllers via Noise Injection	Accept (poster)	Summary: This manuscript presents a hierarchical controller for reinforcement learning-based control. In particular, given a robot that can be decoupled into a high-level controller and a few low level (joint level) controllers, the two modules are learned jointly via a behavior cloning objective. The proposed method further makes the high-level controller robust to noise from the output of the low-level controllers by adding noise to their outputs. The model experiments on a few morphologies (6/12-leg centipede, 6/10-leg worm, etc) to demonstrate its effectiveness over vanilla RL and NerveNet. Strengths: 1. Presents a hierarchical controller that enables sample efficient policy learning under morphological changes 2. Prior methods require training on many morphologies, then generalize to an unseen morphology. The presented method is trained on one morphology and then transferred to a new morphology. 3. Results suggest that the hierarchical formulation provides improved sample efficiency over baseline methods Weaknesses: 1. Opposing strength 2, especially in simulation, many existing works randomize over many morphologies to enable efficient transfer between them or to an unseen morphology. It is unsure how this method can be trained on a dataset that contains more than one morphology. 2. For the experiments, the reviewer is uncertain how significant is the morphology change in affecting the dynamics of the system. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Can the expert policy on 6-leg centipede, and 6-leg worm be executed on 12-leg centipede and the 10-leg worm? The same question applies to the manipulation setting (4-finger to 5-finger claw). 2. How effective is the method at correcting the dynamics of the system? I.e. after pre-training on the 6-leg centipede, if two or more of the pre-trained joints failed (i.e. uncontrollable), how fast can it adapt to the new setting? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Present in the manuscript. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the review. Below we address the reviewer’s concerns and questions. Under "Weaknesses": > Opposing strength 2, especially in simulation, many existing works randomize over many morphologies to enable efficient transfer between them or to an unseen morphology. It is unsure how this method can be trained on a dataset that contains more than one morphology. Our work is designed for the use case where the user does not have available to them a large number of morphologies. On the one hand, if you do have many morphologies, you can easily extend our approach to train the modules across the different morphologies and you may even get better modules; on the other hand, some of our baselines which are explicitly designed for such a use case may then outperform our approach. > For the experiments, the reviewer is uncertain how significant is the morphology change in affecting the dynamics of the system. One piece of evidence to demonstrate the difficulty of the structure transfer task is in Figure 12 (Appendix A.12), where we test the zero-shot generalization of NerveNet-Conv by fixing all weights on the 6 to 12 leg centipede transfer. If the change in morphology did not significantly affect the dynamics, we would expect the pretrained model to still perform relatively well. However, our experiments show that without finetuning, the NerveNet-Conv performance is much worse than before, showing that the coordination required at transfer time is significantly different from pretraining. Under "Questions": > Can the expert policy on 6-leg centipede, and 6-leg worm be executed on 12-leg centipede and the 10-leg worm? The same question applies to the manipulation setting (4-finger to 5-finger claw). In both locomotion and manipulation, our expert policy is a monolithic neural network, so due to the mismatch in the input and output space, it cannot be executed on our transfer morphologies. In addition, NerveNet [1] has MLP baselines in which they transfer weight matrices between hidden layers and they find that these baselines perform worse than their approach. [1] Wang et al. “NerveNet: Learning Structured Policy with Graph Neural Networks” > How effective is the method at correcting the dynamics of the system? I.e. after pre-training on the 6-leg centipede, if two or more of the pre-trained joints failed (i.e. uncontrollable), how fast can it adapt to the new setting? We are not entirely sure how to interpret the question, but if the reviewer means that there is a possibility that due to the narrow interface between the boss and modules, the boss is unable to have fine-grained control over the modules, we find that empirically the boss is still able to learn new dynamics during transfer time. For example, in the 6 to 12 leg centipede transfer, some of the legs in the 12 leg centipede have more limited range of motion than in the 6 leg to avoid collisions. Our results also show that our approach outperforms the baseline where we train the same modular architecture from scratch on the transfer morphology, even though the baseline has more flexibility to learn the appropriate dynamics. Alternatively, if the reviewer is referring to a scenario where during transfer time, some joints may fail to perform as expected even when given the correct control signal, we believe our framework would be able to adapt in this case as well. Our framework learns modules that represent control signals on a low-dimensional manifold with respect to the boss signal. Even though the manifold would change as a result of the pretrained joints behaving differently at transfer time, our pretrained modules still enable us to navigate along a low-dimensional manifold to infer the optimal coordination, enabling the Boss to be quickly adapted at transfer time. If neither response addressed the reviewer’s question, we are happy to provide more clarification. --- Rebuttal Comment 1.1: Comment: Dear authors, Thanks for addressing my concerns regarding your paper and sorry for the belated review. > On the one hand, if you do have many morphologies, you can easily extend our approach to train the modules across the different morphologies and you may even get better modules Can you clarify how you may train the modules across different morphologies? > How effective is the method at correcting the dynamics of the system? I.e. after pre-training on the 6-leg centipede, if two or more of the pre-trained joints failed (i.e. uncontrollable), how fast can it adapt to the new setting? I am referring to the second setting mentioned in the response. It would be great if an experiment could be conducted in such a setting, so that we can see the trained policy is robust to change in dynamics.. For my final rating, I will take into consideration the arguments which arise during the upcoming AC-Reviewers discussion phase. Overall, I lean towards increasing my score. Best regards! --- Rebuttal 2: Title: Official Comment by Submission3440 Authors Comment: > Can you clarify how you may train the modules across different morphologies? Consider an example where the training morphologies consist of the 4 leg centipede and 6 leg centipede, both of which are composed of “leg” and “body” modules, denoted as $\textbf{W}_0$ and $\textbf{W}_1$ respectively. Then the modular architecture for the 4 leg centipede is a composition of the boss $\textbf{B}_1$ and its set of modules $\mathcal{W}_1 =\\{\textbf{W}_0^0, …, \textbf{W}_0^3, \textbf{W}_1^0 \\}$ (4 leg modules and 1 body module). For each module, the subscript corresponds to module type and the superscript denotes different instances of the same module type, which share model parameters. Similarly, the 6 leg centipede is a composition of the boss $\textbf{B}_2$ and its set of modules $\mathcal{W}_2 = \\{\textbf{W}_0^4, …, \textbf{W}_0^9, \textbf{W}_1^1, \textbf{W}_1^2\\}$ (additional 6 leg modules and 2 body modules). While the boss controllers differ across various morphologies, the modules are shared. We can then train both architectures end-to-end with imitation learning while injecting independent noise vectors $\eta_1, \eta_2$ to the output of $\textbf{B}_1$, $\textbf{B}_2$ respectively at each batch. > I am referring to the second setting mentioned in the response. It would be great if an experiment could be conducted in such a setting, so that we can see the trained policy is robust to change in dynamics.. Thanks for the clarification. We have run this experiment on the 6 leg to 12 leg centipede transfer task, where during transfer, we randomly select 7 out of 70 joints in the 12 leg centipede to be uncontrollable (for each seed of the experiment, a different subset of uncontrollable joints is sampled, and we have 3 random seeds). For the uncontrollable joints, we pass in a small random noise instead of the controller's output to the simulator. Below, we compare MeMo’s performance on this transfer task to RL (Modular), the strongest baseline from the 6 to 12 leg centipede transfer. MeMo significantly outperforms the baseline, achieving its final reward in less than half of the timesteps. This demonstrates the potential of our framework to adapt to unforeseen dynamics at transfer time. \| \|4e+6 timesteps \| 8e+6 timesteps\| \| :---------------- \| :------: \| ----: \| \|RL (Modular)\| 1203\|1485 \| \|MeMo\|1591 \|1793 \| Edit: Added table with final rewards.	Summary: The MeMo framework presented in this paper proposes an innovative approach for enhancing the transferability of control systems across robots with varied morphologies by utilizing pre-trained, modular controllers. This method facilitates rapid adaptation to new robot designs by leveraging previously trained modules, thereby significantly improving training efficiency over conventional methods like graph neural networks and Transformers. Strengths: The presentation looks good. The results presented are intriguing and demonstrate potential. Weaknesses: N/A Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The comparative analysis raises some questions regarding the fairness of the evaluation. Specifically, were the control modules of the compared methods also retrained on your newly designed robots, or was it only the master controller that was retrained? If the latter is the case, this might lead to an unfair comparison since these pre-trained models have not been directly exposed to the newly designed robot configurations prior to testing. 2. The transition from simulation to real-world applications is not addressed. The paper would benefit from a discussion on the expected challenges when implementing these modules in actual robotic systems, such as dealing with hardware inconsistencies and environmental variations. For example, how does the framework handle discrepancies in the physical properties between different robot assemblies, such as variations in leg mechanics? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and are glad that they find our results intriguing. Below we address the reviewer’s questions. Under "Questions": > The comparative analysis raises some questions regarding the fairness of the evaluation. Specifically, were the control modules of the compared methods also retrained on your newly designed robots, or was it only the master controller that was retrained? If the latter is the case, this might lead to an unfair comparison since these pre-trained models have not been directly exposed to the newly designed robot configurations prior to testing. For MetaMorph, we finetune the Transformer and the joint modules, as in the original work. For the NerveNet baselines, while parts of the network are fixed, prior work has found that such weight fixing actually improves transfer performance [1]. In the case of NerveNet-Conv, we experiment with both fixing the joint modules and finetuning the entire architecture and find that performance improves by fixing the joint modules. [1] Blake et al. “Snowflake: Scaling GNNs to High-Dimensional Continuous Control via Parameter Freezing” > The transition from simulation to real-world applications is not addressed. The paper would benefit from a discussion on the expected challenges when implementing these modules in actual robotic systems, such as dealing with hardware inconsistencies and environmental variations. For example, how does the framework handle discrepancies in the physical properties between different robot assemblies, such as variations in leg mechanics? We agree that sim-to-real transfer is an important problem; however, the goal of learning controllers that are robust to variations outside of simulation is orthogonal to our problem of learning modules that generalize to different morphologies. We note that our baselines, NerveNet and MetaMorph, also perform their experiments only in simulation. In our revised paper, we will extend our discussion of sim-to-real transfer as an important line of future work. At a high level, we would expect our framework to face the same challenges in adapting to the real world as standard RL policies, including discrepancies in simulated physics vs real dynamics and, as the reviewer mentioned, hardware inconsistencies and environmental variations. --- Rebuttal 2: Title: Thank authors for detailed explanations Comment: Thank authors for detailed explanations. I don't have any other questions and am happy to raise my points to 6. --- Rebuttal Comment 2.1: Title: Thank you for your response Comment: Thank you very much -- we're happy we were able to address your questions!	Summary: This paper introduces a new framework designed to create modular controllers allowing for quicker adaptation of control strategies when building new robots with similar methodology. The MeMo framework employs a novel modularity objective optimized alongside standard behavior cloning loss through noise injection. Experiment results in locomotion and grasping environments, ranging from simple to complex robot morphologies, demonstrate that MeMo improves training efficiency compared to multiple baselines. Additionally, MeMo’s modular approach enhances both structure and task transfer capabilities. Strengths: * The proposed method utilizes noise injection to build meaningful modular controllers that suit the application scenario as described: transfer the controller to a robot with similar but not identical morphology. The proposed method is easy to understand and should be easy to implement * This work provides reasonable grounding of the proposed method in Sec 3.1 * The experiment shows the proposed method outperforms all baselines in locomotion and manipulation tasks as shown in Figure 6/7, in terms of better sample efficiency and better final performance in some environments. * This work provides a decent ablation study on the noise injection to validate the component's importance. including replacing the noise injection with different forms of regulation. * The paper is well-written and easy to follow, a reasonable amount of technical details are included in the method and supplementary materials. * The analysis of the learned module in section 5.4 and Figure 15 looks interesting. Weaknesses: * The tasks evaluated are relatively simple, only basic locomotion tasks and quite simple manipulation tasks are considered. * All tasks involve only a rigid body, which could be simulated much faster with GPU with the recent simulator, which should deliver a much shorter training time compared to the reported time in the appendix. It’s a bit questionable whether the proposed method is necessary if training a new policy from scratch takes a short period. Technical Quality: 3 Clarity: 3 Questions for Authors: No specific questions to add. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: This work has provided reasonable discussion over the limitation and no potential negative social impact of the work needs to be considered. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the review and are glad that they find our method easy to understand and empirical evaluation thorough. Below we address the reviewer’s concerns and questions. Under "Weaknesses": > The tasks evaluated are relatively simple, only basic locomotion tasks and quite simple manipulation tasks are considered. Although relatively simple, our tasks enable us to extensively evaluate the structure transfer capabilities of our framework, which is the primary focus of our work. > All tasks involve only a rigid body, which could be simulated much faster with GPU with the recent simulator, which should deliver a much shorter training time compared to the reported time in the appendix. It’s a bit questionable whether the proposed method is necessary if training a new policy from scratch takes a short period. While we agree that the training time can be faster with recent simulators, the final reward of policies trained from scratch are below that of MeMo in our structure transfer experiments, an issue that cannot be fixed by faster simulation. These advances in simulation would also benefit MeMo, allowing the Boss controller to be retrained more quickly. --- Rebuttal Comment 1.1: Comment: Thank the authors for addressing my concerns, I would like to keep my original evaluation.	Summary: This paper presents a method for learning modular controller that can be transferred and adapted to different morphology of robots and tasks. The high level idea is to decompose control of each motor through distilling a learned hierarchical RL policy and then uses these primitive policies as building blocks to additional motors added. Strengths: This paper proposes a way to transfer low-level control policies to similar but different morphology of robots with a learned structure. The problem is novel and interestingly challenging. The proposed method is conceptually intuitive. Experiments in simulation demonstrate that the proposed algorithm can enable policy transfer across different robots and tasks. Weaknesses: The proposed method only trains the master policy at adaptation stage, which assumes the change of morphology does not induce change in dynamics of the system such that drastically different low-level policies will be needed to solve the task. Technical Quality: 3 Clarity: 3 Questions for Authors: Will this method generalize to multi-agent RL setting where the modular policy might be more complex than controlling 1-DoF motor? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: see weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the review and are glad that they find our problem novel and our proposed framework intuitive. Below we address the reviewer’s concerns and questions. Under "Weaknesses": > The proposed method only trains the master policy at adaptation stage, which assumes the change of morphology does not induce change in dynamics of the system such that drastically different low-level policies will be needed to solve the task. This is correct, and we will be happy to make this clearer in the final version of the paper; we expect our method to be most effective in situations like the ones in our experiments, where the new morphology is performing a similar task to the old morphology, and therefore there is a reasonable expectation that the low-level controllers learned with the initial morphology will be useful in the new morphology. In contrast, a situation where our technique may not be expected to work as well would be one where a module is used in a drastically different context; for example, pretraining a module as a leg for a crawling robot and later using it as a finger in a hand, which may require a very different range of motions even if it reuses the same parts. That said, our task transfer experiments show that the low-level controllers learned through our technique can be useful when transferring to a different but similar task. Even when the task is sufficiently different as to require different local controllers, e.g. the modules used to walk over a flat terrain are finetuned when transferring to a terrain with steps, using the pretrained controllers as a starting point can reduce the training cost. Under "Questions": > Will this method generalize to multi-agent RL setting where the modular policy might be more complex than controlling 1-DoF motor? To clarify, each local module does control subassemblies with more than one DoF -- for instance, the leg module of the centipede controls 5 joints. However, we agree that multi-agent RL would likely require more complex low-level controllers. At an abstract level, the problem of multi-agent RL is quite similar to what we address in our work -- the multi-agent policy can be decomposed into a higher-level controller that coordinates lower-level controllers corresponding to the individual agents. So while we have not experimented in this setting, we envision that our method could be applied to learn an appropriate division of labor between the boss that coordinates multiple agents and the more complex modules. --- Rebuttal 2: Title: thanks for clarification Comment: I have read the authors' response and appreciate the clarification. It would be interesting to see extending this method to multi-agent RL settings!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Energy Rank Alignment: Using Preference Optimization to Search Chemical Space at Scale	Reject	Summary: The paper proposes a new method, called Energy Rank Alignment (ERA) to finetune large language models (LLMs) for molecular generation in a similar fashion to Reinforcement Learning from Human Feedback (RLHF). The paper first introduces how the alignment task in LLMs is very similar to creating property-conditioned molecules from SMILES strings, which are token-based generation techniques. In the introduction, the paper distinguishes ERA from common RLHF methods, such as PPO and DPO, by stating that it has a minimization objectives and leverages a reward function. Next, the paper describes related work for using LLMs for molecular generation and RLHF for language models and reiterates the differences of ERA compared to PPO and DPO. In Section 2, the paper outlines the definition of ERA which mostly center on the derivation of relevant loss functions that the algorithm aims to minimize. In its definition, the ERA loss makes use of the KL divergence to arrive at the final formulation at the end of Section 2 leading up to the on-policy loss formulation for ERA. Section 3 provides a theoretical analysis of the ERA loss and its gradients, as well as its connections to the regularized entropy objective. Section 4 describes the experiments for molecular generation using ERA, including unprompted and prompted generation. The paper also includes a sub-section on general alignment settings of LLMs related to IMDB movie reviews. The results generally show a distribution shift between models finetuned with ERA and those that were not. The paper subsequently ends with a conclusion and discussion of limitations. Strengths: The provides proposed an interesting method and finetuning objectives that is useful for conditioned molecular generation and LLM alignment. The strengths include: * A novel method for designing property conditioned molecules that is also applicable to LLM alignment. [Originality, Significance] * A detailed derivation of the ERA loss, as well as a theoretical analysis on relevant properties. [Quality, Clarity] * Experiments that generally support the distribution shift induced by the ERA method. Weaknesses: The weaknesses of the paper mostly center on expanding relevant related work and baselines for experiments: * The authors do not discussion related work to training of transformer models and LLMs using reinforcement learning to arrive at molecules with desired properties. Some examples include [1] [2] * The experiments do not include baseline evaluation of DPO and PPO, which would have provided relevant details for how ERA performs compared to established baselines. * The paper could be strengthened by providing additional details related to experimental settings (see questions) [1] Ghugare, Raj, Santiago Miret, Adriana Hugessen, Mariano Phielipp, and Glen Berseth. "Searching for High-Value Molecules Using Reinforcement Learning and Transformers." In The Twelfth International Conference on Learning Representations. [2] Blaschke, Thomas, Josep Arús-Pous, Hongming Chen, Christian Margreitter, Christian Tyrchan, Ola Engkvist, Kostas Papadopoulos, and Atanas Patronov. "REINVENT 2.0: an AI tool for de novo drug design." Journal of chemical information and modeling 60, no. 12 (2020): 5918-5922. Technical Quality: 3 Clarity: 3 Questions for Authors: * How did you choose the objectives to optimize? Is there a reason you did not choose docking scores as shown in [1]. * In line 150, did you mean you leave off-policy objectives to future work? * Can you provide more details on how you do multiple properties at the same time for the experiments in Figure 3? Is the formulation of two beta basically a vector concatenation? How are gradients calculated? * Are you only using SMILES strings? Can you discuss other text-based representations, such as [2] [3] [4]? * Why are chemically invalid molecules sampled for your experiments? Does the pretraining not help? [1] Ghugare, Raj, Santiago Miret, Adriana Hugessen, Mariano Phielipp, and Glen Berseth. "Searching for High-Value Molecules Using Reinforcement Learning and Transformers." In The Twelfth International Conference on Learning Representations. [2] Krenn, Mario, Florian Häse, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. "Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation." Machine Learning: Science and Technology 1, no. 4 (2020): 045024. [3] Cheng, Austin H., Andy Cai, Santiago Miret, Gustavo Malkomes, Mariano Phielipp, and Alán Aspuru-Guzik. "Group SELFIES: a robust fragment-based molecular string representation." Digital Discovery 2, no. 3 (2023): 748-758. [4] Noutahi, Emmanuel, Cristian Gabellini, Michael Craig, Jonathan SC Lim, and Prudencio Tossou. "Gotta be SAFE: a new framework for molecular design." Digital Discovery 3, no. 4 (2024): 796-804. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors briefly discuss limitations at the end of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed feedback and insightful questions, which we address below. ## Discussion of related work We compare to state-of-the-art methods and widely used methods such as REINVENT on two benchmark tasks that focus on small-molecule drug design and find that we are able to generate novel and diverse compounds with desired properties more efficiently than existing methods (see global response). We also include a short discussion of related work and better contextualize our method. ## Comaprison to DPO/PPO and other baselines Due to the computational complexities of PPO, we focus only on gradient-based objectives here. We compare to DPO for the task of generating molecules with a high QED (see rebuttal Figure 2). We observe that with DPO, we are initially able to generate high-QED molecules, but that the chemical validity is low (20%) relative to ERA (80%). Furthermore, we find that the chemical validity degrades to 0% in subsequent checkpoints but that for ERA the chemical validity only marginally declines. A notable limitation of DPO is that its regularization is not effective in the finite data regime [1]. While IPO offers solutions around this limitation, it is still not immediately apparent how to independently tune regularization and sample diversity. With ERA, we are able to straightforwardly modulate sample diversity (via β) and regularize towards a reference policy (modulating γ), the latter of which we find helps in increasing chemical validity (main text Figure 4). Finally, we note that this idea has been also investigated elsewhere [2] and increases in regularization have been shown to limit declines in online evaluation metrics. We will further expand on this point in a revised version. [1] A General Theoretical Paradigm to Understand Learning from Human Preferences, Azar et al. [2] Preference Learning Algorithms Do Not Learn Preference Rankings, Chen et al. ## Additional experimental details ### Re: Chosen Objectives For our molecular generator experiments, we chose objectives that were computationally evaluable (e.g. using RDKit). Additionally, we carry out multi-property optimizations that correspond to nontrivial chemical searches. For example, searching for molecules with high LogP results in molecules with high hydrophobicity (e.g. more C's) and this constrains the search space when simultaneously maximizing LogP. As a result, a multi-property optimization of both high QED and LogP results in a more challenging chemical search than simply maximizing QED and LogP independently. We now include two additional experiments that better mimic real-world design of small-molecules for biological targets in the global response. ### Re: Line 150 In line 150, we did intend to write "leave the on-policy objectives to future work." By on-policy objectives, we refer to the loss in Eq. (11) and/or an approach that would include iteratively updating the reference policy and reasmpling a corresponding dataset during the alignment procedure. This second idea has been recently been investigated in another work (see Iterative Committee in [1]). ### Re: Multi-property optimization and gradient calculation For multi-property alignment, we define the multi-property, $\beta$-scaled energy to be a weighted sum of the property-specific energies weighted by the property-specific $\beta$ (see Line 232 in text). Here, each βproperty is a scalar representing the relative weight for that individual property in the overall multi-property optimization. We emphasize that with ERA we do not need to compute gradients of the energy w.r.t the model parameters. Because of this, minimizing Eq. 10 is straightforward for multi-property alignment and we incur no additional training cost for training on additional properties. ### Re: SMILES strings In this work, we only use SMILES as the textual representation for our molecules. While SELFIES is a possible alternative representation, recent work has found that models trained on SELFIES strings perform comparably to those trained on SMILES, and in some cases are worse than SMILES [2, 3, 4]. The fragment-based approach of SAFE is an interesting suggestion and seems promising, but it is not immediately clear what the best tokenization scheme would be for molecules under this representation, and what the vocabulary size would be when considering large datasets. ### Re: Chemically invalid generation Due to resource constraints, the model architecture and data that we used for pretraining was relatively small. With a larger model and/or more data, we expect the validity to increase (see [5]). However, we note that filtering out the small number of invalid SMILES strings has a marginal computational cost especially compared to the increase in training and inference costs for a larger model trained on more data. [1] Apple Intelligence Foundation Language Models, Apple [2] Invalid SMILES are beneficial rather than detrimental to chemical language models, Skinnider [3] Chemical language models enable navigation in sparsely populated chemical space, Skinnider et al. [4] Language models can learn complex molecular distributions, Flam-Shepherd et al. [5] Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction, Schwaller et al. --- Rebuttal Comment 1.1: Comment: Thank you for the additional details. Much of my feedback has been addressed and I have some additional questions: > Due to resource constraints, the model architecture and data that we used for pretraining was relatively small. Do you think your proposed method would also benefit model training at larger scales? This relates to models, data and search space. I am not asking for new experiments, but think a discussion of this would be beneficial. > For our molecular generator experiments, we chose objectives that were computationally evaluable (e.g. using RDKit) Similar to the question above, it seems that your method was mostly applied on problems with smaller compute budgets. Do you think it could translate to problems that require higher compute budgets. Would things like a smaller number of samples be a potential challenge? --- Reply to Comment 1.1.1: Comment: 1. Yes, we believe the performance across the board would improve with larger scale pre-trained models and there is evidence to this effect [1]. We also test ERA on a 13B parameter language model, and find that we are able to align the model towards desired outcomes. Based on the neural scaling laws for chemical models and the success of ERA on a large language model, we expect ERA to benefit model training at larger scales for chemical tasks. 2. If the cost of evaluating an oracle increases, the overall cost goes up for our method, but it will also similarly increase for all other methods. Based on the benchmarks conducted for the rebuttal, we think our methodology is competitive at fixed number of oracle evals, and hence will remain competitive even if the cost of evals increases. We expect the oracle evaluation to be expensive for many chemical tasks, especially those carried out on experimental observations. One strategy that may be useful in this setting is an iterative alignment one. With this strategy, we would first carry out ERA with a small number of samples and then generate further samples with the newly aligned model. Upon oracle evaluation, we would carry out another round of alignment---where the new reference model is the previously aligned model---and iteratively repeat the previous steps. This is a strategy that has had success elsewhere [2], and we anticipate it will be useful in the setting where oracle evaluation is expensive. We will add discussion of these ideas in the limitations section. [1] Neural scaling of deep chemical models, Frey et al. [2] Apple Intelligence Foundation Language Models, Apple	Summary: The authors introduce “Energy Rank Alignment”, a novel alternative to PPO and DPO for policy optimization when an explicit reward model is available. ERA is shown to work for enriching chemical libraries for proxy objectives that are fast and easy to compute, and has clear benefits in the simplicity of tuning the strength of regularization to a reference and entropy of samples with two decoupled parameters. This controllability allows ERA to avoid greedy policies and the sort of mode collapse often observed using DPO. Strengths: The ERA approach is interesting and clearly defined. It is well-suited for many preference optimization settings, where an explicit reward model is available and alternative methods do not take advantage of this. The authors show results on multi-objective optimization to illustrate that the approach is not limited to greedy optimization of single objectives. Weaknesses: The main weakness of the paper is the evaluation with respect to lead optimization of small molecules. This is a notoriously difficult kind of evaluation to make meaningful with purely in silico experiments. One clear opportunity for the authors to improve their evals, while respecting the constraints imposed by easily-computable reward functions, is to incorporate some kind of online evaluation. Comparing DPO and ERA in an online setting would be informative and more relevant for the chemistry community. Technical Quality: 3 Clarity: 3 Questions for Authors: While true that many objectives in chemistry are naturally continuous, binning is a simple solution that solves this problem for applying DPO. However, avoiding mode collapse is a significant problem, and performing direct gradient-based policy optimization is a well-motivated goal. I would suggest emphasizing these points rather than being able to handle continuous signals. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Partially Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed feedback and insightful questions, which we address below. ## Evaluation of lead optimization We agree that this is a weakness of the paper and have carried out additional experiments to design small-molecules with high activity against biological targets as predicted by computationally evaluable oracle functions. We report the results in the global response and find that we are able to generate novel and diverse compounds with desired properties more efficiently than existing state-of-the-art methods (see global response). We recognize that an *in silico* measurement of biological activity does not perfectly reflect the true activity but emphasize that it is straightforward to use ERA with experimentally measured properties. ## Comparison of DPO and ERA in an online setting We carry out an online evaluation of DPO and ERA on the task of generating small-molecules with high QED (main text Figure 2). We align DPO using the same dataset and hyperparameters as ERA with βDPO=0.1 and train for thousands of checkpoints (over 72 GPU-hours). We load intermediate checkpoints for both the DPO and ERA (βERA=20.0, γ=0.0) runs and carry out inference (see rebuttal Figure 2). We observe that at the first saved checkpoint of the DPO alignment run, the model generates molecules with high QED scores but with low validity (~20%). However, upon further training, the chemical validity of further checkpoints drops to 0% for the remaining runs, despite the overall DPO training and validation losses still dropping. With ERA, we see that we are able to similarly sample high QED small-molecules with reasonably high chemical validity (~85%). While the validity does drop over subsequent checkpoints, it does not do so precipitously. Moreover, the ERA-based alignment had no regularization (γ=0), and in the paper, we document how increasing γ can enable increases in chemical validity (main text Figure 4). Finally, we note that we did not extensively tune the hyperparameters for DPO , and it is possible that a different set of hyperparameters would elicit a more desired outcome; however, the lack of meaningful regularization in DPO [1], and its performance degradation in online metrics has been well-documented [2]. [1] A General Theoretical Paradigm to Understand Learning from Human Preferences, Azar et al. [2] Preference Learning Algorithms Do Not Learn Preference Rankings, Chen et al. ## Binning This is an interesting suggestion. Given two samples (y, y') from our reference model, we desire that the relative weights under our policy converges to the true relative Boltzmann weight. If we rely on binning and assume that we have a single ranking of (y, y') in our dataset, the relative weights of y and y' under a model trained with DPO will not converge to the Boltzmann weight and will instead go to ∞. However, with ERA, we can ensure that the relative weights of y and y' will converge to the Boltzmann weight. As the reviewer states, one strength of ERA compared to DPO is that we can more easily avoid mode collapse (by tuning β), and also independently tune regularization towards a reference policy (γ), which is not possible with DPO. This ensures that we can promote desired sample diversity and similarity to a reference policy. We will reinforce this point in a revised version. --- Rebuttal Comment 1.1: Comment: I thank the authors for their response and will maintain my score.	Summary: The authors study an important problem about searching through chemical space, where the number of possible molecules grows combinatorially with the number of atoms. They focus on aligning large autoregressive models trained on chemical compound databases to generate molecules. The energy rank alignment (ERA) algorithm is proposed to use an explicit reward function to produce a gradient-based objective for optimizing autoregressive policies. The authors offer theoretical insights into the relationship between energy rank alignment (ERA) and proximal policy optimization (PPO), direct preference optimization (DPO). Their experiments show that ERA is scalable, does not require reinforcement learning, and performs well compared to DPO when preference observations per pairing are limited. Strengths: 1. The authors study a significant problem about generating molecules with desired properties based on autoregressive models by proposing the energy rank alignment (ERA) algorithm. 2. This paper is well written. 3. The proposed methods work reasonably well. Weaknesses: 1. Diversity, novelty and uniqueness are all important properties for drug discovery as discussed in previous works. To verify whether the models can be used to improve the process of drug discovery, the paper may benefit from comparing the aligned models with the reference model based on these metrics. 2. Missing the discussion of the related works which also focus on molecule optimization and drug discovery for both traditional and state-of-the-art methods, such as [1] [2] and so on. 3. The authors propose using reinforcement learning for drug optimization, a well-established method frequently employed in prior works, such as [3,4]. Additionally, advantage-based and multi-objective policy optimization are well-known in the reinforcement learning literature. A more comprehensive analysis of the limitations of this approach, along with a comparison to other existing methods, would have been beneficial. [1] Drugassist: A large language model for molecule optimization. [2] Automatic chemical design using a data-driven continuous representation of molecules. [3] Optimization of molecules via deep reinforcement learning. Scientific Reports. 2019. [4] Multi-constraint molecular generation based on conditional transformer, knowledge distillation and reinforcement learning. Nature Machine Intelligence. 2021. Technical Quality: 3 Clarity: 2 Questions for Authors: Please see above Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and suggestions and for pointing us to additional works. ## Diversity Novelty and Uniqueness We investigate ERA on two tasks that mimic a drug-discovery effort and find that we are able to efficiently generate novel, diverse, and unique compounds that have high predicted biological activity according to *in silico* oracle functions (see global response). ERA consistently has the highest diversity compared with existing state-of-the-art methods. ## Discussion of and comparison to related works The approach described in [1] uses a Gaussian Process model to optimize molecular properties; however, this necessitates optimizing in a low-dimensional latent space, obtained with a VAE. With ERA and other RL methods described in the global response, we do not need a low-dimensional representation. We will add discussion of this method. We compare our approach to existing state-of-the-art methods (including some suggested by the reviewer) on the two tasks considered challenging and find that we are comparable to or better than existing methods, including on metrics related to diversity and sample efficiency (see global response). [1] Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules, Gomez-Bombarelli et al.	null	null	Rebuttal 1: Rebuttal: We appreciate the careful reviews and detailed feedback of our paper and address shared concerns and suggestions in this response. ## Simulated lead optimization and comparison to competitive approaches The reviewers raised concerns about the complexity of the benchmarks included in the paper. To address this concern, we now show the performance of ERA on two challenging tasks designed to mimic real-world lead optimization of small-molecules for specific biological targets. We find that our performance on these two tasks is competitive with or better than existing state-of-the art methods such as REINVENT [1] and MolRL-MGPT [2], which are both RL-based strategies. We consider two targets, the kinases JNK3 and GSK3β, and aim to design small-molecules that are biologically active against each. For each of these targets, we use an *in silico* oracle that predicts bioactivity, ranging from 0 to 1, where a higher value corresponds to stronger activity [3]. Using only data from ChemBL, we first carry out a short supervised fine-tuning step on all molecules in ChemBL with an oracle score above 0.5 (7386 molecules for JNK3 and 43381 for GSK3β). Using this fine-tuned model as our reference policy, we then align ($\beta$=100 and $\gamma$=0) as in Section 4.1, where we use a comparably high $\beta$ to target molecules with high activity. From the aligned models, we sample 20k molecules (see rebuttal Figure 1a), and tabulate metrics of the top-100 novel molecules (see Table 1). We emphasize that the molecules in the top-100 are filtered to only include molecules that are distinct from any molecule in the ChemBL dataset and additionally that there are no repeated molecules in the top-100. For GSK3β, our mean score is marginally lower than the best performing method but the diversity in sampled molecules is significantly higher (i.e. lower IntDiv). For JNK3 our mean score is significantly higher than the best performing method and the diversity in sampled molecules is higher than any method. The inference costs are low for our approach; sampling 20k molecules and filtering steps takes only minutes. We measure sample efficiency using the area under the curve (AUC) of top-K average property value versus the number of oracle calls top-K AUC [4]. We plot the top-10 average property value versus the number of oracle calls (see rebuttal Figure 1b) and report the top-10 AUC (see rebuttal Table 2). We only include novel molecules in this analysis; any sampled molecule that is in ChemBL, that has already been sampled, or that is invalid is discarded and additionally does not count towards an oracle call as these are filtered out before oracle evaluation. We find that the top-10 AUC metric is higher than any competing method, demonstrating that our method can efficiently sample molecles with desired behavior. We note that once we reach a top-10 average of 1.0, we do not make futher oracle calls as subsequent oracle calls will not change the top-10 average and will artifiically inflate the AUC. Our method is demonstrably better than existing state-of-the-art methods. With ERA, we generate both novel and diverse molecules with high predicted bioactivity, We also sample molecues with a high-oracle score compared to state-of-the-art methods more efficiently, ensuring that desired molecules can be generated with both a low inference cost and a low evaluation cost, the latter of which is important in settings where evaluation is expensive (e.g. wet-lab experiment). [1] Reinvent 4: Modern AI-driven generative model design, Loeffler et al. [2] De novo Drug Design using Reinforcement Learning with multiple GPT Agents, Hu et al. [3] Excape-db: an integrated large scale dataset facilitating big data analysis in chemogenomics, Sun et al. [4] Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization, Gao et al. ## Comparison to Related Works We will include more extensive discussion of related approaches suggested by the reviewers in our revision. We comment briefly here on methods that employ language models or transformers for molecular generation and use RL to optimize molecules and what distinguishes ERA. DrugAssist [1] uses human-machine dialogue with assistance from RL with human feedback (RLHF) and approaches multi-property optimization by incorporating diverse data streams. The ChemRLformer algorithm [2] leverages a text-based policy network and optimizes properties using a policy-gradient RL approach; however, in this framework, tuning sample diversity or regularization to a reference policy is challenging. The MGMG method [3] uses a knowledge-distilled conditional transformer and RL for multi-constraint optimization but involves many components and is cumbersome to optimize. The widely-used REINVENT method [4, 5] employs SMILES and multi-stage RL to tasks with mutliple reward models that vary in computational cost and accuracy. Finally, MolRL-MGPT [6] uses a multi-agent RL framework to promote sample diversity and has state-of-the-art performance, but training multiple GPT agents incurs significant training costs. [1] DrugAssist: A Large Language Model for Molecule Optimization, Ye et al. [2] Searching for High-Value Molecules Using Reinforcement Learning and Transformers, Ghugare et al. [3] Multi-constraint molecular generation based on conditional transformer, knowledge distillation and reinforcement learning, Wang et al. [4] REINVENT 2.0: An AI Tool for De Novo Drug Design, Blaschke et al. [5] Reinvent 4: Modern AI-driven generative model design, Loeffler et al. [6] De novo Drug Design using Reinforcement Learning with Multiple GPT Agents, Hu et al. Pdf: /pdf/c06a206e1c5a10950a554becf03b7cb584cf8689.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unsupervised Modality Adaptation with Text-to-Image Diffusion Models for Semantic Segmentation	Accept (poster)	Summary: The authors extend the traditional image domain adaptation to modality adaptation for unsupervised semantic segmentation in real-world multimodal scenarios. They use text-to-image diffusion model with strong generalization capabilities and proposed Diffusion-based Pseudo-Label Generation (DPLG) and Label Palette and Latent Regression (LPLR) to correct the pseudo-labels and obtain high-resolution features. SOTA performance on depth, infrared, and event modalities proves the effectiveness of the method. Strengths: 1. The paper is well-organized and easy to understand. 2. For the first time, the authors extend the adaptation between image domains to the adaptation between modalities. 3. The proposed MADM significantly outperform existing methods in the adaptation of three different modalities. 4. Figures 3 and 5 clearly demonstrate the effect of DPLG. They visualize the influence of noise-adding on the generation of pseudo-labels. Weaknesses: 1. The paper lacks the visualization results obtained by feeding potential features into the VAE Decoder. These results will provide a better understanding of LPLR. 2. The specific layers of the three multiscale features extracted from the UNet decoder are not clearly stated in the paper. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Add the LPLR visualization results and analyze them. 2. Details of the framework implementation need to be supplemented. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: 1. Add the LPLR visualization results and analyze them. 2. Details of the framework implementation need to be supplemented. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: To Reviewer Qg6E Thank you for the insightful and very positive comments. In the following, we provide our point-by-point response and hope our response helps address your concerns. We also look forward to the subsequent discussion which further helps to solve the current issues. >Q1: Add the LPLR visualization results and analyze them. A1: Thank you for your suggestion. We have visualized LPLR under different iteration steps in Figure 2 of the attached PDF. "Regression" and "Classification" in Figure 2 denote the output of the VAE decoder and segmentation head, respectively. Our proposed LPLR leverages the up-sampling capability of a pre-trained VAE decoder in a recycling manner. As the model converges, the regression results transform from blurry to progressively clearer states, presenting more details compared to the classification results. This assists the segmentation head in producing more accurate semantic segmentation results. We will include these results in the revision. >Q2: The specific layers of the three multiscale features extracted from the UNet decoder are not clearly stated in the paper. A2: Thanks for your carefully proofread. We extract the multiscale features from the denoising UNet decoder at the outputs of the 5th, 8th, and 11th blocks. We will clarify them in the revision. --- Rebuttal Comment 1.1: Comment: The authors have addressed my concerns. I keep my score. --- Rebuttal 2: Comment: We would like to extend our heartfelt thanks for the time and effort you have invested in reviewing our submission. Your insights have been instrumental in enhancing the quality of our work. Following your constructive feedback, we have diligently answered all your questions and presented the LPLR visualization in the rebuttal and attached pdf. We believe these changes have addressed your concerns and further strengthened our research. We understand that the reviewing process is demanding and time-consuming. Thank you once again for your dedication to the review process. We are hopeful for the opportunity to refine our work further based on your feedback.	Summary: This paper introduced text-to-image diffusion models to enhance the generalization across different modalities. The proposed MADM includes two key components: Label Palette and Latent Regression (LPLR) and Diffusion-based Pseudo-Label Generation (DPLG). This method alleviates issues related to pseudo-labeling instability and low-resolution features extraction within TIDMs. Experimental results show that MADM achieves state-of-the-art results across three different modalities. Strengths: 1. The topic is interesting and valuable.By leveraging the powerful pre-trained Text-to-Image Diffusion Models, the method effectively reduces discrepancies across different modalities (e.g., image, depth, infrared, and event). 2. The writing style is concise and easy to understand, and the paper is logically clear and well-organized. Weaknesses: 1. The proposed modules designed of the model is commonly used in generative fields, such as DPLG which converts labels into latent space and DPLG which utilizes the extracted features from TIDM to generate pseudo-labels, thus lacking novelty. 2. The approach is time-consuming, and there is no complexity analysis in Table 1. 3. The paper does not discuss the method's performance with different data volumes. The size of the dataset often significantly impacts performance improvement. It is recommended to test with varying data volumes (e.g., from 500 to 10,000) across different modalities to validate the enhancement and discuss the results, which would be beneficial for practical applications. 4. Minor issues: It should be Table 2 in Lines 260 and 267, Page 8. Technical Quality: 3 Clarity: 3 Questions for Authors: I like the idea of utilizing text-to-image diffusion models to enhance the generalization across different modalities. However, the significant time consumption and absence of experiments testing different data volumes lead me to downgrade the recommendation. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: To Reviewer 4TqZ Thank you for the insightful and positive comments. In the following, we provide our point-by-point response and hope our response helps address your concerns. We also look forward to the subsequent discussion which further helps to solve the current issues. >Q1: The proposed modules lack novelty. A1: Thanks. We address your concerns from two main perspectives. 1) Extension of TIDMs to UMA: For the first time, this work uniquely extends TIDMs to the UMA task across a broader range of visual modalities, introducing a novel perspective to this domain. While TIDMs have been used in various dense prediction tasks, their application has predominantly been limited to the RGB modality. Applying TIDMs directly to our UMA problem encounters significant issues, i.e., unstable and inaccurate pseudo-labeling and the lack of fine-grained feature extraction, as illustrated in Figure 3 and Table 2. These challenges have not been explored, despite their significance. Our work pioneers the extension of TIDMs to unsupervised semantic segmentation in other visual modalities, paving a new path for TIDMs in the UMA problem. 2) Novel Techniques of DPLG and LPLR: Our DPLG and LPLR techniques are novel, and effectively address or alleviate two severe issues that have not been explored before, yielding significant improvements over previous approaches. LPLR addresses the lack of fine-grained features in the UMA task, which has been entirely ignored by previous approaches. Latent-diffusion models pre-trained solely on continuous RGB images cannot handle discrete segmentation labels well. So previous methods [1, 2] discard the VAE decoder and use an additional classifier for features from the low-resolution latent space. However, the spatial size of the low-resolution space is much smaller than the original input size, leading to significant loss of spatial details crucial for semantic segmentation tasks. Our LPLR converts discrete labels into the RGB format, allowing them to be regressed by the VAE decoder to achieve high-resolution features and finer detail. Table 3 shows that LPLR achieves a significant +1.85% average improvement across all modalities over previous methods. DPLG addresses the issue of noisy and unstable pseudo-label generation in previous approaches like [3]. The significant distribution gap between images and other modalities results in noisy and inaccurate vanilla pseudo-labels generated by self-training methods as shown in Figure 5 and Table 3. DPLG utilizes the unique task property by injecting a certain amount of noise into the target modality for accurate pseudo-label generation. This injection aligns the latent space more closely with the data distribution encountered during the pre-training phase, fostering more robust and accurate semantic interpretation and pseudo-label generation. Figure 5 and Table 2 demonstrate that DPLG improves vanilla pseudo-label generation by a significant +3.39% on the infrared dataset. We appreciate your thoughtful comments and will include these clarifications and results in the revision. [1] DDP: Diffusion Model for Dense Visual Prediction. In ICCV, 2023. [2] Unleashing Text-to-Image Diffusion Models for Visual Perception. In CVPR, 2023. [3] DAFormer: Improving Network Architectures and Training Strategies for Domain-adaptive Semantic Segmentation. In CVPR, 2022. >Q2: The approach is time-consuming. A2: We appreciate your feedback on the computational costs of our proposed MADM. The following table presents a detailed comparison of training time per iteration, number of iterations, total training time, parameters, and performance across various methods in the event modality, including our MADM and its distilled variant. While MADM does exhibit a higher training time per iteration, the advanced visual prior derived from TIDMs necessitates fewer iteration for adaptation, presenting a minimum total training time. Moreover, MADM achieves a substantial performance improvement, with an MIoU of 57.34\%, surpassing other methods. Recognizing the trade-off in parameter count, we have leveraged our MADM model as a teacher to perform a secondary self-training. This approach has enabled us to distill the knowledge embedded in MADM into a more compact DAFormer model, MADM (Distilled), which retains a high MIoU of 54.03\% while significantly reducing parameters to 85M and only increasing the training time by 1.3 hours. Our distilled model demonstrates that it is possible to maintain high performance with reduced computational costs, addressing the concerns raised regarding the parameters and efficiency of MADM. \|Method\|Training time/Iter. (seconds)\|Iteration\|Total training time (hours)\|Params (million)\|MIoU\| \|-\|-\|-\|-\|-\|-\| \|DAFormer\|0.36\|40k\|4.0\|85\|33.55\| \|PiPa\|1.12\|60k\|18.7\|85\|43.28\| \|MIC\|0.48\|40k\|5.3\|85\|46.13\| \|Rein\|1.25\|40k\|13.9\|328\|51.86\| \|MADM\|1.38\|10k\|3.8\|949\|56.31\| \|MADM (Distilled)\|0.46\|10k\|1.3\|85\|54.03\| >Q3: The paper does not discuss the performance with different data volumes. It is recommended to test with varying data volumes to validate the enhancement and discuss the results. A3: Thanks. Per your suggestion, we train our method with 10%, 25%, and 50% of the total target samples in the event modality. Here, the "Baseline-100%" column indicates the performance of the MADM model without DPLG and LPLR and trained on the whole target samples. The results in the following table indicate that our proposed MADM consistently outperforms the baseline across all tested data volumes. Additionally, our MADM is robust and effective even when the dataset size is relatively small. We will include these results into revision. \|Method\|Baseline-100%\|MADM-10%\|MADM-25%\|MADM-50%\|MADM-100%\| \|-\|-\|-\|-\|-\|-\| \|MIoU\|52.27\|53.21\|53.69\|54.55\|56.31\| >Q4: Minor issues: It should be Table 2 in Lines 260 and 267, Page 8. A4: Thanks for your carefully proofread. We will fix this typo in revision. --- Rebuttal Comment 1.1: Comment: All my concerns have been addressed. I will raise my score to 6. --- Rebuttal 2: Comment: We would like to extend our heartfelt thanks for the time and effort you have invested in reviewing our submission. Your insights have been instrumental in enhancing the quality of our work. Following your constructive feedback, we have diligently answered all your questions and conducted more convincing experiments (data volumes comparison) in the rebuttal. We believe these changes have addressed your concerns and further strengthened our research. We understand that the reviewing process is demanding and time-consuming. Thank you once again for your dedication to the review process. We are hopeful for the opportunity to refine our work further based on your feedback.	Summary: This paper proposes Modality Adaptation with text-to-image Diffusion Models (MADM). MADM leverages pre-trained text-to-image diffusion models to enhance cross-modality capabilities, comprising two main components: diffusion-based pseudo-label generation to improve label accuracy and a label palette with latent regression to ensure fine-grained features. Experimental results show MADM achieves SOTA performance across various modalities. Strengths: - Overall, the paper is well-written. - Label Palette and Latent Regression are well-designed. - The ablation studies show the effectiveness of the method. Weaknesses: - The use of text-to-image Diffusion Models introduces higher training and inference costs, as well as an increase in model parameters. The authors need to discuss the fairness of these costs compared to other methods. - Overall, using text-to-image Diffusion Models for semantic segmentation is not novel, as many works have applied diffusion models for dense prediction tasks, and pseudo-label generation is also commonly used. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: To Reviewer uNLv Thank you for the insightful and positive comments. In the following, we provide our point-by-point response and hope our response helps address your concerns. We also look forward to the subsequent discussion which further helps to solve the current issues. >Q1: Discuss the fairness of the costs compared to other methods. A1: We appreciate your feedback on the computational costs of our proposed MADM. The following table presents a detailed comparison of training time per iteration, number of iterations, total training time, parameters, and performance across various methods in the event modality, including our MADM and its distilled variant. While MADM does exhibit a higher training time per iteration, the advanced visual prior derived from TIDMs necessitates fewer iteration for adaptation, presenting a minimum total training time. Moreover, MADM achieves a substantial performance improvement, with an MIoU of 57.34\%, surpassing other methods. Recognizing the trade-off in parameter count, we have leveraged our MADM model as a teacher to perform a secondary self-training. This approach has enabled us to distill the knowledge embedded in MADM into a more compact DAFormer model, MADM (Distilled), which retains a high MIoU of 54.03\% while significantly reducing parameters to 85M and only increasing the training time by 1.3 hours. Our distilled model demonstrates that it is possible to maintain high performance with reduced computational costs, addressing the concerns raised regarding the parameters and efficiency of MADM. \|Method\|Training time/Iter. (seconds)\|Iteration\|Total training time (hours)\|Params (million)\|MIoU\| \|-\|-\|-\|-\|-\|-\| \|DAFormer\|0.36\|40k\|4.0\|85\|33.55\| \|PiPa\|1.12\|60k\|18.7\|85\|43.28\| \|MIC\|0.48\|40k\|5.3\|85\|46.13\| \|Rein\|1.25\|40k\|13.9\|328\|51.86\| \|MADM\|1.38\|10k\|3.8\|949\|56.31\| \|MADM (Distilled)\|0.46\|10k\|1.3\|85\|54.03\| >Q2: using text-to-image Diffusion Models for semantic segmentation is not novel, as many works have applied diffusion models for dense prediction tasks, and pseudo-label generation is also commonly used. A2: Thanks. We address your concerns from two main perspectives. 1) Extension of TIDMs to UMA: For the first time, this work uniquely extends TIDMs to the UMA task across a broader range of visual modalities, introducing a novel perspective to this domain. While TIDMs have been used in various dense prediction tasks, their application has predominantly been limited to the RGB modality. Applying TIDMs directly to our UMA problem encounters significant issues, i.e., unstable and inaccurate pseudo-labeling and the lack of fine-grained feature extraction, as illustrated in Figure 3 and Table 2. These challenges have not been explored, despite their significance. Our work pioneers the extension of TIDMs to unsupervised semantic segmentation in other visual modalities, paving a new path for TIDMs in the UMA problem. 2) Novel Techniques of DPLG and LPLR: Our DPLG and LPLR techniques are novel, and effectively address or alleviate two severe issues that have not been explored before, yielding significant improvements over previous approaches. LPLR addresses the lack of fine-grained features in the UMA task, which has been entirely ignored by previous approaches. Latent-diffusion models pre-trained solely on continuous RGB images cannot handle discrete segmentation labels well. So previous methods [1, 2] discard the VAE decoder and use an additional classifier for features from the low-resolution latent space. However, the spatial size of the low-resolution space is much smaller than the original input size, leading to significant loss of spatial details crucial for semantic segmentation tasks. Our LPLR converts discrete labels into the RGB format, allowing them to be regressed by the VAE decoder to achieve high-resolution features and finer detail. Table 3 shows that LPLR achieves a significant +1.85\% average improvement across all modalities over previous methods. DPLG addresses the issue of noisy and unstable pseudo-label generation in previous approaches like [3]. The significant distribution gap between images and other modalities results in noisy and inaccurate vanilla pseudo-labels generated by self-training methods as shown in Figure 5 and Table 3. DPLG utilizes the unique task property by injecting a certain amount of noise into the target modality for accurate pseudo-label generation. This injection aligns the latent space more closely with the data distribution encountered during the pre-training phase, fostering more robust and accurate semantic interpretation and pseudo-label generation. Figure 5 and Table 2 demonstrate that DPLG improves vanilla pseudo-label generation by a significant +3.39\% on the infrared dataset. We appreciate your thoughtful comments and will include these clarifications and results in the revision. [1] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, Ping Luo. DDP: Diffusion Model for Dense Visual Prediction. In ICCV, 2023. [2] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, Jiwen Lu. Unleashing Text-to-Image Diffusion Models for Visual Perception. In CVPR, 2023. [3] Lukas Hoyer, Dengxin Dai, Luc Van Gool. DAFormer: Improving Network Architectures and Training Strategies for Domain-adaptive Semantic Segmentation. In CVPR, 2022. --- Rebuttal 2: Comment: We would like to extend our heartfelt thanks for the time and effort you have invested in reviewing our submission. Your insights have been instrumental in enhancing the quality of our work. Following your constructive feedback, we have diligently answered all your questions and conducted more convincing experiments (complexity analysis) in the rebuttal. We believe these changes have addressed your concerns and further strengthened our research. We understand that the reviewing process is demanding and time-consuming. Thank you once again for your dedication to the review process. We are hopeful for the opportunity to refine our work further based on your feedback.	Summary: This paper proposes an interesting task that adapting image segmentation knowledge to other input modalities, such as depth, infra, and event. This is beneficial for applications at nighttime. Strengths: 1. The task is promising for applications at nighttime. 2. The label palette is novel and can be generalized to more tasks. Weaknesses: 1. The presentation needs to be improved. - The meaning of "Unsupervised Modality Adaptation" is unclear from the introduction section. - Although the authors try to illustrate why they use a pretrained Text2Image Diffusion Model (TIDMs), and why they propose DPLG and LPLR, some motivations are not well supported by experiments or other accepted papers. For example, in Lines 45-47, "Although TIDMs are not trained on other visual modalities, their large-scale samples and unification through texts enable them to adapt to a broader distribution of domains." What is the meaning of "a broader distribution of domains"? Which prior is provided by TIDMs that results in this convenience? 2. Lacking discussions on the motivation of critical components. For example, see the third question below. 3. Experiments do not well support the potential applications (see question 4). Typos: 1. Line 134, Sec.1 --> Fig. 1 Technical Quality: 2 Clarity: 2 Questions for Authors: 1. Since the training dataset of TIDMs mainly consists of RGB images, why can TIDMs robustly extract features across modalities? 2. What's the motivation behind the single-step diffusion operation (Line 146)? 3. Why does injecting noise into the latent code produce more accurate pseudo labels? 4. As stated in Line 28, taking depth or other modalities is valuable in nighttime perception. But there are no experiments showing this. For example, given a nighttime dataset, comparing the performance with the input of RGB images or depth images (adapted by MADM). Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The limitations are briefly stated in Sec. 5. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: To Reviewer x1By Thank you for the insightful and positive comments. In the following, we provide our point-by-point response and hope our response helps address your concerns. We also look forward to the subsequent discussion which further helps to solve the current issues. >Q1: The meaning of "Unsupervised Modality Adaptation" is unclear from the introduction section. A1: Thanks. We apologize for any confusion caused by the unclear explanation. To clarify, "Unsupervised Modality Adaptation" refers to the adaptation of a model from a labeled source image modality to an unlabeled target modality, such as depth, infrared, or event data. This concept was initially explained in lines 132-134 of the manuscript. To further address your concerns, we will provide a more detailed explanation when we first introduce this concept in the revision. >Q2: Some motivations are not well supported by experiments or other accepted papers. What is the meaning of "a broader distribution of domains"? Which prior is provided by TIDMs that results in this convenience? A2: Thanks. The phrase "a broader distribution of domains" refers to the model’s adaptability to different visual modalities beyond its primary design of generating RGB images from text. The prior we refer to is derived from the large-scale pretraining data on which TIDMs are trained. This extensive pretraining provides TIDMs with a robust understanding of high-level visual concepts, enabling their application to various domains, such as semantic matching [1], depth estimation [2], and 3D awareness [3]. As demonstrated in Table 2 of our manuscript, our baseline method performs comparably with SoTA techniques, highlighting the potential of leveraging this advanced visual prior for the Unsupervised Modality Adaptation (UMA) problem. However, two significant challenges remain. First, significant modality discrepancies hinder robust and high-quality pseudo-label generation during self-training. Second, TIDMs extract features in the 8x downsampling latent space which limit the acquisition of high-resolution features. To address these issues, we propose LPLR and DPLG. These methods are designed to mitigate the identified problems, as detailed in our manuscript. In the revised version, we will include the above discussion to enhance the motivations. [1] SD4Match: Learning to Prompt Stable Diffusion Model for Semantic Matching. In CVPR, 2024. [2] Repurposing Diffusion-based Image Generators for Monocular Depth Estimation. In CVPR, 2024. [3] Probing the 3D Awareness of Visual Foundation models. In CVPR, 2024. >Q3: Why can TIDMs robustly extract features across modalities? A3: Thanks. Through extensive pre-training on diverse and large-scale datasets, TIDMs have demonstrated a remarkable capacity to generate various objects with distinct attributes, such as cars of different types and colors. While TIDMs are not explicitly trained on other modalities, they can capture the fundamental characteristics of objects, such as shape and spatial orientation (e.g., a car's position on the road). These characteristics remain consistent across different modalities, enabling TIDMs to achieve high-level visual intelligence. As a result, TIDMs can identify and extract essential characteristics of objects even when encountering modalities that were not part of their pre-training data. This ability to generalize across modalities showcases the strength and versatility of TIDMs on feature extraction in various tasks. >Q4: What's the motivation behind the single-step diffusion operation? A4: Thanks. As evidenced by the experimental results in [1], a single diffusion step effectively removes noise for dense visual prediction. Additionally, single-step diffusion significantly reduces inference costs compared to multi-step diffusion. Considering these factors, we adopt the single-step diffusion operation, following the approach in [2, 3]. [1] DDP: Diffusion Model for Dense Visual Prediction. In ICCV, 2023. [2] Unleashing Text-to-Image Diffusion Models for Visual Perception. In CVPR, 2023. [3] Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. >Q5: Why does injecting noise into the latent code produce more accurate pseudo labels? A5: In the pre-training of TIDMs, the objective is to estimate noise from latent inputs containing various noise levels. By injecting noise into the latent code, we effectively simulate this noisy distribution. This simulation aligns the latent space more closely with the data distribution encountered during the pre-training phase. Such alignment fosters a more robust and accurate semantic interpretation, which, in turn, enhances the quality of the pseudo labels generated. This shares similar spirits in other applications of diffusion models, such as the text-to-3D [1] where injecting extra noise into data can improve the denoising quality of image and yields better pseudo labels. [1] DreamFusion: Text-to-3D using 2D Diffusion. In ICLR, 2023. >Q6: Given a nighttime dataset, comparing the performance with the input of RGB images or depth images (adapted by MADM). A6: The FMB-Infrared dataset includes both image and infrared modalities on daytime and nighttime scenes. We adapt from cityscapes with daytime RGB images to the nighttime image modality and infrared modality by our proposed MADA, respectively. The following table and Figure 1 in the attached pdf show that the infrared modality has a clear advantage in the "Person" class due to obvious thermal differences and a good suppression of light interference. We will include these results into the revision. \|Modality\|Sky\|Build.\|Person\|Pole\|Road\|S.walk\|Veg.\|Vehi.\|Tr.S.\|MIoU(avg)\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|RGB\|88.85\|68.14\|64.79\|25.80\|89.09\|32.43\|70.32\|84.13\|7.27\|58.98\| \|Infrared\|87.94\|82.40\|82.69\|21.50\|76.21\|26.50\|76.61\|83.80\|16.69\|61.59\| --- Rebuttal 2: Comment: We would like to extend our heartfelt thanks for the time and effort you have invested in reviewing our submission. Your insights have been instrumental in enhancing the quality of our work. Following your constructive feedback, we have diligently answered all your questions and conducted more convincing experiments (nighttime comparsion) in the rebuttal and attached pdf. We believe these changes have addressed your concerns and further strengthened our research. We understand that the reviewing process is demanding and time-consuming. Thank you once again for your dedication to the review process. We are hopeful for the opportunity to refine our work further based on your feedback. --- Rebuttal 3: Title: Thanks for your response. Comment: > To further address your concerns, we will provide a more detailed explanation when we first introduce this concept in the revision. UMASS is the task of this paper, which however is first introduced in the Method section. I think it's unfriendly for most of the readers in the Computer Vision community. > This extensive pretraining provides TIDMs with a robust understanding of high-level visual concepts, enabling their application to various domains, such as semantic matching [1], depth estimation [2], and 3D awareness [3]. I briefly read the mentioned references： - SD4Match[1] takes image as the input. - Marigold [2] only use the encoder of VAE to compress depth image, the unet is fine-tuned. - In [3], the authors have demonstrate DINOv2 gets better feature than StableDIffusion. See their introduction: `We find that recent self-supervised models such as DINOv2 [60] learn representations that encode depth and surface normals, with StableDiffusion [69] being a close second. ` Therefore, the question remains unsolved: when the input is depth, or other visual modality, rather than RGB, why the unet of the StableDIffusion can recognize it and provide correct feature? > While TIDMs are not explicitly trained on other modalities, they can capture the fundamental characteristics of objects, such as shape and spatial orientation (e.g., a car's position on the road). Can you show some evidence? How good features can Stable Diffusion learn for some modalities he has never seen before? > Considering these factors, we adopt the single-step diffusion operation, following the approach in [2, 3]. I briefly read VPD [2]. I couldn't find any mention of the use of one-step diffusion in the manuscript. > such as the text-to-3D [1] where injecting extra noise into data can improve the denoising quality of image and yields better pseudo labels. DeamFusion [1] obtains pseudo labels with supervision from StableDiffusion, not by injecting extra noise. I must remind: the author should answer the reviewer's questions carefully and provide the correct references. --- Rebuttal Comment 3.1: Comment: To Reviewer x1By We are truly grateful for your continued engagement and valuable feedback on our submission. Your willingness to communicate with us further is greatly appreciated and provides us with an opportunity to clarify any misunderstandings and enhance our research even more. >Q1: UMASS is the task of this paper, which however is first introduced in the Method section. I think it's unfriendly for most of the readers in the Computer Vision community. A1: We apologize for the oversight in the initial presentation of the UMASS concept. We will revise the manuscript to introduce this key concept earlier in the paper to ensure clarity and accessibility for our readers. >Q2: Therefore, the question remains unsolved: when the input is depth, or other visual modality, rather than RGB, why the unet of the StableDIffusion can recognize it and provide correct feature? Can you show some evidence? How good features can Stable Diffusion learn for some modalities he has never seen before? A2: (1) TIDMs trained on large-scale data can learn very general high-level semantics and concepts, allowing them to generate images in unseen scenarios by combining different semantics/concepts, as demonstrated in many works [e.g., DaLL-E, Parti, SD, etc.]. For instance, SD3 [1] can successfully generate "a hybrid creature that is a mix of a waffle and a hippopotamus" as shown on page 14. Although such specific scenes have never been encountered during training, TIDMs can understand and disentangle the high-level semantics, such as those of a waffle and a hippopotamus, and creatively imagine their combination. This enables them to seamlessly merge these elements into a coherent and realistic image. In our case, while TIDMs may not have seen data from other modalities, such as depth images, they can still grasp the high-level semantics shared across modalities, like object shapes in RGB and depth images, and may also interpret the semantic combinations from different modalities. (2) Importantly, similar to the approach used in Marigold, we fine-tune the UNet for adaptation instead of directly extracting features from frozen TIDMs. In our work, adaptation to other visual modalities is achieved through self-training. This fine-tuning process further enhances the TIDMs' ability to understand high-level semantics and establish their combinations across different modalities. Similar evidence can be seen in ControlNet [2]. TIDMs can also adapt to different modalities by integrating and fine-tuning an additional ControlNet initialized by the same parameters as TIDMs. Specifically, TIDMs can generate corresponding RGB images when text and other modalities are taken as conditions. These reference modality inputs can be sketch, normal map, depth, human pose, etc. that TIDMs have not seen during pre-training. Specifically, the third column of Fig. 7 in [2] demonstrates that after fine-tuning with depth data, TIDMs understands the high-level semantics of depth modality and successfully generates images that meet the depth conditions. [1] Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In ICML, 2024. [2] Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. >Q3: I briefly read VPD [2]. I couldn't find any mention of the use of one-step diffusion in the manuscript. A3: In Section 3.2 of VPD, the authors state, “Note that we simply set t = 0 such that no noise is added to the latent feature map,” and further clarify, “It is also worth noting that our method is not a diffusion-based framework anymore, because we only use a single UNet as a backbone (see Figure 1 to better understand the differences).” Also, `outs = self.unet(latents, t, c_crossattn=[c_crossattn])` in line 102 of https://github.com/wl-zhao/VPD/blob/main/depth/models_depth/model.py indicates the use of one-step diffusion in VPD. >Q4: DeamFusion [1] obtains pseudo labels with supervision from StableDiffusion, not by injecting extra noise. A4: Since TIDMs cannot generate images that are spatially consistent across viewpoints and thus optimize the NeRF model, DeamFusion leverages the generative capabilities of TIDMs for a different purpose: to supervise the training of the NeRF model through its diffusion process, not the final output. In DeamFusion, gaussian noise is added to the 2D images rendered by the NeRF, and the pre-trained TIDM is allowed to predict the noise. Successful prediction, where the noise matches the one added, signifies that the NeRF model has internalized the statistical priors of the TIDMs. Our DPLG shares a conceptual alignment with this approach. TIDMs are essentially used to predict noise from noise-injected inputs, so adding noise to inputs in our proposed DPLG enable TIDMs to provide more robust prior information, thereby enhancing the utilization of pre-trained knowledge.	Rebuttal 1: Rebuttal: We would like to express our sincere gratitude to all of the four reviewers for the valuable and constructive feedback provided on our manuscript. Your insights have been instrumental in enhancing the quality and clarity of our work. In the attached PDF, we include additional visualizations as requested by Reviewers x1By and Qg6E, which we believe will clarify our experimental findings. Thank you for your valuable time and insights. We are open to further discussion and look forward to your continued guidance. Pdf: /pdf/5c9798df2fc660d216e1317ab20425455f8d00cb.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction	Accept (poster)	Summary: The paper presents a novel framework, termed CausalStock, for news-driven multi-stock movement prediction. The authors address two key issues in existing methods: the unidirectional nature of stock relations and the substantial noise in news data. CausalStock introduces a lag-dependent temporal causal discovery module and an LLM-based Denoised News Encoder. The experimental results show that CausalStock outperforms strong baselines on six datasets and provides good explainability. Strengths: The combination of causal discovery and news denoising for stock movement prediction is novel. The authors emphasize the explainability of their model, which is a significant advantage in financial applications. In addition, the paper also includes detailed ablation studies that highlight the contributions of different components of the model. Weaknesses: 1. Regarding the Lag-dependent Temporal Causal Discovery module, since there are already many traditional causal discovery methods such as PC [1] and GES [2], there is a lack of ablation studies comparing this module with traditional causal discovery methods in terms of performance and time efficiency. 2. Regarding the learned causal graph G, I have concerns about the difference between it with a learned correlation matrix. Though it is derived from the causal discovery and Bayesian perspective, in terms of implementation, what is the difference between the correlation matrix? Incorporating a comparison with a correlation matrix in Figure 3b will be convincing. 3. Lack of complexity analysis and cost time comparison. 4. Minor comments: The MIE and the Lag-dependent TCD are not explicitly shown in Figure 2. The detailed structure of each module (DNE, Price Encoder, and FCM) is also lacking. Reference: [1] Estimating high-dimensional directed acyclic graphs with the PC-algorithm (2005) [2] Optimal Structure Identification With Greedy Search (2002) Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How to make sure the learned G is exactly the causal graphs? It seems that the final loss only considers the prediction accuracy. 2. How would the model handle sudden market shifts or unprecedented events that significantly impact stock prices? ### After rebuttal Most of my concerns have been addressed. I raise my score to 5. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discussed the limitations of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: --- Thank you very much for your valuable suggestions! We will try our best to tackle the concerns one by one. (1) Comparision with the traditional causal discovery method and the correlation matrix - We greatly appreciate your insightful comments! Following your suggestion, we compare our discovered causal matrix with the matrix calculated by the traditional causal discovery methods and the correlation matrix. The performance results are shown as follows, which indicates that our causal discovery module has a significant advantage over others. \| Dataset \| ACL18 \| \| CMIN_US \| \| CMIN_CN \| \| \|--\|-\|-\|-\|-\|--\|-\| \| \| ACC \| MCC \| ACC \| MCC \| ACC \| MCC \| \| Correlation Matrix \| 52.24 \| 0.0220 \| 52.85 \| 0.0230 \| 52.52 \| 0.0210 \| \| PC Algorithm \| 51.51 \| 0.0138 \| 52.04 \| 0.0184 \| 51.90 \| 0.0166 \| \| GES Algorithm \| 51.41 \| 0.0135 \| 51.92 \| 0.0152 \| 51.83 \| 0.0145 \| \|CausalStock\| 63.42 \| 0.2172 \| 54.64 \| 0.0481 \| 56.19 \| 0.1417 \| --- - Compared to traditional causal methods, our integrated causal discovery and prediction modules form an end-to-end deep learning framework, which can constrain the discovered causal graph by simulating the data generation process. In comparison to correlation-based methods, causal relationships are more suitable for representing stock relationships. --- (2) Complexity analysis and cost time comparison. - Thanks for pointing this out. We made a rough estimate of the baselines' complexity on the multi-stock movement prediction task from two perspectives: running time and training cost FLOPs corresponding to the time complexity and the number of parameters corresponding to the space complexity. - We leverage an NVIDIA GeForce RTX 3090 and all the parameter settings of baselines follow the original papers. The complexity and prediction accuracy results of our method and baselines are shown in the following table. \| Model \| FLOPs \| Parameters \| Running Time\|Acc on KDD17\|Acc on NI225\|Acc on FTSE100\| \|--\|--\|-\|--\|--\|-\|--\| \| LSTM \| $2.3 × 10^8$ \| $5797$\| $5m58s$\|$51.18$ \| $50.79$\| $50.96$\| \| ALSTM \| $3 × 10^8$ \| $6917$ \| $6m32s$ \|$51.66$\| $50.60$\| $51.06$\| \| Adv-ALSTM \| $3 × 10^8$ \| $6917$ \| $7m02s$ \|$51.69$\| $50.60$\| $50.66$\| \| StockNet \| $5.0 × 10^{10}$ \| $4.4 × 10^6 $ \| $112m$ \|$51.93$\|$50.15$ \|$50.36$ \| \| CausalStock \| $1.4 × 10^9$ \| $5 × 10^5 $\| $7m58s$ \|$56.09$\| $53.01$\| $52.88$\| - From the table above, it can be seen that we can achieve good performance with comparable complexity. --- (3) About Figure 2 - We have shown the main structure and the most contributive part details of our model in the submitted version. - Thank you for your reminder. We have made a more comprehensive model structure in the attached pdf file. --- (4) About the learning of causal graph - The ground truth causal graph of the stock market is unknown, so we need certain theoretical guarantees to approximate the true causal relation. Under the current setup, we can theoretically prove that this learning process can converge to the true causal graph. - Validity of variational objective: - Our model holds the following assumptions: Causal Markov Property, Minimality and Structural Identifiability, Correct Specification, Causal Sufficiency, and Regularity of log-likelihood (see Appendix B for a detailed description). - If we further assume that there is no model misspecification, then the Maximum Likelihood Estimation solution $\theta'$ and the variational posterior distribution of $G$ satisfies $q'_\phi(G)=\sigma(G=G')$ by optimizing the ELBO term with infinite data, where $G'$ is a unique graph. - In particular, $G'=G^$ and $p_\theta'(X;G')=p(X;G^)$, where $G^$ is the ground truth graph and $p(X;G^)$ is the true data generating distribution. Please find reference [1] for a more detailed proof. [1] Wenbo Gong, Joel Jennings, Cheng Zhang, and Nick Pawlowski. Rhino: Deep causal temporal relationship learning with history-dependent noise. arXiv preprint arXiv:2210.14706, 2022. --- (5) About the sudden market shifts or unprecedented events - Thanks for this constructive comment! Our model could handle sudden market shifts or unprecedented events from the following perspectives: - Incorporating News to Capture Market Dynamics: One of the primary purposes of incorporating news into our model is to capture real-time market dynamics. News data can reflect market changes and sudden events promptly. By processing and analyzing news information quickly, our model can rapidly adjust its predictions based on new information. This allows the model to react swiftly and accurately to sudden market shifts. - Causal Relationships to Respond to Sudden Events: Our model's ability to learn causal relationships helps it respond effectively to sudden events. For example, if negative news is detected, the causal relationships in the model can identify which stocks might be affected and react accordingly. This mechanism ensures that the model not only considers the directly impacted stocks but also identifies related stocks that might be influenced, providing a comprehensive view of market dynamics. - Adjusting Priors to Reflect Changes in Causal Relationships: If a sudden event affects the causal relationships between stocks, such as a company ending a partnership, our model can adjust by incorporating the new pre-defined knowledge into the priors. This allows the causal graph to be dynamically updated to reflect the latest market structure and relationship changes, maintaining the accuracy and relevance of the predictions. - Thus, our model could effectively handle sudden market shifts and unprecedented events. --- We would appreciate it very much if you could raise the score if the concerns are solved. Thank you! --- Rebuttal Comment 1.1: Comment: Thanks for your response. My concerns are addressed and I have decided to raise the score to 5. --- Reply to Comment 1.1.1: Comment: I would like to extend my sincere gratitude for your insightful and constructive comments on our manuscript. Your positive evaluation and the time you have invested in thoroughly reviewing our work are greatly appreciated. Thank you once again for your kind and supportive review. We are grateful for the opportunity to improve our work based on your suggestions！	Summary: This work predicts stock movements by inferring the causal relation between stocks and news. News are encoded to structured representation through LLMs to filter out noises. Causal relation is modeled as directional graph and is inferred through variational approach. Strengths: 1. The news denoising module is inspiring and might be extent to other applications. The graph inference module shows promise with interesting design feature such as sparseness and knowledge prior. 2. The paper present comprehensive experiments by evaluating the propose approaches on various datasets and comparing with multiple baselines. The results are promising. 3. Interpretability is more valued in finance applications and the explainability analysis hits on the pain point of many DL based stock prediction system. Weaknesses: 1. The size of graph inference module $O(n^2)$ where $n$ is the number of stocks. The graph is soon becoming unlearnable as the $n$ grows due to limited historical data. The benchmarks used here are small, it is unclear whether the performance can persist on a larger stock set. 2. I think the general framework is not sufficiently novel, neither the graph inference module. I believe the news denoising module is inspiring and worths further investigation. The paper unfortunately doesn't provide such more analysis on the Denoised News Encoder. 3. There are several typos in the paper such as line 42, 44, etc. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Since graph inference is challenging and can be erroneous, I am curious about the performance without graph module, that is, simply using the news representation and historical price as input features for predictions. 2. Table 2 shows that "CausalStock w/o lag-dependent TCD" outperforms the proposed method. How do you reconcile the results? Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: --- Thank you very much for the insightful and constructive comments! We sincerely appreciate the suggestions to improve our submission. We will address all raised concerns one by one. (1) About the performance without graph module - Following your suggestion, we ablate the graph module and compare it with our method to explore the value of our causal discovery module. Besides, we compare our causal discovery module with other traditional causal discovery methods, i.e. PC and GES, and the Spearman correlation matrix. The performance results are shown as follows, which indicates that our causal discovery module has a significant advantage over others. \| \| ACL18 \| \| CMIN_US \| \| CMIN_CN \| \| \|---------------------------\|--------------\|----------\|--------------\|----------\|--------------\|----------\| \| \| ACC \| MCC \| ACC \| MCC \| ACC \| MCC \| \| Correlation Matrix \| 52.24 \| 0.0220 \| 52.85 \| 0.0230 \| 52.52 \| 0.0210 \| \| PC Algorithm \| 51.51 \| 0.0138 \| 52.04 \| 0.0184 \| 51.90 \| 0.0166 \| \| GES Algorithm \| 51.41 \| 0.0135 \| 51.92 \| 0.0152 \| 51.83 \| 0.0145 \| \| No Causal Graph \| 51.08 \| 0.0102 \| 51.48 \| 0.0106 \| 51.37 \| 0.0102 \| \|CausalStock \| 63.42 \| 0.2172 \| 54.64 \| 0.0481 \| 56.19 \| 0.1417 \| --- (2) About the ablation results - Thanks for pointing this out! We have noticed that two ablation results of two datasets were mistakenly swapped in our presentation of the results. We sincerely apologize for this oversight, and we have now corrected it in the following table. \| Ablation Type \| Ablation Variants \| ACL18 \| \| CMIN-CN \| \| CMIN-US \| \| \|--------------------------\|------------------------------------------\|-------------\|-----------\|-------------\|-----------\|-------------\|-----------\| \| \| \| ACC \| MCC \| ACC \| MCC \| ACC \| MCC \| \| Main Framework \| CausalStock w/o lag-dependent TCD \| 59.19 \| 0.1757 \| 52.93 \| 0.0312 \|54.97 \| 0.1298 \| \| \| CausalStock w/o news \| 58.10 \| 0.1421 \| 53.16 \| 0.0375 \|54.16 \| 0.1264 \| \| Traditional News Encoder \| CausalStock with Glove+Bi-GRU \| 60.78 \| 0.1952 \| 53.87 \| 0.0467 \| 55.13 \| 0.1326 \| \| \| CausalStock with Bert \| 61.74 \| 0.2067 \| 53.92 \| 0.0472 \| 55.43 \| 0.1352 \| \| \| CausalStock with Roberta \| 61.81 \| 0.2071 \| 54.06 \| 0.0477 \| 55.58 \| 0.1364 \| \| Denoised News Encoder \| CausalStock with FinGPT \| 61.92 \| 0.2105 \| 54.30 \| 0.0475 \| 55.67 \| 0.1386 \| \| \| CausalStock with Llama \| 62.82 \| 0.2164 \| 54.52 \| 0.0483 \| 55.97 \| 0.1406 \| \| \| CausalStock (with GPT-3.5) \| 63.42 \| 0.2172\| 54.64 \| 0.0481\| 56.19 \| 0.1417\| Thank you for your thorough review again. We will be more diligent in our checks to prevent similar errors in the future. We would appreciate it very much if you could raise the score if the concerns are solved. Thank you! --- Rebuttal 2: Comment: Thanks for the new results and correction. I raise my score accordingly. Please include the ablation study in the new version. --- Rebuttal Comment 2.1: Comment: I would like to extend my sincere gratitude for your insightful and constructive comments on our manuscript. Your positive evaluation and the time you have invested in thoroughly reviewing our work are greatly appreciated. Thank you once again for your kind and supportive review. We are grateful for the opportunity to improve our work based on your suggestions！	Summary: This paper proposes a news-driven multi-stock movement prediction model called CausalStock. The paper introduces a Denoised News Encoder, which utilizes LLMs to evaluate news text and obtain denoised representations. It also presents a Lag-dependent temporal causal discovery module to discover causal relations between stocks. Based on the input market information and learned causal graph distribution, CausalStock employs a FCM to make stock movement predictions. The contributions of the paper include the design of the Denoised News Encoder, the Lag-dependent temporal causal discovery module, and the application of FCM for stock movement prediction. Strengths: 1. The paper is well-written with a complete and coherent structure. 2. The idea of using MIE and Lag-dependent TCD to improve the performance of stock prediction is novel and the effective. Weaknesses: 1. There exist some typos in the paper: a. In line 199, the “or” should be “and”. b. In Figure 2, the abbr of stock APPLE is “AAPL” while in Figure 3 is “APPL”. c. In line 594, the hidden size should be 32 rather than 332. d. In line 603, “it’s” should be “its”. 2. You’d better place the Related Work section into the main manuscript. 3. In Appendix B, the explainations of the assumptions are not very sufficient. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Are the sparseness constraints used in the loss function? It doesn’t appear in the Equation 14. 2. How to calculate the correlation in Section 4.4 and Appendix D? How to obtain the causal strength of a stock in Figure 3(a)? 3. In Appendix B, the author assume their model satisfies the Causal Markov Property. Does it mean the current state is only correlated with the last state? However, the proposed method is based on lag-dependent as in Equation 3. Does the proposed model really satisfy the Causal Markov Property? 4. it’s better for the authors give a case study of the method and provide a visualization of the learned causal graphs, which can show their method can indeed obtain information in news to help the prediction of stocks. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: --- Thank you very much for your valuable suggestions and your encouraging comments! We will try our best to tackle the concerns. (1) If sparseness constraints used in the loss function? - Thank you for the question. We use sparseness constraints in the loss function. These constraints are included in the Evidence Lower Bound (ELBO) term, as shown in Equation 13. The final loss function (Equation 14) includes the ELBO term (Equation 13) and the BCE loss term (Equation 13). --- (2) About the correlation calculation in Section 4.4 - The causal strength graph $\hat{G}$ is designed to evaluate the causal strength. It has the same size as the causal graph $G$, with each position being a learnable parameter. After averaging the causal strength graph across all time lags, we obtain a matrix that indicates the magnitude of the causal influence of each company on other companies. Figure 3(a) is an example of such a matrix. Then the correlation in Section 4.4 and Appendix D is calculated by Spearman’s rank correlation coefficient between this matrix and the market values of the corresponding company to explore the explainability. --- (3) About the Causal Markov Property - The definition of Causal Markov Property [1] is as follows: Given a directed acyclic graph (DAG) $G$ and a joint distribution $p$, this distribution is said to satisfy Causal Markov Property w.r.t. the DAG $G$ if each variable is independent of its non-descendants given its parents. It is different from the traditional temporal Markov Property. - Under our setup, all the historical nodes we considered are served as the parent nodes of the target nodes. Specifically, in our lag-dependent temporal causal mechanism, even though we set the causal nodes are lag-dependent, once all of them are given, the target nodes are independent of the other unmodeled factors. This is a common assumption in the field of causal discovery, and our model satisfies this assumption. [1] Jonas Peters, Dominik Janzing, and Bernhard Scho ̈lkopf. Elements of causal inference: foundations and learning algorithms. The MIT Press, 2017. --- (4) About the visualization of the learned causal graphs and news cases - We have provided a visualization of the learned causal graph and some news cases in Figure 3. - For a more detailed illustration, if negative news is detected, then the causal relationships in the model can identify which stocks might be affected and react accordingly by passing the information extracted from the news, which is empowered by the integration of news and causal discovery. --- We would appreciate it very much if you could raise the score if the concerns are solved. Thank you!	Summary: This paper proposes a stock price prediction system based on the history of stock price features and news features, with a model which models the causal relationships between different stacks throughout window of time. The model is based on a temporal causal graph that determines whether the features of stock $i$ (including news and price features) at time $t$ can affect stock $j$ at a future time point (e.g., $t + \ell$, $\ell \in [L]$, where $L$ is the maximum time lag). Causal discovery is performed by modeling the posterior of the lag-dependent causal graph given the data, $p(\textbf{G} \| \textbf{X}_{<T})$--- this is done by optimizing a variational inference objective. The features of a stock $X_t^i$ capture a combination of its price $P_t^i$ and news $C_t^i$. The system is evaluated on several benchmarks and compared against previous stock prediction systems. Strengths: - The proposed method seems to be a reasonable solution to the problem and incorporates several components that address the multi-faceted nature of such a real-world problem (e.g., temporal causal discovery, encoding different types of features) - The reported results show modest but consistent improvement over previous stock prediction systems. - There is a nice ablation study on the effect of different components of the system (e.g., w/ vs w/o news, different news encoders, w/ vs w/o lag-dependence in the causal graph). Weaknesses: - The procedure for extracting news features is somewhat limited. I.e., it is based on simply prompting an LLM to evaluate a news text across 5 dimensions, obtaining a 5-dimensional feature vector. It may be possible to obtain more information from the news text by more directly accessing an embedding representation, perhaps via a fine-tuned model. However, the ablation results in this paper do demonstrate the usefulness of the news component. - Although reasonable, the temporal causal graph model can likely be extended to better capture more complex aspects of the evolution of stock prices. For example, it appears to model different stock edges as independent of each other. Technical Quality: 3 Clarity: 2 Questions for Authors: - Can you elaborate on the choice of variational approximating class of distributions and discuss any possible limitations? It seems that this is product distribution, making all edges independent (though a particular edge is not independent of that edge at a previous point in time). Is that correct? For example, this wouldn't be able to capture phenomena like "sectors" whose stocks are closely tied to each other. It's fine to make such an assumption for tractability, but it warrants discussion. - Can the model capture a causal graph that varies with time? I.e., perhaps in 2010 stock X is not really causally connected to stock Y, but after 2020 they become causally connected. E.g., this might capture a company entering a new sector, or starting a partnership with another company (e.g., Open AI and Microsoft). - How are news articles selected over time? For example, what if there are multiple articles for a stock within a time period or there are none? - Why do you need variables for both the likelihood of an edge $u_{\ell, ji}$ and the likelihood of no edge $v_{\ell, ji}$? Since this is a binary variable, can't you have a single variable and let $\sigma_{\ell, ji} = \mathrm{sigmoid}(u_{\ell, ji}')$? A minor thing: the use of the word "remarkably" in a few places in the paper is grammatically incorrect (e.g., "Remarkably, we put the Related work in Appendix E"). You are using it to mean something like "we note that", but the word "remarkably" means something more like "suprisingly". Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: There is a short section describing limitations & future work in the appendix, but I would encourage the authors to expand on it and incorporate it into the main paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: --- Thank you very much for your valuable suggestions! We will try our best to tackle the concerns one by one. (1) About the choice reason of Bernoulli distribution - Thank you for the question. The essence of causal relationships is determining whether a change in one variable directly causes a change in another variable. This binary decision process can naturally be modeled using the Bernoulli distribution. For each causal link, we need to determine whether it exists, which is a typical binary problem—either the causal link exists (represented by 1) or it does not exist (represented by 0). The Bernoulli distribution is used to describe this type of binary random variable, making it very suitable for our needs. --- (2) About the independence assumption of causal edges - While each causal link is not completely independent, in our proposed Lag-dependent temporal causal discovery, the causal links are conditionally dependent on temporal lags, as shown in Equation 5 of our paper. However, for different stocks, the causal links are indeed assumed to be independent, which simplifies computational complexity, making the model more efficient to train and infer by parallel computation, especially for large-scale stock sets. We also consider the dependent relationships among stocks, such as sector-specific relationships like you mentioned. We included a plug-and-play option in the prior, as shown in Equation 4, allowing the model to incorporate trusted prior knowledge, such as industry knowledge or company collaboration information. This flexibility enables the model to capture more complex dependencies when such information is available. --- (3) About the time-varied causal graph - Thanks for your constructive comment. The ground truth causal graph of the stock market is unknown, so we need certain theoretical guarantees to approximate the true causal relation. Under the current setup, We can theoretically prove that this time-invaried causal learning mechanism converges to the true causal graph (see Appendix B). - The point you raised is highly valuable in practical applications. In the future, we can consider adopting meta-learning or incremental learning training methods to update the causal graph iteratively. This way, the model can continue to learn and update based on new data in practical applications, thereby reflecting the dynamic causal relation changes in the market. --- (4) About the varied number of news - The number and distribution of news articles are related to the dataset. Below are the statistics of our news datasets: \| Dataset \| #news \| \| \| \| #words \| \| \| \|----------\|-------\|---\|---\|---\|--------\|---\|---\| \| \| Mean \| \| Std \| \| Mean \| \| Std \| \| ACL18 \| 4.14 \| \| 9.92\| \| 43.59 \| \| 50.28 \| \| CMIN_CN \| 5.21 \| \| 4.99\| \| 84.22 \| \| 62.22 \| \| CMIN_US \| 6.23 \| \| 6.16\| \| 42.97 \| \| 39.81 \| - For the absence of news data, CausalStock leverages only the price data to construct the causal graph (see Appendix B for details) and subsequently integrates the discovered causal relations and price data for predictions. Even without news data, our model can effectively perform causal inference and make predictions using price data alone. - When there are multiple news articles for a particular stock within a given look-back window, we employ LLMs to filter and score the articles. From this process, we retain up to a maximum number of 30 most relevant articles per day to ensure that the model receives focused and informative input. These selected articles are then encoded and fed into the network, maintaining the model's effectiveness and accuracy in handling multiple news sources. --- (5) About the simultaneous modeling of the existence and non-existence likelihood - We model the existence and non-existence likelihood simultaneously, which ensures greater flexibility and could avoid the constrained optimization. Besides, we also perform the idea the reviewer mentioned: only model the causal link existence logits and operate the Sigmoid function to obtain the link probability. We compare these two methods and the results are as follows: \| \| ACL18 \| \| CMIN_US \| \| CMIN_CN \| \| \|-------\|--------------\|-----------\|--------------\|-----------\|--------------\|-----------\| \| \| ACC \| MCC \| ACC \| MCC \| ACC \| MCC \| \| only-existence \| 58.21 \| 0.1652 \| 52.32 \| 0.0241 \| 53.96 \| 0.0670 \| \| both existence & non-existence\| 63.42 \| 0.2172 \| 54.64 \| 0.0481 \| 56.19 \| 0.1417 \| --- (6) About limitations and future work Thanks for your constructive suggestions! We have deeply considered it and come up with new thinking results of the limitations. - This paper explores a method that discovers causal relations based on theoretical considerations. In the future, we could try to adopt meta-learning or incremental learning training methods to update the causal graph iteratively, i.e. explore the time-varied causal graph. - While the Bernoulli distribution is suitable for determining whether a causal link exists, if we want to further explore the multi-level nature of causal relationships, more complex distributions might be needed. In the future, we could improve the model in this way. --- We would appreciate it very much if you could raise the score if the concerns are solved. Thank you! --- Rebuttal Comment 1.1: Comment: Thank you for your response. I appreciate the clarifications about the questions and limitations raised in the review. Although your response clarifies some questions, there are no major additions or changes that resolve the two main weaknesses raised in the review. I have decided to maintain my score. --- Reply to Comment 1.1.1: Title: About the weakness. Comment: --- Thank you for your reminder. We have further considered the two points mentioned in the weakness part and have added some experiments and analysis as follows. --- ### (1) About the Traditional News Encoder (text embedding) and LLM-based Denoised News Encoder (5-dimensional representation) In the Ablation Study Section of our manuscript, we have compared the performance of Traditional News Encoders (Glove+Bi-GRU, Bert and Roberta text embeddings) and LLM-based Denoised News Encoders (FinGPT, Llama and GPT 3.5). The results have shown that LLM-based Denoised News Encoders are more effective than Traditional News Encoders. Thanks for pointing this out! In response to your concern, we refine the experiments further to explore the performance of the Traditional News Encoders via some fine-tuned models. Specifically, we employ three classes of fine-tuned and pre-trained text embedding models: i) fine-tuning Bert-base-multilingual-cased and Roberta-base during our CausalStock model training; ii) leveraging two fine-tuned financial text embedding models - FinBert[1] and FinGPT-v3.3. iii) leveraging pre-trained Llama-7b-chat-hf to generate text embedding. \|\|\|ACL18\|\|CMIN-CN\|\|CMIN-US\|\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC\|MCC\|ACC\|MCC\|ACC\|MCC\| \|Traditional News Encoder\|with Glove+Bi-GRU\|60.78\|0.1952\|53.87\|0.0467\|55.13\|0.1326\| \|\|with Bert(Pre-trained)\|61.74\|0.2067\|53.92\|0.0472\|55.43\|0.1352\| \|\|with Bert(Fine-tuned)\|61.26\|0.2033\|53.43\|0.0419\|55.93\|0.1406\| \|\|with Roberta(Pre-trained)\|61.81\|0.2071\|54.06\|0.0477\|55.58\|0.1364\| \|\|with Roberta(Fine-tuned)\|61.75\|0.2065\|54.02\|0.0474\|55.63\|0.1368\| \|\|with FinBert(Pre-trained)\|61.72\|0.2062\|54.01\|0.0471\|55.61\|0.1362\| \|\|with FinGPT(Pre-trained)\|61.69\|0.2060\|54.00\|0.0470\|55.60\|0.1360\| \|\|with Llama(Pre-trained)\|62.20\|0.2130\|54.40\|0.0480\|55.85\|0.1390\| \|Denoised News Encoder\|with FinGPT\|61.92\|0.2105\|54.30\|0.0475\|55.67\|0.1386\| \|\|with Llama\|62.82\|0.2164\|54.52\|0.0483\|55.97\|0.1406\| \|\|CausalStock(with GPT-3.5)\|63.42\|0.2172\|54.64\|0.0481\|56.19\|0.1417\| --- It can be seen that the denoised news representation generally outperforms traditional text embedding. By analyzing some cases, we found that for the news-driven stock movement prediction task, effectively utilizing key information is much more important than retaining the comprehensive information of news (too much noise), and this is why we propose the Denoised News Encoder. [1] Araci D. Finbert: Financial sentiment analysis with pre-trained language models[J]. arXiv preprint arXiv:1908.10063, 2019. --- ### (2) About the stock causal independence assumption. We appreciate your suggestion to model the dependencies between different stock edges, i.e. modeling the variable-dependent causal relations. The initial intention of this assumption is to simplify the complexity of our model. After carefully considering your suggestion, we find that it is feasible to extend our model according to your point. ### Modeling Edge Dependencies Based on the lag-dependent causal mechanism, we propose a variable-dependent causal mechanism that explicitly captures the dependencies among different stock edges. Specifically, each edge $G_{l,ji}$'s probability is conditioned on the states of all other edges at the same time step $l$, and the conditional function is the same as the function in the lag-dependent mechanism (see Equation 6 for details). Formally, we extend to model $q_\phi(G_{l,ji} \mid G_{l,\backslash (ji)})$, where $\backslash (ji)$ indicates the edges except for the $ji$-th edge. We implemented the aforementioned variable-dependent approach. The results are as follows: \|Ablation Variants\|ACL18\|\|CMIN-CN\|\|CMIN-US\|\| \|-\|-\|-\|-\|-\|-\|-\| \|\|ACC\|MCC\|ACC\|MCC\|ACC\|MCC\| \|w/o variable-dependent\|63.42\|0.2172\|54.64\|0.0481\|56.19\|0.1417\| \|with variable-dependent\|63.50\|0.2175\|54.60\|0.0479\|56.25\|0.1419\| --- The results show that incorporating a variable-dependent causal mechanism has the potential to enhance model performance. However, the improvements are not uniform and vary depending on the dataset and may depend on other possible factors (like hyperparameter tuning, dataset size, and stock set size), which emphasizes that further validation is needed. ### Complexity Analysis While the above results show a promising performance of the variable-dependent causal mechanism, it significantly increases the computational complexity. - The original lag-dependent model has a time complexity of $O(\text{lag} \times D^2) $, where $ D $ is the number of stocks. - The extended variable-dependent model increases the complexity to $ O(\text{lag} \times D^4) $ to incorporate the dependency of every link pair. This complexity scales with the fourth power of the number of stocks, making it challenging to apply the model to markets with large numbers of stocks. To sum up, we will conduct further research to comprehensively balance the trade-offs between the modeling flexibility and the computational demands. --- Rebuttal 2: Comment: We sincerely thank you for your encouraging comments! All the additional experimental results will be included in the revision. If our response has addressed your concerns, we would greatly appreciate it if you could kindly consider raising the score to support our work. Thank you for your valuable feedback and supportive encouragement!	Rebuttal 1: Rebuttal: The more detailed model structure. Pdf: /pdf/f41c0fb5a41c57ee286727e6bb42a8a73666103d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
TARSS-Net: Temporal-Aware Radar Semantic Segmentation Network	Accept (poster)	Summary: This paper proposes a novel framework designed to integrate temporal information into radar-based semantic segmentation tasks. To achieve more effective temporal information integration, the framework introduces two key modules: the Target-History Temporal Relation Encoder (TH-TRE), which analyzes the relationships between different time frames, and the Temporal Relation Attentive Pooling (TRAP) module, which aggregates information along the temporal axis. Strengths: This paper introduces a novel algorithm for radar-based semantic segmentation tasks. The manuscript is well-constructed, with the discussion section clearly articulating and comparing the differences between various algorithms. The newly designed algorithm demonstrates a notable improvement in performance without increasing the model size. Weaknesses: Please consult the question section for further information. Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Could the authors provide a comparison in terms of computational time? 2) Although the primary baseline, PKCIn-Net, also incorporates temporal information, the difference in performance between the proposed algorithm and PKCIn-Net appears to be minimal. 3) Additionally, the authors mention that this paper considers compression from both spatial-temporal and depth-temporal dimensions. Therefore, a pertinent question arises: would it be possible to first fuse spatial and depth information, and then subsequently fuse this combined data along the temporal axis? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Please consult the question section for further information. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Questions [Q1. COMPUTATIONAL TIME COMPARISON.] There is real-time performance comparison in `Sec. E.2 of Appendix`, which shows the comparison between TARSS-Net and some other models, involvingmodel size, calculation amount, real-time performance and other information of in details. In addition, we also add the VIT model in`Table S2` as a baseline for comparison, which can also be found in`Table 1 of attached PDF` for this rebuttal. [Q2. MINIMAL DIFFERENCE IN PERFORMANCE COMPARED WITH PKCIN-NET.] PKCIn-net makes the classical radar target detection CFAR principle deeply learnable, and proposes a radar-specific target detection operator PeakConv (PKC). The focus of PKC is not on incorporating temporal information, but we include it in the comparison as it is a landmark work on introducing deep learning to RSS. Under the premise of the same model size, TARSS-Net outperforms PKCIn-Net by an average of 2% (on all data-view and three dataset). First of all, given the particularity and difficulty of RSS tasks, a 2% performance improvement is significant, we also illustrate in our responses to R. MgaH how the RSS performance has improved little by little (if you want to learn more, please see response of `W2 to R. MgaH`). Secondly, PKCIn-Net focuses on making the classical radar detection theory learnable, and TARSS-Net focuses on introducing an efficient temporal relationship modeling method suitable for radar signals. In the later work, we will also consider adapting PKC to TARSS-Net to further improve the performance of the model, since TRAM and PKC are not in conflict due to their different focuses. Both PKCIN-Net and TARSS-Net point the way for the development of the RSS field. That is, compared with directly applying the mature learning method in other field (such as CV, NLP, etc.), designing the learning paradigm suitable for the characteristics of radar signals is more beneficial. [Q3. WOULD IT BE POSSIBLE TO FIRST FUSE SPATIAL AND DEPTH INFORMATION, AND THEN SUBSEQUENTLY FUSE THIS COMBINED DATA ALONG THE TEMPORAL AXIS?] This is a particularly valuable question. When we originally designed the TRAP module to implement temporal relation measurement and weighted-fusion, we intended to incorporate spatio-temporal information at the same time. However, for the space and depth with high dimensions, it is the most economical choice to do dimensionality reduction and fusion processing separately, i.e., Spatio/Depth-TRAP. If we want to fuse the two dimensions at the same time, it is bound to introduce modules with larger parameters, so we did not explore it in this work. Therefore, in the subsequent network optimization process, considering the complexity of RSS network to facilitate its application, we form two versions of Spatio-TRAP and Depth-TRAP on the premise of ensuring their performance advantages. Your idea is very meaningful, and as mentioned in the `limitations Section in Appendix F (L1017)`, deeply exploration on the learning principle and applicable scenarios of TARSS-Net_D and TARSS-Net_S is needed in future work. Based on this, we look forward to working out an exciting solution to your pertinent question. --- Rebuttal Comment 1.1: Title: We appreciate your respect for our hard work Comment: Dear Reviewer kN5s, We have spent a significant amount of time and effort analyzing your concerns and suggestions, and have provided detailed explanations and corresponding modifications. We believe that our response should be sufficient to address your concerns. We sincerely hope that you would take the time to read our response. If our response is adequate, we kindly ask you to give a fair score upgrade; if you still have other concerns, looking forward to further discussions with you. Thank you for your contribution to improving the quality of our paper, and we also appreciate your respect for our hard work. 9831 Authors --- Rebuttal 2: Title: Looking forward to your comments on our reply Comment: We have provided detailed responses to all of your questions. Hope to get your approval on the reply and update the rating. We also look forward to more in-depth communication and discussion with you. Thank you again!	Summary: In this work, the authors created a network called TARSS-Net for radar semantic segmentation. Compared to traditional methods, the authors emphasized the superiority of their approach in the clever utilization of historical information. Specifically, TARSS-Net incorporates a module called TRAM, which is designed similarly to the attention mechanism. This module first learns the relationship between historical frames and the current frame through the TH-TRE module, and then maps this learned relationship to hidden representations via TRAP. TRAM is specifically designed for radar semantic segmentation, ensuring that computational complexity remains manageable while achieving efficient and accurate radar semantic segmentation across three real radar segmentation datasets. Strengths: The main advantages of the paper are as follows: (1) The paper provides a comprehensive review and analysis of the application background, existing methods, and challenges of RSS semantic segmentation. (2) The authors have significantly enhanced the model performance in RSS semantic segmentation tasks by balancing accuracy and efficiency through carefully designed methods, which holds practical significance. Weaknesses: 629/5000 This paper may have the following potential shortcomings: (1) Lack of Clear Hypotheses and Reasoning: The authors designed a novel temporal modeling paradigm for RSS and claimed its effectiveness in the context of RSS semantic segmentation. However, they seem to provide their viewpoint directly without offering clear hypotheses and reasoning about why this paradigm is effective for RSS semantic segmentation, which may confuse the readers. (2) Insufficient Emphasis on the Design Motivation: The authors lack an emphasis on the motivation behind the model design. For instance, regarding the design of the TH-TRE module, the authors aim to use this module to encourage the model to focus more on high-dimensional information of the current time frame. However, Section 3.1 lacks an emphasis on this design motivation and focuses too much on formula statements. Given the considerable complexity of the designed method, it is crucial to introduce the design concept behind the module to the readers. (3) Issues with Paper Formatting and Visualization: There are some issues with the paper's formatting and visualization. For example, the sequence of Figures 4 and 3 does not align well with typical reading habits, leading to some reading difficulty. Additionally, some elements in the figures are not drawn rigorously; for instance, the input of the TH-TRE module should be a 4D tensor, but Figure 3 depicts it more like a 3D tensor. For readers accustomed to visual aids, the authors' lack of rigorous visualization might lead to a misunderstanding of the model details. (4) Potential Issues in the Experimental Section: There may be shortcomings in the experimental section, especially in comparison with the use of the Self Attention (SA) model in the temporal domain. Although the authors mention that applying SA in the temporal domain might lead to excessive computational consumption, the appendix shows that they have implemented a baseline model applying SA in the temporal domain based on ViT. Therefore, a more rigorous experiment should include this baseline model in the accuracy analysis (since accuracy analysis does not involve computational efficiency), as the authors have criticized this method in their analysis. Technical Quality: 3 Clarity: 3 Questions for Authors: Some questions are as follows: (1) What is the core innovation of this paper? Causal dilated convolutions also seem to align with the paradigm proposed by the authors. Compared to these simpler methods, what advantages does the method in this paper have? Why is it more advantageous? (2) Is it meaningful to discuss real-time performance for RSS semantic segmentation? If RSS semantic segmentation in most scenarios demands higher accuracy rather than efficiency, then the criticism of other methods in terms of efficiency in the paper seems less significant. Can the authors provide specific examples of scenarios to analyze and explain this? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors adequately discuss the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses [W1. LACK OF CLEAR HYPOTHESES AND REASONING.] Sorry for the confusing, and you are right that clear hypotheses and reasoning are important. However, at present, the data-driven learning training of AI models allows researchers to conceptuate and design algorithms at a higher level (the functional and motivated level), which means the classification interface (Modeling functions of network) is no longer expressed in an explicit way by formulas, but is learned implicitly in the network parameters. It solves the problem that the classification interface cannot be shown by existing expressions. In this way, the algorithm performance has been further improved and generalized. The writing of TARSS-Net exactly follows this theoretical system of priori interpretability. Of course, TARSS-Net essentially involves a series of theories and derivations of feature engineering, metric learning, differentiable layer design and so on with radar signal as input. Due to the space limitation and ease of understanding for readers with general AI research background, this paper starts from the high level design and implementation, but the necessary derivations in key parts are preserved. In addition, this paper gives a detailed introduction on why TARSS-Net is effective for RSS. We have conducted a comprehensive discussion for existing time series modeling paradigms, and analyze their own drawbacks and the factors that make them unsuitable for RSS (see `Sec. 2`). Based on the above analysis, we illustrate the design motivation of TARSS-Net for RSS task one by one (see `L144-L155`), as well as detailed implementation methods in `Sec. 3`. Also, we further verify the superiority of TARSS-Net over the existing methods that consider temporal relation information. We believe that the confusion you mentioned can be eliminated after readers carefully read the full paper, the Appendix and the code. [W2. INSUFFICIENT EMPHASIS ON THE DESIGN MOTIVATION.] In `Sec. 2`, we have conducted a comprehensive discussion for existing temporal modeling paradigms, and based on the above analysis, we illustrate the design motivation of TARSS-Net for RSS one by one, including the design motivation of TH-TRE module. We believe that after you go back and read the first two sections of this paper again, your questions will be answered. [W3. ISSUES WITH PAPER FORMATTING AND VISUALIZATION.] Sorry for the reading trouble caused by unreasonable formatting and visualization. Due to the space limit of paper submission, we had to make some typography which might make it uncomfortable to read. These will be corrected in next manuscript version, including the order of Fig. 3 and Fig. 4, more rigorously draw for elements in the figures, etc. [W4. POTENTIAL ISSUES IN THE EXPERIMENTAL SECTION.] Due to space limitation, we show the experimental results that can best help to verify the performance of TARSS-Net in the most concise way. Due to the sparsity of radar targets, the dense computation of SA will inevitably introduce redundant computations on irrelevant information thus degrading the RSS performance. The performance of the VIT model is supplemented in `Table 1 of attached PDF` for this rebuttal. We also promise to add it to Table S2 in the Appendix of the revised manuscript. ## Questions [Q1. THE CORE INNOVATION OF THIS PAPER.] The core innovation is to propose an effective temporal modeling method specific to RSS tasks, i.e., the plug-and-play TRAM which combines the advantages of causality, end-to-end learnability, constant model parameters under arbitrary length input, and linear growth of computational complexity with the length of the sequence. These advantages cannot be satisfied in the same time when using other existing temporal modeling methods including Tranformers, 3DConv, RNNs and HMMs. For its significance, innovation and advantages, please read the first two Sections of the paper in details. In terms of causal dilated convolution (CDC), it definitly meets the parallel-computation, larger RF with fixed-size kernels and causal computing mechanism, however the dilation rate should be pre-defined, i.e., if the input length changes, the hyper-parameter, dilation rate should be changed accordingly before training. While TRAM does not require any adjustment when handeling different lengths of input. Morover, as far as we know, CDC is not in the 3D form which has the limitation for handeling temporal-spatial data such as radar RAD sequence. Hence, instead of talking about CDC let's dive into 3DConvs, which are more prefered by the researchers in RSS field. [Q2. IS IT MEANINGFUL TO DISCUSS REAL-TIME PERFORMANCE FOR RSS?] Yes, it is very important to discuss the real-time capability of RSS. As a remote sensing device, radar is applied in many fileds, such as automatic driving, security warning and so on. Taking Ku-band drone surveillance radar as example, the PRT (pulse repetition time) is around 80us, and it has 128 coherent pulses in one CPI (one Range-Doppler frame), then the data rate for detection will be $\frac{1 \times 10^6}{80 \times 128} = 97.66~\text{FPS}$. This requires subsequent signal processing and detection/segmentation algorithms to match this data rate as much as possible. Hence, in order to accurately detect and stably track the target, real-time performance is one of the important indicators of RSS task, which has practical significance at the application level. Taking automatic driving as another example, the moving car needs real-time feedback of detection results in the surrounding environment, otherwise it will lead to unexpected consequences. Therefore, RSS needs to balance accuracy and efficiency. --- Rebuttal Comment 1.1: Comment: Thank you for your response and the detailed explanation of my questions. I believe addressing these issues is crucial for refining TARSS-Net. Regarding the authors' response to [W1. LACK OF CLEAR HYPOTHESES AND REASONING], as a reader, it is beneficial to understand the characteristics of the input data in specific scenarios and the challenges associated with processing it. A clear thought process—such as the analysis and reasoning behind the hypotheses proposed by the authors, followed by the methods tailored to specific scenarios—can help readers with a general AI research background better comprehend the intentions behind the authors' work. In this regard, one of the baseline methods in this paper, TransRSS, presents its ideas more effectively in the Introduction section. --- Reply to Comment 1.1.1: Title: Further response to [W1] Comment: Dear Reviewer WAmw, We appreciate your recognition of our response and further suggestion for modification. TransRSS indeed sets a good example in the radar detection background introduction, with a clear chain of thought. However, we did not choose this writing logic for the following two reasons: 1. TARSS-Net has different methodology from that of TransRSS, the essence of the problem in TARSS-Net's hand is the efficient temporal encoding of radar high-dimensional spatio-temporal tensors. In the initial version of TARSS-Net's introduction, we also began by related background knowledge, including the similarities and differences between radar devices and other visual devices, radar signal processing pipeline, and the characteristics of the radar data in hand. Then, we summarized the development of radar target detection methods along a timeline, ultimately highlighting the importance of temporal relationship modeling. However, unlike TransRSS, which combines Transformer and CNN in a way that is more easily understood and accepted by general AI researchers, such writing style would fail to allow readers to realize the limitations and challenges of existing temporal modeling methods in radar signal processing. That is, given the existence of methods like 3DConv, RNNs, and Transformer, why is there a fundamental need for TARSS-Net? After repeated discussions and revisions, we chose to quickly highlight the current state of research in temporal relationship modeling in RSS in the limited space of the Introduction, point out the research gaps, and then provide a detailed summary of existing temporal modeling methods in Sec. 2. We deeply analyze why existing methods are not suitable for RSS and how to design temporal modeling methods suitable for the RSS field by addressing these limitations. 2. Currently, TARSS-Net has a higher research starting point. While organizing this work, we found that excellent works like TMVA, PKC, and TransRSS already possess the complete chain of thought you expect in their Introductions. Therefore, we boldly omitted some common background knowledge that has been thoroughly discussed, allowing us to take a higher perspective to discuss and analyze the RSS task from the entry point of temporal modeling mechanisms. This represents new cognition and understanding not present in current research work in this field. We believe these contents can bring new insights and more inspiration to readers, which is also why we believe this work is suitable for the NeurIPS community. TARSS-Net, standing on the foundation of existing excellent work, brings fresh perspectives and cognition. However, we must admit that since we placed more emphasis on temporal modeling, we had to omit some background information that does not affect the understanding of this paper under the constraints of the current limited space. This background information is included in the appendix. For example, in Sec. A From Radar Signals to Multi-View Representations, readers can understand the detailed process of millimeter-wave radar data processing in autonomous driving scenarios. In Sec. C Additional Descriptions of TARSS-Net, readers can gain a deeper understanding of the overall thought process of TARSS-Net in conjunction with the Methodology described in the manuscript. However, we agree with your suggestion. *Since additional page will be allowed in the camera-ready version, we promise to supplement the background introduction that cannot be added at this stage in the introduction, including the characteristics of radar data* and the challenges associated with processing it you mentioned, to better present the clear thought process of this paper. This will allow readers to easily grasp our intentions without having to read other materials or appendix as much as possible. Thank you for your contributions and thoughts to improve the quality of this paper. We sincerely express our respect to you and if you are satisfied with our new response to [W1], we also look forward to your higher evaluation, so that more readers can see the higher quality TARSS-Net after the revision! --- Reply to Comment 1.1.2: Title: Please take some time to review our new responses Comment: Dear Review WAmw, Thank you for you hard work! The discussions we have had not only help to improve the quality of this paper but also show respect for our hard work. Thank you again. We have responded to your new concerns accordingly. Since the rebuttal is almost over, please take some time to review our new responses. If you are satisfied with the replies, we especially appreciate you raising the evaluation score. If there are any new concerns, we look forward to continuing the in-depth discussion with you**. Sincerely, 9831 Authors.	Summary: This paper primarily introduces a network model called TARSS-Net, designed for the task of radar semantic segmentation. It effectively utilizes the temporal information of radar signals by introducing a novel temporal modeling mechanism, enhancing radar semantic segmentation performance. Specifically, the paper proposes a module called TRAM for temporal relationship learning between the target and historical frames, integrating it with other core components to construct the TARSS-Net model. Strengths: 1.The author successfully applies time modeling to the radar semantic segmentation task, achieving state-of-the-art performance on the RD-View radar semantic segmentation benchmark. 2.Rigorous ablation studies were conducted, providing solid evidence of the proposed method's efficacy. 3.The paper provides a good review of time modeling, which is helpful to the radar semantic segmentation task. Weaknesses: 1.This paper applies the time modeling to the field of radar semantic segmentation reasonably, but there are few innovations in the paper. 2.The performance of time modeling approaches tends to degrade in crowded scenarios. 3.It is difficult to significantly improve the performance of RA-View under the time modeling method 4.The related works reviewed for the radar semantic segmentation task are not comprehensive. 5.The placement of illustrations within the main text is not particularly reasonable, making re-display in the Appendix an unsatisfactory choice. Technical Quality: 3 Clarity: 3 Questions for Authors: 1.Can you provide a more detailed analysis of the computational complexity? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses [W1. FEW INNOVATIONS.] Thank you for recognizing the rationality of our approach. However, beyond reasonability, the novelty of TARSS-Net is also guaranteed. We feel sorry that we failed to make you see the novelty, as well as the careful thought and extensive experimental validation that went into it. Therefore, we think it is important to reiterate the real value and advantages of TARRS-Net. As we covered in `Sec. 2`, if you want to exploit temporal information in radar data, then you basically have 4 options: HMMs, RNNs, 3DConvs and Transformers. However, when we decide to use them, we encounter some unavoidable problems: i) as shallow probabilistic models, HMMs are very limited in their ability to express primitives for each single time step, and they cannot be trained end-to-end (`L86~L89`); ii) RNNs inherit the causal computing mechanism of HMMs and can be deeply trained end-to-end, but they cannot fully enjoy the computational efficiency of parallelization (`L90~L100`); iii) 3DConv can be fully parallelized, but its calculation nature is non-causal and cannot be adaptively adjusted according to the length of the input sequence, i.e., to process longer sequences, the network needs go wider or deeper, which leads to the introduction of more training parameters (`L102~L114`); iv) the Transformer with SA as the core overcomes the problem of limited local receptive field when 3DConv deals with long sequences, and also overcomes the defect of RNNs that cannot be parallelized. However, its computational complexity will show square level growth with the length of the sequence, resulting in computational redundancy (`L115~L127`). Seeing these problems, we deeply realized that there is NO FREE LUNCH, we have to redesign temporal modeling method for radar data, and RSS model should not be stagnant on these methods, it should have better sequence modeling methods. To this end, we meticulously redesign the temporal modeling method and propose the design principles, resulting in TRAM, which combines the advantages of causality, end-to-end learnability, constant model parameters under arbitrary length input, and linear growth of computational complexity with the length of the sequence. Without careful design, we will not find a model in current methods that can achieve all of these advantages and still have the performance of SoTA. Therefore, we hope you to understand our efforts and recognize our work. [W2. PERFORMANCE IN CROWDED SCENARIOS.] The concern you raised is indeed worth discussing. In real-measured radar data, target signatures often show extremely sparse characteristics (this does not refer to the physical size of the target, but to the characteristics of the target reflected in the radar signal)，so sparse and small target detection is a classic pain point which radar signal researcher dedicated to solving. So, if we understand your concern correctly, the RSS task does not suffer from the problem of model performance tends to degrade in crowded scenarios, which is one of the reasons why TARSS-Net emphasize the importance of learning and exploiting temporal relationships to enhance the representation of sparse target signature for RSS. [W3. DIFFICULTY OF PERFORMANCE IMPROVEMENT ON RA-VIEW.] This is a good question that points out something unique to the RSS domain. The mismatch between the quality of RA-view data (for range and angle measurement) and RD-view data (for range and Doppler measurement) is a common phenomenon. Restricted by the hardware, most radars pay more attention to the ranging accuracy when designing, and the angular accuracy of the commonly used low-cost radar is hard to guarantee. Therefore, the difficulty in improving the segmentation accuracy of RA is more caused by the poor data quality, rather than the failure of the temporal modeling method. [35] notices this problem and modifies annotation quality of RA, resulting CARRADA-RAC dataset. Comparing `Table 1 and Table 2`, it can be seen that annotation correction can improve the performance of RA-view from 51.3% to 58.7 % (TARSS-Net_D). Therefore, it is reasonable to think that there is more room for the temporal modeling method to play a role after further improving the data quality. [W4. NOT COMPREHENSIVE RELATED RSS WORKS.] Temporal modeling is not an emerging topic in other fields. However, effective and necessary modeling of temporal relationships has not been fully emphasized in RSS. Therefore, in order to encourage readers to revisit the temporal relation modeling from the perspective of radar signal processing and understand the motivation of TARSS-Net, we focus on the analysis of existing temporal modeling methods, the obstacles to their application in RSS domain, and the elaborate design of TARSSnet to face those obstacles. It is believed that `Sec. 1 and 2` can bring somehow inspiration for readers in the NIPS community. It is worth noting that in SoTA method comparison (see `Sec. 4.2`), we have included as comprehensive as possible the existing excellent RSS works and made a brief analysis. Limited by the length of this paper, please visit the relevant index to download original paper for details of these works. [W5. PLACEMENT OF ILLUSTRATIONS.] Thanks for your kind reminder and we are sorry for the reading trouble caused by unreasonable placement. Due to the space limit of paper submission, we had to make some typography which might make it uncomfortable to read. These will be corrected in next manuscript version, including the transposition of the order of Fig. 3 and Fig. 4, as far as possible to supplement the main text to reduce the redisplay in Appendix, etc. ## Questions [Q1. MORE DETAILED ANALYSIS OF THE COMPUTATIONAL COMPLEXITY.] There is real-time performance comparison in `Sec. E.2 of Appendix` and `Table 1 of PDF for rebuttal`, including model size, MACs, FPS and metrics. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I appreciate your further explanation of the TARRS-Net and the additional analysis of computational complexity. It is noteworthy that Crowded Scenarios refer to situations with a higher number of objects, and typically, time modeling may experience performance degradation in such scenarios. Overall, the authors have addressed some of the issues raised during the review process, reinforcing the significance of TARSS-Net as an innovative approach in the field of radar signal processing. I look forward to seeing the proposed improvements, and I will maintain my current rating. --- Reply to Comment 1.1.1: Title: Replying to Official Comment by Reviewer awi8 Comment: First, we want to thank you for recognizing the contributions and innovations of our paper. Then thanks for your more detailed explanation about [W2 PERFORMANCE DEGRADES IN CROWDED SCENARIOS]. Your mention of multiple targets has clarified your definition of crowded scenarios. The common multi-target situations in radars are already included in the CARRADA and our self-collected KuRALS datasets (e.g., the simultaneous presence of pedestrians and cars, multiple drones/UAVs or multiple ships). For the crowded scenarios you mentioned, we believe the closest examples are drone swarm detection and bird flock detection: - Taking the drone swarm detection scenario as an example. Typically, the radar will first perform a wide-area search, i.e., in the scanning mode it will emit coherent/incoherent pulse trains in each direction. The use of coherent pulse trains is more common. In each direction, these coherent pulse trains will form what we call RD representations. Considering the flight safety distance of the drones themselves, the number of targets that can be covered in the same direction (i.e., one RD frame) will be much smaller than the entire swarm size. In the end, it is difficult for single RD frame to reflect the crowded scenarios that appear in imaging signals such as camera images and SAR data. Subsequently, radar gets in the tracking mode, the system will form multiple tracking channels with smaller detection range, making each channel cover only the spatial range of one target as much as possible. After accumulating tracks over multiple RD frames, it will perform target identification or other more refined perception tasks. Therefore, whether in the scanning or the tracking mode, it is difficult to obtain RD frame which can reflect so called crowded scenario. Moreover, collecting such radar data for drone swarms is very challenging, requiring the cooperation of very professional flight control technicians and radar technicians to successfully collect this type of data. - In the bird flock detection scenario, the flight distance of birds (about 1~2 meters during the migration of large birds and tens of centimeters when small birds are foraging) is often lower than the radar's range resolution. For example, the Ku-band radar we use for collecting KuRALS dataset has a radial range resolution of about 3 meters. Therefore, the bird flock targets will all appear as a single mass target on one RD frame. At this point, the crowded scenario in vision collapses into a connected domain in the radar perspective. - Finally, let's analyze a more common crowded case, namely heavy traffic. Unlike vision, the form of crowded traffic scenarios in RDs cannot satisfy what you mentioned as crowded scenarios. This is because the R axis means radial distance, and D axis means Doppler or velocity. In heavy traffic, the speeds of the moving targets are similar, which results in a line pattern in parallel with R axis of RD frame. Such situations are included in the dataset involved in our experiments, where TARSS-Net can demonstrate its effectiveness. In summary, the crowded scenarios you refered to may appear more frequently in camera or SAR imaging radar. However, in the pulse Doppler or continuous wave radar discussed in this paper, crowded scenarios are generally diluted across different channels at signal processing level or collapsed into a single "target" due to the radar's physical characteristics. Therefore, visual crowding cannot be directly mapped to the radar scenarios discussed in this paper. The crowded scenarios you mentioned exceed the scope of signal processing algorithms and are a systematic issue. Radar designers will consider multiple aspects to design corresponding solutions, not just relying on signal processing algorithms. TARSS-Net proposes a general method for improving RSS performance. The datasets used in this paper cover typical radar application fields, such as autonomous driving and low-altitude surveillance. The experimental results also demonstrate the effectiveness of TARSS-Net. Hence the impact of crowded scenarios on the performance of radar temporal modeling methods is not within the scope of this paper. Nonetheless, thanks for your valuable concern, which motivates us to do more inspiring discussions and points us in a new research direction of this field. We will specifically consider this situation in subsequent studies. The TARSS-Net proposed in this paper has already demonstrated its effectiveness in improving the performance of general radar detection scenarios. We are pleased that this has been recognized by you and other reviewers. If our response can address your concerns, we hope you can further provide a higher rating to inspire more peers and further improve radar signal processing capabilities.	Summary: This paper proposed a temporal-aware framework, TARSS-Net, to enhance Radar semantic segmentation. The key idea is to propose a Temporal relation attentive module, TRAM (consists of Target-History Temporal Relation Encoding [TH-TRE] and Temporal Relation-Aware Pooling [TRAP] ), to capture the relations between Radar sequences. Comparisons with baseline methods on three datasets (CARRADA, CARRADA-RAC and a self-collected KuRALS) show the advantages of the proposed method Strengths: - The paper is well-written and in details. The motivation of incorporating temporal information is practical. - The proposed method outperforms the SOTA methods in some aspects, especially in RD-View. - The experiments and ablation studies are extensive. Weaknesses: - Since there are many existing works to incorporate spatio-temporal information, the technical contribution is limited without the insights considering the data characteristic of Range-Angle-Doppler data. - The performance increment is marginal, although various modules are designed. - It lacks of runtime statics to evaluate the time consumption introduced by temporal design. Technical Quality: 2 Clarity: 3 Questions for Authors: - How are the results when two frames are considered? - The performance decreases after 6 frames and increases again from 9 frames. Is there any analysis of that? - Can the authors provide to some feature visualization to confirm that whether the designed scheme payed attention to the object (or where to pay attention) in Radar sequences. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors appropriately discussed the limitations in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses [W1. LIMITED TECNICAL CONTRIBUTION.] Sorry to have caused the reviewer such concerns. The core point of this paper is to redesign a better spatio-temporal modeling method for radar data from the perspective of temporal information utilization. Indeed, there are many works in other fields that discuss the utilization of spatio-temporal information. However, as we analyzed in `L34~L44`, there is still a large research gap in the related research of radar-oriented deep learning models: "In terms of temporal information utilization, the common practice is still 3DConv." Hence, we dedicated to exploring an advanced temporal modeling mechanism, whcih is urgently needed for modern radar (alleviating the problem of time-sensitive changes in radar data quality and learning better target representations). For the insights of RAD data handling, it is not the main focus of this paper, for two reasons: i) TMVA-Net [21] had provided a multi-view learning approach that well balances information utilization and efficient processing for RAD data; ii) conducting detection/segmentation on RAD data is not universal in radar systems, for some phased array radar system, it is more efficient to perform the detection directly on RD, e.g., KuRALS dataset used in this paper. In summary, this paper aims at the problem of temporal modeling which is more general in radar systems. From this perspective, this work summarizes the advantages and disadvantages of modern temporal modeling methods in `L82~L127` and the problems that need special attention when processing radar data in `L130~140`. We will reiterate our core research objectives in the Introduction section to highlight our contributions. [W2. MARGINAL PERFORMANCE INCREMENT.] Sorry not to surprise you, but this is actually not a small improvement for RSS tasks. As we mentioned in `L20~L24`, radar data is different from image data. It is susceptible to various interference and does not have semantic information, which make RSS more chanllenging. From the experimental part, it can be seen that the general semantic segmentation model of CV has not achieved good results on radar data (such as FCN, Unet). Then TMVA-Net [21], has improved the performance of RSS by about 2% (taking RD view as an example). The performance improvement of subsequent TransRadar [5], TransRSS [36] and PKCIn-Net [35] also increases by about 2%. To date, from FCN to TARSS-Net, RSS performance has improved from 66% to 75% on benchmark dataset, CARRADA. As you can see, RSS performance is thus advanced bit by bit. That said, the authentic and reproducible 2% performance improvement achieved by TARSS-Net is significant for the progress in RSS field. [W3. LACK OF RUNTIME STATISTICS.] Due to page limitation, the runtime performance comparison is listed in `Sec. E2 of Appendix`. TARSS-Net_D takes 43ms to infer one RAD@5Frames, 9 ms for one RD@5Frames and 9 ms for RA@5Frames (tested on single Nvidia3090). ## Questions [Q1. RESULTS WITH 2 FRAMES.] Thanks for your suggestions to make our experiment more comprehensive and solid. We further tested TARSS-Net_D with 2 input frames on RD view CARRADA: mDice 73.7%, mIoU 62.0%, Precision 71.3%, Recall 76.5%. It is supplymented in `Figure 1 of attached PDF` for this rebuttal. [Q2. ANALYSIS OF PERFORMANCE DECREASING AFTER 6 FRAMES AND INCREASING AGAIN FROM 9 FRAMES.] This is a particularly valuable question. We did do some analysis of the trend of the curve shown in `Fig. 5`, and try to summarize some general conclusions that can guide TARSS-Net usage. Different operating conditions of radar affect the quality of each data frame, so in one RSS dataset, it is hard to generalize how many historical frames are closely related to the target one, or which historical frames are helpful to the current frame (poor quality history frame will introduce a lot of irrelevant noise and lead to performance degradation). Not to mention different RSS datasets. All networks dealing with temporal-sturctured radar data face the problem that they cannot calculate the correspondence between the number of input frames and network performance. TARSS-Net has its own advantages to help ease the choice of input frame length. With the design of TRIC layer, TARSS-Net could accept input with adjustable time length while keeping the number of parameters constant. When using it, readers can try and choose the number of input frames according to constraints of inference performance of the hardware (TFLOPS) without worrying about model size. [Q3. Feature visualization.] Please see `Figure 2 in attached PDF` for this rebuttal. --- Rebuttal Comment 1.1: Title: We appreciate your respect for our hard work Comment: Dear Reviewer MgaH, We have spent a significant amount of time and effort analyzing your concerns and suggestions, and have provided detailed explanations and corresponding modifications. We believe that our response should be sufficient to address your concerns. We sincerely hope that you would take the time to read our response. If our response is adequate, we kindly ask you to give a fair score upgrade; if you still have other concerns, looking forward to further discussions with you. Thank you for your contribution to improving the quality of our paper, and we also appreciate your respect for our hard work. 9831 Authors --- Rebuttal 2: Title: Looking forward to your comments on our reply Comment: We have provided detailed responses to all of your questions. Hope to get your approval on the reply and update the rating. We also look forward to more in-depth communication and discussion with you! Thank you again!	Rebuttal 1: Rebuttal: We would like to express our respect to all reviewers and AC. Thanks for your time and hard work. Based on your professional opinions, we have carefully replied all the high-value questions and supplemented the content accordingly. Pdf: /pdf/8fee510df92fd698c7cc6a3c4be2739378aa624a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Relating Hopfield Networks to Episodic Control	Accept (poster)	Summary: The paper shows that the Differentiable Neural Dictionary (DND) used in the context of reinforcement learning is mathematically equivalent to an Hopfield network with heteroassociative memories. Based on this observation, the paper generalises DND using the formulation of Universal Hopfield Networks (UHN) and studies the effect on 'retrieval' and 'capacity' of different separation and similarity functions. Additionally, the authors introduce a new criterion to measure the retrieval capability of an associative memory, which, they argue is a better proxy for memorisation. Strengths: The paper is generally well written and has a good structure. Particular strengths reside in: - Solid contribution: the characterisation of differentiable neural dictionaries as associative memories can help future research in the field of neural episodic control by applying the advancements of the research on associative memories to the DND implementation. - Soundness of the approach: the authors establish the connection theoretically and then experimentally study the effect of changing the form of the DND associative memory on the memory capacity. They use enough baselines and different datasets, so the results are convincing. Weaknesses: I think the paper is relatively light, meaning that there is little content. In particular, the authors focus most of the experiments on studying the effect of different similarity and separation functions on capacity for storing images, whilst they could have investigated the implications in the reinforcement learning domain more. It's unclear why the study of the 'max' separating function in comparison to softer versions thereof is so important for this paper: it indeed seems that section 3 belongs to another paper, which talks about the characteristics of different separating functions. Another example is the Discussion section, which occupies a page and a half mostly with speculative sentences and arbitrary connections to neuroscience research, which are far from justified in the context of this paper. It would have been much better if the paper actually empirically or theoretically attempted at establishing this connections. For an example see lines 217-219. Technical Quality: 2 Clarity: 2 Questions for Authors: Some minor points and questions: - In Section 2 what are the dimensions of the vectors and matrices, e.g. V, K, q, $\phi$, etc? Please rectify directly in the paper. - The definition of the similarity function used by DND has an unusual $\delta$ in it. That is probably there for numerical stability reasons and should be ignored in the mathematical treatment. - Equations (7-9) are obvious and not needed. - The dynamic and differentiable way of defining the threshold (Equations 15-16) is clever but not necessary since you can just zero-out the non top-K ones? (which you do in the experiments if I understood correctly) - It is not really clear to me the difference between capacity and retrieval, which is central to section 3 so you might want to state it somewhere at the beginning of it. The line between them seems very thin to me. - Line 134-135 maybe you wanted to say the opposite? -In section 3 you are testing an hypothesis which has not been clearly spelled out before. If I understand correctly, your hypothesis is that the k-max function is useful for generalisation or memorisation depending on the value of k chosen. I believe you have tried to mention this in the introduction of Section 3 (lines 109-112) but I think it's not clear and confusing at the moment. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: I can see no major limitations, but the paper could have been developed more. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer w6Cb for the constructive feedback and for acknowledging the strengths of our paper, including the solid contribution and soundness of our approach. We appreciate the recognition of our work in characterizing Differentiable Neural Dictionaries (DNDs) as associative memories and its potential to advance research in neural episodic control. We also thank you for pointing raising valid comments. We directly included explicitly changes you proposed (e.g. clarification of matrix dimensionality), and we address additional concerns and questions that call for a response below. 1. Comment: "I think the paper is relatively light [...]" - Answer: We agree that further exploring reinforcement learning (RL) implications is valuable, and we are actively pursuing this direction in ongoing work. The primary focus of this manuscript was to describe the theoretical connection between DNDs and Hopfield networks and to evaluate the memory capabilities of DNDs. Due to the page limit and dense supplementary information, we concentrated on establishing this foundational work. We look forward to sharing new results on RL applications in future publications. 2. Comment: "It's unclear why the study of the 'max' separating function in comparison to softer versions thereof is so important for this paper: it indeed seems that section 3 belongs to another paper, which talks about the characteristics of different separating functions." - Answer: Thank you for your feedback. All experiments, including the one you mentioned, evaluate the k-max (not only max) function to demonstrate its unique properties and effectiveness in different contexts. We believe this comparison provides valuable insights into optimizing memory retrieval strategies. More precisely, it suggests that NEC could be improved by replacing k-max by softmax. Does this clarification address your concern? 3. Comment: "Another example is the Discussion section, which occupies a page and a half mostly with speculative sentences and arbitrary connections to neuroscience research" - Answer: Thank you for pointing that out. We do think the connection to neuroscience can be made and "adds an interesting interdisciplinary perspective" (Reviewer 4Gfy). However, we replaced speculative content by referenced claims. 4. Question: "The dynamic and differentiable way of defining the threshold (Equations 15-16) is clever but not necessary since you can just zero-out the non top-K ones? (which you do in the experiments if I understood correctly)" - Answer: You are right. This new definition was introduced to derive the energy functions for neural episodic control, but it was not used in the experiments. 5. Question: "It is not really clear to me the difference between capacity and retrieval, which is central to section 3 so you might want to state it somewhere at the beginning of it. The line between them seems very thin to me." - Answer: Thank you for bringing this to our attention. We adhered to the distinction of Millidge et al. (2022). Capacity refers to the maximum number of images the DND can store while maintaining accurate memory representation. Retrieval, on the other hand, measures the model’s ability to accurately recall stored images when subjected to increasing levels of noise. We have added a definition and explanation of these terms in the manuscript to clarify their distinction. We appreciate the constructive feedback provided by Reviewer w6Cb, which has been instrumental in enhancing the clarity and depth of our manuscript. We have addressed the concerns and questions raised, and we believe these revisions have strengthened our work. We are committed to further exploring the implications of our findings in reinforcement learning and look forward to sharing these insights in future publications. Thank you again for your thoughtful and valuable comments, which have helped us improve our contribution to the field.	Summary: This paper introduces a formulation of an energy function which involves retrieving top-k memories while making a connection to both Neural Episodic Control and Associative Memory. The novel energy function utilized in this work demonstrates superior performance in the retrieval setting which involves image in-painting where the top half of an image is masked out. Additionally, the work explores the efficacy of the model under the retrieval setting in which queries are perturbed with a range of noise values. Strengths: The proposed energy function is novel and well thought. The formulation follows the Neural Episodic Control and Universal Hopfield Network (UHN) paradigms. Moreover, the experiments are interesting, and the model is shown to have great performance, specifically the illustrations of improved memorization capacity across a variety of k-Max functions demonstrated on MNIST. When it comes to denoising, the proposed model is able to recover patterns very well as k increases. Weaknesses: Although the experiments are good, many of the experiments are ablation studies of the introduced energy function, while there is one experiment section contrasting the proposed energy function and softmax based energy function, which illustrates a worse performance of the new function. Moreover, when dealing with RGB images, the function performs badly as k-neighbors increases. Finally, the connection to RL is fascinating but it seems too brief in the paper. Technical Quality: 4 Clarity: 2 Questions for Authors: Why is the range of $\beta$, 0 to 50000, chosen for figure 5? Such values seem to be too large, while the range of k makes sense. Looking at figures 6 and 8, it seems that when k = 1, the model performs best on CIFAR10 and Tiny ImageNet which is an opposite trend to MNIST. What is the explanation behind this fact? For figure 7, as k increases, the performance of the model increases given the setting of denoising. What is the number of images that the model is being evaluated on? Could a visual example of denoising Tiny ImageNet images be provided? What is the threshold utilize to determine whether a recovered pattern is memorized or not for each of the dataset? "The novel criterion is such that a trial is correct if and only if the squared pixel difference between the truth and the output is lower than it is between the output and any other memory" --- Why is there no equation which describes this important criterion? General comment: Since the maximum chosen k value is set to 50 in the tables, I think it would be good to include k = 100 in such tables as it is demonstrated in figure 5. Confidence: 4 Soundness: 4 Presentation: 2 Contribution: 2 Limitations: The connection to Neural Episodic Control and RL is too brief. Additionally, the function does not beat the softmax based energy function in terms of retrieval while as the number of neighbors k increases, the performance is worsen on CIFAR10 and TinyImageNet datasets. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: [concise because hitting char lim] Thank you for your constructive feedback and insights on our work. 1. Comment: "Although the experiments are good, many of the experiments are ablation studies of the introduced energy function, while there is one experiment section contrasting the proposed energy function and softmax based energy function, which illustrates a worse performance of the new function." - Answer: Yes, and we think these are important findings. The Manhattan similarity function consistently outperforms the Euclidean one, even with softmax, as noted by Millidge et al. (2022). This suggests that NEC can be improved through our connection between the DND and UHN. We also argue that softmax could improve the flexibility and performance of NEC. Conversely, k-max constitutes a strong alternative due to its superior search complexity, especially when implemented with a k-d tree, as we now argue in the new version of the manuscript. 2. Comment: "Moreover, when dealing with RGB, the function performs badly as k increases." - Answer: Several factors could indeed influence how robust performance is to changes in k, such as the use of RGB images, the higher dimensionality and more naturalistic nature of complex images compared to MNIST. 3. Comment: "The connection to RL is fascinating but it seems too brief in the paper." - Answer: Thank you for finding the connection fascinating. Our paper establishes a comprehensive theoretical link between NEC and associative memory models. Additionally, we present a set of new empirical results demonstrating the applicability of NEC to associative memory tasks. We are now working on exploring how insights from associative memory can in turn enhance NEC. 4. Question: "Why is the range of beta, 0 to 50000, chosen for figure 5? Such values seem to be too large, while the range of k makes sense." - Answer: The experiment for Figure 5 is computationally expensive, so we aimed to identify peak performance across a wide range for all datasets and similarity functions. We found that peak performance does not always occur with low beta values (fig. 11, 10). Finally, the range already demonstrates that softmax most often outperforms k-max. Please let us know if you have any more concerns. 5. Question: "Looking at figures 6 and 8, it seems that when k = 1, the model performs best on CIFAR10 and Tiny ImageNet which is an opposite trend to MNIST. What is the explanation behind this fact?" - Answer: We believe your comment may be a reference to Figures 6 and 7 instead. We indeed observe that for lower-dimensional datasets like MNIST (and CIFAR-10 to a lesser extent), values of k>1 can lead to improved performance. This finding contradicts the statement by Millidge et al. (2022) that k=1 is always optimal and motivates our introduction of a new performance criterion where k=1 becomes optimal. These points are discussed in our abstract (lines 8 to 14) and throughout the manuscript. 6. Question: "For figure 7, as k increases, the performance of the model increases given the setting of denoising. What is the number of images that the model is being evaluated on? Could a visual example of denoising Tiny ImageNet images be provided?" - Answer: Thank you very much for highlighting this crucial detail. We apologize for the oversight. Similar to the approach used by Millidge et al. (2022), we evaluated the model using sets of 100 images. We have now included this information in the revised manuscript. We also appreciate your suggestion to provide a visual example of denoising, and have added such examples for Tiny ImageNet in the supplementary material. 7. Question: "What is the threshold utilize to determine whether a recovered pattern is memorized or not for each of the dataset?" - Answer: The error threshold is used for the absolute criterion and is set to 50, consistent with the approach used by Millidge et al. (2022), who utilized the same datasets. We mentioned this criterion in line 125 of the manuscript, and we have now made it clearer in the revised version to ensure this important detail is easily accessible. 8. Question: "Why is there no equation which describes [the generalization] criterion?" - Answer: Thank you for pointing this out. Upon review, we realized that our previous description was incorrect. The correct criterion allows the sum of squared pixel differences to be equal, not just less, for a trial to be correct. We appreciate your suggestion to include an equation, as it prompted us to clarify this aspect. The retrieval output is denoted as $z$, and the matrix of stored memories is $K$. A trial is considered correct if the sum of squared pixel differences between $z$ and the correct memory $K_{\text{correct}}$ is equal to the minimum difference between $z$ and any memory $K_i$, i.e. $z - K_{\text{correct}} \\|^2 = \min_{i} \\| z - K_i \\|^2$. This adjustment to the phrasing does not affect our results, as our implementation already adhered to this correct criterion. We have included this equation and corrected the phrasing in the revised manuscript to ensure clarity and accuracy. 9. Comment: "Since the maximum chosen k value is set to 50 in the tables, I think it would be good to include k = 100 in such tables as it is demonstrated in figure 5." - Answer: Thank you for your suggestion regarding the inclusion of k=100 in the tables, but we do not have such data. For tables, the max is set to k=50 like in the original NEC paper. Other experiments have consistently shown that performance peaks before k=50, as evidenced by the results in Figures 5, 10, and 11. Let us know if you have any more concerns. We appreciate the time and effort that Reviewer ebPC has put into evaluating our paper. Your constructive feedback has helped us clarify key aspects of our manuscript. --- Rebuttal Comment 1.1: Comment: I appreciate the detailed response and hard work from the authors. Based on the response and clarifications provided by the authors, I will be raising my score from 4 to 6. My judgement for the score being --- I think the work is fascinating and experiments are good but the connection to RL is rather brief and the performance of the introduced model is decent but not competitive. But I wish the authors best of luck.	Summary: This paper establishes a direct connection between two important frameworks: neural episodic control and Universal Hopfield Network. It further derives Lyapunov functions for the dynamics and explores the ability of Neural Episodic Control to function as a memory system. Strengths: The idea of the paper is original and the manuscript is clear. The new theoretical connection is significant and important in relating previously unrelated frameworks. Weaknesses: While the contribution is relatively important, it is solely based on a mathematical equivalence between the two frameworks. It is a bit unclear to me, whether such an equivalence runs deep and might affect the way memory and RL systems co-function together. In a way, it feels like this paper could dig deeper into an underlying unifying framework. Technical Quality: 4 Clarity: 3 Questions for Authors: - Can the authors further clarify how does the proposed connection between Hopfield Networks and reinforcement learning algorithms advance the current understanding of episodic memory in neural networks? - In deriving the energy functions, specific modifications were made to the separation function κ. What are the theoretical justifications for these modifications, and how do they impact the overall model performance? - The paper suggests that the Manhattan distance kernel improves performance over the Euclidean distance kernel. Can you provide a detailed analysis of why this might be the case? - The introduction of a new criterion to disentangle memorization from generalization is an interesting contribution. How does this criterion compare to existing methods of evaluating associative memory models? Confidence: 2 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer 9C8q for the constructive feedback and positive remarks on our work. We appreciate the recognition of the originality and clarity of our manuscript, as well as the significance of the theoretical connections made. Below, we address the specific questions and points raised. 1. Comment: "While the contribution is relatively important, it is solely based on a mathematical equivalence between the two frameworks. It is a bit unclear to me, whether such an equivalence runs deep and might affect the way memory and RL systems co-function together. In a way, it feels like this paper could dig deeper into an underlying unifying framework." - Answer: Thank you for highlighting the importance of our contribution. To better highlight the connection between DND and UHN, we have added a new figure illustrating this equivalence. Our empirical results also provide clear insights into how RL methods based on episodic control could be improved by leveraging this connection. We are currently testing these hypotheses in RL settings, which we believe will further validate and deepen our understanding of this unifying framework. 2. Question: "Can the authors further clarify how does the proposed connection between Hopfield Networks and reinforcement learning algorithms advance the current understanding of episodic memory in neural networks?" - Answer: The proposed connection allows us to leverage the well-studied dynamics of Hopfield Networks to improve the efficiency and effectiveness of neural episodic control in RL. By understanding how memories are stored and retrieved within this framework, we can design RL systems that better integrate episodic memory, leading to improved decision-making and faster learning. This connection enhances our understanding of episodic memory by providing a unified theoretical framework that explains memory dynamics in both associative memory models and RL systems. 3. Question: "In deriving the energy functions, specific modifications were made to the separation function [k-max]. What are the theoretical justifications for these modifications, and how do they impact the overall model performance?" - Answer: The modifications to k-max were made to derive the energy functions for neural episodic control. We ensured the new function has nonzero gradient while preserving the functional properties of k-max. As beta and beta_k grow to infinity, the new function approaches k-max. Hence, given sufficently large beta and beta_k, the overall model performance is not affected. We have included this justification in the revised manuscript. 4. Question: "The paper suggests that the Manhattan distance kernel improves performance over the Euclidean distance kernel. Can you provide a detailed analysis of why this might be the case?" - Answer: Similar to the findings of Millidge et al. (2022) in their Universal Hopfield Network paper, we observe that the Manhattan distance kernel outperforms the Euclidean distance kernel. The selection of similarity functions is crucial because it significantly impacts the ranking of different memories. While the exact reason why the Manhattan distance yields better results is not entirely clear, one possible explanation is that the Euclidean distance involves squaring the differences, which can introduce distortion, whereas the Manhattan distance preserves linearity. 5. Question: "The introduction of a new criterion to disentangle memorization from generalization is an interesting contribution. How does this criterion compare to existing methods of evaluating associative memory models?" - Answer: We thank the reviewer for highlighting out contribution. Our new criterion provides a more nuanced evaluation by explicitly distinguishing between an associative memory model's ability to memorize specific instances and its ability to generalize across similar instances. We show that the k-max separation function (k>1) can yield state of the art performance with the standard generalization performance. This contradicts the prediction of Millidge et al. (2022) that 1-max is the optimal separation function. We also point out that 1-max is optimal when performance is assessed with the new memorization criterion. Hence, our work challenges the canonical criterion and allows for a richer evaluation of associative memory models. We hope these clarifications address your concerns and enhance the understanding and impact of our work. Thank you again for your valuable feedback, which has helped improve our manuscript significantly.	Summary: The paper introduces the differentiable neural dictionary, which uses template-based memory storage, relating it mathematically to Hopfield Networks within the UHN framework. This novel model is shown to be capable of storing memories through operations such as similarity, separation, and projection, thus demonstrating high capacity and adaptability. The model employs different separation functions like k-nearest neighbor and Max function, with the latter transitioning sharply between memory states upon noise increments. It also discusses how the dictionary outperforms traditional models in associative memory tasks by using different distance kernels (Euclidean, Manhattan), which aids in better generalization over memorization. The discussion extends to the potential applications of DND in episodic control within reinforcement learning, suggesting that DND can speed up learning by reducing the bottleneck in decision processes and integrating generalization into episodic memory. The text ends by suggesting further research into the biological basis of DND and its relationship with neural mechanisms of memory. Strengths: 1. The article is clearly articulated and provides detailed experimental procedures. 2. This paper presents a novel approach to integrating associative memory with reinforcement learning. Specifically, it re-derives mathematical formulations, transforming the form of the Differentiable Neural Dictionary (DND) into the Universal Hopfield Network framework (UHN). This transformation leads to the derivation of corresponding energy functions from the UHN. 3. The paper demonstrates the capacity of the DND model on MNIST, CIFAR-10, and Tiny ImageNet datasets using various functions. It evaluates the model’s retrieval capability against noise, performance based on the memorization criterion, and the relationship between k-Max and Softmax functions. Weaknesses: The article starts from the Hopfield Network framework, and my concern lies in the fairness of introducing the k-nearest neighbor for comparisons in associative memory. This introduction seems to bring in additional memory information, potentially skewing the comparisons. Furthermore, the paper suggests that evaluating retrieval should involve comparisons with the entire dataset. This approach deviates from the essential nature of memory retrieval tasks. If there are enough elements in the dataset for comparison, such as with the 50-Max criterion, selecting the most likely candidate from a set, even in the presence of significant noise, seems justifiable. However, this method may not accurately reflect the true performance of memory retrieval under more typical conditions where fewer comparison points are available. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the "Weaknesses" section. Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The article extensively discusses its limitations towards the end. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer UfDJ for the constructive feedback and positive remarks on our work. We appreciate the recognition of our novel approach and detailed experiments. Below, we address the specific concerns raised. 1. Comment: "The article starts from the Hopfield Network framework, and my concern lies in the fairness of introducing the k-nearest neighbor for comparisons in associative memory. This introduction seems to bring in additional memory information, potentially skewing the comparisons." - Answer: It is not clear to us what additional memory information is added by introducing the k-nearest neighbor function. If the concern is about memory complexity, it is not impacted by this new separation function. We added information about the computational complexity of the different UHN instances and highlight that the search complexity of k-max is better than other separation functions when implemented with a k-d tree. If the concern is about additional hyperparameters, k is equivalent to softmax's beta, so k-max uses the same number of hyperparameters as softmax. 2. Comment: "The paper suggests that evaluating retrieval should involve comparisons with the entire dataset. This approach deviates from the essential nature of memory retrieval tasks." - Answer: We do not mean that retrieval should always involve comparisons with the entire dataset. Our intent is to highlight that the standard criterion, which evaluates performance in absolute terms, differs from a criterion that assesses how close the retrieval is to the correct image relative to other images in the training set. This distinction is important as it influences whether the evaluation favors memorization or generalization strategies. We rephrased key sentences in the Discussion section. We hope these clarifications address your concerns and enhance the understanding and impact of our work. Thank you again for your valuable feedback. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I have carefully reviewed the other reviewers' comments as well as your replies. I believe I will maintain my current evaluation score.	Rebuttal 1: Rebuttal: We thank all reviewers for their detailed and constructive feedback on our submission. We are pleased that the reviewers recognize our significant theoretical contribution of linking Differentiable Neural Dictionaries (DNDs) to the Universal Hopfield Network (UHN) framework (R.4Gfy) and the extensive empirical evaluation we conducted (R.4Gfy). The clarity and articulation of our manuscript were also noted positively (R.UfDJ, R.4Gfy, R.9C8q, R.w6Cb, R.ebPC). Additionally, our novel approach to integrating associative memory with reinforcement learning was appreciated (R.UfDJ), along with the introduction of a new "memorization criterion" (R.4Gfy) and the originality and significance of relating previously unrelated frameworks (R.9C8q). The value of deriving new energy functions for DNDs was also recognized, both theoretically (R.4Gfy) and empirically (R.ebPC). In response to the reviewers' comments, we have enhanced the clarity, organization, and practical relevance of our manuscript. We addressed concerns about the focus on image reconstruction tasks, provided a clearer discussion of practical implications and computational complexity, and added visual aids to improve accessibility. In individual rebuttals, we detail our responses to each of the reviewers' comments and outline the revisions made to address their concerns.	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper establishes a novel connection between differentiable neural dictionaries (DNDs) used in episodic control for reinforcement learning and Hopfield networks used as associative memory models. The authors show that DNDs can be formulated within the Universal Hopfield Network (UHN) framework, derive energy functions for DND recall, and conduct experiments comparing DND performance to other associative memory models on image reconstruction tasks. I believe that this paper presents a valuable theoretical contribution by connecting DNDs to the UHN framework, supported by extensive empirical evaluation. The work has the potential to impact both reinforcement learning and associative memory research. However, the paper would benefit from a bit clearer organization, a more explicit discussion of practical implications, and ideally some exploration of the impact on RL tasks. Finally, the addition of a conceptual figure or schematic would further strengthen the paper by making its key ideas more accessible and memorable. With these improvements, I believe this paper would be a strong contribution to NeurIPS. Strengths: - Despite some organisation issues I mention below, I still find this paper relatively easy to follow and with good quality of English, plots, and articulation of ideas. - The theoretical contribution linking DNDs to Hopfield networks and the UHN framework is significant and well-argued. This connection opens up interesting avenues for cross-pollination between reinforcement learning and associative memory research. - The derivation of energy functions for DND recall, including a novel continuous approximation, is mathematically sound and adds to our theoretical understanding of these models. - The experimental comparisons are cover three datasets (MNIST, CIFAR10, Tiny ImageNet) and evaluating different similarity and separation functions. - The introduction of a new "memorization criterion" for evaluating associative memory performance is thoughtful and helps distinguish between memorization and generalization capabilities. Weaknesses: - While the experiments are comprehensive, they focus solely on image reconstruction tasks. Given the paper's motivation from reinforcement learning, it would have been valuable to include experiments or discussion on how these findings might impact episodic control in RL settings. - The paper is quite dense and could benefit from clearer organization. Some important findings, like the superior performance of the Manhattan distance similarity function, are buried in the results sections and could be highlighted more prominently. - The discussion, while interesting, feels somewhat speculative and disconnected from the main technical contributions. This section could be tightened or more clearly linked to the paper's core findings. - The paper lacks a clear discussion of computational complexity trade-offs between different similarity and separation functions. This would be particularly relevant for practical applications in RL or large-scale associative memory tasks. Technical Quality: 3 Clarity: 3 Questions for Authors: ### I have the following questions/suggestions: - How do you expect the choice of similarity and separation functions to impact sample efficiency and performance in RL tasks using episodic control? - Have you considered how the memorization vs. generalization trade-off might be dynamically adjusted in a learning system? Could this provide benefits in certain RL scenarios? - How does the computational cost of DND retrieval with k-Max separation compare to other UHN instances, particularly for large memory sizes? - Have you considered the relation between this model and other episodic memory-like approaches, including for instance generative models for video, LLMs, and model-based RL? ### Suggestion for improvement: I think the paper would greatly benefit from the addition of a high-level schematic or conceptual figure that visually illustrates the connection between DNDs and the UHN framework. Such a figure could perhaps 1. Show the parallel structures and operations in DNDs and UHNs side by side. 2. Illustrate how the similarity, separation, and projection operations map between the two frameworks. 3. Visualize the proposed continuous approximation of the k-Max function. 4. Demonstrate how the energy functions relate to the model's dynamics. A clear, well-designed figure with any (or all) of these, would significantly enhance the paper's accessibility, especially for readers less familiar with either DNDs or UHNs. It would also help to crystallize the paper's main theoretical contribution and make the work more memorable and impactful. ### Minor points: - The discussion of biological plausibility in relation to hippocampal models adds an interesting interdisciplinary perspective, but it’s very brief. I would extend it if possible. - The paragraph in line 180 is quite long and difficult to follow. I know you were trying to keep the 9 page limit but in the camera ready version I would recommend you to break it into two paragraphs and rephrase a few sentences. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: ### Overview: The authors have made an effort to address some limitations of their work, which is commendable. However, there is room for improvement in this area: 1. Scope of experiments: The authors acknowledge that they did not explore the evaluation of similarity and separation functions in reinforcement learning tasks due to time constraints. This is a good start, but they could expand on why this limitation is important and how it might impact the broader applicability of their findings. 2. Theoretical vs. practical implications: While the paper provides a strong theoretical foundation, it could benefit from a more explicit discussion of the limitations in translating these theoretical insights into practical applications, especially in reinforcement learning contexts. 3. Scalability: The paper doesn't adequately address potential limitations in scaling the proposed methods to very large datasets or complex, high-dimensional state spaces that might be encountered in real-world reinforcement learning tasks. 4. Computational resources: There's no discussion of the computational requirements for the various methods compared, which could be a significant limitation in certain applications. 5. Negative societal impact: The authors do not explicitly discuss potential negative societal impacts of their work. While this research is primarily theoretical, it would be beneficial to consider and address potential misuse or unintended consequences of more efficient episodic control in AI systems, in the camera ready version of this paper. ### Suggestions: 1. Include a dedicated "Limitations and Future Work" section. 2. Discuss how to address these limitations in future research. 3. Consider potential societal impacts (e.g., privacy implications, risks of enhanced AI memory systems). 4. Briefly address computational resource requirements and scalability. Addressing these points would demonstrate a more comprehensive understanding of the work's broader implications and potential impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank Reviewer 4Gfy for the detailed and constructive feedback on our submission. We appreciate the recognition of our significant theoretical contribution and extensive empirical evaluation. Below, we address each of your comments and suggestions in detail. 1. Comment: "While the experiments are comprehensive, they focus solely on image reconstruction tasks. Given the paper's motivation from reinforcement learning, it would have been valuable to include experiments or discussion on how these findings might impact episodic control in RL settings." - Answer: We understand your concern. We used three different image datasets, including CIFAR10 and Tiny ImageNet, which are naturalistic and high-dimensional, to evaluate the performance of memory models comprehensively. Image reconstruction tasks provide a robust testbed for assessing memory models due to their complexity and interpretability. We are now working on evaluating the models on RL tasks, which will hopefully give rise to a new manuscript. We have added a paragraph in the Discussion section to outline these future directions. 2. Comment: "The paper is quite dense and could benefit from clearer organization. Some important findings, like the superior performance of the Manhattan distance similarity function, are buried in the results sections and could be highlighted more prominently." - Answer: We have reorganized the results section to highlight key findings, including the superior performance of the Manhattan distance similarity function, and improved the overall clarity of the manuscript. 3. Comment: "The discussion, while interesting, feels somewhat speculative and disconnected from the main technical contributions. This section could be tightened or more clearly linked to the paper's core findings." - Answer: We have revised the discussion section to more clearly link it to our core technical contributions, replacing speculative content related to neuroscience by referenced claims and focusing on practical implications and future research directions. 4. Comment: "The paper lacks a clear discussion of computational complexity trade-offs between different similarity and separation functions. This would be particularly relevant for practical applications in RL or large-scale associative memory tasks." - Answer: We thank the reviewer for raising this very interesting point. We have now provided a comparative analysis of the computational costs for DND retrieval with k-Max separation versus other UHN instances. Neural episodic control implements k-Max using a k-d tree, whose average search complexity is O(log n). In comparison, the complexity of the softmax function as used in modern Hopfield networks is O(n). This difference in complexity can have significant implications for large memory sizes, where the k-Max approach can disregard significant portions of the memory. 5. Question: "How do you expect the choice of similarity and separation functions to impact sample efficiency and performance in RL tasks using episodic control?" - Answer: We are very much looking forward to finding out how the choice of similarity and separation functions impacts sample efficiency and performance in RL tasks. We believe that functions promoting better memory, such as the Manhattan distance, can enhance sample efficiency by enabling more effective recall of relevant past experiences. We have elaborated on this in the Discussion section and are currently working on experiments to validate these hypotheses in RL settings. 6. Question: "Have you considered how the memorization vs. generalization trade-off might be dynamically adjusted in a learning system? Could this provide benefits in certain RL scenarios?" - Answer: Yes, gradually shifting from a strategy that memorizes specific examples to one that generalizes aligns with the complementary roles of episodic and semantic memory in cognitive science. We discuss how parameters such as k and beta can be dynamically adjusted during training to balance memorization and generalization. This approach can provide significant benefits by adapting the model to different phases of learning and varying task requirements in reinforcement learning scenarios. 7. Question: "How does the computational cost of DND retrieval with k-Max separation compare to other UHN instances, particularly for large memory sizes?" - Answer: See 4. 8. Question: "Have you considered the relation between this model and other episodic memory-like approaches, including generative models for video, LLMs, and model-based RL?" - Answer: We have expanded the related work section to discuss how our model relates to other episodic memory-like approaches, such as generative models for video, large language models (LLMs), and model-based RL, highlighting unique advantages and potential synergies. 9. Suggestion: "The paper would greatly benefit from the addition of a high-level schematic or conceptual figure." - Answer: We agree and have added a high-level figure that visually illustrates the connection between DNDs and the UHN framework, including parallel structures, operations, and mapping of similarity, separation, and projection functions. 10. Comment: "The discussion of biological plausibility in relation to hippocampal models is very brief." - Answer: See 3. 11. Comment: "The paragraph in line 180 is long and difficult to follow." - Answer: We have revised the paragraph, breaking it into two shorter paragraphs and rephrasing sentences for better readability. 12. Suggestions: "Include a dedicated "Limitations and Future Work" section." and "Consider potential societal impacts" - Answer: We included a "Limitations and Future Work" section and briefly consider societal impacts. We hope these revisions address your concerns and enhance the quality and impact of our paper. Thank you again for your valuable feedback. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed and informative response. The revisions you suggest address most of my concerns. I'd prefer if the paper was more complete in terms of including comprehensive RL experiments but I do understand that this topic is so interesting that it might deserve its own paper. After carefully reviewing your responses and the other reviewers' comments, I decided to increase my score to 7. Good luck!	null	null	null	null	null	null
Adaptable Logical Control for Large Language Models	Accept (poster)	Summary: This paper proposes an approach called Ctrl-G to control LLM generation, specifically, constraining LLM’s output to deterministically follow certain logical constraints, such as maintaining a certain keyword in the generated text. The approach has two main parts, the first is a Hidden Markov Model (HMM) that serves as a “prediction” model to guide the generation. The data used to train the HMM is sampled from the LLM. Essentially, the HMM is like a distillation model of the LLM that tries to capture the next-token generation “search space” of the LLM on a specific task. The second part is a series of deterministic finite automata (DFAs), each designed to represent a certain logical constraint $\alpha$. The DFAs serves as a “checker” to determine the acceptance or rejection of an output from LLM. Combining the HMM and the DFAs, the proposed approach computes the conditional probability $P_hmm(\alpha \| x_t, x_<t)$, i.e., how likely $x_t$ is leading to $\alpha$ being satisfied, to sample the next token. Strengths: The idea of using a HMM model to approximate a white-box model for exploring the LLM’s token generation space is novel. Despite that the idea of using HMM was published in a previous work [Zhang et al. Tractable control for autoregressive language generation], this paper proposes to improve by using DFAs for modeling various other task-specific logical constraints. Weaknesses: The evaluation requires more details and/or more experimental results to justify the contribution of the proposed approach. The limitation of the proposed approach is not clearly laid out and requries more discussions. Please see questions and limitations below. Technical Quality: 3 Clarity: 2 Questions for Authors: In section 4, the proposed approach is compared with FUDGE [Yang and Klein, 2021], NeuroLogic A*esque decoding [Lu et al., 2022], and ILM model [Donahue et al. 2020]. The background and motivation of comparing with these baseline methods is not clear. Why choose these models for comparison? In section 5, the proposed approach is compared with plain GPT-3.5-Turbo and GPT4-Turbo. The comparison should also be made using GPT models and the proposed DFAs method as an output checking method. In fact, there are other simpler constrained parsing techniques, for example, the method used in [Constrained Language Models Yield Few-Shot Semantic Parsers](https://aclanthology.org/2021.emnlp-main.608) (Shin et al., EMNLP 2021). In section 5, the HMM is distilled from TULU-2-7B model. Is there a specific reason for choosing this model? A 7B model seems relatively small. The llama-2 series and Google Gemma has open-sourced models with larger (> 20B) parameter sizes. Since distillation of a HMM is the key in this approach, the HMM’s training quality should be a concern. The paper mentioned that the HMM was trained on millions of samples generated from a baseline LLM (not fine-tuned on the task-specific data), but if those samples are of bad-quality, how can the quality of the obtained HMM be assured? The paper claimed a 100% constraint satisfaction achieved with the proposed approach. Following the previous three questions, it is unclear if this 100% is achieved with the HMM model or the constrained parsing with DFAs. The proposed approach would be more valuable if the author can demonstrate the contribution is mainly from the HMM. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: As mentioned in the questions, the experiment only shows the HMM being trained on a 7B model. Is it possible to generalize the approach to larger models, eventually to be used with models with trillions of parameter like GPT? For example, is the sampling amount a barrier for distillation from larger models? If so, what is the size limit here? More discussion should be made on the benefits of the proposed approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback and questions. We would like to clarify that Ctrl-G does not use DFAs to post-check the generated text; instead, the generated text is always guaranteed to satisfy the constraint. Specifically, Ctrl-G achieves constrained generation by generating from $p_{ctrlg}(x_{t} \| x_{<t}, \alpha) \propto p_{lm}(x_{t} \| x_{<t}) p_{hmm} (\alpha \| x_{\leq t})$, where the LM $p_{lm}(x_{t} \| x_{<t})$ is responsible for generating high-quality text while the HMM $p_{hmm} (\alpha \| x_{\leq t})$ is responsible for guiding the LM to satisfy the constraint $\alpha$. Here, we assume that $\alpha$ can be represented as a DFA so this marginal probability $p_{hmm} (\alpha \| x_{\leq t})$ can be computed by the algorithm proposed in Sec. 3.2. ### Choice of baselines: For the CommonGen benchmark, we chose FUDGE as a representative for the classifier-guided constrained generation methods; we chose NeuroLogic A\esque as a representative for the search-based decoding methods. We have further included NADO and GeLaTo as the baselines and we show that Ctrl-G beats them by large margins. Please refer to the global response for more details.* We have also released the outputs of Ctrl-G via an anonymous link shared to the AC. For the task of text infilling, ILM is the only available baseline that can generate infillings for multiple masks of arbitrary length; all baselines for CommonGen cannot be immediately adapted to this particular task. Nevertheless, given that Ctrl-G is an unsupervised approach, ILM, which is finetuned with full supervision on the task of text infilling, is a very strong baseline. ### Quality of the distilled HMMs: By training an HMM on examples sampled from the base LM, our goal is to effectively minimize the KL-divergence between $p_{hmm}$ and $p_{lm}$ and the actual quality of the samples should not matter. In the ideal case, we would have $p_{hmm} = p_{lm}$ and using Ctrl-G would be equivalent to sampling from $p_{lm} (x_{1:n} \| \alpha)$. Note that Ctrl-G controls the generation from the base LM by approximating $p_{lm} (x_{t} \| \alpha, x_{< t})$ but is not expected to extrapolate beyond the distribution of the base LM. Luckily, given the recent advancement of LLMs, the quality of the base LM should not be a bottleneck. ### Generalize to larger LLMs: Though there are larger open-source LLMs, we chose a 7B-parameter model due to the limit of computation. However, this is not a weakness of this work: - Note that prior controllable generation approaches like NADO, FUDGE, GeLaTo and NeuroLogic Aesque were only evaluated on LMs with < 0.7B parameters; by demonstrating the effectiveness of Ctrl-G on a 7B-parameter LLM, we have provided strong evidence for scaling it up to even larger models with potentially hundreds of Billions of parameters. - It would be more computationaly expensive to distill an HMM from larger LLMs but it would not be a bottleneck. (1) The distillation process is independent from the constraints: once the HMM is trained, it can be used to enforce whatever constraints that can be represented as DFAS. (2) According to our experience with GPT2-large and TULU2-7B, it turns out that we don't really need to sample a lot more examples from larger LLMs: 128 million tokens for GPT2-large and 320 million tokens for TULU2-7B are good enough. - We show that Ctrl-G allows an LLM with only 7B parameters* (coupled with a 2B-parameter HMM) to beat GPT3.5 and GPT4. This constitute an even stronger argument demonstrating the effectiveness of our approach, compared to showing that Ctrl-G allows a 70B-parameter LLM to beat the GPT models. We will add a more detailed discussion to the paper. --- Rebuttal Comment 1.1: Comment: Thank you for the reply. The rebuttal addresses some of the questions. I understand that Ctrl-G does not use DFAs to explicitly post-check the generated text. However, it remains unclear how Ctrl-G compares to simpler constrained decoding methods, such as those described in "Constrained Language Models Yield Few-Shot Semantic Parsers" (Shin et al., EMNLP 2021). This is especially the case when these simpler methods are combined with more powerful LLMs like GPT-4. As noted in the review, the evaluation only compares Ctrl-G to plain GPT-4 models. Given this, the contribution of the proposed approach could be questioned. The added value of achieving comparable (or even inferior) performance to GPT-4 with a simpler decoding method, while requiring significantly more training effort to develop a HMM, is not entirely clear. I agree with reviewer N2WB that the section describing DFAs is too lengthy and could be condensed. As noted in my review, the discussion on distilling an HMM seems to be the major contribution, while the DFAs provide only incremental improvements. And training details such as using the KL divergence should be briefly mentioned to give more context about the unsupervised learning. However, this also may make the paper somewhat incremental compared to the GeLaTo work. It would be beneficial if the authors could clarify the differences between GeLaTo and Ctrl-G in the methodology section. Specifically, highlight the contributions of introducing DFAs and demonstrate clearly in the evaluation results. The main concern regarding the contributions of the work has not been fully addressed, so I will maintain my current rating. --- Rebuttal 2: Title: Further Clarification Comment: Thank you for following up with our discussion. Here are some further clarifications to your questions. ---- ### Ctrl-G vs. GPT4. We would like to first clarify that Ctrl-G (with TULU2-7B and HMM-2B) actually beats GPT4 by large margins. Please refer to the pdf shared in our global response for details: the Overall section of Table 3 measures the percentange of examples where the models' outputs (1) satsifies the given constraints (keyphrase inclusion, word count range) and (2) attains an average quality score higher than 3 (out of 5) in human evaluation. Here Ctrl-G beats GPT4 by 30% - 70% in Overall Satisfaction rate in all settings. ---- ### Why do we need an HMM? Thank you for mentioning [1]. To the best of our understanding, [1] mainly leverages SCFGs to prune away the next-token that would violate the constraints. From this perspective, [1] is actually similar to Outlines [2] and guidance [3], which also use CFGs/DFAs to prune away the next-tokens. As discussed in Sec. 3.3, [1, 2, 3] is subsumed by Ctrl-G in the sense that [1, 2, 3] only decide whether $p_{hmm}(\alpha \| x_{\leq t})$ is 0 or not. [1, 2, 3] would not work for many applications because they do not have any probabilistic information. We added the following example to Sec. 3.3 to illustrate the distinction: consider the task of generating a sentence that ends with the phrase " in the park", where we compare Ctrl-G with guidance, both applied to the same model. ``` guidance: silhouette of suspected ... an heavily secured.in the park Ctrl-G: A man and a woman are walking in the park ``` Even though both generations end with " in the park", it is clear that the output from guidance is not desirable as it unnaturally appends the phrase to some irrelevant text. The reason is that guidance, by performing pure logical reasoning over the DFA, only discard the next tokens that would make the constraint unsatisfiable, while the probabilities of the other next tokens remain unchanged; in contrast, Ctrl-G performs probabilistic reasoning by estimating $p_{lm}(\alpha \| x_{\leq t})$; i.e., we estimate how likely each next token would eventually lead to $\alpha$ being satisfied. Hopefully this would answer your question why an HMM is needed. We have added [1] as a reference to Sec. 3.3. Thank you for your suggestion. ---- ### HMM training details: The goal of the HMM training is to minimize $D_{KL}(p_{lm} \|\| p_{hmm}) = E_{x_{\leq n} \sim p_{lm}} \log p_{lm}(x_{\leq n}) - E_{x_{\leq n} \sim p_{lm}} \log p_{hmm}(x_{\leq n})$, which is equivlant to maximizing the second term (log-likelihood of the data sampled from the LM) because the first term (entropy of the LM) is a constant. We will add a more detaile discussion to the methodology section. Thank you for your suggestion. ---- ### Ctrl-G significantly generalizes GeLaTo: GeLaTo proposes the idea of using a distilled HMM to approximate $p_{lm}(\alpha \| x_{\leq t})$ via $p_{hmm}(\alpha \| x_{\leq t})$, which is indeed significant, but GeLaTo is not applicable to many down-stream applications: it only derived the algorithm for computing the marginal probability $p_{hmm}(\alpha \| x_{\leq t})$ for $\alpha$ being the keyword constraints. i.e., GeLaTo could not handle the keyword exclusion, text infilling or word count control because the algorithm for computing such marginal probability in general is not known in existing literature. Compared to Ctrl-G, the main technical contribution of Ctrl-G is deriving a polynomial-time and GPU-parallelizable algorithm for computing $p_{hmm}(\alpha \| x_{\leq t})$, as long as $\alpha$ can be represented as a DFA (see Sec. 3.2). This contribution is not trivial. We will add more technical details (e.g., deriviation of Eq. 4) and show how the algorithm is tensorized in the revised paper. Besides, we have also cut down the DFA content by over a half to highlight our technical contribution here. Thank you for your suggestion. In summary, GeLaTo only supports the down-stream application of keyword-constrained generation, and Ctrl-G significantly generalizes GeLaTo by allowing this approach to be applicable to arbitrary constraints that can be represented as DFAs. We will add a more comprehensive comparison between GeLaTo and Ctrl-G to the methodology section. Thank you for your suggestion. Please let us know if you have further questions or concerns. --- [1] Shin, Richard, et al. "Constrained Language Models Yield Few-Shot Semantic Parsers." Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. [2] Brandon T Willard and Rémi Louf. Efficient guided generation for large language models. arXiv e-prints, pages arXiv–2307, 2023. [3] Scott Lundberg, Marco Ribeiro, Richard Edgar, and Harsha-Nori. Guidance: a guidance language for controlling large language models., 2024.	Summary: This paper proposes Ctrl-G to enable flexible lexically constrained generation with high accuracy and efficiency. Ctrl-G first distills an HMM from an unconditioned language model and then formulates the lexical/logical constraints with deterministic finite automata (DFAs). The inference algorithm takes both the HMM and the DFA as input and computes the desired conditional probability for guiding 48 LLM generation towards satisfying the given constraints. Experiments in keyphrase generation, text infilling, interactive text editing, and mathematical reasoning indicate that Ctrl-G achieves competitive performances for various sizes of LMs. Strengths: 1. The proposed method is flexible and manageable for constrained generation. It achieves competitive performance on various tasks, and it can scale up to control larger models with high efficiency. 2. The proposed method is theoretically sound and is clearly presented. The illustrations help in understanding the DFA formulation. Weaknesses: 1. The experimental comparison for small model tasks is relatively weak. The paper discussed several related works in the introduction and method part (e.g., GeLaTo and NADO), but the works are not included in the experiments. The evaluation metrics BLEU and ROUGE-L measure text overlaps and cannot comprehensively represent the generation quality. Qualitative results are not provided. 2. The experiment part of the paper is a little bit hard to follow. Section 5 and 6 represent key improvements, but the experimental results are displayed in the appendix (Table 4 and 5). Some sentences/sections are not finished. Technical Quality: 3 Clarity: 2 Questions for Authors: See weakness. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors discuss limitations in the experiment sections but do not discuss potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback and questions. - For the CommonGen benchmark, we did not include NADO because they only published results on the dev set and their released code is not reproducible due to lack of documentation; In the revised paper, we have added the results from NADO and GeLaTo on both dev & test sets, along with more evaluation metrics, for a direct comparison. Please see the global response for the updated results. - Regarding qualitative examples, for a quick reference, here are the outputs of Ctrl-G on the first 10 examples of the dev set of CommonGen; each line is a list of concepts followed by the output: ``` field stand look: a woman stands in a field looking at flowers kid room dance: The kids are dancing in the living room. pet couch cat: A pet cat sleeping on the couch. climb side building: A man climbs up the side of a building. talk climb wall: A man climbs a wall while talking on a cell phone snow car drive: A car drives through the snow. phone talk wear: A man wearing sunglasses talks on a cell phone. rink hockey team: hockey player on rink during team practice surfer surf ocean: surfers surf in the ocean off the coast flight stair jump: A dog jumps down a flight of stairs. ``` We have further released the outputs of Ctrl-G for all three tasks, i.e., CommonGen, Text Infilling and Interactive Text Editing. Following the instructions for authors, we uploaded them to an anonymous link and shared the link with the AC. - For the text infilling task, ILM, as a model trained with full supervision, is already a very strong baseline for Ctrl-G, which is not trained with any supervision on infilling. Furthermore, all baselines for CommonGen cannot be adapted to this task. In addition, to the best of our knowledge, the diffusion-based LMs cannot be used to fill in multiple blanks of unknown length thus cannot be adapted to this task either. - In the revised paper, we cut down the contents about DFAs by more than a half and moved all main results, pseudocode for the algorithm, and the runtime analysis from the appendix back to the main paper. We have also cleaned up the unfinished sentences and greatly improved the writing. We would be more than happy to share the revised version upon AC’s approval. - Potential social impact: as we discussed in Sec. 6, Ctrl-G could potentially be used to detoxify language models by preventing inappropriate phrases from appearing. At the same time, it is also possible that people can use Ctrl-G to make LM generate toxic contents by enforcing the occurrence of inappropriate contents. The ability to control the generation from LMs is indeed a double-edged sword and we will add a more detailed discussion on the potential social impact of our work. --- Rebuttal Comment 1.1: Comment: I appreciate the authors' detailed response. The rebuttal addresses my concerns and I will keep my rating.	Summary: It is a challenge to control LLMs' generation to adhere to logical constraints. Ctrl-G integrates LLMs with a Hidden Markov Model (HMM) and deterministic finite automata (DFAs) for flexible and tractable control, such as keyword and length constraints. Ctrl-G combined with TULU-2-7B model outperforms GPT-3.5 and GPT-4 in human evaluations for interactive text editing, achieving a 30% higher satisfaction rate and 100% constraint satisfaction. Additionally, Ctrl-G improves performance on the Grade School Math (GSM) dataset by guiding the reasoning process with logical constraints. Strengths: 1. The paper presented an interesting method on logically-controlled text generation. 2. Comprehensive experimental results and analysis are provided. The results are strong and the findings are backed by well-designed human evaluation. Runtime analysis on efficiency of the method is provided. 3. The paper is well-written and presented. Weaknesses: There are no major weaknesses. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Can the framework potentially handle highly creative tasks that might require more fluid and less structured outputs? 2. Can Ctrl-G potentially generalize to domains outside of those tested, such as scientific text generation or code synthesis? No need to provide results, but insights on this would be helpful. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Yes, the authors adequately addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback and questions. - We believe that better control in general will substantially benefit creative tasks. For example, according to the experience of our collaborators, when LLMs are asked to generate generic lyrics, they often tend to generate lyrics for love songs even when they are instructed not to do so. We can potentially avoid such a phenomenon by using Ctrl-G to prevent the phrases commonly used by love songs from appearing. - The effectiveness of Ctrl-G mainly relies on how well the assumption $p_{hmm}(\alpha \| x_{\leq t}) \approx p_{lm}(\alpha \| x_{\leq t})$ holds. Distributions over scientific writing or code could be potentially harder to approximate but there should not be a fundamental difference compared to generic writing: for the more challenging domains we can always further scale up the HMMs.	Summary: The paper tackles the problem of generating sequences from llms while following a constraint \alpha, providing a solution for the case where \alpha can be represented as a regular language (e.g., a constraint on containing certain substrings). The method allows application of any regular-language constraint, and is somewhat similar to/inspired by a previously proposed method, Gelato, in which the LLM is distilled into an HMM. The method is evaluated with domain-finetuned TULU-7B language models, and compared to some other constrained sequence generation methods, FUDGE and Neurologic A* esque decoding: these results are favourable. It is also compared to baselines such as directly using very large language models (e.g. GPT4), with the constraint given simply as part of the input. These results are unfortunately missing (the appendix is missing), but apparently GPT4 struggles to follow constraints (whereas ctrl-G always follows the constraints, by design). It is not clear to me whether the method was compared to Gelato. A small investigation is done to see whether constrained decoding can help when solving math problems, specifically by forcing the solution to contain all numbers from the question, and the method is deployed on the GSM benchmark. The performance is improved relative to not using the method, but not compared to other methods. Strengths: 1. A method for constrained generation that always follows the constraints, by construction 2. The method possibly allows richer constraints than a previous one? Unfortunately I am not entirely sure about this, it is not clearly enough stated in the work 3. Comparison with some previous methods Weaknesses: 1. Weak evaluation (see questions/comments), in particular no comparison to gelato (?), no inference time comparison ( I get the impression that inference with this method is very slow), key results put in appendix, GSM experiment more a demonstration of usefulness of constrained decoding than of ctrl-G in particular (no comparison to other other constrained decoders) 2. Some hard to follow details and overview. Missing in particular: 1) a complete description of how inference is done in this framework, and 2) more detailed description of differences and similarities between this method and Gelato (both use an HMM distilled from the LLM, but Gelato then uses only that while this method still uses the LLM as well - why does this method have both an LLM and its distilled HMM - I will appreciate more intuition on what is happening here). Also, it is not clear how/whether this method differs from gelato with respect to inference complexity (time) or potential constraint space (i.e. type of constraints that can be encoded) 3. Confusing amount of attention given to problem of expressing "regular" constraints (e.g., contains/doesnt contain substring, has maximum length 3) as DFAs, I would expect this is trivial, but somehow the authors treat it as not, referencing multiple algorithms and dedicating several paragraphs and figures (e.g. 3: example of a DFA, 4: operations on DFAs) to it. 2 figures and over 40 lines (127-135, 176-196, 222-235) dedicated to question of creating DFAs for rather simple constraints, all would be better spent on describing main alg and bringing in main results (which are currently in missing appendix). Technical Quality: 2 Clarity: 1 Questions for Authors: questions/comments: 0. several references made to materials in appendix, but no appendix! even if it were there, main materials should not be in appendix - pseudo code and greater description of ctrl g would be much more appreciated in main, as well as several results referred to from main text! 1. figure 2 never referenced in main text 2. lines 38-39 unfinished/mangled sentence 3. line 39-40: possibly unclear sentence? it reads like gelato provides multiple algorithms in general, but for the case of keyword constraints it only provides one, in which case it is unclear what this alg is, what the situation is for other types of constraints, and whether ctrl-G covers the same or a larger set of constraints as can be represented with gelato. generally, the comparison between ctrl-g and gelato is not clear enough on the similarities and differences and in fact on what gelato is exactly. end of this sentence (in line 41) also unclear. 4. lines 87-91: it seems (also from lines 39-41) that the paper faults Gelato for only providing one algorithm for handling certain constraints. but doesn't this paper also have only one algorithm? this criticism doesnt make sense 5. lines 87-91: gelato is critiqued for possibly not being efficiently computable, but my impression is that this is true for ctrl-g too (with complexity nmh^2, where h is the distilled HMM size and reaches 32k even in this paper). can you compare the two explicitly? 6. lines 87-91: the claim that it is unclear whether gelato would scale - why is that? is it because of the distillation to an hmm, and an assumption that this will become a worse approximation as the original model grows larger? if so, does that drawback not hold also for this method? 7. lines 99-111: the description says a DFA is used to encode the constraint, but the DFA does not seem to appear in the equations, and it is not clear how it factors in 8. line 121 typo: "fo" 9. line 125 consider making it clearer that q0 here is "the" q0 10. 127-135 this feels overly mystical for the problem of designing a dfa to express simple constraints. Unless I am missing something (in which case I would appreciate explanation), I think the whole discussion of designing DFAs for the constraints does not need more details than: "this method allows for any constraint that can be expressed with a DFA", with maybe a reference to textbooks on dfas if someone is not experienced. similarly figs 3 and 4 are not necessary. 11. line 138 is this DFA only partial? if not, surely m=k * \|\Sigma\| - why not just write that? 12. 146 what subscript? 13. 146-151 here i could do with being walked through what is happening in simpler terms/more detail/justifications/something - this bit is harder for me 14. thrm 3.2 - give some justificution. did you compute this? (if so, prove). is it taken from a paper? if so, refer. 15. line 166: "we set": i get what youre saying but rephrase because you cant set this distribution as you please. maybe you mean used 0/1 or only evaluated whether or not phmm(a\|..)>0 16. section 4 again bewildering amount of space dedicated to design of dfas for ultimately very simple constraints. line 170, and then lines 176-196 (also, 186-189 seem to be a definition not a theorem), all for this simple task. Also, if must refer to various algorithms (e.g. Aho-Corasick, line 182), then need to also provide relevant sources (ie cite). 17. lines 202-204 im a bit confused: if each epoch has its own samples, what separates epochs? i dont understand. i see similar phrasing used in gelato but i dont follow this. are you changing anything between epochs? if not, this sounds like 1 epoch over 4m samples? if yes, explain what is being changed 18. line 226 "inflection of keywords" - huh? 19. 222-235 (and 235-241 kind of): more space spent on needless discussion of dfa operations. i respect that maybe sometimes these operations are computationally expensive but it seems all constraints considered in this paper do not bump against this anyway. it would make more sense to mention this only if it somehow limited the experiments or practical applications of the method. 20. line 231: "dead states" not been defined 21. line 259: seems like a fairer comparison against ILM would have it trained with the right percentage of masked out tokens each time, or even a varying percentage if want to see how it performs when only allowed one model 22. line 300 "word range" - maybe you mean sentence length range? word count range? 23. line 331: saying ctrl-G is a 7B param model: you are ignoring the parameter count of the distilled HMM, which seems to me quite large (32k states!). What is the vocabulary size of Tulu2? If I assume 50k (i.e. similar to GPT2), that's 32 * 50M = 1.6B parameters for the emission matrix, and another 32 * 32 M = ~1B parameters for the transition matrix. 24. line 339-341 hard to understand what being compared to what, rephrase 25. 350-352: everything following "we can" - it is possible, but this is speculation, and not backed by the results, so be more careful in presentation. Confidence: 3 Soundness: 2 Presentation: 1 Contribution: 3 Limitations: weak eval - missing comparisons and results Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed feedback. Following some prior years’ tradition, our appendix was mistakenly uploaded as the supplementary material and we apologize for the confusion. We have greatly improved the presentation of the paper and would love to share the revision upon AC’s approval. - We cut down the DFA content by more than a half and we will further move Fig. 3 & 4 to the appendix as you suggested. - The pseudocode, the main experiment results and the runtime analysis is moved from the appendix to the main text. Please see the global response for the main results. ### Efficiency of Ctrl-G: Despite the time complexity result $O(nmh^2)$ from Theorem 3.2., Ctrl-G is actually quite fast due to its GPU-based implementation, where the $mh^2$ part is fully tensorized. For example, on the task of interactive text editing, Ctrl-G, when applied to the TULU2-7B model, is able to generate infillings under multiple constraints within a few seconds, enabling a user interface for real-time interaction. On the CommonGen benchmark, in the unsupervised setting, Ctrl-G not only generates outputs of higher quality but also runs 30 - 100 times faster than NeuroLogic A\esque and 6 - 16 times faster than GeLaTo. Please refer to the global response for details.* ### Comparison between GeLaTo and Ctrl-G: GeLaTo is really a special case of Ctrl-G. GeLaTo is much more limited in the sense that it only supports the keyword constraint, that is, no support for text infilling, word count control or arbitrary DFAs. Both GeLaTo and Ctrl-G follow the same formulation for controllable autoregressive generation: they sample from $p_{ctrlg}(x_{t} \| x_{<t}, \alpha) \propto p_{lm}(x_{t} \| x_{<t}) p_{hmm} (\alpha \| x_{\leq t})$, where $p_{hmm} (\alpha \| x_{\leq t})$ approximates $p_{lm} (\alpha \| x_{\leq t})$; specifically, both GeLaTo and Ctrl-G use the base LM and the distilled HMM. The intuition is that $p_{lm}(x_{t} \| x_{<t})$ is responsible generating fluent text while $p_{hmm}(\alpha \| x_{\leq t})$ is responsible for guiding the LM to satisfy the constraint $\alpha$. We revised line 87 - 93 as: > - To control LM generation with an HMM, we need to compute $p_{hmm}(\alpha \| x_{\leq t})$. GeLaTo only shows how to compute it for $\alpha$ being the keyword constraint. A polynomial-time algorithm for computing this probability for logical constraints in general is not known. > - GeLaTo assumes that $p_{hmm}(\alpha \| x_{\leq t}) \approx p_{lm}(\alpha \| x_{\leq t})$ and have verified its effectiveness on GPT2-large. It remains to be explored whether this assumption would hold for the more recent LLMs (e.g. Llama2) with over 10 times more parameters. > (1) In Ctrl-G, we propose a polynomial-time algorithm that computes $p_{hmm}(\alpha \| x_{\leq t})$ for any $\alpha$ that can be represented as a DFA. (2) In addition to LMs on the scale of GPT2-large, we further verify the effectiveness of Ctrl-G in controlling LLMs with 7B parameters, demonstrating its scalability to even larger models. ### Derivation of Eq. (4) and Theorem 3.2: The derivation of Eq. (3) and (4) relies on the Markov property of HMMs and DFAs, and the fact that s_{t} is determined by x_{\leq t}. We derived Theorem 3.2 and here is a sketch analysis: > The computation cost of Ctrl-G is dominated by the computation of $p(S_{n} \in F \| z_{t}, s_{t})$ for all $t$, $z_t$ and $s_t$ following Eq. (4). Since $\sum_{x_{t+1} \in \text{edge}(s_{t}, s_{t+1})} p(x_{t+1} \| z_{t+1})$ does not depend on t, we can precompute and cache it, resulting a cost of $O(mh\|\Sigma\|)$. > Then, note that for $s_{t}$, we only need to consider the $s_{t+1}$ where $\text{edge}(s_t, s_{t+1}) \neq \emptyset$. Hence, fixing $t$ and $z_t$, when we compute $p(S_{n} \in F \| z_{t}, s_{t})$ for all $1 \leq s_t \leq k$, we only need to (1) enumerate through $1 \leq z_{t+1} \leq h$ and (2) for each $z_{t+1}$, we only need to visit each edge exactly once; there are $m$ edges in total, so it follows that cost is $O(n\cdot h\cdot h \cdot m) = O(nmh^2)$. > The total time complexity is $O(nmh^2 + mh\|\Sigma\|)$, which simplifies to $O(nmh^2)$ given that $\|\Sigma\| < nh$ in practice. ### Comparison with ILM: For the task of text infilling, the base LM used by Ctrl-G is not finetuned with any supervision on this task; it is only finetuned to adapt to the style of ROC stories. In contrast, the ILM baseline is explicitly trained with full supervision on text infilling. Hence, it is remarkable for Ctrl-G to match the performance of ILM on the training distribution of ILM (13% ratio). The evaluation on other masking ratios is meant to show that Ctrl-G is able to generalize well to different distributions. ### Answers to Other Questions: Line 99 - 111. We assume that $\alpha$ can be represented as a DFA so that $p_{hmm} (\alpha \| x_{\leq t})$ can be computed by the algorithm proposed in Sec. 3.2. Line 138. We merge all transitions from one state to the other into one edge and the edges of a DFA denote the pair of states $(s_1, s_2)$ connected by transitions. Definition added to Sec. 3.1. Line 146. Changed to “we omit the subscript “hmm” for simplicity.” Line 166. Changed to “the other approaches are subsumed by Ctrl-G in the sense that they only evaluate whether $p_{hmm}(\alpha \| x_{\leq t})$ is 0 or not” Line 202-204. “one epoch” is actually “one step of EM parameter update;” we will rephrase to avoid confusion. Line 226. “Inflections of keywords” refers to CommonGen where “swims” “swimming” are inflections of “swim”; for the task of text infilling we don’t need to handle such cases. We will remove this reference. Line 231. Dead states now defined in Sec. 3.1. Line 300. Changed to “word count range” Line 331. The vocab size of TULU2-7B is 32K; changed to “we show that Ctrl-G, where a 7B-parameter model is coupled with a 2B-parameter HMM, …” Line 350 - 352. “we can” changed to “In future work, it is possible to” --- Rebuttal Comment 1.1: Title: thank you Comment: thank you for your comments and for sharing some missing results. i am surprised by how much faster ctrl-g is than gelato, this was not at all clear to me from the original manuscript! i raise my score. please do make all the changes needed for clarity as mentioned in my review. regarding latency, please clarify why ctrl-g can use the gpu but gelato cannot (if indeed it cannot. if it can, and just hasn't been implemented to take advantage of the gpu, please share this fact).	Rebuttal 1: Rebuttal: Thank you all for your feedback. Please refer to the attached PDF for some of our main evaluation results as well as an additional runtime comparison: (1) Evaluation results on CommonGen (dev & test) for FUDGE, NADO, NeuroLogic A\esque, GeLaTo and Ctrl-G. (2) Runtime analysis on CommonGen (dev) for NeuroLogic A\esque, GeLaTo and Ctrl-G. (3) Evaluation results on Interactive Text Editing for TULU2-7B, GPT3.5, GPT4 and Ctrl-G. (4) Runtime analysis on Interactive Text Editing for TULU2-7B and Ctrl-G. Upon your request, we have also released the output of Ctrl-G via an anonymous link. The link contains the output files as well as an html-based data visualization tool for you to browse the outputs. The link is sent to the AC following the instrutions for authors. Pdf: /pdf/2b4c679609feb4f5a7216db3742916c456981724.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment	Accept (poster)	Summary: This paper proposed Cal-DPO, a variation of DPO to address the issue of decreasing rewards of chosen answers. In addition to DPO loss, Cal-DPO add a pair of calibration terms, which aim to match the rewards induced by language model with some absolute ground truth reward value. Theoretical analysis shows that, during the training process of Cal-DPO, the likelihood of chosen responses will likely to increase and the likelihood of rejected responses will likely to decrease. The authors also prove that DPO loss can be upper bounded by Cal-DPO loss. Experiment on various benchmarks shows that Cal-DPO outperforms DPO and other baselines. Strengths: The strengths of the paper are listed below: 1. The decrease in chosen answers' reward is a notable issue in DPO. In this paper, the authors focus on this issue and proposed Cal-DPO. This observation and corresponding mitigation is interesting and deserves attention. 2. The paper provides theoretical analysis of Cal-DPO. The analysis looks right to me. 3. The paper is well developed. Motivations, method and analysis are clearly presented. Weaknesses: This paper does not have specific weakness. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Have the authors conducted experiment with models other than zephyr-7b-sft-full since zephyr-7b-sft-full demonstrate a relatively inferior performance on OpenLLM Leaderboard (which includes all reasoning benchmarks considered in the paper). Also, have the author conducted experiment on the other two benchmarks in OpenLLM Leaderboard (MMLU, TruthfulQA)? Confidence: 2 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The limitation is adequately clarified by the authors Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer yAWB, we appreciate your great summarization and recognition of our contributions and your positive comments on our work: "interesting," "solid theoretical analysis," and "motivations, method and analysis are clearly presented ." Please find our responses to your comments below: --- Q1. Have the authors conducted experiments with models other than zephyr-7b-sft-full since zephyr-7b-sft-full demonstrate a relatively inferior performance on OpenLLM Leaderboard. Also, has the author conducted experiments on the other two benchmarks in OpenLLM Leaderboard (MMLU, TruthfulQA)? A1. We greatly appreciate the reviewer's insightful suggestion and question. We would like to kindly remind the reviewers that Pythia-2.8b and Zephyr-7b-SFT in our paper are two of the most widely used LLMs for alignment. We believe our experiments on these LLMs can corroborate the effectiveness of Cal-DPO. Nevertheless, we completely agree with the reviewer that including an empirical comparison on better LLMs and more benchmarks would be beneficial. Thus, we ran additional experiments on the recent Llama-3-8B across more benchmarks, including MMLU and TruthfulQA. For all baselines, we conducted a hyperparameter search over a range of $\beta \in [0.001, 0.01, 0.1, 0.5, 1.0]$. The following results show that Cal-DPO significantly outperforms the baselines. Together with the results in our paper, these findings further validate the effectiveness of Cal-DPO. We will include these results in the final version. \| Llama-3-8B-Instruct \| MMLU \| ARC \| TruthfulQA \| Winograde \| GSM8K \| \| --- \| --- \| --- \| --- \| --- \|--- \| \| IPO \| 64.40 \| 62.88 \|   54.20 \| 72.22 \| 22.67 \| \| KTO \| 64.42 \| 63.14 \|   55.76 \| 76.09 \| 38.97 \| \| R-DPO \| 64.19 \| 64.59 \|   53.41 \| 75.93 \| 39.27 \| \| DPO \| 64.31 \| 64.42 \|   53.48 \| 76.32 \| 38.67 \| \| Cal-DPO \| 64.92 \| 65.58 \|   59.34 \| 77.53 \| 47.59 \| --- As the reviewer also noticed, Cal-DPO presents a simple yet effective approach to align LLMs, offering interesting findings and theoretical contributions for future work. As the reviewers' comments are not fatal to the major contributions of our manuscript, and involving the above results in our paper does not lead to a major revision. We would sincerely appreciate it if you would consider raising your score in light of our response. Thank you again for your time. --- Rebuttal 2: Title: Dear NeurIPS Reviewer yAWB, discussion period is ending soon Comment: Dear NeurIPS Reviewer yAWB, We gratefully appreciate your time in reviewing our paper and your comments. We have made extensive efforts to address your comments and believe that they adequately address all your concerns. The reviewer's comments are mainly about some clarifications and are not fatal to the contributions of our manuscript; we believe that the reviewer's insightful comments can be easily and effectively addressed in the final version. We would like to confirm whether there are any other clarifications and would be grateful if the reviewer could increase the score. Many thanks for your time; we are extremely grateful.	Summary: This paper proposes a simple yet effective change to the DPO objective that acts as a regularizer on the implicit reward, $\beta \log \frac{\pi(y \mid x)}{\pi_\mathrm{ref}(y \mid x)} + \beta \log Z(x)$, that is maximized by the LM. Specifically, the implicit reward is encouraged to be appropriately scaled relative to the desired reward $r(x, y)$. The main change is to add an additional squared loss term to the loss function that encourages the implicit reward for positive examples to concentrate around $\frac{1}{2 \beta}$ and the implicit reward for negative examples to concentrate around $\frac{-1}{2 \beta}$. The paper also includes some theoretical analysis, and positive empirical results. Strengths: The proposed change is nice in that it is simple to implement, and shows favorable empirical results. The motivation and presentation are a bit unclear to me (see next section), however, for the most part it seems like a practical, sound result that would be of interest to practitioners in RLHF. Weaknesses: My main difficulty is in the presentation of the motivation behind the method. In particular, I don't think that "calibration" is the right term to be using here --- it's a bit of a misnomer as it doesn't seem to have much to do with the standard way calibration is referred to (e.g., one might say that the BT model is calibrated if it produces calibrated preference probabilities). See also questions below. I also think that the theoretical results can benefit from being stated a bit more precisely---in particular the comment following Theorem 2 (i.e., "Theorem 2 also implies that DPO, RLHF, and Cal-DPO asymptotically converge to the same global optimal policy ....") should be stated clearly and proven. There are also a number of other grammatical / writing errors (line 197, line 355, etc) that can be cleaned up. From the empirical side of things, I think that it would be good to also compare to IPO, as the style of loss is quite similar (e.g., this loss also enforces the margin between the implicit reward of positive / negative examples to be a constant, though without constant differences). It would be good to show the effect. Technical Quality: 3 Clarity: 2 Questions for Authors: If I understand the main motivation correctly, Eq. (9) is essentially saying that the log partition term for the optimal policy should be 0 in this "calibrated" version? Or, in other words, one is simply defining a new, translated reward via the equality: $$\pi^(y \mid x) = \frac{1}{Z(x)} \pi_\mathrm{ref}(y \mid x) \exp( \frac{1}{\beta} r(x, y)) \Rightarrow \pi^(y \mid x) = \pi_\mathrm{ref}(y \mid x) \exp( \frac{1}{\beta} r'(x, y)),$$ where $r'(x, y) = r(x, y) - \beta \log Z(x)$. This can then be arranged in the usual way to get $\log \frac{\pi^(y \mid x)}{\pi_\mathrm{ref}(y \mid x)} = r'(x, y) / \beta$. Setting $r'(x, y) = \pm \frac{1}{2}$ is then equivalent to adding $\beta \log Z(x)$ to the original reward (assuming it was $\pm \frac{1}{2}$), which still preserves $\pi^(y \mid x)$, as that term does not depend on $y$ . However, enforcing this equality (instead of just the difference equality), then removes the underdetermined aspect of the BT model, which I can imagine is helpful (in addition to the theoretical properties w.r.t. gradient dynamics analyzed). Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer f8MD, we appreciate the reviewer's perception of our contributions to both empirical and theoretical analysis, and we thank the reviewer for their insightful questions. Please find our detailed responses below: --- Q1. My main difficulty is in the presentation of the motivation behind the method. In particular, I don't think that "calibration" is the right term to be using here. A1. Thank you for your suggestions. We agree with the reviewer that "calibration" typically indicates that the model produces calibrated probabilities to accurately reflect the uncertainty, i.e., “uncertainty calibration”, where “calibration” is performed to produce confidence intervals. We want to highlight that our scale calibration technique aims to align the learned implicit reward produced by the LLMs with some external scale (scale-calibrated), for example, the absolute ground truth reward. Readers should not confuse it with uncertainty calibration, where “calibration” is performed to produce confidence intervals. Actually, this "scale calibration" is similarly used in many works in reinforcement learning [1] and learning to rank [2,3]. We apologize for any confusion and will clarify this more explicitly in the revision. --- Q2. I also think that the theoretical results can benefit from being stated a bit more precisely---in particular the comment following Theorem 2 (i.e., "Theorem 2 also implies that DPO, RLHF, and Cal-DPO asymptotically converge to the same global optimal policy ....") should be stated clearly and proven. A2. Thank you for your valuable suggestion! We apologize for your confusion. Theorem 2 essentially shows that minimizing our proposed Cal-DPO is equivalent to minimizing the original RLHF objective (reverse KL divergence) as Cal-DPO theoretically serves as its upper bound. Theorem 2 proves that our Cal-DPO and RLHF encourage mode-seeking behavior while DPO is mode-covering because it can be shown to minimize the forward KL-divergence as shown by [4]. Sorry for the confusion. We will clarify and refine this sentence. --- Q3. There are also a number of other grammatical / writing errors (line 197, line 355, etc). A3. Thanks for pointing out these typos! We have corrected them. --- Q4. From the empirical side of things, I think that it would be good to also compare to IPO, as the style of loss is quite similar. A4. We greatly appreciate the reviewer's insightful suggestion. We believe there are some misunderstandings. - Actually, we have already conducted a comparison with the IPO. The reviewer can find the results in Figure 3 (right) in our paper. It shows that our calibrated objective exhibits clear advantages over vanilla IPO. - In addition, our proposed calibration objective function fundamentally differs from that used in IPO. Specifically, IPO enforces a fixed constant margin between the implicit rewards associated with the chosen and rejected responses. However, it does not guarantee that the estimated reward of the chosen response will increase and that of the rejected response will decrease. In contrast, it can be seen that our calibration loss not only pushes the gap to be constant, but also attempts to push the rewards of chosen responses to positive values and those of rejected responses to negative values, effectively and theoretically preventing the rewards of the chosen responses from decreasing. --- Q5. If I understand the main motivation correctly, Eq. (9) is essentially saying that the log partition term for the optimal policy should be 0 in this "calibrated" version? Or, in other words, one is simply defining a new, translated reward via the equality... A5. Thanks for your insightful comments! We would like to clarify the following point and will highlight the following discussions in the revision. - The reviewer's comment is astute. This is another interesting perspective to understand our objective. By setting the rewards for chosen and rejected actions to 1/2 and -1/2 respectively, we are indeed making an implicit assumption about the normalizing partition function being 1, meaning the log partition term is zero. This essentially assumes that the learned optimal policy is self-normalized. This is a reasonable and widely accepted assumption [5,6,7], as the model is typically rich enough to incorporate the dependent partition function, making it approximately self-normalized [7]. - Furthermore, as demonstrated in our Theorem 2, our Cal-DPO with Eq. (9) is equivalent to minimizing the original RLHF objective, as it theoretically serves as an upper bound for the RLHF objective. Importantly, Theorem 2 holds true regardless of the specific rewards assigned to the chosen and rejected actions, although we simply set the rewards for chosen and rejected actions to 1/2 and -1/2 in practice. --- We appreciate the efforts from the reviewer and also sincerely hope our posted responses can address your questions. We also believe your comments can also be easily addressed in the revision. In light of these responses, we sincerely hope you could consider increasing your score. Please feel free to let us know if there are any remaining questions. Thank you for your efforts! [1] Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning. NeurIPS 2023 [2] Scale Calibration of Deep Ranking Models. KDD 2022 [3]Regression Compatible Listwise Objectives for Calibrated Ranking with Binary Relevance. CIKM 2023 [4] Towards Efficient Exact Optimization of Language Model Alignment. ICML 2024 [5] Residual Energy-Based Models for Text Generation. ICLR 2020 [6] Noise-Contrastive Estimation: A New Estimation Principle for Unnormalized Statistical Models. AISTATS 2010 [7] Noise Contrastive Estimation and Negative Sampling for Conditional Models: Consistency and Statistical Efficiency. EMNLP 2018 --- Rebuttal 2: Title: Dear NeurIPS Reviewer f8MD, discussion period is ending soon Comment: Dear NeurIPS Reviewer f8MD, We gratefully appreciate your time in reviewing our paper and your comments. We have made extensive efforts to address your comments and believe that they adequately address all your concerns. The reviewer's comments are mainly about some clarifications and are not fatal to the contributions of our manuscript; we believe that the reviewer's insightful comments can be easily and effectively addressed in the final version. We would like to confirm whether there are any other clarifications and would be grateful if the reviewer could increase the score. Many thanks for your time; we are extremely grateful. --- Rebuttal Comment 2.1: Comment: Thanks for the thorough response to my comments. I have raised my score from 6 to 7 as I do believe my concerns can be addressed as discussed in another revision. --- Reply to Comment 2.1.1: Title: Dear NeurIPS Reviewer f8MD, thank you very much for reading our rebuttal. Comment: Dear NeurIPS Reviewer f8MD, Thank you very much for reviewing our paper and reading our rebuttal. We sincerely appreciate your recognition of our contribution! We are truly grateful for your time and your reply.	Summary: The authors propose Cal-DPO, a preference-tuning algorithm that modifies the DPO loss by adding two MSE terms that aim to "calibrate" the log-likelihood ratios of y_w and y_l to their respective reward values. They claim that this is advantageous because the DPO loss only maximizes the reward ratio, and does not constrain the respective reward values. They also claim that Cal-DPO exhibits a "negative gradient" (it pushes down the likelihood of undesirable outputs) and mode-seeking behavior. They evaluate their methods by training the Zephyr 7B model on Cal-DPO and DPO, and comparing the results on a few different benchmarks. Cal-DPO generally seems to outperform DPO and other related variants. Strengths: - Cal-DPO has some solid motivations -- it is true that DPO does not constrain the reward margin or the absolute scale of the rewards, which often leads to issues with over-optimization. IPO was similarly motivated in that it explicitly constrains the reward margin. - The concept of the "negative gradient" is a well-established motivation for designing training objectives that decrease the likelihood of y_l. - Cal-DPO is easy to implement and can be straightforwardly combined with other training objectives. Weaknesses: - Unfair comparisons to other methods: The authors show in Figure 4 that the best $\beta$ value for Cal-DPO is 0.001, and they choose to use this same value for their comparisons to all other methods (described in Appendix B.1, where they state that "all the hyperparameters are set to be the same"). However, this implies that hyperparameter tuning was conducted for Cal-DPO, but not for any of the other methods. In fact, past work indicates that the optimal $\beta$ value can vary widely depending on the combination of model and algorithm, in some cases reaching as high as 0.6 (see https://huggingface.co/blog/pref-tuning). The difference in accuracy/performance can be vast, depending on the hyperparameter configuration. In general, the fairer way to compare methods is to conduct a hyperparameter search separately for every method, rather than to fix a set of hyperparameters. This is especially important in this case, since the $\beta$ was already selected to be the optimal one for Cal-DPO. Without this procedure, it is unclear whether Cal-DPO truly outperforms the other methods, or if the other methods were sub-optimally trained. - The authors train the model on a set of training datasets covering safety behaviors, summarization, and sentiment detection, but the evaluation benchmarks in Table 2 primarily cover other types of capabilities, such as math and abstract reasoning. These are not preference learning tasks, as they have exact answers, rather than subjective ratings of human preference. Even if a couple other papers have done this evaluation, this does not make as much sense as evaluating on more related benchmarks, such as MT-Bench. The Cal-DPO accuracy numbers in this table are also quite close to the other methods' -- and given the hyperparameter tuning issues, it is unclear how much of this difference is noise or is meaningful. - The authors predicate much of their rationale on the claim that other techniques like DPO reduce the likelihood of the chosen response. However, since DPO also reduces the likelihood of the rejected response, it is unclear whether this effect is actually undesirable -- to answer this question, one would have to know where the probability mass gets moved to instead. Since we know that DPO does in fact improve win rate, it seems likely that there is probability mass moved to more preferred outputs. The authors also do not cite any evidence that this is necessarily undesirable. - Additionally, not all y_w's in preference datasets are high-quality -- it is often the case that both y_w and y_l are low-quality, but y_w is slightly less low quality than y_l. In this case, it is not necessarily desirable for the model to increase the probability mass on y_w. - Similarly, the authors state that "an undesirable consequence of this behavior is that the learned policy increases the likelihood of unknown out-of-distribution responses," but this is also unfounded. If there is past work or experimental evidence that supports this claim, it would be helpful to cite it here. - The authors also claim that mode-seeking behavior is more desirable here -- however, one can also see cases where this is undesirable. Since mode-seeking encourages more probability mass to be placed on high-reward outputs, and DPO/Cal-DPO rewards are based off the current model's likelihood, this promotes a positive feedback cycle where the algorithm continuously places more mass on outputs that are already high-likelihood under the current model. This may cause fast reward over-optimization, which has been a frequently observed issue in RLHF-based algorithms. - The experiments are only conducted on the Zephyr 7B model, and it would be much more convincing if the results reproduced to a couple other sizes and types of LLMs as well. - The paper also does not seem to list the generation parameters used for any of the evaluation experiments. Minor nits: - The stated IPO objective in Eq. 18 is incorrect -- the constant term should be $\frac{1}{2\tau}$, not $\frac{1}{2}$. Also, there is no $\beta$ coefficient. Alternatively, one could rename $\tau$ as $\beta$ instead and write $L_{IPO}=(h_{\theta}(x,y_w,y_l)-\frac{1}{2\beta})^2$. - What are the "Value"s on the y-axis in the left two plots of Figure 3? Technical Quality: 1 Clarity: 3 Questions for Authors: - Do the authors have results where the baseline methods have also been hyperparameter tuned? Confidence: 4 Soundness: 1 Presentation: 3 Contribution: 2 Limitations: The stated limitations about Cal-DPO being limited to offline methods is reasonable. An additional limitation is that the experiments have only been conducted on the Zephyr 7B model, and evaluated for preference learning on only a couple of small datasets (e.g. TL;DR summarization and Anthropic HH). Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer s9Bq, we appreciate the reviewer's perception of our contributions to both empirical and theoretical analysis. We believe that there are some important misunderstandings --- Q1. Unfair comparisons: The authors show in Figure 4 that the best $\beta$ value for Cal-DPO is 0.001, and they choose to use this same value for their comparisons to all other methods. However, this implies that hyperparameter tuning was conducted for Cal-DPO, but not for any of the other methods. A1. We apologize for your confusion and we believe there are indeed some misunderstandings. Our comparisons to other methods are indeed fair. Please see our detailed responses in Global Response 1, where we provide important clarification and results of the hyperparameter tuning, respectively. --- Q2. The authors train the model on a set of training datasets covering safety behaviors, summarization, and sentiment detection, but the evaluation benchmarks in Table 2 primarily cover other types of capabilities, such as math and abstract reasoning. These are not preference learning tasks, as they have exact answers, rather than subjective ratings of human preference. A2. We appreciate your suggestions and believe there are some misunderstandings. In addition to the reasoning benchmarks in Table 2, we have tested AlpacaEval 2, as shown in Figure 2, which is the instruction-following benchmark for preference learning similar to MT-Bench [1]. Thus, we believe our experiments on safety behaviors, summarization, sentiment detection, and instruction-following benchmark AlpacaEval 2 can corroborate the effectiveness of Cal-DPO for preference learning. Moreover, following your suggestions, we also tested Cal-DPO on MT-Bench and obtained a better average MT-Bench score of 7.4, while DPO achieved 7.3. MT-Bench exhibits poor separability across different methods, as also shown in recent work [1]. --- Q3. The authors predicate much of their rationale on the claim that other techniques like DPO reduce the likelihood of the chosen response. However, since DPO also reduces the likelihood of the rejected response, it is unclear whether this effect is actually undesirable. A3. Thank you for your comments. Please see our detailed responses in Global Response 2. --- Q4. The authors also claim that mode-seeking behavior is more desirable here -- however, one can also see cases where this is undesirable. This may cause fast reward over-optimization. A4. Thank you for your comments. We believe there are some misunderstandings. - Our work theoretically demonstrates that Cal-DPO also encourages mode-seeking behavior. Recent studies [2,3] have empercially and theorically already shown to be more desirable for alignment as it can sharpen the probability mass on certain high-reward regions, thereby leading to aggressive reorganization of probability mass [2,3]. - Additionally, Cal-DPO does not lead to rapid reward over-optimization. As shown in concurrent work [3], standard DPO exhibits nonlinear over-optimization dynamics due to decreasing chosen likelihoods. In contrast, Cal-DPO maintains increasing and positive chosen likelihoods, thus actually avoiding the rapid reward over-optimization seen with standard DPO. --- Q5. The experiments are only conducted on the Zephyr 7B model, and it would be much more convincing if the results reproduced to a couple other sizes and types of LLMs as well. A5. Thanks for your suggestion. Please see our responses in Response to Reviewer yAWB, where we provide additional results on LLama 8B. --- Q6. The paper also does not seem to list the generation parameters used for any of the evaluation experiments. A6. Thank you for your comments. For the benchmarks in the LLM HuggingFace leaderboard, we use the default greedy decoding for all settings and methods in Table 2. Except for the HuggingFace leaderboard, we use a sampling decoding strategy to generate responses, with a temperature of 0.7 following zephyr-7b-beta. --- Q7. The stated IPO objective in Eq. 18 is incorrect. A7. Thank you for your comments. We believe this is a misunderstanding. The stated IPO objective in Eq. (18) in the paper i.e, $\mathcal{L}=(\beta h_\theta\left(\mathbf{x}, \mathbf{y}_{w}, \mathbf{y}_l \right)-\frac{1}{2})^{2}$ has also been used in [4] and is correct and mathematically equivalent to the original IPO objective $\mathcal{L}=(h_\theta\left(\mathbf{x}, \mathbf{y}_{w}, \mathbf{y}_l \right)-\frac{1}{2\beta})^{2}$ by multiplying by a constant $\beta^{2}$. We apologize for your confusion and will use the IPO original form as you suggested. --- Q8. What are the "Value"s on the y-axis in the left two plots of Figure 3? A8. "Value" refers to the calculated rewards of chosen and rejected responses, as well as their difference (margin). --- Q9. Do the authors have results where the baseline methods have also been hyperparameter tuned? A9 Thank you for your question. Please refer to our detailed responses in Global Response 1, where we provide further clarification and results of the hyperparameter tuning, respectively. --- In light of these responses, we hope we have addressed your misunderstanding and sincerely hope you consider raising your score. As noticed by the reviewer, our work presents some interesting empirical and theoretical findings and a simple yet effective framework. Your key concerns are indeed misunderstandings and are not fatal to our contributions to our manuscript. We truly appreciate your time spent reviewing our paper. [1] SimPO: Simple Preference Optimization with a Reference-Free Reward. Arxiv 2024. [2] Towards Efficient Exact Optimization of Language Model Alignment. ICML 2024 [3] Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms. Arxiv 2024. [4] Provably Robust DPO: Aligning Language Models with Noisy Feedback. ICML 2024 --- Rebuttal 2: Title: Sincere Comments and Clarifications for Reviewer s9Bq: Our comparisons are indeed fair. Comment: We appreciate the reviewer's perception of our novelty and thank the reviewer for the insightful comments. ### The key and important concern from the reviewer is about the unfair hyperparameter search on Cal-DPO. *As this is an important misunderstanding, We would like to first clarify that our comparisons to other methods are indeed fair using Comments. The detailed responses to reviewer's other comments and misunderstandings will be presented with Rebuttal* later before the rebuttal deadline. --- Q1. Unfair comparisons to other methods: The authors show in Figure 4 that the best value for Cal-DPO is 0.001, and they choose to use this same value for their comparisons to all other methods (described in Appendix B.1, where they state that "all the hyperparameters are set to be the same"). A1. We apologize for your confusion and we believe there are some misunderstandings. ### Our comparisons to other methods are indeed fair. We did not perform a hyperparameter search for our Cal-DPO to cherry-pick parameters for better performance, and then set the same hyperparameters for other methods. Instead, we directly set its hyperparameters based on its respective base method, such as DPO, following thorough searches for these base methods. Therefore, comparisons with other methods are fair since the base methods are optimally configured. Specifically, the value $\beta = 0.001$ is the optimal hyperparameter for DPO through an extensive search across $\beta \in [0.001, 0.01, 0.1, 0.5, 1.0]$ on Zephyr-7b-SFT. We chose to use $\beta = 0.001$ for Cal-DPO without conducting further hyperparameter searches. When we state that "all the hyperparameters are set to be the same," we mean that the $\beta$ of Cal-DPO is set to match the optimal $\beta$ of the base DPO directly (not inversely as you mentioned). Similarly, for Cal-IPO and Cal-SLiC, we set their hyperparameters to match the optimal values found for IPO ($\beta=0.5$) and SLiC ($\beta=0.01$) through extensive searches of IPO and SLiC across $\beta \in [0.001, 0.01, 0.1, 0.5, 1.0]$. ### We sincerely apologize for any confusion and we will clearly state this in the revision. Nevertheless, we believe this is indeed a misunderstanding and not fatal to the major contributions of our manuscript. The responses to reviewer's other comments and misunderstandings will be presented soon. --- Rebuttal 3: Title: Dear NeurIPS Reviewer s9Bq, there are indeed misunderstandings and discussion period is ending soon Comment: Dear NeurIPS Reviewer s9Bq, We sincerely appreciate your time and the insightful comments you have provided during the review of our paper. The concerns highlighted in your comments relate primarily to misunderstandings on the setting of hyperparameters. We have made extensive efforts to address the misunderstandings you pointed out in our responses. As the discussion period is drawing to a close, we would like to confirm if there are any further clarifications you require. We would be grateful if you could consider revising your score upward. Thank you once again for your time and attention; we truly appreciate it. --- Rebuttal 4: Comment: Thank you for your responses to my review. I am still concerned by the fact that the rebuttal contradicts what is stated in the paper. In L585-586, the text says "All the hyperparameters are set to be the same for DPO and Cal-DPO for a fair comparison" and in the same paragraph, "Unless specified otherwise, the default parameterization coefficient $\beta$ is 0.001, the batch size is 64, and we use the RMSprop optimizer with a learning rate of 5e-6." Nowhere are there listed other hyperparameters for Cal-IPO and Cal-SLiC, and Figure 4 clearly shows that Cal-DPO was tried with multiple different $\beta$ values, which contradicts the part of the rebuttal that says "We chose to use $\beta=0.001$ for Cal-DPO without conducting further hyperparameter searches." "In contrast, Cal-DPO maintains increasing and positive chosen likelihoods, thus actually avoiding the rapid reward over-optimization seen with standard DPO."** Reward over-optimization in direct alignment algorithms does not refer to the trend of the model's likelihoods on the offline dataset -- it refers to the trend of the on-policy win-rate during training, as shown in Fig. 1 of the source that you cited ([3]). There is no evidence provided that Cal-DPO indeed avoids this issue. I still have strong concerns related to both the motivations and soundness of this work, but due to the new experiments provided, I will increase my score 3->4. (Concerning the motivations, some motivations make sense to me, such as adding information about the magnitude of the rewards, and most of the other ones appear to be unfounded / not supported by evidence.) --- Rebuttal 5: Title: Thank you very much for your reply Comment: Dear NeurIPS Reviewer s9Bq, Thank you very much for reading our response. Thank you also for your additional comments to facilitate further discussion! Q1. I am still concerned by the fact that the rebuttal contradicts what is stated in the paper. A1. We would like to provide further responses and clarifications on your misunderstandings: - The main and only purpose of Figure 4 is to demonstrate the sensitivity of Cal-DPO to hyperparameters; importantly, we indeed did not conduct a hyperparameter search for Cal-DPO to cherry-pick parameters that would enhance performance in the main results table. Furthermore, we have provided the source code to facilitate reproducibility. - Moreover, as shown in the Table 9 in SimPO [2]—which states that they conducted extensive searches on baselines (DPO, IPO, SLiC)—our reported baseline results are even better than theirs. - As indicated by the results in the Global Response, $\beta=0.001$ is the optimal parameter for DPO on the Zephyr-7b-beta-sft model with the ultrafeedback-binarized dataset,. This is consistent with previous study [1], which demonstrate that a small $\beta$ leads to better results. Thus our comparisons with other methods are fair since the base methods are optimally configured. Q2. Reward over-optimization in direct alignment algorithms does not refer to the trend of the model's likelihoods on the offline dataset -- it refers to the trend of the on-policy win-rate during training, as shown in Fig. 1 of the source that you cited ([2,3]). There is no evidence provided that Cal-DPO indeed avoids this issue. A2. We would like to provide further responses and clarifications. The observed over-optimization in Figure 1 in [4] suggests that an additional increase in the KL budget leads to decreased model performance. We refer to the original sentences in [4]: "This indicates that under the standard DAA training pipeline, decreasing likelihoods are not necessarily an issue for performance and are even necessary for improvement, but they exhibit non-linear over-optimization dynamics." Thus, we believe there is no evidence to support that continually increasing and choosing positive likelihoods will result in over-optimization; rather, it is the decreasing likelihoods that exhibit non-linear over-optimization dynamics. We agree with the reviewer that addressing the challenge of over-optimization is a promising future direction. However, this topic indeed extends beyond the scope of one paper. We will ensure to include this discussion in our final version. [1] https://huggingface.co/blog/pref-tuning. [2] SimPO: Simple Preference Optimization with a Reference-Free Reward. In Arxiv. [3] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models. In ICML. [4] Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms. Arxiv 2024. --- Rebuttal 6: Title: Further response and important clarifications to Reviewer s9Bq Comment: Dear NeurIPS Reviewer s9Bq, Here, we would like to provide a detailed and further response to your comments for our rebuttal. ### R1. Our reported DPO results are indeed optimal, and the comparison to our Cal-DPO is fair. - The main and only purpose of Figure 4 is to demonstrate the sensitivity of Cal-DPO to hyperparameters; importantly, we indeed did not conduct a hyperparameter search for Cal-DPO to cherry-pick parameters that would enhance performance in the main results table. Furthermore, we have provided the source code to facilitate reproducibility. - Moreover, as shown in many recent works [2,5,6]—which state that extensive hyperparameter searches were conducted—our reported DPO results are close to, and even better than, their reported results (as the results may still vary due to differences in hardware configurations and CUDA versions, etc. as shown in SimPO [10]). There is strong evidence that our reported DPO is near optimal, although there is a trade-off between different benchmarks. \| \| GSM8K \| ARC \| Winogrande \| HellaSwag \| \| --- \| --- \| --- \| --- \|--- \| \| DPO 1.0 \| 25.34 \| 57.96 \|   71.64 \| 81.28 \| \| DPO 0.5 \| 27.12 \| 58.41 \|   73.59 \| 81.95 \| \| DPO 0.1 \| 33.51 \| 60.34 \|   74.11 \| 83.10 \| \| DPO 0.01 \| 34.36 \| 61.53 \|   75.18 \| 83.67 \| \| DPO 0.001 (reported in our paper) \| 35.41 \| 62.02 \|   76.22 \| 84.51 \| \| DPO (reported in SimPO [2]) \| 21.76 \| 61.26 \|   76.80 \| 83.59 \| \| DPO (reported in Zephyr [5] \|   - \| 62.03 \|   - \| 84.52 \| \| DPO (HuggingFaceH4/zephyr-7b-beta reported in LLM Leadboard [6] \| 29.04 \| 62.03 \|   77.74 \| 84.36 \| \| Cal-DPO 0.001 \| 40.34 \| 64.34 \|   78.54 \| 85.33 \| --- ### R2. We would like to provide further responses and clarifications on reward over-optimization. First, we agree with the reviewer that addressing the challenge of over-optimization is a promising future direction. However, this topic indeed extends beyond the scope of one paper. We will ensure to include this discussion in our final version. In addition, the observed over-optimization in Figure 1 in [4] suggests that an additional increase in the KL budget leads to decreased model performance. We refer to the original sentences in [4]: "This indicates that under the standard DAA training pipeline, decreasing likelihoods are not necessarily an issue for performance and are even necessary for improvement, but they exhibit non-linear over-optimization dynamics." Thus, we believe there is no evidence to support that continually increasing and choosing positive likelihoods will result in over-optimization; rather, it is the decreasing likelihoods that exhibit non-linear over-optimization dynamics. --- ### R3. We would like to provide further responses and clarifications on motivation. As shown in many recent and concurrent works [2,7,8,9], DPO and other preference optimization methods can not effectively increase the likelihood of preferred sequences despite increasing the reward margin. This phenomenon generally decreases downstream task performance, particularly on reasoning-heavy tasks [2,7,8,9]]. The key intuition behind our Cal-DPO is very simple yet effective: If the implicit reward estimates from preference data are well calibrated relative to the actual ground-truth rewards, we can prevent the reward (likelihood) of chosen responses from continually decreasing while ensuring that the learned policy theoretically converges to the optimal. Specifically, Cal-DPO pushes chosen rewards to be as large as 1/2 and rejected rewards to be as small as −1/2. --- [1] https://huggingface.co/blog/pref-tuning. [2] SimPO: Simple Preference Optimization with a Reference-Free Reward. In Arxiv. [3] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models. In ICML. [4] Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms. Arxiv 2024. [5] Zephyr: Direct Distillation of LM Alignment. Arxiv 2023. [6] https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard. [7] Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive. Arxiv 2024 [8] 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward. Arxiv 2024 [9] Iterative Reasoning Preference Optimization. Arxiv 2024 [10] https://github.com/princeton-nlp/SimPO	Summary: This paper proposes a simple yet effective method called calibrated direct preference optimization (Cal-DPO), which addresses the limitation of ignoring the actual values of implicit rewards. The authors demonstrate the theoretical advantages of Cal-DPO over existing approaches and show the effectiveness on a variety of standard benchmarks. Strengths: 1. The proposed method Cal-DPO can support the motivation and claim theoretically. 2. Experimental results on different benchmarks show the effectiveness of the proposed method. Weaknesses: 1. It is still not intuitive for me why the scale of the reward's actual value plays an important role in the generation performance. The proposed method seems to control this scale via an additional regularization term (Equation 10) and improve the empirical generation performance. But the relationship between them lacks explanations. 2. The authors should further justify why they choose the squared loss to constrain the learned implicit reward theoretically or empirically. Although they analyze some theoretical properties of the loss function, I believe that the squared loss is not the only form to possess these properties. 3. Equation 10 is over-simplified when directly assigning $r(x,y_w)%$ and $r(x,y_l)%$ to 1/2 and -1/2, respectively. The authors should provide intuitions and show whether the method is robust by assigning other values. Technical Quality: 3 Clarity: 3 Questions for Authors: I have included my questions in the weaknesses part. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer EJ4r, we appreciate your efforts and detailed comments very much! However, we believe that there are some misunderstandings. Therefore, we would like to provide a point-by-point response to your comments. --- Q1. It is still not intuitive for me why the scale of the reward's actual value plays an important role in the generation performance. A1. Thank you for your comments. As shown in many recent and concurrent works [1,2,3,4], DPO and other preference optimization methods can not effectively increase the likelihood of preferred sequences despite increasing the reward margin. This phenomenon generally decreases downstream task performance, particularly on reasoning-heavy tasks [1,3,4]. The key intuition behind our Cal-DPO is very simple yet effective: If the implicit reward estimates from preference data are well calibrated relative to the actual ground-truth rewards, we can prevent the reward (likelihood) of chosen responses from continually decreasing while ensuring that the learned policy theoretically converges to the optimal. Specifically, Cal-DPO pushes chosen rewards to be as large as 1/2 and rejected rewards to be as small as −1/2. --- Q2. The authors should further justify why they choose the squared loss to constrain the learned implicit reward theoretically or empirically. Although they analyze some theoretical properties of the loss function, I believe that the squared loss is not the only form to possess these properties. A2. Thank you for your comments. We believe there are important misunderstandings: The squared loss is indeed the only form to possess these theoretical properties, as our Theorem 2 holds only for the squared loss (please see Eq. (47) and (48) in our paper). Thus, we indeed have provided a strong theoretical guarantee of choosing the squared loss to constrain the learned implicit reward. --- Q3. Equation 10 is over-simplified when directly assigning rewards of chosen and rejected to 1/2 and -1/2, respectively. The authors should provide intuitions and show whether the method is robust by assigning other values. A3. Thank you for your suggestion. The reason we assign rewards of 1/2 and -1/2 to chosen and rejected, respectively, is precisely due to its simplicity. Simplicity is our key motivation, and we found it works very well. Moreover, setting the rewards of chosen and rejected to 1/2 and -1/2 is the most reasonable and straightforward assumption without any prior under the general preference model as shown in [5,6]. Following your suggestion, we also conducted experiments on Cal-DPO with assigning other values (1,-1) and (1/4,-1/4). The following results shows that all variants of Cal-DPO outperforms vanilla DPO and Cal-DPO is indeed robust to by assigning other values. \| Method \| GSM8K \| ARC \| Winograde \| HellaSwag \| \| --- \| --- \| --- \| --- \| --- \| \| DPO \| 35.41 \| 62.02 \| 76.22 \| 84.51 \| \| Cal-DPO (-1/2,-1/2) \| 40.34 \| 64.34 \| 78.54 \| 85.33 \| \| Cal-DPO (1,-1) \| 41.71 \| 64.52 \| 77.92 \| 85.45 \| \| Cal-DPO (1/4,-1/4) \| 39.55 \| 64.17 \| 77.21 \| 85.95 \| --- We sincerely hope that our responses can address your comments. Moreover, as noticed by the reviewer, our work presents some interesting findings, a simple yet effective framework, and some theoretical contributions. The reviewer's suggestions can be easily and effectively addressed, and we genuinely hope that the reviewer can consider increasing the score. Thank you very much for your time! [1] Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive. Arxiv 2024 [2] 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward. Arxiv 2024 [3] Iterative Reasoning Preference Optimization. Arxiv 2024 [4] Simpo: Simple Preference Optimization with a Reference-free Reward. Arxiv 2024 [5] A General Theoretical Paradigm to Understand Learning from Human Preferences. AISTATS 2024 [6] Nash Learning from Human Feedback. ICML 2024 --- Rebuttal 2: Title: Official Comment Comment: We sincerely hope that our responses can address your comments. Moreover, as noticed by the reviewer, our work presents some interesting findings, a simple yet effective framework, and some theoretical contributions. The reviewer's suggestions can be easily and effectively addressed, and we genuinely hope that the reviewer can consider increasing the score. Thank you very much for your time! [1] Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive. Arxiv 2024 [2] 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward. Arxiv 2024 [3] Iterative Reasoning Preference Optimization. Arxiv 2024 [4] Simpo: Simple Preference Optimization with a Reference-free Reward. Arxiv 2024 [5] A General Theoretical Paradigm to Understand Learning from Human Preferences. AISTATS 2024 [6] Nash Learning from Human Feedback. ICML 2024 --- Rebuttal 3: Title: Thank you for your comments Comment: Dear NeurIPS Reviewer EJ4r, We gratefully appreciate your time in reviewing our paper and your insightful comments. We made our greatest efforts to address your concerns in the rebuttal. The reviewer's comments are mainly about some clarifications and misunderstandings and are indeed not fatal to the contributions of our manuscript; We would appreciate it if you could consider increasing your score. Thank you very much once again; we are extremely grateful. Best regards --- Rebuttal 4: Title: Dear NeurIPS Reviewer EJ4r: we understand that you maybe busy, so we would greatly appreciate it if you could check out our rebuttal. Comment: Dear NeurIPS Reviewer EJ4r Regarding the initial review from reviewer EJ4r, we just want to reiterate that there are very clear-cut answers to every question and misunderstandings that was raised, and our rebuttal has carefully addressed each point-by-point. The reviewer's comments are mainly about some clarifications and misunderstandings and are indeed not fatal to the contributions of our manuscript; we believe that the reviewer's insightful comments can be easily and effectively addressed in the final version. We would be grateful if the reviewer could increase the score. Many thanks for your time; we are extremely grateful. The authors of "Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment"	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for their insightful comments and helpful suggestions. We deeply appreciate the numerous positive comments on our work, such as describing it as "simple and effective," "solid motivations," and "solid theoretical and empirical analysis". We have made our greatest efforts to prepare a point-by-point response to each reviewer. Here, we provide this Global Response to address important misunderstandings from reviewers. --- (Global Q1). Unfair comparisons to other methods: The authors show in Figure 4 that the best $\beta$ value for Cal-DPO is 0.001, and they choose to use this same value for their comparisons to all other methods (described in Appendix B.1, where they state that "all the hyperparameters are set to be the same"). However, this implies that hyperparameter tuning was conducted for Cal-DPO, but not for any of the other methods. (Global A1). The key and important concern from the Reviewer s9Bq is about the unfair hyperparameter search on Cal-DPO. We would like to first clarify and emphasize that our comparisons to other methods are indeed fair. We apologize for the confusion of Reviewer s9Bq and we believe there are some misunderstandings. Our comparisons to other methods are indeed fair. We did not perform a hyperparameter search for our Cal-DPO to cherry-pick parameters for better performance, and then set the same hyperparameters for other methods. Instead, we directly set its hyperparameters based on its respective base method, such as DPO, following thorough searches for these base methods. Therefore, comparisons with other methods are fair since the base methods are optimally configured. Specifically, the value $\beta = 0.001$ is the optimal hyperparameter for DPO through an extensive search across $\beta \in [0.001, 0.01, 0.1, 0.5, 1.0]$ on Zephyr-7b-SFT. We chose to use $\beta = 0.001$ for Cal-DPO without conducting further hyperparameter searches. When we state that "all the hyperparameters are set to be the same," we mean that the $\beta$ of Cal-DPO is set to match the optimal $\beta$ of the base DPO directly (not inversely as you mentioned). Similarly, for Cal-IPO and Cal-SLiC, we set their hyperparameters to match the optimal values found for IPO ($\beta=0.5$) and SLiC ($\beta=0.01$) through extensive searches of IPO and SLiC across $\beta \in [0.001, 0.01, 0.1, 0.5, 1.0]$. In the following table, we also provide the DPO performance with different hyperparameters, demonstrating that our Cal-DPO comparisons with DPO are fair since the base methods are optimally configured. \| \| GSM8K \| ARC \| Winogrande \| HellaSwag \| \| --- \| --- \| --- \| --- \|--- \| \| DPO 1.0 \| 25.34 \| 57.96 \| 71.64 \| 81.28 \| \| DPO 0.5 \| 27.12 \| 58.41 \| 73.59 \| 81.95 \| \| DPO 0.1 \| 33.51 \| 60.34 \| 74.11 \| 83.10 \| \| DPO 0.01 \| 34.36 \| 61.53 \| 75.18 \| 83.67 \| \| DPO 0.001 \| 35.41 \| 62.02 \| 76.22 \| 84.51 \| \| Cal-DPO 0.001 \| 40.34 \| 64.34 \| 78.54 \| 85.33 \| ### We sincerely apologize for any confusion and we will clearly state this in the revision. Nevertheless, we believe this is indeed a misunderstanding and not fatal to the major contributions of our manuscript. --- (Global Q2). The authors predicate much of their rationale on the claim that other techniques like DPO reduce the likelihood of the chosen response. However, since DPO also reduces the likelihood of the rejected response, it is unclear whether this effect is actually undesirable…. (Global A2) . Thank you for your insightful comments. We believe there are some misunderstandings and would like to clarify the following points: - We agree that this may be a non-issue when aligning with general human values, where preferences are “relative” and multiple valid answers can exist for the same input. However, many recent studies [1,2,3,4,5] have observed that preference optimization algorithms often lead to decreased performance on downstream tasks, particularly those requiring significant reasoning, such as math and coding, where the space of correct answers is much smaller than that of incorrect ones, due to the decrease in chosen rewards. Our Cal-DPO can significantly improve DPO on these tasks, while maintaining other capability of LLMs as DPO. - Yes, if not all y_w's in preference datasets are high-quality, increasing the chosen reward may not help much. However, as shown in many works [2,3,4], increasing the chosen reward in many datasets results in significantly better performance on downstream tasks. Thus, addressing the issue of decreasing the chosen reward in DPO is indeed an important and worthwhile problem in certain real-world scenarios. - Recent works [6,7] have shown that decreasing the chosen reward may increase the likelihood of unknown out-of-distribution responses, resulting in poor performance on challenging tasks such as code generation. Sorry for your confusion; we will cite these works in the paper. ### In summary, addressing the issue of decreasing the chosen reward in DPO is indeed important and worthwhile, as evident and shown in many recent works [1,2,3,4,5]. Our simple and effective Cal-DPO has considerable potential and impact in various real-world scenarios. --- [1] SimPO: Simple Preference Optimization with a Reference-Free Reward. Arxiv 2024. [2] Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive. Arxiv 2024. [3] 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward. Arxiv 2024 [4] Iterative Reasoning Preference Optimization. Arxiv 2024 [5] Simpo: Simple Preference Optimization with a Reference-free Reward. Arxiv 2024 [6] Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment. Arxiv 2024 [7] Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study. ICML 2024	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Provably Transformers Harness Multi-Concept Word Semantics for Efficient In-Context Learning	Accept (poster)	Summary: This paper provides a theoretical analysis of in-context learning (ICL) in transformer models. The authors prove exponential convergence of the 0-1 loss for a three-layer transformer (attention + MLP) trained on a concept-specific prompt distribution. They also demonstrate how the model can leverage multi-concept semantics for out-of-distribution generalization. The analysis connects the geometric properties of concept-encoded representations to ICL capabilities. Strengths: - First work to show exponential convergence of 0-1 loss for a realistic transformer model: The authors prove exponential convergence for a three-layer model with softmax attention and ReLU MLP, which is more practical than previous simplified models. This is a significant theoretical advancement. - Solid mathematical analysis: The proof techniques seem rigorous and well-developed. The authors leverage advanced techniques to handle challenges like softmax attention and logistic loss, which have been difficult to analyze theoretically. - Experiments support theoretical findings: The simulations in Section 6 validate key aspects of the theory, like the evolution of attention weights and MLP coefficients. - Good motivation and connections to empirical observations: The authors motivate their work well by connecting to empirical findings on LLM representations (e.g., within-concept positive inner products). This grounds the theory in practical observations. Weaknesses: - Limited training setup: The paper only trains $W_Q, W_K, W_O$, keeping other weights fixed. This is quite restrictive compared to full transformer training and may limit the applicability of the results. - Overly specific data model: The concept-specific prompt distribution seems quite structured and simple. It's not clear how well this captures the complexity of real language data. A linear hyperplane can solve all tasks, including OOD ones, in this setup. - Lack of non-linear task consideration: The analysis focuses on essentially linear classification tasks. It would be more compelling to see how the model handles tasks requiring non-linear decision boundaries or composition of multiple concepts. Technical Quality: 3 Clarity: 2 Questions for Authors: - How sensitive are the results to the specific data distribution assumptions? Would the exponential convergence hold under more general conditions? - Have you considered extending the analysis to training all weights in the network? What challenges would this introduce? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: In Appendix J, the authors acknowledge that the data model may need refinement to better align with practical scenarios. They also note that adding more attention layers could make the model more realistic. These are fair limitations to point out. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the insightful review. We greatly appreciate your acknowledgment of our theoretical advancements, rigorous analysis, well-grounded connections to empirical observations, and fair assessment of the limitations we address. Q: Limited training setup & challenges of training all weights. A: Thank you for raising this point. The reason we left some matrices fixed (e.g., zero entries, $W_V$ and $a$) was to simplify the coupled gradient dynamics and enable rigorous theoretical analysis, which is a common practice in deep learning theory [1-2]. We remark that this restriction does not fundamentally limit the functionality of the transformer with sufficient sample complexity in our learning problem (e.g. one $W_V$ here already can take the full job of $W_O W_V$), which is validated by experiments. We appreciate your encouragement to consider training all weights. Indeed, it’s known that adding more trainable layers could lower the sample complexity, especially for harder learning targets that exhibit hierarchy [3-4]. Extending the analysis to training all weights would certainly pose additional challenges since there’re many gradient updates of various matrices simultaneously. This would require simplification techniques such as layer-wise training and reparameterization [3], as well as special structural assumptions on the concept classes (e.g., information gap) and algorithm configurations (e.g., different regularizers at different phases) [4]. Q: Alignment of the data model to real-world & applicability to general conditions. A: Thanks for raising the points. We would like to address your concerns as follows. - Empirical Relevance. Mathematical theories often simplify models to reveal intrinsic capabilities. To better align with the real world, we leverage a sparse coding approach, proposed as a suitable theoretical setup for capturing language polysemy [5], and in our case, seen as a special prompt version of the recognized LDA [6]. Our purpose is to understand how transformers leverage latent linear geometry for certain OOD tasks, modeled after empirical observations on LLM latent structure [7-9]. The practical meanings of the OOD samples are partially validated in [9-10], where different settings show that the combination of concepts forms new meaningful semantics. While there is room for improvement, such as considering hierarchy, we believe our theory stands out and provides an initial positive response to the research question in [11], which asks whether the observed LLM latent geometry can explain their OOD extrapolation abilities. - Applicability. The exponential convergence result hinges on the hard low-noise condition [12]. Intuitively it suggests that when the target pattern is sufficiently “low-noise”, the exponential convergence with a suitable learning rate can be achieved. We note that strict assumptions are common in theoretical studies to enable rigorous analysis. While our current theory holds under Condition 1, we believe there is scope to relax these assumptions and develop theories for more general data distributions. Looking ahead, a key future direction is to explore analytical techniques that can handle more general conditions. Q: Lack of non-linear task & composition of multiple concepts A: Thank you for this insightful feedback. We would like to address your concerns as follows: - Linearity of LLM Representations. Recent work has shown that high-level concepts in language models tend to be encoded linearly [7-8]. Our theory aims to connect this observed linear structure to the transformer's capability on certain OOD tasks. - OOD Tasks with Multiple Concepts. We note that our proposed OOD tasks allow for the composition of multiple co-concepts, as stated in the second point of Proposition 1. - Handling Non-linear Tasks. Given that our transformer model includes non-linear MLPs, which can theoretically handle non-linear tasks [13], it is feasible to consider non-linear tasks. In this context, we expect that the attention mechanisms would still assign more weight to words in demonstrations that share the most similar "non-linear patterns" with the query to complete ICL tasks. We believe our focus on empirically grounded linear modeling has allowed us to provide the first theories linking multi-concept geometry to the OOD capabilities of transformers, marking an important step forward. However, we agree that exploring whether and how transformers can excel in certain non-linear tasks is another crucial direction for our future endeavors. We appreciate you highlighting this insightful point. Reference [1] Jelassi et al. Vision Transformers provably learn spatial structure. NeurIPS 2022 [2] Huang et al. In-Context Convergence of Transformers. ICML 2024 [3] Tian et al. JoMA: Demystifying Multilayer Transformers via Joint Dynamics of MLP and Attention. ICLR 2024 [4] Allen-zhu and Li. Backward feature correction: How deep learning performs deep (hierarchical) learning. COLT 2023 [5] Arora et al. Linear algebraic structure of word senses, with applications to polysemy. TACL 2018 [6] Blei et al. Latent Dirichlet Allocation. NIPS 2001 [7] Park et al. 2023: The linear representation hypothesis and the geometry of large language models. ICML 2024 [8] Jiang et al. On the origins of linear representations in large language models. ICML 2024 [9] Yamagiwa et al. Discovering universal geometry in embeddings with ICA. EMNLP 2023 [10] Park et al. The Geometry of Categorical and Hierarchical Concepts in Large Language Models. ICML Workshop MI 2024 [11] Reizinger et al. Position: Understanding LLMs Requires More Than Statistical Generalization. ICML 2024 [12] Massart and Nedelec. Risk Bounds for Statistical Learning. AISTATS 2006 [13] Kou et al. Matching the Statistical Query Lower Bound for k-sparse Parity Problems with Stochastic Gradient Descent. arXiv:2404.12376 --- Rebuttal Comment 1.1: Title: Thank you and raise confidence score from 3 to 4 Comment: Thank you for your response. I have checked all rebuttals, and they have addressed my concerns. Thus, I raised my confidence score from 3 to 4. The discussion about the Linearity of LLM Representations is interesting. The PDF Figure 1 looks nice. It would be good to include them in the main body. --- Reply to Comment 1.1.1: Title: Thank You for Your Feedback and Our Planned Revisions Comment: Dear Reviewer Bq3t, Thank you for your positive feedback. We are delighted and encouraged that our rebuttals have addressed all your concerns and that you have raised your confidence score. It is an honor to receive your recognition of the value Figure 1 adds to our presentation. We will appropriately include both the discussion and Figure 1 in the main body based on your constructive suggestion. Once again, we truly appreciate your support and insightful comments. Warm regards, Authors of Submission 15661	Summary: This paper tries to understand the mechanisms that explain the emergence of in-context learning.The starting point of their approach is the observation that the embeddings have geometric properties to encode within-concept and cross-concept relationships. Their goal is to connect this geometric properties with the ability to conduct efficient in-context learning. They first describe a data distribution that is sparse coding prompt distribution and then describe their model which is a 3-layer transformer. They show that such a model trained with SGD converge in an exponential speed to the 0-1 Bayes optimal test error. Lastly, they show that their learned model can utilize the multi-concept semantics to conduct unseen ICL tasks. Strengths: I think that the results are pretty strong and I liked the proposition 1 which shows how the transformer can use its multi-concept semantics knowledge to perform in-context learning on unseen tasks. Weaknesses: I do not have any knowledge on the topic and so this is why my evaluation is going to be very shallow. Even though the theorems look sound to me, I think the authors could improve the presentation of their results by giving some proof sketch (+drawing) in the introduction. Technical Quality: 3 Clarity: 2 Questions for Authors: I think the authors should improve the presentation of their theory by giving more insights and intuition + simplified proof sketch. Confidence: 1 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors did not mention the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive feedback and recognition of the strength of our theoretical results, particularly Proposition 1 which demonstrates how transformers can leverage their multi-concept semantic knowledge to perform effective unseen ICL tasks. We appreciate your insightful assessment of the significance of our work. Q: knowledge on the topic & more insights and intuitions for presentation. A: We sincerely appreciate your feedback and consider it an honor to share more background on our topic with you. We also value your advice and aim to provide additional insights and intuitions as follows: - Understanding LLM. In recent years, the remarkable power of LLMs has prompted both practitioners and theoreticians to explore their underlying mechanisms. Practitioners validate their claims through large-scale experiments [1-2], while theoreticians use simplified models for analysis [3-6]. A natural practice is that the theoreticians employ mathematics to explain the phenomena observed by practitioners. - Research Gap and Our Contributions. On the one hand, empirical findings suggest that the emergent power of LLMs mainly stems from their ICL ability [1], which is partially attributed to data properties and transformer structures [2]. On the other hand, existing theories are often based on unrealistic settings (e.g., noise-free orthogonal data, simplified transformer architectures like linear/QK-combined/ReLU/MLP-free/infinite-dimensional models, and square or hinge loss) [3-6]. This leaves room for providing explanations for emergent ICL capabilities in more realistic settings. To advance this understanding, we theoretically demonstrate that transformers leverage latent linear geometry for certain OOD tasks. We model the data based on empirical observations of LLM latent structures [7-9] and consider realistic non-linear transformer architectures and losses (softmax attention, ReLU MLP, and cross-entropy (logistic) loss). Using advanced analytical techniques, we showcase an exponential convergence rate for our non-trivial setup, going beyond the linear or sublinear rates achievable in previous simplified setups [3-5] due to their technical limitation. Interestingly, our work addresses a research question raised in Question 5.1.4 of the ICML 2024 position paper [10], which inquires whether the observed linear latent geometry of LLMs can explain their OOD extrapolation capabilities. We believe our results provide an initial positive response to this question. Q: provide (simplified) proof sketch(+ drawing) in introduction. A: Thank you for the valuable feedback. We appreciate your suggestion to enhance the accessibility of our proof sketch with visual aids. In response, we have prepared Figure 1 for inclusion in the introduction, which would be collaborated with a simplified proof sketch in our revised manuscript. Simplified Poof Sketch. Please refer to our illustration (Figure 1) in the new PDF. The Idempotent Operator Techniques depicted allow us to rigorously analyze the model's learning dynamics by examining matrix lengths across orthogonal components, whose practical meanings are partially validated in [11]. In Sections 5.2-3, we extend expectation-variance reduction techniques [12-13]. By treating conditional expectations as Doob martingales, we exploit exponential convergence properties, deriving the rate under low-noise conditions. Our theory suggests that the transformer's learned knowledge, characterized by specific lengths and cross-concept orthogonality, is essential for OOD extrapolation, particularly in prioritizing words that share components with queries. Q: The authors did not mention the limitations of their work. A: We would like to clarify that we have a Limitations Section in Appendix J. There is potential to enhance our modellings to better align with real-world scenarios, such as increasing attention layers or considering the Markovianity of the prompt distribution. We will highlight these limitations in the main body and move Appendix J to Appendix A in our revisions. Summary We are sincerely grateful for the reviewer’s constructive feedback. Should the reviewer have any further suggestions or wish to discuss any points in more detail, we would be delighted to continue this productive exchange. Once again, we deeply appreciate the reviewer’s time and valuable comments. Reference [1] Lu et al. Are emergent abilities in large language models just In-Context Learning? ACL 2024 [2] Chan et al. Data Distributional Properties Drive Emergent In-Context Learning in Transformers. NeurIPS 2022 [3] Zhang et. al. Trained transformers learn linear models In-Context. JMLR 2024 [4] Kim and Suzuki. Transformers learn nonlinear features In-Context: nonconvex mean-field dynamics on the attention landscape. ICML 2024 [5] Li et al. How do nonlinear transformers learn and generalize in In-Context Learning? ICML 2024 [6] Chen et al. Training dynamics of multi-head softmax attention for In-Context Learning: emergence, convergence, and optimality. COLT 2024 [7] Park et al. 2023: The linear representation hypothesis and the geometry of large language models. ICML 2024 [8] Jiang et. al. On the origins of linear representations in large language models. ICML 2024 [9] Li et. al. How do transformers learn topic structure: towards a mechanistic understanding. ICML 2023 [10] Reizinger et al. Position: Understanding LLMs Requires More Than Statistical Generalization. ICML 2024 [11] Yamagiwa et al. Discovering universal geometry in embeddings with ICA. EMNLP 2023 [12] Nitanda and Suzuki. Stochastic gradient descent with exponential convergence rates of expected classification errors. AISTATS 2019 [13] Yashima et. al. Exponential convergence rates of classification errors on learning with SGD and random feature. AISTATS 2021 --- Rebuttal Comment 1.1: Title: Post rebuttal response Comment: I thank the reviewers for their rebuttal that has addressed most of my concerns. I appreciate the diagram (Figure 1) that allows to get a better sense of the proof sketch and I hope it will be included to the paper. I increase my score by one point. --- Reply to Comment 1.1.1: Title: Grateful Response to Reviewer's Positive Feedback Comment: Dear Reviewer L2Bx, We're delighted that you've found our rebuttals have met your satisfaction, and we're encouraged that you've raised the score. We're happy that you found the attached Figure 1 useful, and we would be sure to include it appropriately in the main body of our manuscript. Once again, we appreciate your invaluable time and comments. Warmest regards, Author of Submission 15661	Summary: This paper performs an in-depth analysis of the optimization dynamics of a simplified Transformer on a sparse coding prompt model. It manages to show the exponential 0-1 loss convergence on this non-convex loss. Experiments on synthetic data verify the convergence. Strengths: 1. This paper is very original and investigates the interesting question of how a Transformer learns shared concepts within context. 2. This paper shows the clear OOD generalization of the Transformer model when trained on the sparse coding distribution. Weaknesses: 1. The notation in the proof sketch is very heavy and not very easy to track. It would improve the paper if more explanation is given. Also, some of the statements seem counterintuitive. For example: a) Definition 4 states that 'with prob 1 - $\delta$, $L_{D*}(\Phi) = 0$'. What is the source of randomness here? b) Proposition 4 states minimization of the loss to 0, which seems highly surprising at first pass. This seems like a specialty due to 0-1 test loss considered here and should be highlighted. 2. Some of the legends in Figure 1 is left unexplained and in general the graph may need more explanations on why it matches Lemma 1. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. It seems that the proof considering mini-batch SGD is showing a concentration to the population gradient descent by Lemma 3. As mentioned in the intro, one of the key properties of the theoretical claim here is that it doesn't require a large batch size. Why is the technique here so robust to batch size? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: The limitation is adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are immensely grateful for your thoughtful and insightful feedback on our work. Your recognition of the originality, soundness, and excellent contributions of our paper is deeply encouraging. We found your comments to be highly professional and constructive, and we would like to respond as follows. Q: more explanation for the notation of the proof sketch. A: Thanks for your suggestions. We would be happy to add more insights and illustrations (Figure 1 in the PDF file) to interpret the notations. The key definitions are: - $a_k$ is the mean vector of $\mu_k^+$ and $\mu_k^-$, which is denoted as the "concept vector"; $b_k$ is denoted as the "task-specific vector", where $b_k=\mu_k^+ - a_k$. $c_k$ and $d_k$ are defined similarly with regard to $q_k^+$ and $q_k^-$. - The coefficients $\alpha_{Q, k}$ and $\beta_{Q, k}$ are defined through the idempotent decomposition of $W_Q$. This allows us to show that ${(W_K\mu_k^{\pm e})}^{T}W_Q\mu_k^e = \alpha_{Q,k}\cdot\alpha_{K,k}\pm\beta_{Q,k}\cdot\beta_{K,k}$ by simple calculations, which pave the way for us to examine the evolution of attention weights by tracking the dynamics of the coefficients. $\alpha_O$ is defined similarly. Our theory then builds up the relationship between the test error of the expected model and the expected coefficients. This would facilitate us to establish the 0-1 loss exponential convergence by closely examining the evolution of these coefficients. Q: source of randomness. A: The primary source of randomness here stems from two factors: the uncertainty in the model's initialization, and the tail behavior of the Gaussian noise, as detailed in Appendix B.1. Our overparameterization conditions ensure that we can control, with high probability 1-δ, the volumes and directional lengths of the matrix parameters, as well as the influence of the noise. Without careful control of these preliminary properties, our theoretical results would not be able to hold rigorously. We appreciate you noting this important detail. Q: test error considered in Proposition 4 should be highlighted. A: We appreciate your intuition about our results. You are correct that, similar to [1], the consideration of the 0-1 loss and treating cross-entropy loss as a surrogate is crucial to establishing the convergence in our Proposition 4. We will certainly highlight this aspect in our revised manuscript. Another key point is that our proposition examines the expected model behavior at each time step, where the randomness here stems from the stochastic batch sampling in SGD. Given the isotropic noise and balanced positive/negative labels in our setting, the expected model can be viewed as being trained by noise-free gradient descent on a balanced dataset. From this perspective, the expected model's test error can be shown to converge to zero very rapidly, as long as the lengths of the neural network matrices grow sufficiently along the critical directions. We appreciate your reminder and would highlight the points with interpretations in our revised manuscript. Q: more explanation for Figure 1. A: Thank you for the helpful suggestion. We will replace the existing Figure 1 with a new one (Figure 2 in the supplementary pdf), and ensure the caption provides detailed descriptions. Specifically, the new Figure 2 includes: 1. A plot demonstrating the exponential convergence of the test error, validating our theory. 2. A plot showing that the correct attention weights for both concepts' classification tasks converge to 1, aligning with the last conclusion of Lemma 1. 3. A plot verifying our Lemma 1, as it shows the products $\alpha_{Q, s}^{(t)}\cdot\alpha_{K, s}^{(t)}$ are all non-increasing, while the values of $\beta_{Q, s}^{(t)}\cdot\beta_{K, s}^{(t)}$ can grow sufficiently. 4. A plot supporting the claims in Lemma 1 regarding the sufficient growth of the$\lvert\beta_{O_{(i, \cdot)},k}^{(t)}\rvert$ and $\alpha_{O_{(i, \cdot)}, k}^{(t)}$, as well as their relationships. As our analysis demonstrates the convergence between the expected model and the SGD-updated model, the new Figure 2 clearly illustrates how the empirical results align with and validate the theoretical claims our theories. Additionally, we include supplementary Figure 3 of OOD scenarios to further collaborate our theories. In the final manuscript, we will provide extended experiments and descriptions in the appendix. Q: robustness to batch size. A: The key reason that our technique is robust to batch size, compared to prior work [2], is that we introduce standard techniques from the theoretical literature on exponential convergence [1] to our settings. Specifically, rather than directly using the batch size (at least $\varepsilon^{-2}$, where $\varepsilon$ is the test error) to bound the gradient difference between the expected model and the SGD-updated model via Hoeffding's inequality, as done in [2], our analysis first considers the test error convergence of the expected model. We then examine the rate of convergence for the test error difference between the expected model and the SGD-updated model at each iteration, which could be exponential due to the hard low-noise condition [1]. Crucially, our analysis considers the extreme case where the batch size can be as small as 1, as in [1], to provide an upper bound on the test error. This allows our technique to be robust to the batch size, in contrast with the larger batch size requirements in prior work. Reference [1] Nitanda and Suzuki. Stochastic gradient descent with exponential convergence rates of expected classification errors. AISTATS 2019 [2] Li et al. How do nonlinear transformers learn and generalize in In-Context Learning? ICML 2024 --- Rebuttal Comment 1.1: Comment: Thank the authors for the detailed response. Regarding Weakness 1.a, I still find the argument confusing, if I understand correctly $\Psi$' refers to a specific weight of the model, and $\Phi$ contains all the weights that can minimize the 0-1 loss. Why does this have anything to do with noise or initialization? --- Rebuttal 2: Title: Clarification on Definition 4 and the Source of Randomness Comment: Dear Reviewer dQGe, Thank you very much for your follow-up. We would like to provide an edited* version of our response for further clarification. In the original Definition 4, $\Phi^*$ represents all the estimators that can achieve zero 0-1 loss with high probability at least $1-\delta$. For clarity, we propose removing Definition 4 and directly stating in subsequent results (Proposition 2 and Lemma 1-2) that with high probability at least $1-\delta$, certain results will hold. The reason for these high-probability statements is to avoid extreme cases of initialization and noise, under which our subsequent results would not rigorously hold. The successful evolution of the estimators (whether the expected $E[\Psi']$ or the SGD-updated $\Psi'$) relies on a non-extreme, non-zero initialization. We apologize for any confusion and appreciate your continuous patience and constructive feedback. We will include more clarifications in our revisions to make this clearer. Please let us know if this addresses your concern. Warmest regards, Authors of Submission 15661	Summary: The paper investigates how transformer-based large language models (LLMs) leverage their in-context learning (ICL) capabilities through the lens of multi-concept semantics. The authors cited limitations in existing theoretical work, which uses oversimplified models and unrealistic loss functions, leading to only linear or sub-linear convergence rates. To address this, the authors present an analysis of a three-layer transformer model, consisting of one attention layer and a ReLU-activated feedforward network, trained with logistic loss. The analysis included an examination of the learning dynamics of the transformer model and proved exponential convergence of the 0-1 loss in this complex setting, demonstrating that the transformer can achieve Bayes optimal test error with a logarithmic number of iterations. The paper also demonstrated how multi-concept encoded semantic geometry enables transformers to perform efficiently on out-of-distribution ICL tasks, explaining their success in leveraging the polysemous nature of words for diverse, unseen tasks. The paper included empirical simulations to support these theoretical findings. Strengths: The paper brings a new theoretical perspective linking multi-concept word semantics with in-context learning Weaknesses: The paper is sometimes hard to follow. More concise and clearer presentation, especially the theoretical content, would be helpful. Lack of empirical experiments with more realistic datasets to show relevance in real-world setting. Technical Quality: 2 Clarity: 2 Questions for Authors: NA Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for acknowledging our work for linking the multi-concept word semantics to transformer’s ICL capability. Q: More concise and clearer presentation, especially the theoretical content, would be helpful. A: We very much appreciate your suggestion. We will provide more explanations and figures to improve the overall readability. The following are some explanations (which will be extended in the final version). - Definitions and Technical Challenge. As mentioned in the paper, we leverage a sparse coding approach suitable for capturing language polysemy [1-2] and incorporate observed latent geometric properties of LLMs [3-5], distinguishing our work from previous studies that assume idealistic, noise-free orthogonal settings [6]. Our learning problem considers practical non-linear transformer architectures and losses (softmax attention, ReLU MLP, and cross-entropy loss), contrasting with the oversimplified settings (linear/QK-combined/ReLU/MLP-free/infinite-dimensional transformers with square or hinge loss) in prior work [7-8]. - Theoretical Achievement. Our theorem demonstrates an exponential convergence rate for our non-trivial setup, going beyond the linear or sublinear rates achievable in previous simplified setups [6-7] due to their technical limitation. Intuitively, our theory suggests the transformer's learned knowledge after training, with certain lengths and cross-concept orthogonality, enables its capability for certain OOD extrapolation, such that the test prompts can enjoy different lengths, various distributions of latent concepts, and even shifts in word semantics. - Proof Sketch. We prepare Figure 1 to visualize our Idempotent Operator Techniques in the new PDF file. This technique enables us to rigorously analyze the learning dynamics by conducting scale analyses on different orthogonal components of the data, whose practical meanings are partially validated in [9]. In addition, in Sections 5.2-3, we extend standard expectation-variance reduction techniques [10] to our setting. We treat the conditional expectations of the NN matrices as Doob martingales, and by constructing martingale difference sequences and exploiting the exponential convergence property of the tails, we derive the exponential convergence rate under low-noise conditions. Q: Alignment to Real-world. A: Thank you for your feedback. We would like to address your concerns as follow. 1. Empirical Relevance. Our setting of non-orthogonal dictionaries is inspired by practical observations [4-5, 9], where the LLM's latent geometry exhibits within-concept positive inner products and cross-concept orthogonality, which is also theoretically validated by [3]. Furthermore, according to [1-2], the sparse coding approach is a suitable setup for modelling the polysemy of language. 2. Data Modelling in Feature Learning Theory. Mathematical theories often deal with simplified modelling, aiming to reveal certain intrinsic capabilities of models. There is a rich literature that adopts modelling similar to ours, such as studying self-supervised contrastive learning [2], or classification upon orthogonal dictionaries [6]. 3. Alignment to the Real-World Setting. In contrast to prior theories that considered oversimplified settings like noise-free orthogonal feature dictionaries [6] or linear/QK-combined/ReLU/MLP-free/infinite-dimensional transformers [7-8], our setup with non-orthogonal dictionaries, non-linear transformers, and cross-entropy loss is more aligned with practical scenarios. Our goal is to demonstrate the capability of transformers to utilize latent linear geometry for certain OOD tasks. The practical meanings of the OOD samples are partially validated in [9] and Theorem 8 in [11], where different settings show that the combination of concepts forms new meaningful semantics. We believe this provides an initial positive response to Question 5.1.4 in the ICML 2024 position paper [12], which asks whether the observed latent geometry of LLMs can explain their OOD extrapolation abilities. In summary, grounded in practical observations, we study this non-trivial learning problem with comparatively realistic settings, and is an important step forward to reveal the ICL emergent capability of transformer [13]. Looking ahead, one of our important future directions include exploring more realistic model settings, such as incorporating additional attention layers, as well as considering the hierarchy of language in our theoretical analysis. Reference [1] Arora et. al. Linear algebraic structure of word senses, with applications to polysemy. TACL 2018 [2] Wen and Li. Toward understanding the feature learning process of self-supervised contrastive learning. ICML 2021 [3] Li et al. How do transformers learn topic structure: towards a mechanistic understanding. ICML 2023 [4] Park et al. 2023: The linear representation hypothesis and the geometry of large language models. ICML 2024 [5] Jiang et al. On the origins of linear representations in large language models. ICML 2024 [6] Li et al. How do nonlinear transformers learn and generalize in In-Context Learning? ICML 2024 [7] Chen et al. Training dynamics of multi-head softmax attention for In-Context Learning: emergence, convergence, and optimality. COLT 2024 [8] Zhang et al. Trained transformers learn linear models In-Context. JMLR 2024 [9] Yamagiwa et al. Discovering universal geometry in embeddings with ICA. EMNLP 2023 [10] Nitanda and Suzuki. Stochastic gradient descent with exponential convergence rates of expected classification errors. AISTATS 2019 [11] Park et al. The Geometry of Categorical and Hierarchical Concepts in Large Language Models. ICML Workshop MI 2024 [12] Reizinger et al. Position: Understanding LLMs Requires More Than Statistical Generalization. ICML 2024 [13] Lu et al. Are emergent abilities in large language models just In-Context Learning? ACL 2024 --- Rebuttal Comment 1.1: Title: Thank you Comment: I thank the authors for their response. The authors have edit the manuscript to address my core concerns, I have increased the score accordingly. --- Reply to Comment 1.1.1: Title: Thank you for your recognition Comment: Dear Reviewer qjJG, We're greatly encouraged that our rebuttals have been effective in addressing all your core concerns and that you have raised the scores accordingly. We are honored by your recognition and look forward to incorporating the necessary changes. Thank you for your invaluable time and comments throughout this process. Warmest regards, Authors of Submission 15661	Rebuttal 1: Rebuttal: Dear ACs and Reviewers, Thank you again for all of your positive and constructive feedback! We are truly encouraged to see so many well-recognized comments on our work, such as the innovative angle (Reviewer qjJG, Reviewer dQGe), advanced techniques (Reviewer dQGe, Reviewer Bq3t), significant theoretical achievement (Reviewer L2Bx, Reviewer Bq3t), practical modelings (Reviewer Bq3t), sound analysis (Reviewer L2Bx, Reviewer Bq3t), and empirically-grounded motivation (Reviewer Bq3t). Your insights have greatly helped us strengthen our manuscript. Alongside addressing your comments point-by-point, we have attached a new PDF featuring illustrations of our proving technique (Figure 1) and additional experiments with detailed descriptions (Figures 2-3). We particularly thank Reviewer dQGe for the reminder to provide more interpretation for the notations and Reviewer L2Bx for suggesting visual aids. Your feedback assures us that we have successfully achieved our main goal: understanding how transformers leverage latent linear geometry for certain out-of-distribution tasks, modeled after the empirical observations on LLM latent structure [1-3]. Interestingly, the research gap we’re addressing aligns with a research question raised in the ICML 2024 position paper [4], regarding whether the observed linear latent geometry of LLMs can be leveraged to explain their OOD extrapolation abilities. By emphasizing your points in our revisions, we believe we can excel as an important step forward and continue contributing to the community in future directions inspired by your valuable comments. Should the reviewer have any further suggestions or wish to discuss any points in more detail, we would be more than delighted to continue our productive exchange. Once again, we deeply appreciate the reviewer’s time and valuable comments. Warmest regards, Authors of Submission 15661 Reference [1] Yamagiwa et al. Discovering universal geometry in embeddings with ICA. EMNLP 2023 [2] Park et al. 2023: The linear representation hypothesis and the geometry of large language models. ICML 2024 [3] Jiang et. al. On the origins of linear representations in large language models. ICML 2024 [4] Reizinger et al. Position: Understanding LLMs Requires More Than Statistical Generalization. ICML 2024 Pdf: /pdf/c8c55590111682e50d01c069aa784e5d1d9d084e.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Molecule Generation with Fragment Retrieval Augmentation	Accept (poster)	Summary: The paper introduces a novel fragment-based molecule generation framework called Fragment Retrieval-Augmented Generation (f-RAG). This framework aims to address the limitations of existing fragment-based molecule generation methods, which often struggle to explore beyond the existing fragments in their databases. f-RAG utilizes a pre-trained molecular generative model to propose additional fragments from input fragments, completing and generating a new molecule. It retrieves two types of fragments: hard fragments, which are explicitly included in the newly generated molecule, and soft fragments, which serve as references to guide the generation of new fragments through a trainable fragment injection module. To further extrapolate beyond the existing fragments, f-RAG updates the fragment vocabulary with generated fragments through an iterative refinement process. This process is enhanced with post-hoc genetic fragment modification, allowing f-RAG to maintain a pool of fragments and expand it with novel and high-quality fragments through a strong generative prior. This approach enables f-RAG to achieve an improved exploration-exploitation trade-off in fragment-based molecule generation. Strengths: 1. f-RAG uses a pre-trained molecular generative model to propose new fragments, allowing new fragments to be generated 2. The approach retrieves both hard fragments that are directly incorporated into the new molecule and soft fragments that guide the generation process, enhancing the diversity and effectiveness of the generated molecules. 3. f-RAG updates the fragment vocabulary with generated fragments through an iterative process, continuously improving the quality and novelty of the fragments used for molecule generation. Weaknesses: 1. The improvement of f-RAG compared to Genetic GFN is not considered large enough. In many tasks, Genetic GFN actually performs much better than f-RAG, which raises the concern of generalization of the proposed method. 2. Some notations are not clear enough. For example, in Table 2, arrows could be used to illustrate lower-better or higher-better. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the authors explain the potential reason that f-RAG falls behind genetic GFN? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the limitation is well-discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your comments. We appreciate your positive comments that our paper proposes an effective strategy to enhance the quality, diversity and novelty of molecules. We address your concerns and questions below. --- Comment 1 The improvement of f-RAG compared to Genetic GFN is not large enough. Could the authors explain the potential reason that f-RAG falls behind genetic GFN? Response 1 Drug discovery is a comprehensive problem and optimization performance alone is not meaningful without consideration of other factors such as diversity, novelty, and synthesizability. Genetic GFN [1] showed competitive results with $f$-RAG in terms of optimization performance (Table 1), but significantly worse results compared to $f$-RAG in terms of diversity, novelty, and synthesizability (Figure 1, Table 2 and Figure 7). As described in Lines 232-235, the essential considerations in drug discovery often conflict with each other, and the results of Genetic GFN illustrate these trade-offs. On the contrary, our proposed $f$-RAG effectively improves these trade-offs by utilizing existing fragments while dynamically updating the fragment vocabulary. To further compare $f$-RAG and Genetic GFN, we evaluated the performance of Genetic GFN in the experiments of Section 4.2 in the below table. $f$-RAG outperforms Genetic GFN by a very large margin, again demonstrating its superiority as a universal method applicable to various drug discovery tasks. We will include this result in the revised paper. Table: Novel top 5% docking score (kcal/mol) results. The results are the means and standard deviations of 3 independent runs. Lower is better. \| Method \| parp1 \| fa7 \| 5ht1b \| braf \| jak2 \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| Genetic GFN \| -9.227 ± 0.644 \| -7.288 ± 0.433 \| -8.973 ± 0.804 \| -8.719 ± 0.190 \| -8.539±0.592 \| \| $f$-RAG (ours) \| -12.945 ± 0.053 \| -9.899 ± 0.205 \| -12.670 ± 0.144 \| -12.390 ± 0.046 \| -11.842 ± 0.316 \| --- Comment 2 In Table 2, arrows could be used to illustrate lower-better or higher-better. Response 2 We appreciate the suggestion to improve the readability of our paper. We will include the arrows in the revised paper. --- References [1] Kim et al., Genetic-guided gflownets: Advancing in practical molecular optimization benchmark, arXiv, 2024. --- Rebuttal Comment 1.1: Title: Reviewer Response Comment: Thanks for clarifying my concerns. Considering that my score is already supportive enough, I decide to maintain my scores. Have a good luck~	Summary: The paper proposed a fragment retrieval-augmented generation for molecule discovery, namely f-RAG. f-RAG retrieves two types of fragments, i.e., hard fragments and soft fragments. Hard fragments serve as build blocks that are explicitly included in the newly generated molecules, while soft fragments guide the generation of new fragments through a trainable module. Strengths: The work is well written. The experiments show the effectiveness of the proposed method. Weaknesses: 1. The novelty of the work is limited. Very similar works have been published, such as [1], [2], and [3]. They basically follow the same pipeline, which generates new molecules through LLMs, GAs, or both. Compared to these works, this work does not seem to bring new insights. 2. Moreover, the so-called soft and hard fragments are not new either. In reference [1], retrieved exemplar molecules are used as inputs to guide the generation of new molecules through trainable networks. Therefore, the concept of soft fragments is not novel. 3. In addition, reference [2] also used GAs and LLMs to generate molecules. The author needs to further clarify the differences between the two works. 4. The baselines used by the authors are not the latest. The authors should consider incorporating LLMs related molecular generation methods into the comparison. 5. Why did the author use SAFE-GPT instead of other chemical language models? Can this method be extended to other chemical large language models? 6. Some key details in genetic algorithms are missing, such as how to use mutation and crossover to generate new molecules. [1] Wang, Zichao, et al. "Retrieval-based controllable molecule generation.", ICLR, 2023. [2] Lee, Seul, et al. "Drug Discovery with Dynamic Goal-aware Fragments." Forty-first International Conference on Machine Learning, 2023 [3] Wang, Haorui, et al. "Efficient Evolutionary Search Over Chemical Space with Large Language Models." arXiv preprint arXiv:2406.16976 (2024). Technical Quality: 2 Clarity: 2 Questions for Authors: See weakness. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 1 Limitations: No potential negative societal impact Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your comments. We appreciate your positive comments that our paper is well-written and the experiments show the effectiveness of our method. We address your concerns below. --- Comment 1 The novelty of the work is limited compared to [1,2,3]. In [1], retrieved exemplar molecules are used as inputs to guide the generation of new molecules through trainable networks. [2] also used GAs and LLMs. Response 1 Compared to RetMol [1], $f$-RAG (1) retrieves fragments instead of molecules, (2) utilizes two types of retrieval--hard and soft, and (3) generates molecules in a one-shot manner instead of multiple iterations. For a more detailed explanation of the differences, please see Global Rebuttal. Compared to GEAM [2], $f$-RAG is better at exploring chemical space. GEAM does not use an LLM. GEAM uses an RL model that only reassembles the given fragments, therefore unable to generate novel molecular structure and solely relies on the modification of the GA to generate new fragments. On the contrary, $f$-RAG applies fragment-level retrieval augmentation to an LLM to propose novel, high-quality fragments rather than simply reassembling existing fragments, greatly enhancing exploration beyond known fragments. Due to these differences, $f$-RAG outperforms GEAM by a large margin (Table 3 and Table 7). Compared to MolLEO [3], $f$-RAG applies RAG to an LLM, therefore is better at generating high-quality drug candidates. MolLEO does not use RAG and is therefore less effective at utilizing chemical knowledge. In addition, MolLEO focuses on replacing one of the crossover or mutation of Graph GA [4] by LLM-guided operations, so it is basically a GA and suffers from the common limitation of GAs--low diversity [5]. On the contrary, $f$-RAG shows an improved balance between optimization performance, diversity, novelty, and synthesizability through the proposed RAG with the fragment injection module (Figure 1). Also, we would like to kindly inform you that this work was released in arXiv on June 23, 2024, after the NeurIPS submission deadline, and thus should be considered as a concurrent work. --- Comment 2 The baselines used by the authors are not the latest. The authors should consider incorporating LLM-related molecular generation methods into the comparison. Response 2 We used extensive drug discovery tasks--23 PMO benchmark [5] tasks and 5 docking score tasks--that simulate various real-world drug discovery scenarios for a solid evaluation. In these tasks, we compared $f$-RAG against a large number of baselines. In Section 4.1, we employed the top-7 methods reported by the PMO benchmark and two latest SOTA methods, Genetic GFN and Mol GA, as our baselines (Line 213). Since there are 25 baselines reported in the PMO paper, this is equivalent to showing superiority of $f$-RAG over 27 baselines. In Section 4.2, we compared $f$-RAG with 14 baselines. For LLM-related methods, Wang et al. [3] was released in arXiv after the NeurIPS submission deadline and there is no publicly released codebase. RetMol [1] uses an LLM (a BART model) and is included as our baseline. We have made every effort to faithfully demonstrate the superiority of $f$-RAG through extensive experiments, and if you kindly suggest any strong baselines, we would be happy to compare $f$-RAG to them to make our experiments more robust. --- Comment 3 Why did the author use SAFE-GPT instead of other chemical language models? Response 3 We adopted Sequential Attachment-based Fragment Embedding (SAFE), non-canonical SMILES that represents molecules as a sequence of fragments, as the molecular representation. We chose SAFE because it is well-suited for fragment-based molecule generation, and it enables $f$-RAG to easily include the hard fragments in a generated molecule by simply providing them as an input sequence to SAFE-GPT to complete the rest of the sequence. Other language models besides GPT are also compatible with the $f$-RAG framework as long as they are trained on a SAFE dataset, and SAFE-GPT is chosen as an example since the pre-training of chemical language models is not the main interest of our paper. We emphasize that our proposed strategy of combining fragment retrieval from a goal-aware vocabulary with the SAFE representation to build a molecular optimization framework is simple but powerful, demonstrating its effectiveness in a wide range of drug discovery tasks. --- Comment 4 Some key details in genetic algorithms are missing. Response 4 As described in Line 181, we adopted the mutation/crossover of Graph GA [4] to further improve exploration in the chemical space. We appreciate your comment and will include the details about the genetic operations in the revised paper for the completeness of the paper. Specifically, in the crossover operation, parents are cut at random positions at ring or non-ring positions with a probability of 50%, and random fragments from the cut are combined to generate offspring. In the mutation operation, bond insertion/deletion, atom insertion/deletion, bond order swapping, or atom changes are performed on the offspring molecule with a predefined probability. We hope our response addresses your concerns and that you consider upgrading your rating. We are happy to elaborate further if there are any remaining concerns. --- References [1] Wang et al., Retrieval-based controllable molecule generation, ICLR, 2023. [2] Lee et al., Drug discovery with dynamic goal-aware fragments, ICML, 2024. [3] Wang et al., Efficient evolutionary search over chemical space with large language models, arXiv, 2024. [4] Jensen, A graph-based genetic algorithm and generative model/monte carlo tree search for the exploration of chemical space. Chemical science, 10(12):3567-3572, 2019. [5] Gao et al., Sample efficiency matters: a benchmark for practical molecular optimization, NeurIPS Datasets and Benchmarks, 2022.	Summary: The paper introduces f-RAG, a novel framework for fragment-based molecular generation that integrates hard and soft fragment retrieval and genetic fragment modification. It aims to improve the exploration-exploitation trade-off in drug discovery by leveraging existing molecular fragments and exploring beyond the existing chemical space. On each generation, two hard fragments are sampled which will be ensured to appear in the generated molecule. Then several soft fragments are sampled to derive an embedding via a pretrained chemical language model as the guidance for generation. Another round of genetic algorithm is also implemented to further explore the neighborhood of the generated high-scoring molecules. The authors have conducted extensive experiments on various molecular optimization tasks, demonstrating f-RAG's effectiveness in generating molecules with improved optimization performance, diversity, novelty, and synthesizability. Strengths: 1. The paper tries to tackle an important problem of fragment-based molecular generation, namely exploring further chemical spaces beyond known fragments. During generations, novel and high-scoring fragments will be dynamically updated into the vocabulary so that this kind of "novelty" will be passed on to further generations, enlarging the explorable chemical space. 2. The retrieval-augmented generation provide explainability to some extent, therefore has some ensurance on the quality of the generated molecules. Weaknesses: 1. Integrating both language models and genetic algorithms into generation might result in huge computational cost, which will affect the practical applications of generative models. The authors should discuss and analyze the computational efficiency as well as its trade-off with the performance. 2. Commonly the RAG process will retrieve related information from large-scale database (vocabulary) through some techniques (e.g. Vectorize the data and implement query-key matching). However, in the proposed method, both the hard and soft fragments are randomly sampled from the vocabulary, which leaves great burdens on the construction of high-quality and relevant databases for each task or property. 3. The implementation details are missing, especially for the baselines which incorporate genetic algorithms, whose performance is closely related to the values of parameters like population size and number of cycles. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why can all the molecules be decomposed into an arm-linker-arm form? Does this formalization impose additional biases on the chemical space which can be explored? For example, what will happen if I want to generate benzene with small substituents (e.g. toluene). Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The property-related fragment vocabulary can only be built if the property score can be decomposed into fragment-level sums, which limits the applications of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your comments. We appreciate your positive comments that our dynamic vocabulary update strategy expands the explorable chemical space and that our retrieval-augmented generation strategy provides chemical explainability. We address your concerns and questions below. --- Comment 1 Integrating both language models and genetic algorithms into generation might result in huge computational costs. The authors should analyze the computational efficiency. Response 1 First of all, we note that GAs generally do not require GPUs and are very fast to execute. For example, the GA adopted in our paper, Graph GA, takes only 3 minutes for a single run in the PMO benchmark (Table 7 in the benchmark paper [1]), while $f$-RAG takes 1~2 hours as described in Section D.4. During generation with $f$-RAG, the fragment retrieval augmentation and GA parts are very fast to run, and the forward pass of the backbone language model takes up most of the runtime. The slow runtime is a common limitation of autoregressive models, including the backbone model used in our paper, SAFE-GPT. Here, SAFE-GPT was chosen as an example because the pre-training of the backbone model is not the main interest of our paper, and the use of a non-autoregressive chemical language model can improve the runtime of the $f$-RAG framework. Furthermore, we also emphasize that the runtime of $f$-RAG of 1~2 hours to generate 10,000 drug candidates is sufficiently fast to be practical for real-world drug discovery problems, especially given its effectiveness in a wide range of drug discovery tasks. --- Comment 2 Both the hard and soft fragments are randomly sampled from the vocabulary, which leaves great burdens on the construction of high-quality and relevant databases for each task or property. Response 2 Instead of querying for relevant fragments every time a new molecule is generated, we proposed to construct an initial high-quality fragment vocabulary only once before generation. This strategy of confining the pool (or vocabulary) itself and randomly retrieving information is also commonly used in GAs. Through extensive experiments, we have demonstrated the initial vocabulary construction procedure based on the target property (Eq. (1)) is very simple yet effective, and universally applicable to any task or target property. Moreover, we proposed to dynamically refine the initial vocabulary with newly proposed fragments during generation. This gives $f$-RAG the ability to explore beyond the best of the database, as shown in Figure 6. --- Comment 3 The implementation details are missing, especially for the baselines which incorporate genetic algorithms. Response 3 For the experiments in Section 4.1 and Section 4.2, we did not reimplement any of the baselines. For the baselines in Section 4.1, the results of Genetic GFN [1] and Mol GA [2] are taken from the respective original papers and the results of other baselines are taken from the PMO benchmark paper [3]. For the baselines in Section 4.2, the results of RationaleRL, PS-VAE, RetMol, and GEAM are taken from Lee et al. [4] and the results of other baselines are taken from Lee et al. [5]. Implementation details for our $f$-RAG are included in Section D.2 and Section D.3. --- Comment 4 Why can all the molecules be decomposed into an arm-linker-arm form? Response 4 As described in Lines 117-118, arms are defined as fragments that have one attachment point and linkers are defined as fragments that have two attachment points in our paper. Therefore, all molecules with two or more breakable bonds (i.e., non-ring bonds in the arm-linker-arm slicing algorithm of Noutahi et al [6] we adopted) can be decomposed into arm-linker-arm forms, and we ignored the small number of molecules in the training set that cannot be decomposed. We will include this detail in the revised paper. However, we would like to mention that any other molecular decomposition algorithm is equally compatible with our $f$-RAG framework. --- Comment 5 The property-related fragment vocabulary can only be built if the property score can be decomposed into fragment-level sums. Response 5 The initial vocabulary construction procedure is based on the target property at the molecular level, and does not require the property score to be decomposed to the fragment level. $y$ in Eq. (1) is the target property value of the whole molecule, not a fragment. As explained in Lines 124-126, this scoring function evaluates the contribution of a given fragment to the target property of the molecule of which it is a part. We emphasize that the proposed fragment vocabulary construction scheme is very simple and universally applicable to any target property. --- References [1] Kim et al., Genetic-guided gflownets: Advancing in practical molecular optimization benchmark, arXiv, 2024. [2] Tripp et al., Genetic algorithms are strong baselines for molecule generation, arXiv, 2023. [3] Gao et al., Sample efficiency matters: a benchmark for practical molecular optimization, NeurIPS Datasets and Benchmarks, 2022. [4] Lee et al., Drug discovery with dynamic goal-aware fragments, ICML, 2024. [5] Lee et al., Exploring chemical space with score-based out-of- distribution generation, ICML, 2023. [6] Noutahi et al., Gotta be safe: a new framework for molecular design, Digital Discovery, 3(4):796–804, 2024. --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: Thank the authors for the detailed response, which has alleviated some of my concerns. I would like to maintain my recommendation of weak acceptance, and hope the authors good luck in addressing the concerns of the other reviewers.	Summary: This study proposes a fragment retrieval-augmented generation framework for molecular designs based on language models. The arm and linker vocabulary are constructed by fragments that have top average contribution to the given property. The hard fragments and a pool of soft fragments are retrieved from the vocabulary whose embeddings are fused and used for the generation. The generated molecules and the vocabulary are iteratively refined and augmented to further enhance the performance. The authors report competitive performance on multiple properties and generative tasks, including both single and multiple objectives. Strengths: This work formulates a novel framework for molecule designs and serves as a useful platform for future explorations. Though most of the techniques used in this study are not novel, the authors demonstrate effective ways of combining them to improve the performance. Specifically, the use of soft fragment pools and the genetic refinement make a good balance between exploration and exploitation. The authors provide comprehensive evaluation, comparison and ablation results, and also show competitive performance in multi-objective optimization, which is more relevant to real-world drug discovery applications. Weaknesses: The work is overall solid and I don't have major concerns except some minor questions. Technical Quality: 4 Clarity: 4 Questions for Authors: 1\. As the vocabularies are defined on the target property, is the fragment injection module also trained for each property? Would a universal model also work for the scenario? 2\. Line 242: the multi-objective optimization is performed using a unified score which is the product of all objectives. This may cause some problems such as high scores in one objective overshadowing others. How would this compare with, for example, selecting fragments that have higher scores for all three objectives? 3\. Table 3: as shown in Table 1, Genetic GFN does rather well in other tasks, so it would be better to also compare with it here. 4\. Line 275 and Table 8: the no-GA setting has substantially higher diversity than all other settings. This is somewhat counterintuitive as GA expands the fragment vocabulary and could potentially lead to more diverse and novel molecules. Thus, the no-GA setting should have lower diversity just like it has lower novelty. What could be the cause of this observation? 5\. Fig 5c: what are the actual values of the results? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have properly addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your positive comments that our paper demonstrates an effective strategy for balancing exploration and exploitation, provides a comprehensive evaluation in many drug discovery tasks simulating real-world scenarios, and is overall solid. We address your questions below. --- Comment 1 As the vocabularies are defined on the target property, is the fragment injection module also trained for each property? Would a universal model also work for the scenario? Response 1 No, the training of the fragment injection module is target property-agnostic. As described in Section 3.2, we proposed to train the fragment injection module using a self-supervised objective that predicts the soft fragment that is most similar to the input fragment. The fragments used in training are independent of the target property, while the fragment vocabulary used in generation is constructed using the scoring function of Eq. (1). The purpose of the training fragment injection module is not to solve a particular drug discovery task, but to teach the whole model how to retrieve and fuse useful information to guide its generation. We will clarify this in the revised paper. --- Comment 2 The multi-objective optimization is performed using a unified score which is the product of all objectives (line 242). How would this compare with selecting fragments that have higher scores for all three objectives? Response 2 This is a good point! First of all, we used the unified score (Eq. (4)) to strictly follow the setting in previous works [1,2] for a fair evaluation. Second, in the experiment, we found that the two settings, (1) using the unified score and (2) first filtering out fragments that do not meet the QED and SA constraints and scoring them, yielded similar fragment vocabularies (out of top-50 fragments, 49, 50, 48, 50, and 50 were identical for the parp1, fa7, 5ht1b, braf, and jak2 tasks, respectively). Therefore, we believe that there is little difference between the two scoring schemes. --- Comment 3 It would be better to compare with Genetic GFN in Table 3. Response 3 We additionally compare our $f$-RAG with Genetic GFN [3] in the below table. $f$-RAG outperforms Genetic GFN by a very large margin, demonstrating its superiority as a universal method applicable to various drug discovery tasks. We appreciate your suggestion and will include this result in the revised paper. Table: Novel top 5% docking score (kcal/mol) results. The results are the means and standard deviations of 3 independent runs. Lower is better. \| Method \| parp1 \| fa7 \| 5ht1b \| braf \| jak2 \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| Genetic GFN \| -9.227 ± 0.644 \| -7.288 ± 0.433 \| -8.973 ± 0.804 \| -8.719 ± 0.190 \| -8.539 ± 0.592 \| \| $f$-RAG (ours) \| -12.945 ± 0.053 \| -9.899 ± 0.205 \| -12.670 ± 0.144 \| -12.390 ± 0.046 \| -11.842 ± 0.316 \| --- Comment 4 The no-GA setting has higher diversity (Table 9). Response 4 We note that optimization performance and molecular diversity are conflicting factors, as methods that generate non-optimized molecules can naturally easily have high diversity [3,4] (as described in Line 284). As shown in Table 8, $f$-RAG without GA showed AUC top-10 sum of 14.048, significantly worse than $f$-RAG’s 16.928. Therefore, the high diversity with low optimization performance of $f$-RAG without GA does not mean that the model can generate high-quality, diverse molecules, but rather that the optimization is poor. As you mentioned, GA helps $f$-RAG explore the chemical space beyond the initial fragment vocabulary, leading $f$-RAG to find better chemical optima and generate optimized molecules, whereas $f$-RAG without GA performs worse at finding chemical optima and generates diverse but low-quality molecules. --- Comment 5 What are the actual values of the results in Fig 5c? Response 5 We report the actual values of Figure 5(c) here. We will include the below table in the revised paper. Table: PMO AUC top-10, top-100 diversity, top-100 novelty, and top-100 SA score results with different values of $\delta$ of the similarity-based fragment filter. \| Metric \| $f$-RAG \| $f$-RAG ($\delta$=0.8) \| $f$-RAG ($\delta$=0.6) \| $f$-RAG ($\delta$=0.4) \| \| --- \| --- \| --- \| --- \| --- \| \| Sum AUC \| 16.928 \| 16.648 \| 16.262 \| 15.765 \| \| Average diversity \| 0.532 \| 0.606 \| 0.681 \| 0.724 \| \| Average novelty \| 0.800 \| 0.778 \| 0.751 \| 0.796 \| \| Average SA score \| 2.026 \| 3.836 \| 3.825 \| 3.852 \| --- References [1] Lee et al., Exploring chemical space with score-based out-of- distribution generation, ICML, 2023. [2] Lee et al., Drug discovery with dynamic goal-aware fragments, ICML, 2024. [3] Kim et al., Genetic-guided gflownets: Advancing in practical molecular optimization benchmark, arXiv, 2024. [4] Gao et al., Sample efficiency matters: a benchmark for practical molecular optimization, NeurIPS Datasets and Benchmarks, 2022. --- Rebuttal 2: Comment: I appreciate the authors for the detailed response and updated results.	Rebuttal 1: Rebuttal: Dear reviewers, we sincerely appreciate your constructive comments. There were a number of comments that will help us strengthen our paper, and we will be sure to incorporate them into the revision. We have had a few questions about the difference of our proposed $f$-RAG compared to RetMol [1], so we would like to clarify this in the global response. Compared to RetMol, our proposed $f$-RAG is critically different in three aspects: (1) $f$-RAG retrieves fragments instead of molecules, enabling much more fine-grained generative guidance. There is a strong correlation between molecular structures and their activity, referred to as the structure-activity relationship (SAR) [2], which means that fragments are building blocks of molecules that critically contribute to their target chemical property. Therefore, utilizing fragments instead of whole molecules enables compositionally in generation and results in more effective and chemically intuitive guidance. (2) $f$-RAG utilizes two types of retrieval, i.e., hard and soft fragment retrieval, while RetMol only performs soft retrieval (of molecules). In this way, $f$-RAG can effectively balance between exploitation of current chemical knowledge and exploration in the chemical space. (3) $f$-RAG generates molecules in a one-shot manner, while RetMol relies on iterative refinement that uses retrieved guidance to refine noise over multiple iterations (80 iterations in the paper). This is a significant drawback for many drug discovery problems where oracle calls are expensive and oracle budgets must be considered. Due to these differences, $f$-RAG outperforms RetMol by a very large margin (Table 3). --- References [1] Wang et al., Retrieval-based controllable molecule generation, ICLR, 2023. [2] Crum-Brown et al., The connection of chemical constitution and physiological action. Trans R Soc Edinb, 25(1968-1969):257, 1865.	NeurIPS_2024_submissions_huggingface	2,024	Summary: Fragment-based drug discovery methods are limited in their exploration beyond existing database fragments, as they primarily reassemble or slightly modify the given fragments. This paper introduces a new approach, fragment retrieval-augmented generation (f-RAG), which retrieves two types of fragments—hard fragments and soft fragments—from a fragment vocabulary to achieve an improved exploration-exploitation to address this limitation. Strengths: This paper introduces a novel molecular generative framework that combines fragment-based drug discovery (FBDD) and retrieval-augmented generation (RAG). This paper proposes a retrieval augmentation strategy that operates at the fragment level, utilizing two types of fragments to provide fine-grained guidance. This approach aims to achieve a better exploration-exploitation trade-off and generate high-quality drug candidates. Weaknesses: 1. This paper claims the f-RAG approach improve the exploration-exploitation trade-off. However, there is no experiment demonstrate this point. 2. No limitation is discussed. 3. Novelty and contribution is a concern, from the RAG part, it seems the main difference compared to Want et al. [42] is f-RAG dealing with fragment instead of molecule. 4. In addition, the critical part, SAFE-GPT[34], is a previous work. 5. f-RAG is built on a pre-trained backbone molecular language model, and it relies heavily on the generation performance of this backbone. This also means that the method delegates the challenging task of molecule generation to a large model. Technical Quality: 2 Clarity: 3 Questions for Authors: see weakness. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: f-RAG is built on a pre-trained backbone molecular language model, and it relies heavily on the generation performance of this backbone. This also means that the method delegates the challenging task of molecule generation to a large model. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your comments. We appreciate your positive comments that our paper introduces a novel molecular generative framework that combines fragment-based drug discovery (FBDD) and retrieval-augmented generation (RAG). We address your concerns and questions below. --- Comment 1 This paper claims the f-RAG approach improves the exploration-exploitation trade-off, but there is no experiment demonstrating this point. Response 1 The balance between exploiting chemical knowledge and exploring in the chemical space is essential to achieve good performance in molecular optimization problems, and we have demonstrated the superiority of our proposed $f$-RAG through extensive molecular optimization experiments (Table 1 and Table 3). Furthermore, in addition to optimization performance, we also evaluated diversity, novelty, and synthesizability, other essential considerations in drug discovery. Here, diversity and novelty measure the ability of exploring the chemical space, and synthesizability measures the ability of exploiting chemical knowledge. As shown in Figure 1 and Table 2, $f$-RAG achieves the best optimization performance, diversity, and synthesizability, and the second best novelty. Overall, we can conclude that $f$-RAG exhibits the best balance between exploration and exploitation across these essential considerations, demonstrating its applicability as a promising tool for drug discovery. --- Comment 2 No limitation is discussed. Response 2 We discussed the limitations in Section A of the appendix. --- Comment 3 Novelty is a concern. In the RAG part, the main difference from Wang et al. [1] is that f-RAG deals with fragments instead of molecules. Response 3 Compared to RetMol, our proposed $f$-RAG is critically different in three aspects: (1) $f$-RAG retrieves fragments instead of molecules, enabling much more fine-grained generative guidance. There is a strong correlation between molecular structures and their activity, referred to as structure-activity relationship (SAR) [2], which means that there are important fragments in a given molecule that critically contribute to the target chemical property. Therefore, utilizing fragments instead of whole molecules results in more effective and chemically intuitive guidance. (2) $f$-RAG utilizes two types of retrieval, i.e., hard and soft fragment retrieval, while RetMol only performs soft retrieval (of molecules). In this way, $f$-RAG can effectively balance between exploitation of current chemical knowledge and exploration in the chemical space. (3) $f$-RAG generates molecules in a one-shot manner, while RetMol relies on iterative refinement that uses retrieved guidance to refine noise over multiple iterations (80 iterations in the paper). This is a significant drawback for many drug discovery problems where oracle calls are expensive and oracle budgets must be considered. Due to these differences, $f$-RAG outperforms RetMol by a very large margin (Table 3). --- Comment 4 f-RAG is built on a pre-trained backbone molecular language model, SAFE-GPT [3], and it relies heavily on the generation performance of this backbone. This also means that the method delegates the challenging task of molecule generation to a large model. Response 4 As we have mentioned in Limitations (Section A), our proposed $f$-RAG is built on a pre-trained backbone molecular language model, SAFE-GPT. This design choice that utilizes a pre-trained LLM is a very popular strategy across many domains [1,3]. This strategy is a large advantage rather than a disadvantage, as it lets the lightweight fragment injection module take care of the relatively easy task of fragment retrieval augmentation. As shown through the extensive experiments in Section 4, this strategy makes $f$-RAG a simple but powerful method to solve various drug discovery tasks. Furthermore, it also enables very efficient and fast training of $f$-RAG. As described in Section D.1, the fragment injection module, the only part of $f$-RAG that requires training, is very lightweight. The module has 2,362,368 trainable parameters, which correspond to only 2.64% of the total parameters of 89,648,640. As described in Section D.4, this allows us to train $f$-RAG in less than 4 hours using a single GeForce RTX 3090 GPU, while training of SAFE-GPT takes 7 days using 4 NVIDIA A100 GPUs [4]. Moreover, we emphasize that the high performance of $f$-RAG cannot be achieved by the backbone large model alone. For example, $f$-RAG without our proposed fragment retrieval showed AUC top-10 sum of 15.395 while the full $f$-RAG showed 16.928 (Table 8). Overall, we believe that combining the generative power of a large pre-trained model and a novel fragment retrieval augmentation strategy is an important contribution that can have practical impact. We appreciate your detailed feedback. We hope our response addresses your concerns and that you consider upgrading your rating. We are happy to elaborate further if there are any remaining concerns. --- References [1] Wang et al., Retrieval-based controllable molecule generation, ICLR, 2023. [2] Crum-Brown et al., The connection of chemical constitution and physiological action. Trans R Soc Edinb, 25(1968-1969):257, 1865. [3] Yu et al., Enzyme function prediction using contrastive learning, Science 379.6639: 1358-1363, 2023. [4] Noutahi et al., Gotta be safe: a new framework for molecular design, Digital Discovery, 3(4):796–804, 2024. --- Rebuttal Comment 1.1: Comment: Thank you for your response. My concerns regarding the contribution and novelty of this paper still remain, so I will keep my score unchanged.	null	null	null	null	null	null
Graph-enhanced Optimizers for Structure-aware Recommendation Embedding Evolution	Accept (poster)	Summary: In this paper, the authors propose a novel optimization algorithm that is talored for recommender systems. It incorporates graph structural information into the optimization process, aleviating the burden of performing GNN for RS. The convergence of the algorithm is theoretically demonstrated. Besides, it could be incorporated into existing well-performed optimizers like AdamW. The experiments are conducted to test its effectiveness on different types of recommendation models and consistent performance improvements are observed. Strengths: 1. This paper is innovative in its consideration of utilizing graph information the algorithm level, not model level, for recommender systems. The proposed new algorithm is theoretically guaranteed to converge. 2. The application of the proposed algorithm is widely discussed, including its incorporation into existing popular optimizers and its combination with knowledge distillation for recommender systems. 3. The experiments are extensively conducted, including different size of datasets, different types of recommendation backbones and baselines. The results demonstrate both the effectiveness and efficiency of the proposed algorithm for improving existing recommendation models. Weaknesses: 1. The technique in Section 2.4 is a little too specific for AdamW. 2. Table 5 seems to have an error. In the last row, 0.568 is not the best result. Is it just a typo? Or does +DKD not further improve the performance in this case? Technical Quality: 3 Clarity: 4 Questions for Authors: In Table 6, the proposed optimization algorithm even improves the GNN-based recommendation models. What is the reason for this phenomenon? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors have addressed the limitations in Section 5 and Appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for finding our work innovative. Each comment (presented in italics) is followed by its corresponding response. > W1: The technique in Section 2.4 is a little too specific for AdamW. Thank you for your feedback. AdamW was chosen as an example because it is the most widely used optimizer in recommender systems other than Adam. Similar modifications can also be implemented for other optimizers if they incorporate an exponential moving average momentum term (e.g., Adamax [R1], AMSGrad [R2], AdaBelif [R4]) and can decouple the weight decay regularization [R3]. > W2: Table 5 seems to have an error. In the last row, 0.568 is not the best result. Is it just a typo? Or does +DKD not further improve the performance in this case? \|Seed\| HR@1 \| HR@5\| HR@10\| NDCG@5\| NDCG@10\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|0 \|0.01744 \|0.04109 \|0.05755 \|0.02956 \|0.03489\| \|1 \|0.01646 \|0.04087 \|0.05643 \|0.02907 \|0.03404\| \|2 \|0.01655 \|0.04092 \|0.05666 \|0.02900 \|0.03406\| \|3 \|0.01641 \|0.04092 \|0.05612 \|0.02878 \|0.03370\| \|4 \|0.01619 \|0.03971 \|0.05746 \|0.02811 \|0.03381\| Thank you for pointing out this mistake. We reviewed the results (for the 5 seeds shown above) and found that '+DKD' indeed does not improve the HR@10 performance. Furthermore, the overall comparison in Table 2 indicates the common conclusion: the application of SEvo is primarily beneficial for the top-ranked targets. This is understandable as SEvo encourages the related nodes to be closer; conversely, it may adversely affect the targets being less related to the historical items. To address this problem, graphs describing multiplex relations should be introduced, which is left as future work as discussed in Section 5. We will further analyze this interesting observation in the revised manuscript. Thank you once again for your meticulous review. > Q1: In Table 6, the proposed optimization algorithm even improves the GNN-based recommendation models. What is the reason for this phenomenon? We believe that it is challenging to exploit both structural and sequential information for these GNN-based recommenders. LESSR and MAERec have developed sophisticated architectures for this purpose; however, they remain suboptimal in effectively utilizing either structural or sequential information alone. To be specific, the local session graph used by LESSR is insufficient to model the necessary structural and sequential information. In contrast, MAERec has improved its performance in this regard due to the use of two separate modules; however, effectively fusing these modules remains a challenging task. [R1] Kingma D. P., et al. Adam: A method for stochastic optimization. ICLR, 2015. [R2] Reddi S. J., et al. On the convergence of adam and beyond. ICLR, 2018. [R3] Loshchilov I., et al. Decoupled weight decay regularization. ICLR, 2019. [R4] Adabelief optimizer: Adapting stepsizes by the belief in observed gradients. NeurIPS, 2020.	Summary: This paper proposes Structure-aware Embedding Evolution (SEvo) to improve recommender systems by directly integrating graph structural information into embeddings. Unlike traditional methods, authors propose guide embedding update momentum with graph smoothing regularization. The proposed method can be integrated with a wide range of optimizers for neural networks, e.g., AdamW. The proposed method significantly increases performance metrics on several recommender datasets for standard models and outperforms GNNs. Strengths: * Novel intriguing view on the problem of preserving structural information for recommender systems * Plug-n-play design of the method so that it can be easily transferred to any recommendation architecture * Solid performance gains * Faster than training GNNs Weaknesses: * The paper is difficult to follow. It would be nice to see a simple summary of the SEvo pipeline in the form of a scheme or algorithm. * The SEvo formulae use the normalized adjacency matrix, so propagation of the gradients over the sampled node neighborhood is required. Suppose we have a large graph with a relatively high degree of each node. Mini-batch may have poorly correlated nodes, so many node embeddings should be updated simultaneously. This can lead to memory consumption issues and a notable increase in training time. * The tables report only the average time across different datasets. The detailed computational (or non-aggregated graphs with time) and space complexity is required to understand the ability of the method to scale * SEvo accelerates momentum only over first-order neighbors. However, for some graph-related tasks, it is critical to handle long-range dependencies. Technical Quality: 3 Clarity: 3 Questions for Authors: * How does SEvo work in mini-batch fashion? Does it require specific batch preparations, or should it be smaller on average? * How does the model training time scale with the size of the graph? Could you provide a graph epoch time vs. graph size e.g. for a SASRec? * Can high-order proximity / long-range dependencies be incorporated using SEvo? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: * The time and space complexity analyses of the method are required to understand its scalability. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for finding our work intriguing. Each comment (presented in italics) is followed by its corresponding response. > W1: The paper is difficult to follow. It would be nice to see a simple summary of the SEvo pipeline in the form of a scheme or algorithm. Thank you for your constructive feedback. The SEvo pipeline can be summarized as follows (the algorithms for SEvo-enhanced SGD/Adam/AdamW are detailed in Appendix B.1): 1. Compute gradients for embeddings; 2. Update moment estimates (with certain modifications discussed in Section 2.4); 3. Smooth the variations according to Eq. (7); 4. Update the embeddings. > W2: The SEvo formulae use the normalized adjacency matrix, so propagation of the gradients over the sampled node neighborhood is required. Suppose we have a large graph with a relatively high degree of each node. Mini-batch may have poorly correlated nodes, so many node embeddings should be updated simultaneously. This can lead to memory consumption issues and a notable increase in training time. As discussed in Appendix D.4, the complexity required for SEvo is comparable to that of the simplest GCN solely with neighborhood aggregation. Therefore, SEvo can be readily used for large graphs to which other GCNs can be applied. > W3/Q2: The tables report only the average time across different datasets. The detailed computational (or non-aggregated graphs with time) and space complexity is required to understand the ability of the method to scale. How does the model training time scale with the size of the graph? Could you provide a graph epoch time vs. graph size e.g. for a SASRec? We present the computational and memory costs below (more training and inference times have been detailed in Appendix D.4). The graph size increases from Tools to Clothing. \| \| Tools \| Beauty \| Electronics \| Clothing \| \|:-: \|:-: \|:-: \|:-: \|:-: \| \| #Users \| 16,638 \| 22,363 \| 728,489 \| 1,219,337 \| \| #Items \| 10,217 \| 12,101 \| 159,729 \| 376,378 \| \| #Edges \| 134,476 \| 198,502 \| 6,737,580 \| 11,282,445 \| \| Method (second/epoch) \| Tools \| Beauty \| Electronics \| Clothing \| \|:-: \|:-: \|:-: \|:-: \|:-: \| \| SR-GNN \| 63.03s \| 86.13s \| 1,653.53s \| 2,909.38s \| \| LESSR \| 36.65s \| 65.62s \| 1,846.30s \| 3,615.53s \| \| MAERec \| 169.20s \| 239.56s \| 15,464.64s \| 14,017.92s \| \| SASRec \| 1.76s \| 2.23s \| 19.94s \| 25.20s \| \| SASRec+SEvo \| 1.89s \| 2.35s \| 22.16s \| 33.47s \| \| Method (GPU memory) \| Tools \| Beauty \| Electronics \| Clothing \| \|:-: \|:-: \|:-: \|:-: \|:-: \| \| SR-GNN \| 1,214M \| 1,212M \| 2,056M \| 2,328M \| \| LESSR \| 1,618M \| 1,660M \| 19,868M \| 20,772M \| \| MAERec \| 1,952M \| 1,972M \| 4,664M \| 7,478M \| \| SASRec \| 2,064M \| 2,036M \| 2,510M \| 3,282M \| \| SASRec+SEvo \| 2,074M \| 2,046M \| 2,908M \| 4,080M \| We have the following observations: 1) Compared to vanilla SASRec, the implementation of SEvo incurs minimal computational and memory costs. 2) Due to the high sparsity of recommendation datasets, the additional cost increases acceptably as the graph size increases. 3) The costs associated with SEvo are negligible compared to other GNN-based sequence models, including SR-GNN, LESSR, and MAERec. Surprisingly, the training time for these models on Tools (the smallest dataset) greatly exceeds the time required for SEvo on Clothing (the largest dataset). This limitation hinders their application in real recommendation scenarios. > W4/Q3: SEvo accelerates momentum only over first-order neighbors. However, for some graph-related tasks, it is critical to handle long-range dependencies. Can high-order proximity / long-range dependencies be incorporated using SEvo? Thanks for the comment but there is a misunderstanding about this. SEvo can intrinsically handle long-range dependencies as reflected in Eq. (7). Moreover, as investigated in Appendix D.3 and Table 7, SEvo can benefit from long-range dependencies when $L > 1$, with better results when $L=$ 2 or 3. > Q1: How does SEvo work in mini-batch fashion? Graph sampling methods [R1, R2] can be applied in a manner similar to that employed for other GCNs. Furthermore, it is worth noting that graph partition approaches [R3, R4] are more appropriate here, as the items/entities in a mini-batch are not necessarily included in the graph utilized for SEvo. When the original graph is too large for practical applications, it can be pre-sliced into multiple subgraphs for subsequent training. Moreover, if these smaller subgraphs are mutually exclusive, parallel updates can be performed for better acceleration and accuracy. > Q1.5: Does it require specific batch preparations, or should it be smaller on average? Do you mean the sampled/sliced sub-graphs should be smaller on average? We think the bigger the better, if possible. > L1: The time and space complexity analyses of the method are required to understand its scalability. As discussed in Appendix D.4, the time complexity of SEvo is mainly determined by the arithmetic operations of $\mathbf{\tilde{A}}^l \Delta \mathbf{E}, l=1,2, \ldots, L$. Assuming that the number of non-zero entries of $\mathbf{\tilde{A}}$ is $S$, the complexity required is about $\mathcal{O}(LSd)$. On the other hand, the additional space complexity of SEvo is $\mathcal{O}(S)$ for the storage of the normalized adjacency matrix. Because the recommendation datasets are known for high sparsity (i.e., $S$ is very small), the actual overhead can be reduced to a very low level. [R1] Hamilton W. L., et al. Inductive representation learning on large graphs. NeurIPS, 2017. [R2] Zou D., et al. Layer-dependent importance sampling for training deep and large graph convolutional networks. NeurIPS, 2019. [R3] Chiang W., et al. Cluster-GCN: An efficient algorithm for training deep and large graph convolutional networks. KDD, 2019. [R4] Liu X., et al. Survey on graph neural network acceleration: An algorithmic perspective. IJCAI, 2022. --- Rebuttal Comment 1.1: Title: Acknowledgment of Clarifications and Updated Final Rating Comment: Thank you for your detailed and thoughtful rebuttal. I appreciate the time and effort you have taken to address my concerns. Having carefully considered your responses, I am satisfied with the clarifications and additional insights provided. Your explanations have resolved the issues I initially raised, and I now have a clearer understanding of the contributions and significance of your work. I appreciate your efforts and will be revising my final rating to reflect the improvements made. --- Reply to Comment 1.1.1: Comment: We are delighted to learn that our responses have addressed your concerns. We sincerely appreciate the time you spent reviewing our paper!	Summary: The paper introduces Structure-aware Embedding Evolution (SEvo), a novel embedding update mechanism for recommender systems. SEvo directly integrates graph structural information into embeddings, ensuring that related nodes evolve similarly with minimal computational overhead. This approach differs from traditional Graph Neural Networks (GNNs), which typically serve as intermediate modules. SEvo is designed to enhance existing optimizers, particularly AdamW, to improve recommendation performance by incorporating moment estimate corrections. Theoretical analysis confirms the convergence properties of SEvo, and experiments demonstrate consistent improvements across various models and datasets. Strengths: 1. The paper proposes a new method to enhance over smoothing during the backward pass. 2. The method can be naturally integrated with momentum-based optimizers Weaknesses: 1. Uncleared relationship related to recommender system 2. Key points need further explanation. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. I do not see the relationship between your method and recommendation task. Your method aims to change gradient direction based on graph topology. It is more suitable to study it under more general graph datasets. Recommendation task has no relationship with your method. 2. In line 116, the author mentioned “These two criteria inherently conflict to some extent”. Why are structure-aware and direction-aware inherently in conflict? Do you have any explanation on this point? 3. We have already enhanced smoothness during the forward-pass during neighborhood aggregation. Why do authors think it is necessary to further enhance it during the backward pass? Will it lead to more severe over smoothing problem? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: See weakness and questions Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for finding our work novel. Each comment (presented in italics) is followed by its corresponding response. > W1/Q1: Uncleared relationship related to recommender system. It is more suitable to study it under more general graph datasets. Recommendation task has no relationship with your method. Thank you for pointing out the lack of a clear statement of the motivation in this manuscript. We agree that technically SEvo can be applied to a wider range of graph datasets; however, it is more appropriate to discuss its application in a recommendation scenario for two practical considerations. 1. Embedding is particularly important in modern recommender systems and its quality directly affects the subsequent decisions. However, due to data sparsity [R2], millions of item embeddings cannot be consistently updated through a simple recommendation-driven loss function. To this end, SEvo-enhanced AdamW introduces a graph regularization framework for consistent embedding evolution while simultaneously modifying AdamW to effectively address the challenges posed by extremely sparse gradients. 2. Compared to embedding learning for general graph datasets [R1], a major challenge for recommendation is how to effectively injecting structural information while leveraging other types of information (e.g., sequential information). SEvo excels in this aspect as it has minimal impact on the forward process (this can be empirically demonstrated through comparisons with other GNN-based sequence recommenders). In summary, SEvo-enhanced AdamW is specifically designed to address the challenges of data sparsity and injecting multiple types of information, which rarely co-exist in general graph datasets. We will further emphasize these challenges in the revised manuscript. > Q2: Why are structure-aware and direction-aware inherently in conflict? Do you have any explanation on this point? Given an adjacency matrix $\mathbf{\tilde{A}}$, the smoothest direction is along its principal eigenvector $\mathbf{D}^{1/2} \mathbf{1}$, ensuring that $\mathcal{J}_{smoothness}(\mathbf{D}^{1/2}) = 0$. However, the region around this smoothest direction tends to be an infeasible descent direction. Therefore, we have to resort to Eq. (6) for a trade-off. > Q3: We have already enhanced smoothness during the forward-pass during neighborhood aggregation. Why do authors think it is necessary to further enhance it during the backward pass? Will it lead to more severe over smoothing problem? Thank you for your insightful comment. We agree that if the forward process already involves neighborhood aggregation, re-enhancing smoothness during the backward pass may exacerbate the over-smoothing issue. However, as stated in the introduction, SEvo is not intended to replace those sophisticated GNNs but rather to offer an easy-to-use and plug-and-play alternative for structural information learning. In addition, if the base model fails to exploit the structural information adequately during the forward pass (e.g., LESSR and MAERec), SEvo can still facilitate the learning (the comparison on Beauty as an example): \| \| HR@5 \| HR@10 \| NDCG@5 \| NDCG@10 \| \|:-: \|:-: \|:-: \|:-: \|:-: \| \| LESSR \| 0.0322 \| 0.0506 \| 0.0205 \| 0.0264 \| \| +SEvo \| 0.0405 \| 0.0625 \| 0.0267 \| 0.0338 \| \| Improv. \| 26.0% \| 23.5% \| 30.4% \| 27.9% \| \| MAERec\| 0.0424 \| 0.0662 \| 0.0269 \| 0.0346 \| \| +SEvo \| 0.0441 \| 0.0677 \| 0.0283 \| 0.0358 \| \| Improv. \| 4.0% \| 2.3% \| 4.9% \| 3.6% \| [R1] Chami I., et al. Low-dimensional hyperbolic knowledge graph embedding. arXiv preprint, 2020. [R2] Chen Z., et al. A systematic literature review of sparsity issues in recommender systems. TORS, 2024. --- Rebuttal Comment 1.1: Title: Keep score unchanged Comment: I have acknowledged the rebuttal from authors and keep my score unchanged. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your time and valuable comments. If you have any further questions please let us know.	Summary: This paper proposes SEvo, an embedding updating mechanism that directly injects the graph information into the optimization process. This paper points out two critical criteria for directly injecting graph structure information into the embedding updating process for recommendation. Based on the proposed two criteria, this paper makes efforts to derive a solution named SEvo for injecting the graph structure information directly. SEvo is model-agnostic and can be implemented in various optimizers. The experiments are detailed, and the algorithm is theoretically guaranteed. Strengths: 1. This paper is well-motivated and well-organized, making this paper easy to understand. This paper first proposes two criteria and derives the final form of SEvo. I appreciate the efforts of the authors to make this process so clear. 2. The experiments are detailed. Experiments on various datasets showcase the effectiveness of SEvo, with detailed ablation studies. 3. The proposed method is easy to implement. SEvo can be integrated into various optimizers without complex modification. Weaknesses: The reason behind the success of SEvo on large-scale datasets remains unclear. I am extremely curious about this. The improvement is unbelievably huge, making me doubt the reported results. A level of 5% in practice industrial application is extremely huge. However, the experiment results show that SASRec equipped with SEvo performs twice as well as the vanilla SASRec. I believe that if the reported results are true, this performance even exceeds the SOTA method by a large margin since SASRec is still a strong baseline in practice. Has the author carefully tuned the base model? Technical Quality: 3 Clarity: 4 Questions for Authors: I am curious about the differences in gradient descent between SEvo and methods that explicitly model the neighborhood relationship with graph structure (e.g., LightGCN). The modified embedding updating mechanism in SEvo also includes components of the adjacent matrix, which looks similar to the graph propagation mechanism in GNN-based methods. Could you please give an example showcasing the differences in gradient descent (e.g., comparing the gradient descent processes of LightGCN and MF-BPR+SEvo)? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: See weaknesses and questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for finding our work well-motivated and well-organized. Each comment (presented in italics) is followed by its corresponding response. > W1: The reason behind the success of SEvo on large-scale datasets remains unclear. I am extremely curious about this. The improvement is unbelievably huge, making me doubt the reported results. A level of 5% in practice industrial application is extremely huge. However, the experiment results show that SASRec equipped with SEvo performs twice as well as the vanilla SASRec. I believe that if the reported results are true, this performance even exceeds the SOTA method by a large margin since SASRec is still a strong baseline in practice. Has the author carefully tuned the base model? Thank you for your constructive feedback. We did carefully tune the base model and the following hyper-parameters were grid-searched: \| Parameter \| Range \| \|:-: \|:-: \| \|Learning rate \| \{1e-4, 5e-4, 1e-3, 5e-3\} \| \| Weight decay \| [0, 1e-4] \| \| Dropout rate \| [0, 0.4] \| \| Batch size \| \{1024, 2048, 4096\} \| The best checkpoint, determined by the validation metric, was used for the final comparison. In the attached PDF within the global response, we illustrate partial results in terms of the Electronics dataset. We were as surprised as the reviewers by SEvo's success with these large-scale datasets. At this point, the following observations may provide valuable insights. - The training process of the base model exhibits increased instability on larger-scale datasets. This can be attributed to the sampling randomness: only a small fraction of items are sampled for training within a mini-batch, resulting in the remaining embeddings receiving zero gradients and being updated along the outdated and inconsistent directions. This problem gets worse as the dataset grows larger and sparser. SEvo and the specific modifications to AdamW (Section 2.4) have a positive impact on this issue. In this scenario, even the embeddings of highly inactive items will be updated appropriately throughout the training process. - In real industrial applications, this performance gap may not be as large because the embeddings are typically trained more adequately beyond a link prediction task [R2]. Hence, the data sparsity problem can be greatly alleviated. Moreover, the rich side information (e.g., attributes [R1], behaviors [R3]) can further minimize the impact of poor embedding quality. > Q1: Could you please give an example showcasing the differences in gradient descent (e.g., comparing the gradient descent processes of LightGCN and MF-BPR+SEvo)? Thank you for your insightful question. We investigate the gradient descent process of LightGCN and find an interesting connection to SEvo. For a $L$-layer LightGCN, it can be formulated as follows $$ \mathbf{F} = \psi (\mathbf{E}) := \sum_{l=0}^L \alpha_l \mathbf{\tilde{A}}^l \mathbf{E}, $$ where $\alpha_l (l=0, \ldots, L)$ represent the layer weights. According to the linear nature of the gradient operator, it can be obtained that $$ \nabla_{\mathbf{E}} \mathcal{L} = \psi (\nabla_{\mathbf{F}} \mathcal{L}). $$ Hence, denoted by $\zeta(\cdot)$ the gradient processing procedure of an optimizer, we can establish that LightGCN is identical to the following system ($\mathbf{F}(t) := \mathbf{F}_t$ due to OpenReview's inability to recognize '\_' in very long formulas): $$ \mathbf{F}(t) = \psi( \mathbf{E}(t) ) = \psi( \mathbf{E}(t-1) - \eta \Delta \mathbf{E}(t-1) ) = \psi( \mathbf{E}(t-1) ) - \eta \psi( \Delta \mathbf{E}(t-1) ) = \mathbf{F}(t-1) - \eta \psi \circ \zeta \circ \psi ( \nabla_{\mathbf{F}} \mathcal{L} ). $$ When $\zeta(\cdot)$ is an identity mapping (i.e., standard gradient descent), LightGCN is equivalent to MF-BPR with SEvo being applied twice at each update. However, when $\zeta(\cdot)$ is not an identity mapping (e.g., an optimizer with momentum or weight decay is integrated), they cannot be unified into a single system. Compared to explicit GNNs, SEvo is easy-to-use and has minimal impact on the forward pass, making it more suitable for assisting recommenders in simultaneously utilizing multiple types of information. These connections in part justify why SEvo can inject structural information directly. We will emphasize this point in the revised manuscript. [R1] Chen Q., et al. Behavior sequence transformer for e-commerce recommendation in Alibaba. DLP-KDD, 2019. [R2] Bai T., et al. A contrastive sharing model for multi-task recommendation. WWW, 2022. [R3] Yang Y., et al. Multi behavior hypergraph-enhanced transformer for sequential recommendation. KDD, 2022. --- Rebuttal Comment 1.1: Title: My concern has been addressed. Comment: Thank the author for addressing my concern. I will keep my rating. --- Reply to Comment 1.1.1: Comment: We thank you for the engagement with our work and for your effort during the rebuttal.	Rebuttal 1: Rebuttal: We would like to thank all reviewers for reviewing our paper and providing us with their insightful comments. We are excited that the reviewers found our work novel and well-written, and pleased that they were satisfied with both our theoretical analysis and experimental results. We do our best to respond to reviewers' comments so that the work is of a higher standard. Some of the major critiques are listed below; each reviewer's comments will be addressed point by point. 1. Reviewer Um97 expressed concerns regarding the substantial success of SEvo on large-scale datasets. We have thoroughly examined our reproducibility results and ensured that the comparisons are conducted fairly. 2. Reviewer DLFs thought that SEvo may apply to general graph datasets. We believe that both the motivation and design are specifically tailored to address the two major challenges of recommendation. 3. Reviewer feMV had concerns about the computational and space complexity of SEvo. We have carried out a complexity analysis supported by experimental evidence, which shows that SEvo is scalable and efficient compared to the simplest GCN. 4. Reviewer 9GVe carefully reviewed the paper and found a few typos. We have made modifications in the revised manuscript. Please check our rebuttal and let us know if you have any further questions. Pdf: /pdf/fac9e81be4c5e02f700585ae8b58ce89db998211.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback	Accept (poster)	Summary: The paper presents T2V-Turbo, a training strategy where additional reward models are introduced during the consistency distillation process to enhance the T2V consistency model's quality. With such enhancement, the trained model achieves favorable results in both VBench and human evaluations. Strengths: 1. The paper is presented clearly and easy to follow; 2. The supplementary material provides abundant background information and results; 3. According to the figures, the proposed method significantly boosts the video quality. Weaknesses: 1. Regarding the technical contribution, the proposed method seems like a straightforward extension of the video consistency model. The idea of adding direct supervision on the clean samples in consistency distillation has also been explored in previous works, e.g., [1]; 2. It's hard to evaluate the motion quality improvements according to the static frames in the paper. It's highly recommended to include the corresponding mp4 files; 3. I wonder which reward model contributes more to quality improvement from Fig.4 row 1 and Fig.4 row 3? This cannot be inferred from Tab.2 and it's recommended to present some qualitative results in the ablation study; 4. The paper lacks in-depth discussions on the choice and influence of the image/video reward models. For instance, would it be better to CLIP/GAN-discriminator compared to the human preference score? Why do MS and VC2 perform differently when combined with InternVidS2? Section 4.3 only provides empirical results. [1] Adversarial Diffusion Distillation Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why use different video reward models for ModelScope and VideoCrafter? 2. What leads to different hyperparameter choices in L164-165? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback! > Regarding the technical contribution, the proposed method seems like a straightforward extension of the video consistency model. The idea of adding direct supervision on the clean samples in consistency distillation has also been explored in previous works, e.g., [1] Adversarial Diffusion Distillation (ADD). We would like to clarify that ADD is not based on consistency distillation. Instead, ADD's distillation loss corresponds to score distillation sampling [2]. Moreover, adversarial training is prone to instability [3] and might require massive hyperparameter tuning [4]. In contrast, our methods optimize towards differentiable RMs enjoying a stable training process. To the best of our knowledge, we are the first to add direct supervision on the clean samples when distilling from a video diffusion generator. Additionally, we are the first to learn video generators from the feedback of video-text models. [2] Poole et al. "Dreamfusion: Text-to-3d using 2d diffusion." ICLR 2023. [3] Yue et al. "On the algorithmic stability of adversarial training." NeurIPS 2021 [4] Pang et al. "Bag of tricks for adversarial training." ICLR 2021 > It's highly recommended to include the corresponding mp4 files We appreciate the suggestions! We promise to include all corresponding mp4 files in our revised manuscript and create a website to better present the videos generated by our methods. > I wonder which reward model contributes more to quality improvement from Fig.4, row 1, and row 3? For Fig. 4, the image-text RM HPSv2.1 contributes more to our T2V-Turbo's (row 3) quality improvement over the baseline VCM (row 1). In the attached PDF, we empirically show that incorporating feedback from the $\mathcal{R}\_\text{vid}$ improves the video quality. > This cannot be inferred from Tab.2 and it's recommended to present some qualitative results in the ablation study First, we would like to clarify that Table 2 provides strong quantitative evidence for the effectiveness of each RM. For both VC2 and MS variants, leveraging feedback from $\mathcal{R}\_\text{img}$ alone (VCM + $\mathcal{R}\_\text{img}$) matches the Quality Score of our T2V-Turbo, indicating the similarly high visual quality of the generated videos. However, VCM + $\mathcal{R}\_\text{img}$ still falls behind our T2V-Turbo in terms of Semantic Score. Further incorporating feedback from $\mathcal{R}\_\text{vid}$ can bridge this gap, leading to better text-video alignment. The attached PDF corroborates the results in Table 2 with video examples. Specifically, it compares videos generated by VCM (VC2) + $\mathcal{R}\_\text{img}$ and our T2V-Turbo (VC2). The results show that while the visual quality of both methods is generally similar, additional feedback from $\mathcal{R}_\text{vid}$ significantly enhances the text-to-video alignment in T2V-Turbo. > The paper lacks in-depth discussions on the choice and influence of the image/video RMs. For instance, would it be better to CLIP/GAN-discriminator compared to the HPS? We thank the reviewer for the insightful question! Firstly, we would like to remind the reviewer that we have experimented with additional image-text RMs, including PickScore and ImageReward, as detailed in Appendix E. We qualitatively show that incorporating reward feedback from any of these image-text RMs leads to quality improvement over the baseline VCM (VC2). It is worth noting that HPSv2.1 and PickScore are fine-tuned from CLIP with human preference data. Therefore, learning from CLIP might not lead to better performance compared to learning from image-text RMs. We did not experiment with a GAN-discriminator, as training a GAN-discriminator can suffer from instability and may require substantial hyperparameter tuning. However, given the success of ADD, we acknowledge the potential benefit of learning from a GAN-discriminator. Consequently, we consider this a promising direction for future work. Lastly, we promise to include more in-depth discussions in our revised manuscript on the choice and influence of image/video RMs. > Why do MS and VC2 perform differently with InternVidS2? Firstly, we would like to emphasize that when training T2V-Turbo (MS) with InternVid2 S2, T2V-Turbo (MS) still shows improvement over both VCM (MS) and VCM (MS) + $\mathcal{R}\_\text{img}$. However, we acknowledge that ViCLIP is more effective in enhancing T2V-Turbo (MS) compared to InternVid2 S2, as demonstrated in Tables 2 and 3. In contrast, ViCLIP and InternVid2 S2 work similarly well for T2V-Turbo (VC2). We conjecture that this difference might be due to the quality disparity between the teacher models ModelScopeT2V and VideoCrafter2. For example, ModelScopeT2V suffers from low resolution, and the generated videos contain watermarks, which may limit VCM (MS)'s ability to learn from InternVid2 S2. Nonetheless, further tuning of the $\beta_\text{vid}$ parameter might help improve the performance of T2V-Turbo (MS) with InternVid2 S2. > Why use different video RMs for ModelScope and VC2? Our main purpose is to examine our methods with a diverse set of $\mathcal{R}\_\text{vid}$. As we did not have access to a video-text RM trained to reflect human preference on video, we instead chose to experiment with video foundation models, including ViCLIP and InternVid2 S2. Note that we perform an ablation study and report results with both ViCLIP and InternVid2 S2 in Table 3. > What leads to different hyperparameter choices in L164-165? Compared to VideoCrafter2, ModelScopeT2V has not been trained on high-quality video or image datasets. Additionally, ModelScopeT2V suffers from low resolution, and the generated videos contain watermarks. Thus, ModelScope suffers from lower video quality. As a result, T2V-Turbo (MS) requires larger weighting parameters $\beta_\text{img}$ and $\beta_\text{vid}$ to get stronger external supervision from the RMs to improve its generation quality. --- Rebuttal 2: Title: Follow-up the discussion Comment: Dear Reviewer Kq5y, Your feedback has been invaluable in helping us clarify, improve, and refine our work. We have diligently addressed your comments in our response, made every effort to dispel misunderstandings, and provided various video examples for qualitative comparisons to demonstrate the effectiveness of our method. We kindly ask you to revisit our paper in light of our response and consider whether the changes and clarifications we have provided might warrant a reconsideration of your rating. Best regards, The Authors --- Rebuttal 3: Title: New Results for Your Review! Comment: Dear Reviewer Kq5y, We would like to kindly invite you to review the results we have submitted during the rebuttal period. We have thoroughly addressed your concerns and included videos corresponding to the static frames in our paper. Thanks to the support of our Area Chair, we are able to share an [anonymous website](https://spangled-blanket-128.notion.site/Qualitative-results-of-T2V-Turbo-1290f9ec0fb34685918438d7ba590e83#69781d1bb2a048c69ce6c51441ecde50) containing all these videos. We hope that you will reconsider your ratings in light of these new results : ) Best regards, The Authors --- Rebuttal 4: Title: Re: Rebuttal Comment: I appreciate the authors' detailed rebuttal and the additional results. They address some of my concerns. However, my major concern about the technical contribution still exists. Especially when comparing Fig. 2 in this paper and Fig. 2 in VideoLCM, the major technical contribution is adding a reward loss on the clean samples, which seems straightforward without too many challenges. Therefore, I tend to maintain my original score. --- Rebuttal Comment 4.1: Title: Clarification on the technical contribution Comment: Dear Reviewer Kq5y, Thank you for your response. We want to further clarify our technical contributions. First, we regret not emphasizing the technical challenge of learning from a video-text reward model (RM) sufficiently in our original manuscript. Unlike learning from an image-text RM, obtaining feedback from a video-text RM $\mathcal{R}\_\text{vid}$ demands significantly more memory. For instance, using models like ViCLIP and InternVid2 S2 requires sampling a batch size of 8 frames from video clips. Since we work with a latent video generator, we must enable gradients during the decoding of these video frames from latent vectors to allow for passing from the $\mathcal{R}\_\text{vid}$ to the video generator. Consequently, this computational process becomes nearly impossible to fit within a 40GB A100 GPU if we also need to pass gradients through an iterative sampling process. To tackle this challenge, our method cleverly takes advantage of the single-step generation arising from consistency distillation, which is crucial for learning from $\mathcal{R}_\text{vid}$. Notably, even by operating with single-step generation, we almost fully utilize the 40GB memory of the A100 GPU. Additionally, we want to emphasize that the technical simplicity of our method should not be seen as a drawback. On the contrary, being technically simple yet highly effective is a distinct advantage of our approach. In this paper, we address the core challenges of video generation: 1) improving generation quality, 2) reducing inference time, and 3) alleviating the intensive memory and computational cost. Empirically, we achieve significant results by outperforming SOTA video-generation systems on standard benchmarks. We hope the reviewer can evaluate our paper based on our core scientific contribution. Looking forward to your response. Best regards, The Authors	Summary: This paper presents a distillation method for text-to-video models. In short, it builds upon latent consistency models (more specifically, adapting the paper "Reward Guided Latent Consistency Distillation" from image to video models). The method involves the usual consistency model objective, in addition to that feedback from both an image-text reward model (e.g. HPSv2.1) and a video-text model (e.g. InternVideo) is used to enhance the quality of the generated videos. Overall, the proposed approach not only is a able to few-step video generation (4/8), but is also outperforming existing open-source video generators such as ModelScope and VideoCrafter. Strengths: The biggest strength of the paper would definitely be the strong results. Having a 4 step model that is able to achieve results comparable/better to 50 step models is indeed commendable. In addition to this, the evaluation seems quite thorough: there are quantitative results on V-Bench with several categories, user studies comparing against standard video models, and also ablation studies which makes the paper quite sound. Finally, the presentation of the paper is also quite clear, everything is easy to grasp and understandable. Weaknesses: The most obvious weakness about this paper is that it is a direct application of Li et al. ("Reward Guided Latent Consistency Distillation") to video models. As far as I can understand, the changes involve replacing SD2.1 with VideoCrafter/ModelScope and incorporating an additional video-text RM in addition to the image-text RM. Similarly, the main difference I can see with Wang et al. 2023 (VideoLCM) is the utilization of the reward objectives. Given that previous works have shown the efficacy of Latent Consistency Models for video generation, and the benefits of reward models for LCM distillation in images, it's not entirely novel to combine these 2 ideas into a single distilled video model. [Minor] I don't find too many qualitative examples, and I couldn't find more videos in the supplementary (apart from the code). It would definitely be nicer if there were more examples that were provided. Technical Quality: 4 Clarity: 4 Questions for Authors: Overall, I definitely lean towards accepting the paper. Despite being "obvious", I think that having an open-weight model+implementation for this would definitely help research in the field. However, I would really like the authors to point out if I have missed out details in their contributions or any other novel aspects of their work, since that is the major shortcoming I see currently. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 2 Limitations: Limitations are sufficiently discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for the reviewer's positive feedback on our work. Please find our detailed response below. > I would really like the authors to point out if I have missed out details in their contributions or any other novel aspects of their work. We appreciate the opportunity to clarify our contributions. We would like to emphasize the importance of our mixture of RMs design, which enables us to achieve significant empirical results even without access to a $\mathcal{R}_\text{vid}$ trained to reflect human preference. Specifically, our method leverages feedback from both an image-text RM $\mathcal{R}\_\text{img}$ and a video foundation model $\mathcal{R}\_\text{vid}$, such as ViCLIP and InternVid2 S2. This combination allows our T2V-Turbo to break the quality bottlenecks in the video consistency model, resulting in both fast and high-quality video generation. Notably, the 4-step generations from our T2V-Turbo achieve the SOTA performance on VBench, surpassing proprietary models, including Gen-2 and Pika. Additionally, our ablation study in Table 2 empirically provides valuable scientific insights into the effectiveness of different RMs. While leveraging feedback from $\mathcal{R}\_\text{img}$ alone (VCM + $\mathcal{R}\_\text{img}$) is sufficient to match the visual quality (Quality Score) of our T2V-Turbo, the additional feedback from $\mathcal{R}\_\text{vid}$ further enhances text-video alignment, resulting in higher Semantic Score. Qualitative evidence is provided in the attached PDF. To the best of our knowledge, we are the first to improve video generation using the feedback from video-text RMs. We believe that future advancements in video-text RMs will further enhance the performance of our methods. Our joint training pipeline is computationally efficient. For example, InstructVideo requires over 40 hours of training to align a pretrained T2V model. In contrast, our method reduces training time to less than 10 hours, achieving a fourfold increase in efficiency in wall-clock training time. Lastly, our method is broadly applicable to a diverse set of prompts. We have empirically shown that our approach outperforms existing methods on comprehensive benchmarks, including VBench and EvalCrafter. In contrast, previous methods, such as InstructVideo, conduct experiments on a limited set of user prompts, most of which are related to animals. > I don't find too many qualitative examples, and I couldn't find more videos in the supplementary (apart from the code). It would definitely be nicer if there were more examples that were provided. Thank you for your suggestions! We have now included more videos in the attached PDF. We promise to include more videos in our revised manuscript and create a website to better present the videos generated by our methods. --- Rebuttal Comment 1.1: Title: Thanks for the clarification Comment: I thank the authors for providing additional videos, and also clarifying the contributions of the method. Based on this, I am now raising my score. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer 61hZ, Thank you for raising your score! Your feedback has been invaluable in helping us improve the presentation of our work. We are pleased that our rebuttal has clarified the contributions of our work. Thanks and best regards, The Authors	Summary: This paper aims to achieve a video consistency model with both fast and high-quality generation. Specifically, the authors introduce T2V-Turbo, which integrates feedback from a mixture of differentiable reward models into the consistency distillation (CD) process of a pre-trained T2V model. The differentiable reward models consists of an image-text reward model and a video-text reward model. Experiment results verify the effectiveness of proposed method. Strengths: - The paper is easy to follow and understand. - The idea is reasonable and simple. - Experiment results are good. Weaknesses: - The novelty is limited. The paper seems to simple combine the video consistency model and the differentiable reward models. Different from previous works, the authors utilize a mixture of reward models, where a video-text reward model is additionally added to encourage the diffusion model to better model the temporal dynamics. However, to me, this contribution is quite incremental. - For the paper presentation, I think it is better to show what the text prompt is in Fig. 1 to help reader better understand the difference between different models. - For the quantitative comparison, it can be seen that the proposed method actually doesn't perform the best for most of the evaluation metrics. I am not sure if comparing total score (i.e., a weighted sum of Quality Score and Semantic Score) only is enough to show the superiority of proposed method. Technical Quality: 2 Clarity: 3 Questions for Authors: See the weakness part. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive comments! >The paper seems to simple combine the video consistency model (VCM) and the differentiable reward models We emphasize that our method is NOT a simple combination of VCM and differentiable RM. Previous works focus on aligning a pretrained DM to the preference given by RMs. In contrast to previous methods, e.g., InstructVideo, aligning a pretrained T2V model by backpropagating gradients through the memory-intensive iterative sampling process, our method cleverly leverages the single-step generation that naturally arises from consistency distillation. By optimizing the rewards of the single-step generation, our method avoids the highly memory-intensive issues associated with passing gradients through an iterative sampling process. Additionally, our joint optimization technique is notably computationally efficient. Previous approaches, such as InstructVideo, require over 40 hours of training to align a pretrained T2V model. In contrast, our method reduces training time to less than 10 hours, achieving a fourfold increase in efficiency in wall-clock training time. Empirically, we have shown that our approach outperforms existing methods on comprehensive benchmarks, including VBench and EvalCrafter. In contrast, previous reward learning methods for video generation, such as InstructVideo, conducted experiments on a limited set of user prompts, most of which were related to animals. > the authors utilize a mixture of reward models, where a video-text reward model is additionally added to encourage the diffusion model to better model the temporal dynamics. However, to me, this contribution is quite incremental. We would like to clarify the contribution of our mixture of RMs design, which enables us to break the quality bottleneck in VCM without access to a $\mathcal{R}\_\text{vid}$ trained to mirror human preference. By learning from an image-text RM $\mathcal{R}\_\text{img}$ and a video foundation model $\mathcal{R}\_\text{vid}$, such as ViCLIP and InternVid2 S2, we achieve both fast and high-quality video generation. Notably, the 4-step generations from our T2V-Turbo achieve the SOTA performance on VBench, surpassing proprietary models, including Gen-2 and Pika. Additionally, our ablation study in Table 2 further provides valuable scientific insights into the effectiveness of different RMs. While leveraging feedback from $\mathcal{R}\_\text{img}$ alone (VCM + $\mathcal{R}\_\text{img}$) is sufficient to match the visual quality (Quality Score) of our T2V-Turbo, the additional feedback from $\mathcal{R}\_\text{vid}$ further enhances text-video alignment, resulting in higher Semantic Score. Qualitative evidence is provided in the attached PDF. To the best of our knowledge, we are the first to improve video generation using feedback from video-text RMs. We believe that future advancements in video-text RMs will further enhance the performance of our methods. > I think it is better to show what the text prompt is in Fig. 1 to help reader better understand the difference between different models. We thank the reviewer for the suggestion. We will include the corresponding prompts in our revised manuscript. The prompt for the top two and bottom two rows are 1. With the style of low-poly game art, A majestic, white horse gallops gracefully across a moonlit beach. 2. Kung Fu Panda posing in cyberpunk, neonpunk style > For the quantitative comparison, it can be seen that the proposed method actually doesn't perform the best for most of the evaluation metrics. I am not sure if comparing total score (i.e., a weighted sum of Quality Score and Semantic Score) only is enough to show the superiority of proposed method. First, we emphasize that our automatic evaluation results in Table 1 are corroborated by the human evaluations in Fig 3, where the 4-step generation from both T2V-Turbo (VC2) and T2V-Turbo (MS) are preferred over the 50-step generations from their teacher VideoCrafter2 and ModelScopeT2V. Second, we highlight that the Total Score, Quality Score, and Semantic Score are sufficient proxies for human preference. Consider the comparison between VideoCrafter2 and our T2V-Turbo (VC2). Figure 3 indicates that human annotators favor the 4-step generation from our T2V-Turbo (VC2) in terms of Visual Quality, Text-Video Alignment, and General Preference. These preferences align with the higher Quality Score, Semantic Score, and Total Score of our T2V-Turbo (VC2), thereby validating these metrics' effectiveness in reflecting our method's superiority. Lastly, we address why our T2V-Turbo does not perform the best across all evaluation metrics yet still achieves the highest overall scores. We extracted the performance of VideoCrafter2 and our T2V-Turbo (VC2) from Table 1 of our paper. Although VideoCrafter2 scores slightly higher on 5 out of 7 dimensions constituting the Quality Score and 4 out of 9 dimensions constituting the Semantic Score, the differences are not significant. In contrast, our T2V-Turbo (VC2) significantly outperforms VideoCrafter2 in metrics such as `Dynamic Degree`, `Image Quality`, and `Multiple Objects`, contributing to its overall superiority. \| Models\|Total Score\|Quality Score\|Subject Consist.\|BG Consist.\|Temporal Flicker\|Motion Smooth.\|Aesthetic Quality\|Dynamic Degree\|Image Quality\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\| \| VideoCrafter2\| 80.44\|82.20\|96.85\|98.22\|98.41\|97.73\|63.13\|42.50\|67.22\| \| $\texttt{T2V-Turbo}$ (VC2)\|81.01\|82.57\|96.28\|97.02\|97.48\|97.34\|63.04\|49.17\|72.49\| \| Models\|Semantic Score\|Object Class\|Multiple Objects\|Human Action\|Color\|Spatial Relation.\|Scene\|Appear. Style\|Temporal Style\|Overall Consist.\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|VideoCrafter2\|73.42\|92.55\|40.66\|95.00\|92.92\|35.86\|55.29\|25.13\|25.84\|28.23\| \|$\texttt{T2V-Turbo}$ (VC2)\|74.76\|93.96\|54.65\|95.20\|89.90\| 38.67\|55.58\|24.42\|25.51\| 28.16\| --- Rebuttal Comment 1.1: Title: Follow-up the discussion Comment: Dear Reviewer MMwk, We greatly appreciate your insightful feedback, which has significantly contributed to the clarity and enhancement of our work. We have carefully addressed your comments in our response, worked to resolve any misunderstandings, and included multiple video examples for qualitative comparisons to illustrate the effectiveness of our method. We kindly request that you revisit our paper in light of our response and clarifications, and consider whether these updates might lead to a reevaluation of your rating. Best regards, The Authors --- Rebuttal Comment 1.2: Title: Re: Rebuttal Comment: Thanks for the authors' detailed response. However, the response still does not clarify my concerns about the novelty. To me, the method proposed in this paper is a combination of VCM and differentiable RM. I understand that compared with differentiable RM methods, it is different as it introduces VCM instead of using conventional diffusion models. I also understand that combining the two techniques does bring benefits and achieve superior results. However, I didn't see this paper solving any novel technical issues when combining the two techniques. To me, the engineering significance of this paper is greater than its technological innovation. Thus, I tend to keep my original rating. --- Reply to Comment 1.2.1: Title: Clarification on our technical significance Comment: Dear Reviewer MMwk, Thank you for your response. We want to clarify the technical issues we solved in this paper and emphasize our technical significance. 1. Traditional reward learning methods for video generation, such as InstructVideo, suffer from intensive memory costs due to backpropagating gradients through a diffusion model (DM)'s iterative sampling process. Our method solves this issue by leveraging the single-step generation arising from consistency distillation. 2. Our method significantly reduces the training time required to align a video generation model, achieving a 4x increase in efficiency in wall-clock training time compared to InstructVideo. We also wish to address a major understanding. Our approach does NOT simply replace the DM with a VCM from a conventional reward learning method. A straightforward combination of VCM and differentiable RM requires either 1) distilling a VCM from a pretrained DM and subsequently aligning the VCM with a differentiable RM, or 2) aligning a DM with a differentiable RM before distilling a VCM from the aligned DM. Neither of these sequential methods achieves the same memory reduction as our approach because they still involve backpropagating gradients through the iterative sampling processes of either the DM or the VCM. Furthermore, these methods are not as computationally efficient as ours due to their two-phase training pipelines. We sincerely appreciate your feedback and hope that this clarification will encourage you to reconsider your evaluation of our work. Best regards, The Authors --- Rebuttal 2: Title: Further clarification Comment: Dear Reviewer MMwk, Thanks for your response. Firstly, we would like to emphasize that the technical simplicity of our method should not be seen as a drawback. On the contrary, being technically simple yet highly effective is a distinct advantage of our approach. In this paper, we address the core challenges of video generation: 1) improving generation quality, 2) reducing inference time, and 3) alleviating the intensive memory and computational cost when aligning a video generator. Current state-of-the-art proprietary video generation systems, such as Gen-2 and Pika, require several minutes to produce a short video clip. In contrast, our T2V-Turbo can generate high-quality videos within 5 seconds, making it significantly more suitable for real-time applications. T2V-Turbo achieves both fast and high-quality video generation. As validated on VBench, its 4-step generation process surpasses the performance of proprietary systems like Gen-2 and Pika, which rely on extensive resources. To the best of our knowledge, we are the first to simultaneously address these two contradictory aspects—speed and quality—within the same video-generation framework. Secondly, we would like to reiterate the significance of our mixture of RMs design, which allows us to break the quality bottleneck in VCM without access to a $\mathcal{R}\_\text{vid}$ trained to mirror human preference. Our ablation study in Table 2 empirically provides valuable scientific insights into the effectiveness of different RMs. While leveraging feedback from $\mathcal{R}\_\text{img}$ alone (VCM + $\mathcal{R}\_\text{img}$) is sufficient to match the visual quality (Quality Score) of our T2V-Turbo, the additional feedback from $\mathcal{R}\_\text{vid}$ further enhances text-video alignment, resulting in higher Semantic Score. We provide qualitative evidence in the attached PDF. To the best of our knowledge, we are the first to improve video generation with feedback from video-text RMs. We believe that future advancements in video-text RMs will further improve the performance of our methods. Thank you again for your time and effort in providing valuable feedback on our work. We hope our clarifications will lead to a more favorable evaluation of our contribution. Best, The Authors --- Rebuttal Comment 2.1: Title: On the technical difficulty of learning from a video-text RM Comment: Dear Reviewer MMwk, We would like to bring your attention to the technical difficulty of learning from a video-text RM. We regret not emphasizing this challenge sufficiently in our original manuscript. Unlike learning from an image-text RM, obtaining feedback from a video-text RM $\mathcal{R}\_\text{vid}$ demands significantly more memory. For instance, using models like ViCLIP and InternVid2 S2 requires sampling a batch size of 8 frames from the generated video clips. Since we work with a latent video generator, we must enable gradients while decoding these video frames from latent vectors to allow for passing from $\mathcal{R}\_\text{vid}$ to the video generator. Consequently, this computational process becomes nearly impossible to fit within a 40GB A100 GPU if we also need to pass gradients through an iterative sampling process. Our method addresses this challenge by cleverly leveraging single-step generation from consistency distillation. This approach is crucial, as even with single-step generation, we nearly max out the 40GB memory of the A100 GPU—let alone handling gradients through an iterative sampling process. As the discussion period draws to a close, we kindly ask the reviewer to reconsider their rating of our paper by focusing on our core scientific contributions. Best regards, The Authors	Summary: The paper introduces T2V-Turbo to enhance the quality of video consistency models in text-to-video generation. The authors address the slow sampling speed of diffusion-based T2V models and the low quality of generated video by integrating feedback from a mixture of differentiable reward models into the consistency distillation (CD) process of a pre-trained T2V model. This integration allows for the optimization of single-step generations, bypassing the memory constraints of backpropagating gradients through iterative sampling processes. T2V-Turbo demonstrates significant improvements in both speed and quality, achieving high performance on the VBench benchmark and surpassing leading models such as Gen-2 and Pika. Strengths: - Efficiency and Quality: T2V-Turbo achieves impressive results with 4-step generations, offering a tenfold acceleration in inference speed while improving video quality compared to 50-step DDIM samples from teacher models. - Comprehensive Evaluation: The authors conduct extensive experiments, including automatic evaluations on VBench and human evaluations with 700 prompts from EvalCrafter, to validate the effectiveness of T2V-Turbo. Weaknesses: - Technical contribution: The proposed approach is simply a mixture of previous works, i.e., consistency distillation and reward feedback learning, which deflates the technical contribution of the paper. However, I acknowledge that this is one of the pioneering approaches in text-to-video generation. - Reward models: While the proposed method heavily depends on the reward models, the reward models that the paper used are actually not designed to function as reward models. This is because of the absence of video-text reward models in comparison to image generation (as the author mentioned in their limitations). However, it is a more right approach to first design a good reward model for video generation than developing a video generation model to align with such reward models. Technical Quality: 3 Clarity: 4 Questions for Authors: - Regarding reward models for video generation. Could the author provide more details about how the reward models should be designed in order to obtain better video generation or evaluating video generation models? - It seems like only consistency distillation for VC2 degrades the performance (Table 1), while the score remains the same for ModelScope. What if we do not use consistency distillation, and only use reward feedback learning for VC2 or MS? Then this may have a higher score than T2V-Turbo? Suppose it is true, then what if we distill after the feedback learning? To summarize my points, does the joint training of feedback learning and consistency distillation is better than sequential fine-tuning? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: The paper highlights the limitation of relying on existing video-text reward models that are not explicitly trained to reflect human preferences on video-text pairs. Instead, the authors use video foundation models like ViCLIP and InternVid S2 as substitutes. While incorporating feedback from these models has enhanced T2V-Turbo’s performance, the authors acknowledge that the development of more advanced video-text reward models could further improve the results. Additionally, the complexity of integrating mixed reward feedback into the CD process could hinder the broader adoption and scalability of the approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback on our work! Please find our detailed response below. > Q1: The proposed approach is simply a mixture of previous works, i.e., consistency distillation and reward feedback learning We emphasize that our method is NOT simply combining the video consistency model and reward feedback learning. Previous methods, such as InstructVideo [1], require backpropagating gradients through an iterative sampling process, which can lead to substantial memory costs. In contrast, our method cleverly leverages the single-step generation arising from consistency distillation. By optimizing the rewards of the single-step generation, our method avoids the highly memory-intensive issues associated with passing gradients through an iterative sampling process. Additionally, we would like to emphasize the contribution of our mixture of RMs design, which enables us to break the quality bottleneck in VCM without access to a $\mathcal{R}\_\text{vid}$ trained to mirror human preference. By learning from an image-text RM $\mathcal{R}\_\text{img}$ and a video foundation model $\mathcal{R}\_\text{vid}$, such as ViCLIP and InternVid2 S2, we achieve both fast and high-quality video generation. Notably, the 4-step generations from our T2V-Turbo achieve the SOTA performance on VBench, surpassing proprietary models, including Gen-2 and Pika. To the best of our knowledge, we are the first to improve video generation using feedback from video-text RMs. We believe that future advancements in video-text RMs will further enhance the performance of our methods. [1] Yuan et al., InstructVideo: Instructing Video Diffusion Models with Human Feedback. CVPR 2024 > Q2: However, it is a more right approach to first design a good reward model for video generation than developing a video generation model to align with such reward models. We fully recognize the importance of designing a good video-text RM $\mathcal{R}\_\text{vid}$. However, creating an effective $\mathcal{R}\_\text{vid}$ might require long-term efforts. A video is more complex than an image due to its additional temporal dimension. We envision that it requires multiple iterations to derive an effective $\mathcal{R}\_\text{vid}$. Specifically, one iteration involves 1) Training $\mathcal{R}\_\text{vid}$ by collecting new preference data from a video generator and 2) Training the video generator to align with $\mathcal{R}\_\text{vid}$. Our method provides an efficient way to accomplish the second step. On the other hand, even without a $\mathcal{R}_\text{vid}$ trained to reflect human preferences on video-text pairs, we highlight that our method can still enhance the generation quality of a T2V model by aligning it with video-text foundation models, such as ViCLIP and InternVid2 S2. > Q3: Could the author provide more details about how the reward models should be designed in order to obtain better video generation or evaluating video generation models? We thank the reviewer for the question. In addition to our response to Q2, we conjecture a $\mathcal{R}_\text{vid}$ effective for training video generators might need to be finetuned from existing video foundation models, such as ViCLIP and InternVid2 S2. This approach mirrors the success seen with image-text RMs, including HPSv2, ImageReward, PickScore, and AestheticScore, which have proven effective for training image generators. These models are finetuned from image-text foundation models, such as CLIP and BLIP, using human preference data. Moreover, learning a multi-dimensional reward to reflect fine-grained human preferences might also be helpful. For example, each dimension of the reward vector could be trained to reflect visual quality, transition dynamics, text-to-video alignment, etc. > Q4: Does the joint training of feedback learning and consistency distillation is better than sequential fine-tuning? First, sequential fine-tuning is computationally more expensive than joint training. According to the InstructVideo paper, their reward feedback learning phase costs more than 40 hours. If we add the distillation time cost further, the total time cost for sequential fine-tuning can easily exceed 50 hours. Conversely, our joint training only requires less than 10 hours, representing a fivefold reduction in terms of training wall-clock time. Second, as mentioned in Sec. 1 of our paper, finetuning a diffusion T2V model towards a differentiable RM requires backpropagating gradients through the diffusion model's iterative sampling process. Therefore, calculating the full reward gradient is prohibitively expensive, resulting in substantial memory costs. Conversely, our method leverages the single-step generation that arises naturally from computing the CD loss, effectively bypassing the memory constraints. > Q5: It seems like only consistency distillation for VC2 degrades the performance (Table 1) In terms of the performance degradation of distilling VC2, we conjecture it can be alleviated by performing full model training, or further optimizing the hyperparameters related to learning the LoRA weights. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed rebuttal and the extensive qualitative examples provided. I acknowledge the computational burden involved in T2V generation and the meticulous nature of evaluating video generation models. This paper contributes valuable strategies for improving T2V generation and offers a robust framework for evaluating video generation models, which will be beneficial for other researchers in the field. After careful consideration, I will maintain my original score. --- Reply to Comment 1.1.1: Comment: Dear Reviewer x94N, Thank you again for your positive feedback on our work! Best regards, The Authors	Rebuttal 1: Rebuttal: We appreciate the reviewers for their time and constructive feedback on our work. We have responded to individual reviews below and would like to highlight additional qualitative results in the attached PDF. Please download and open it with Adobe Acrobat to click and play the videos. The attached PDF corroborates the results in Table 2 with video examples. To demonstrate the effectiveness of our mixture of RMs, we compare 8 pairs of videos generated by our $\texttt{T2V-Turbo}$ and $\text{VCM}$ + $\mathcal{R}\_\text{img}$. Due to space constraints, we focus on the VC2 variants and did not include results for $\text{VCM}$ and $\text{VCM}$ + $\mathcal{R}_\text{vid}$, as our $\texttt{T2V-Turbo}$'s results are significantly better. Additionally, we include videos corresponding to Figure 4 of our paper to better showcase the superiority of our $\texttt{T2V-Turbo}$ over its teacher model and the baseline $\text{VCM}$. We would also like to re-iterate the contributions of our methods: 1. To the best of our knowledge, we are the first to improve video generation using feedback from video-text RMs. 2. We emphasize the importance of our mixture of RMs design, which enables us to break the quality bottlenecks in VCM without access to a $\mathcal{R}\_\text{vid}$ trained to mirror human preference. The 4-step generations from our $\texttt{T2V-Turbo}$ set the SOTA performance on VBench, surpassing proprietary models like Gen-2 and Pika. 3. Our training pipeline is NOT a simple combination of VCM and differentiable reward. Our method cleverly leverages the single-step generation arising from consistency distillation, avoiding the need to backpropagate gradients through the memory-intensive iterative sampling process required by traditional methods, such as InstructVideo. 4. Our method is notably computationally efficient. For example, InstructVideo requires over 40 hours of training to align a pretrained T2V model. In contrast, our method reduces training time to less than 10 hours, achieving a fourfold increase in efficiency in wall-clock training time. 5. Our method is broadly applicable to a diverse set of prompts. We have empirically shown that our approach outperforms existing methods on comprehensive benchmarks, including VBench and EvalCrafter. In contrast, InstructVideo's experiments are conducted on a limited set of user prompts, most of which are related to animals. Pdf: /pdf/7c6b98e633a8a6e291ba7dee8352ce6363735785.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper proposes T2V-Turbo, a model aiming to achieve both fast and high-quality text-to-video generation by breaking the quality bottleneck of a video consistency model (VCM). It integrates mixed reward feedback from one image and one video reward model into the consistency distillation process of a teacher T2V model. The 4-step generations from T2V-Turbo outperform SOTA methods on the VBench benchmark and are favored by humans over the 50-step DDIM samples from the teacher model, achieving over ten-fold inference acceleration with quality improvement. Strengths: - The paper is well-structured and clearly presents the problem, the proposed method, the experimental setup, and the results. The use of figures and tables helps to illustrate the concepts and results. - The automatic evaluation results on the VBench benchmark and human evaluation results with the 700 prompts from EvalCrafter demonstrate the effectiveness of the proposed method, with T2V-Turbo outperforming baseline methods and proprietary systems in terms of total score and human preference. - The ability to generate high-quality videos quickly has significant impacts in various fields, such as digital art and visual content creation, and sets a new benchmark for future research in T2V synthesis. Weaknesses: - The method mainly combines the consistency distillation in the Video Consistency Model (VCM) with multiple reward models. Although it has achieved good results in solving the existing problems of the T2V model, this combination is relatively conventional and may be somewhat lacking in originality. - The citation format of this paper should adhere to the NeurIPS standard, the authors misuse \citet throughout the paper, especially in Sections 2 and 5, making it hard to read smoothly. Technical Quality: 4 Clarity: 3 Questions for Authors: - In Line 228, the authors write “InternVid2 S2 outperforms ViCLIP in several zero-shot video-text retrieval task”. While in Table 1, we know that ModelScopeT2V (MS) has less total score than VideoCrafter2 (VC2), which means that MS is a relatively weaker model than VC2, then why choose ViCLIP as the video RM for MS rather than using a stronger InternVid2 S2? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations of the work in the paper. They acknowledge the lack of an open-sourced video-text reward model trained to reflect human preferences on video-text pairs and discuss the potential use of a more advanced video reward model in the future. They also mention the concerns about misinformation and deepfakes raised by the ability to create highly realistic synthetic videos and commit to installing safeguards when releasing the models, such as requiring users to adhere to usage guidelines. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Title: Rebuttal by Authors Comment: We thank the reviewer for the positive feedback on our work! Please find our detailed feedback below. > The method mainly combines the consistency distillation in the Video Consistency Model (VCM) with multiple reward models. Although it has achieved good results in solving the existing problems of the T2V model, this combination is relatively conventional and may be somewhat lacking in originality. We would like to highlight that our method cleverly leverages the single-step generation arising from consistency distillation. By optimizing the rewards of the single-step generation, our method avoids the highly memory-intensive issues associated with passing gradients through an iterative sampling process. In contrast, previous methods, such as InstructVideo [1], require backpropagating gradients through diffusion model's iterative sampling process, resulting in substantial memory costs. Additionally, we emphasize the importance of our mixture of RMs design, which enables us to break the quality bottleneck in VCM without access to a $\mathcal{R}\_\text{vid}$ trained to mirror human preference. By learning from an image-text RM $\mathcal{R}\_\text{img}$ and a video foundation model $\mathcal{R}\_\text{vid}$, such as ViCLIP and InternVid2 S2, we achieve both fast and high-quality video generation. Notably, the 4-step generations from our T2V-Turbo achieve the SOTA performance on VBench, surpassing proprietary models, including Gen-2 and Pika. To the best of our knowledge, we are the first to improve video generation using feedback from video-text RMs. We believe that future advancements in video-text RMs will further enhance the performance of our methods. > The citation format of this paper should adhere to the NeurIPS standard, the authors misuse \citet throughout the paper, especially in Sections 2 and 5, making it hard to read smoothly. We thank the reviewer for pointing out the issues! We will make sure to fix the citation issues in our revised manuscript! > why choose ViCLIP as the video RM for MS rather than using a stronger InternVid2 S2? Our main purpose is to examine our methods with a diverse set of $\mathcal{R}\_\text{vid}$. As we did not have access to a $\mathcal{R}\_\text{vid}$ trained to reflect human preference on video, we instead chose to experiment with video foundation models, including ViCLIP and InternVid2 S2. Notably, we conduct a comprehensive ablation study in Table 3, presenting results for our T2V-Turbo (VC2) and T2V-Turbo (MS) when setting $\mathcal{R}\_\text{vid}$ to both ViCLIP and InternVid2 S2. This approach allows for a thorough assessment of our methods across different $\mathcal{R}\_\text{vid}$.	null	null	null	null	null	null
Improving Context-Aware Preference Modeling for Language Models	Accept (poster)	Summary: The paper focuses on fine-tuning LLMs to improve their ability to handle context-aware preference modeling. The authors address the challenge of the underspecified and ambiguous nature of natural language preferences by introducing a two-step modeling process. This includes selecting a context and evaluating preferences within that context. The approach is backed by the introduction of new datasets, named RPR, which are designed to test the effectiveness of context-specific preference modeling. The study provides extensive experimental evidence showing how context-aware models can surpass traditional methods in handling ambiguous and context-specific scenarios. Strengths: 1. The paper tackles a crucial issue in the realm of language modeling by enhancing the LLM's ability to understand and process user preferences in context-dependent scenarios. This is particularly important as LLMs are increasingly used in diverse real-world applications. 2. The introduction of the RPR datasets is a notable contribution, as these datasets specifically aim to disentangle context-specific preferences from general preferences, offering a valuable resource for further research. 3. The paper provides thorough experimental results that not only demonstrate the effectiveness of the proposed method but also explore various aspects of context-aware preference modeling, showing improvements over existing models like GPT-4 and Llama 3. Weaknesses: 1. The paper could benefit from testing the proposed method across a broader range of general benchmarks, such as MMLU and AGI-Eval, to assess how fine-tuning for context-aware preferences might affect the LLM's original capabilities or general applicability. 2. It remains unclear how applicable the proposed method is to other LMs, especially those of different sizes or from different series. Addressing this would help validate the robustness and versatility of the method. Technical Quality: 3 Clarity: 4 Questions for Authors: Please see my concerns in weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: n/a Flag For Ethics Review: ['Ethics review needed: Data privacy, copyright, and consent', 'Ethics review needed: Data quality and representativeness'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and feedback. Please find our responses below.   > The paper could benefit from testing the proposed method across a broader range of general benchmarks, such as MMLU and AGI-Eval, to assess how fine-tuning for context-aware preferences might affect the LLM's original capabilities or general applicability. We agree that training generative models with context-aware rewards is a great direction, and part of our broader agenda. In particular, we believe this could improve sensitivity to the system prompts and diverse users, and to prompts with multiple instructions or constraints, which is something we have observed state-of-the-art models struggle with. The design decisions here are not obvious though (where does context for MMLU come from, for example, especially given its more objective nature) and add another layer of complexity that we believe would be better suited for future work.   > It remains unclear how applicable the proposed method is to other LMs, especially those of different sizes or from different series. This is a good point. We took the opportunity to use RPR to finetune one of the stronger 2B parameter reward models according to the Reward Bench leaderboard, which is based on Gemma (hf:Ray2333/Gemma-2B-rewardmodel-baseline). We used the same hyperparameters as used for the 7B Mistral RM in the paper. We obtained the following context-conditioned results, which we will include the paper (Mistral results included for reference, best Gemma model bolded): \| \|\| Gemma-2B-NC \| Gemma-2B-CTX \| Gemma-CARM \| \\| \|Mistral-7B-CTX \| Mistral-CARM \| \| --------\|--------\|:-: \| :-: \| :-: \| :-: \| :-:\| :-: \| \| RPR Criteria \|\| 0.511 \| 0.761 \| 0.968 \| \\| \|0.867 \| 0.985 \| \| RPR Scenarios \|\| 0.511 \| 0.655 \| 0.909 \| \\| \|0.749 \| 0.962 \| \| Multifaceted Bench \|\| 0.508 \| 0.597 \| 0.681 \| \\| \|0.679 \| 0.787 \| \| Preference Bench \|\| 0.852 \| 0.861 \| 0.849 \| \\| \|0.915 \| 0.919 \| \| HHH (CTX) \|\| 0.751 \| 0.751 \| 0.760 \| \\| \|0.905 \| 0.919 \| \| Rewardbench (CTX)\| \| 0.718 \| 0.735 \| 0.786 \| \\| \|0.833 \| 0.871 \| \| Chatbot Arena (CTX) \|\| 0.745 \| 0.806 \| 0.873* \| \\| \|0.859 \| 0.909 \| Note: Multifaceted Bench is a new context-conditioned preference dataset — see general response. We observe that finetuning the Gemma RM shows similar patterns to finetuning the Mistral RM, and that the finetuned model is in many cases competitive with the larger Mistral base model.	Summary: This paper divides preference modeling into two steps: first estimating the user's intent then evaluate the generated text within the context of this intent. The paper makes the Reasonable Preference Reversal Datasets, which encompass criteria and scenarios for preference data. Experiments find that models can achieve higher performance when providing intent contexts during evaluation. Strengths: - The approach of first estimating user intent before evaluation may be promising. - This work constructs the first open-source context-conditioned preference datasets, which may be useful for future work. Weaknesses: 1. Lack of citations and comparisons to related work. For example, Li et al. [1] generate the criteria for evaluation and then get the final answer according to the criteria, which is similar to this work. 2. In Table 2, the author aims to demonstrate that fine-tuning a context-aware reward model significantly enhances context-specific performance. However, the observed improvement in Table 2 could be attributed to the model being trained on data from the same distribution as the test set. Even for new datasets that are not context-specific, training the model on such datasets will likely improve its performance on its test sets. 3. The lack of experiments illustrating that two-step preference modeling achieves better performance compared to traditional reward modeling on general preference datasets is a concern. In Table 4, the model is prompted by context generated by GPT-4, leading to unfair comparisons with similar models. To demonstrate that two-step preference modeling is superior, it is essential to show its advantages for models of equal ability. [1] Li J, Sun S, Yuan W, et al. Generative judge for evaluating alignment[J]. arXiv preprint arXiv:2310.05470, 2023. Technical Quality: 2 Clarity: 3 Questions for Authors: Why is two-step preference modeling necessary? The user's intent is inherently contained in the query, and traditional preference modeling implicitly evaluates whether the response aligns with the user's intent. Therefore, why should the estimated intent be treated independently? Additionally, ambiguity exists in estimating intent, as different individuals may have inconsistent interpretations of intent for the same query. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Please see Weakness part. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and feedback. Please find our responses below.   > citations / Li et al. [1] Thank you for bringing Li et al. to our attention. We agree it is closely related and will add it to the related work, and we will revisit our literature search for other recent work we may have missed. As part of our rebuttal, we have run their Auto-J model, but found that both Ultra RM and the base Mistral RM outperform it on every dataset/benchmark we use, both unconditioned and conditioned. We note that Auto-J, including the datasets used in the Auto-J paper, seeks to directly address the unconditioned preference modeling problem, even if Auto-J generates criteria as part of its judging process.   >Table 2 could be attributed to the model being trained on data from the same distribution as the test set. We agree this is true for RPR and will make this more clear in our paper (see, e.g., the stars in the new data ablation table of the PDF attachment). However, we note that our finetuned CARM also improves performance on all other context-specific datasets, including the newly added Multifaceted Bench (see General Response), as is now more clearly shown in Figure 1 of the PDF attachment.   > In Table 4, the model is prompted by context generated by GPT-4, leading to unfair comparisons with similar models We disagree that these comparisons are unfair. In particular, all models—not just our CARM—have access to the same additional context, hence their improved performance relative to the “NC” column.   > The lack of experiments illustrating that two-step preference modeling achieves better performance compared to traditional reward modeling on general preference datasets is a concern. We agree it would be ideal to show that two-step preference modeling can achieve better performance on general preference datasets. However, this requires two parts: (1) better context-specific modeling, and (2) strong context inference or specification. While we show in this work how we can improve the context-specific modeling problem, the context inference problem remains difficult. We have tried to generate context with respect to unconditioned queries as part of our work as well, and found that getting current models to generate useful context is quite challenging (for evidence of this, see the Auto-J results noted above), which is why we believe one of the next steps for future work is to investigate how to best obtain strong context supervision (as we allude to at L186-187). We think supervision of some kind is critical in order to go beyond the “user's intent is inherently contained in the query” as you wrote. Furthermore, we believe two-step preference modeling to be useful even if it could not improve on current unconditioned preference datasets such as Reward Bench. See next response.   > Why is two-step preference modeling necessary? … why should the estimated intent be treated independently? We think your last sentence here captures our motivation: “ambiguity exists in estimating intent, as different individuals may have inconsistent interpretations of intent for the same query.” Part of our argument is that intent need not be estimated solely from the prompt—indeed, it can be: - specified for annotators (see L84-91), or via default rules similar to Constitutional AI, or via the system prompt - inferred from past interactions in cases of persistent context (e.g. user profiles, L287; see revised Table in the attached pdf), or - learned from data via some context supervision system. To the extent that there are inconsistent interpretations of the same query, any such context will usefully disambiguate the query, which provides several advantages: - (a) Given context, we reduce reliance on unstated, implicit assumptions made by annotators, thereby increasing overall agreement (see Ziegler at al., quoted at L88). - (b) Using context we can improve steerability and pluralistic alignment (see Sorensen et al.). - (c) Making the context explicit may also be useful to diagnose errors made by models. - (d) Notably, making context explicit allows us to change the aggregation rule from a Borda count (see Siththaranjan et al.) to a more flexible Social Welfare Aggregation (as in Bakker et al., who find this to be effective as a consensus mechanism). In the last case, (d), agreement with respect to unconditional preferences gathered from a group (and thus, performance on general preference datasets) seizes to become a good measure, as Equation (2) [L158-159] (or modifications of Equation 2 for non-EU SWFs) will not hold, so that explicit contextualized aggregation (of human preferences) becomes a better alignment target than overall preference (see Siththaranjan et al. for a concrete example of this). --- Rebuttal 2: Title: Response to Authors Comment: Thank you for the authors' detailed response that solves some of my questions through additional experiments and explanations. However, my core question still exists, which is ``whether two-stage modeling is better than traditional reward modeling''. I understand that using experiments to illustrate this can be difficult, but it should be an important part. In real conversation and preference annotation scenarios, the intent of a query is ambiguous and has multiple possibilities. The two-stage approach may lead to better modeling results through decomposition, but it may also result in error accumulation. Moreover, as Weakness1 pointed out, previous work has already explored developing evaluation criteria before conducting evaluations. This results in a limited contribution of this paper. I think the author's research direction is promising, but there are still many points that can be improved. So, I keep my rating.	Summary: This paper points out that the preference label can be reversed by inserting additional context into the prompt. Based on this observation, the authors build a paired preference dataset. The authors also try to provide some theoretical analysis. Strengths: * This paper studies a specific and interesting problem in preference data. * The proposed data augmentation method may benefit preference optimization of LLMs and inspire further research. Weaknesses: My major concern lies in the experiments. The current experiments do not demonstrate the general advantage of using a paired dataset with reversed preference labels. On widely used preference datasets, the model trained on the proposed dataset does not outperform the baselines. It only performs better on the test set, which is built in the same way as the training set. Note that the preference labels on this test set are not verified by humans and could be unreliable. Technical Quality: 2 Clarity: 2 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: This paper does not have a limitation section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and feedback. Please find our responses below.   > The current experiments do not demonstrate the general advantage of using a paired dataset with reversed preference labels. Thank you. This is a good criticism. We realized this post submission and have prepared a data ablation that finetunes the base Mistral reward model on context-conditioned data from other distributions, including: - Preference Collection (PC) (hf:prometheus-eval/Preference-Collection), in which the context is highly correlated with unconditioned preference, and - a “one-sided” version of our RPR datasets, where kept only one of the two sides by having GPT-4 pick one of the two criteria or scenarios that it thought was more suitable for the given prompt. The results are shown in Table 1 in the PDF attachment, and demonstrate the advantage provided by our RPR dataset. We will add this to our paper. Notably, the Multifaceted Bench (MF), released after the NeurIPS submission deadline, introduced an additional context-conditioned preference benchmark that is outside of our training distribution, where you can see that training on RPR strongly improves context-conditioned performance over the base model (and the PC model).   > On widely used preference datasets, the model trained on the proposed dataset does not outperform the baselines. If you are referring to the no context versions of HHH, Reward Bench, and Chatbot Arena, these are unconditional preference benchmarks, where it is not expected that our model, trained for context-conditioned preference queries, would do better than the base model. We report results here mainly to provide a reference figure, and also to show that fine tuning on RPR does not hurt the unconditioned preference prediction of the base model.   > It only performs better on the test set, which is built in the same way as the training set. This is not true. Our model not only performs better on the RPR test sets, but also: Multifaceted Bench (MF, see general response), HHH with context, Reward Bench with context, and Chatbot Arena with Context*. This is more clearly shown in Figure 1 in the PDF attachment, which we will add to our paper as Figure 2. Please note the errata for GPT-4 results in the general response.   > Note that the preference labels on this test set are not verified by humans and could be unreliable. Human validation of the dataset labels was done by the authors (blind / response orders and criteria/scenarios randomized). As the authors, we are the closest to the data, and are able to provide high quality labels. In our original submission we reported 100 total labels (50 for RPR Criteria and 50 for RPR Scenarios; different prompts for each) in Table 2 (pdf page 7). Since submission, we have added an additional 100 labels (50 for each Criteria and Scenarios), achieving a total agreement of 97% for RPR Criteria and 95% for RPR Scenarios (100 labels each). This gives 95% confidence intervals for human-RPR agreement as follows: \| RPR Criteria \| RPR Scenarios \| \| :--------: \| :-------: \| \| $$(0.937, 1)$$ \| $$(0.907, 0.993)$$ \| Some noise in human agreement is unavoidable. As noted in our paper (L73) there is significant inter-human disagreement on unconditioned queries, e.g., humans agree with each other only 65.7% of the time on AlpacaEval. Even the carefully curated MT-bench shows has agreement of only 81-82% on strict preference queries (Zheng et al., Table 5). Thus, our 90+% agreement between the human authors and the synthesized labels is quite high, likely as a result of the added context and multiple levels of filtering.	Summary: The motivation behind this paper is to address the critical challenges of finetuning language models (LMs) from pairwise preferences due to the underspecified nature of natural language. Direct preference feedback is often uninterpretable, inconsistent, and difficult to provide, especially when multidimensional criteria are involved. These issues arise from incomplete instructions or the diverse backgrounds of the individuals providing the feedback. To tackle these challenges, the authors propose a two-step preference modeling approach: first, selecting a context to resolve under-specification, and second, evaluating preference with respect to the chosen context. There are mainly 4 contributions the authors have made: Decomposition of Reward Modeling Error: The paper introduces a method to decompose reward modeling error into two components: context inference error and context-specific reward modeling error. This decomposition supports the idea that supervising both context and context-specific preference could align models more effectively with diverse human preferences. Context-Conditioned Preference Datasets: The authors contribute several novel datasets designed to investigate the ability of LMs to evaluate context-specific preferences. These datasets, termed "preference reversal" datasets, isolate context-specific capabilities by disentangling them from general preferences. Context-Aware Reward Model: The paper demonstrates the development and finetuning of a context-aware reward model, showing that this model achieves performance comparable to or exceeding that of state-of-the-art models like GPT-4 and Llama 3 70B. The context-aware model also shows improved performance in context-specific tasks. Experiments and Benchmarking: The authors conduct experiments to benchmark the context-specific performance of various models, highlighting that current models benefit from additional context but often fail to fully utilize it. Finetuning with the preference reversal datasets significantly enhances the models' context-specific performance. Strengths: This paper presents a novel approach to preference modeling in LMs by integrating context-specific evaluation. The key contributions include the introduction of context-conditioned preference datasets, the decomposition of reward modeling error, the development of a context-aware reward model, and comprehensive experiments to validate the approach. Weaknesses: ## Under-supported claims In section 3.3, the author included various claims that I believe to be interesting but only so when they are better-substantiated. Assumption about Cardinality The hypothesis that the cardinality of the space of contexts is smaller than that of the space of possible completions given a prompt is stated without empirical or theoretical justification. This assumption is critical to the argument about data efficiency in context annotation versus preference annotation. Without evidence or a rationale to support this assumption, the argument remains speculative. ## Details of the dataset curation The reviewer appreciate the author's efforts in curating the dataset. Section 4 briefly mentions that the dataset was generated using GPT-4 Turbo with a series of prompts designed to maximize validity. However, it lacks specific details about the prompts, the criteria for selection, and the methodology for ensuring the validity of the samples. This ambiguity makes it difficult to assess the soundness of the curation process. ## Clarity While the paper presents interesting ideas, there are some areas in writing and presentation that could benefit from improvement: - The introduction and related work sections are detailed, but they could be more concise to better highlight the key contributions and their significance. - The figures and tables, although informative, could be optimized to more effectively convey the main points. Some elements might appear redundant or unclear. - Enhancing the narrative flow would help in making the logical progression of the arguments clearer and more accessible to the reader. Technical Quality: 2 Clarity: 3 Questions for Authors: ## Annotation and intent inference I particularly do not understand the assumption related to Bernoulli distribution: Section 3.1 states that an annotator implicitly infers intent i from prompt x and samples a preference from a Bernoulli distribution. This process of intent inference by annotators is critical yet highly under-explained. There is no empirical evidence or studies cited that demonstrate annotators' ability to accurately infer intents from prompts. Additionally, the variability in annotators' interpretations and the potential biases they introduce are not addressed. ## Intent distribution The paper asserts that both users and annotators may possess or infer a distribution of intents and that annotation for most preference queries involves a distribution rather than a specific intent. This is a substantial claim that needs empirical support. ## Data curation process I would appreciate it if the authors can provide further details on data curation. In particular, how the quality of the synthetic data is maintained. ## Human validation What is the scale of this validation, and how are the human validators selected? Is the sample size sufficient to generalize the results to the entire dataset? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed review and feedback. Please find our responses below.   > Under-supported claims / Assumption about Cardinality The framing of this in the discussion is not as “claims” but rather a “conjecture” (L182) and a “hypothesis” (L188,L193). As noted in the text, we do find “some” (L181,L193) empirical support for each, in Tables 2 and 5. We leave a more detailed exploration of this to future work, as the focus of this work is on improving context-specific preference modeling.   > Data curation process / “lacks specific details about the prompts, the criteria for selection, and the methodology for ensuring the validity of the samples” As noted in the main text at L211 (right after introducing the datasets), the full details of the dataset curation process, including prompts, criteria for selection, and methodology for ensuring validity (e.g. L794), are detailed in Appendix B (L766-L815). If there is some shortcoming in the presentation there, we would appreciate a more specific critique.   > Clarity / Enhancing the narrative flow We have added a Figure 2 and revised Figure 1 (see pdf attachment) to clarify the framework and main results. We believe the narrative flow is strong, but would appreciate and happily consider any more specific critiques the reviewer has on this point.   > Assumption related to Bernoulli distribution / Intent distribution Modeling paired preference as sampling from Bernoulli distribution is standard; e.g., in the widely used Bradley-Terry model. Allowing for an intent distribution (as opposed to a single intent) seems like a natural modeling choice that broadens the scope of the model (it gets us Equation (3) in addition to the single intent Equation (4)). We do not believe these modeling choices require empirical justification, as we are not claiming in Sections 3.1-3.2 that real human annotators explicitly infer an intent or distribution of intents, or that real human annotators explicitly model their label as a Bernoulli distribution. Indeed, at L148-151 we adopt the (commonly used) Expected Utility model, which is used to arrive at Equations (3) and (4); but it is well known that humans systematically deviate from expected utility. Our formalism is Sections 3.1-3.2 is best understood as any other idealized model: it offers analytical insights. In our case, our model is used to arrive at the bounds in Equation (3) and (4), which expands on and strengthens the motivation in Section 2, and suggests the discussion in Section 3.3. As with any model, it is up to the reader to determine whether the model in question is a reasonable, and useful enough, approximation to reality.   > There is no empirical evidence or studies cited that demonstrate annotators' ability to accurately infer intents from prompts. Additionally, the variability in annotators' interpretations and the potential biases they introduce are not addressed. This is a core motivation for our work, as set out at L82-85: “If we train models using non-basic preference annotations, the contextual biases and assumptions underlying those judgments may be implicitly embedded into the model [35, 48] … Rather than rely on annotators to integrate the correct distribution of contextual assumptions … ”, wherein it is suggested that we might not want to rely on annotators to “accurately infer intents” [i.e., contexts], as this is subject to “the variability in annotators' interpretations and the potential biases”. A descriptive study of the biases and assumptions made by real human annotators is beyond the scope of the present work.   > Human validation The human validation of the dataset labels was done by the authors (blind / response orders and criteria/scenarios randomized – we will add this detail to the paper). As the authors, we are the closest to the data, and are able to provide high quality labels. In our paper we reported 100 total labels (50 for RPR Criteria and 50 for RPR Scenarios; different prompts for each). Since submission, we have added an additional 100 labels (50 for each Criteria and Scenarios), achieving a total agreement of 97% for RPR Criteria and 95% for RPR Scenarios (100 labels each). This gives 95% confidence intervals as follows: \| RPR Criteria \| RPR Scenarios \| \| -------- \| ------- \| \| $$(0.937, 1)$$ \| $$(0.907, 0.993)$$ \|	Rebuttal 1: Rebuttal: We thank the reviewers for their time, consideration, and numerous comments that will help us improve the manuscript. We have responded to each reviewer individually. If you find our rebuttal to be responsive to your concerns, we kindly ask you to consider recommending “accept” --- we believe context-specific modeling and pluralistic alignment is an important direction that has received recent interest from a number of different research groups, and that the present submission is complete and provides interesting results and contributions in this area. We have the following general comments and revisions to note:   Added new Figure, and updated Table 6 to better present the results See PDF attachment (Figure 1 and Table 2). Figure 1 of the PDF attachment will appear as Figure 2 in our revised manuscript.   Updated to include results on Multifaceted Bench and discuss the concurrent work by Lee et al. [1] Post submission, the work [1] by Lee et al. was released, which includes a synthesized dataset of diverse system prompts for finetuning generative models. Although formed for a different purpose, their dataset includes multiple system prompts for the same user prompt, which allows it to be used in a similar fashion as our RPR datasets. We run context-specific evaluation on their dataset. The main results are now shown in the PDF attachment (Figure 1), and we will include a discussion of their work in the Related Works section. [1] Lee, Seongyun, et al. "Aligning to thousands of preferences via system message generalization." arXiv preprint arXiv:2405.17977 (2024).   Updated to include a Data Ablation Reviewer mLs1 made the excellent point that “The current experiments do not demonstrate the general advantage of using a paired dataset with reversed preference labels.” In response, we have prepared a data ablation that finetunes the base Mistral reward model on context-conditioned data from other distributions, including: - Preference Collection (PC), in which the context is highly correlated with unconditioned preference by Kim et al. (2024) (Prometheus 2), and - a “one-sided” version of our RPR datasets, where we keep only one of the two sides by having GPT-4 pick the criteria that it thinks is best suited for the given prompt. The results are shown in Table 1 in the PDF attachment, and demonstrate the advantage provided by our RPR dataset. We will add this to our paper. Notably, on Multifaceted Bench (MF), which is outside of our training distribution, training on RPR strongly improves context-conditioned performance over the base model (and the PC model).   Errata re: GPT-4 results After submission, we noticed that the GPT-4 results appeared high (e.g. our submission reported a score of 93.5% on Reward Bench, which is far higher than what it obtains on the Reward Bench leaderboard). This was due to a bug in our code that was counting ties issued by GPT-4 as being correct. Only the results reported for GPT-4 Turbo were affected—all other originally reported results are correct. Here are the corrected numbers for GPT-4, which are also reflected in the attached PDF Figure 1. \| \| RPR-C (CTX) \| RPR-S (CTX) \| PB (CTX) \| MF (CTX) \| HHH (NC) \| HHH (CTX) \| RB (NC) \| RB (CTX) \| CBA (NC) \| CBA (CTX) \| CBA (CTX) \| \| --------\| :--------: \| :-------: \| :--------: \| :-------: \| :--------: \| :-------: \| :--------: \| :-------: \| :--------: \| :-------: \| :-------: \| \| Original GPT-4 \| 0.899 \| 0.725 \| 0.868 \| - \| 0.950 \| 0.964 \| 0.935 \| 0.904 \| 0.825 \| 0.833 \| 0.930\| \| Corrected GPT-4 \| 0.901 \| 0.748 \| 0.860 \| 0.640 \| 0.871 \| 0.873 \| 0.824 \| 0.821 \| 0.720 \| 0.771 \| 0.858 \| After this correction, our finetuned Mistral CARM is the best performing model not only on RPR (for which it is in distribution), but also: Multifaceted Bench, HHH (CTX), Reward Bench (CTX), and Chatbot Arena (CTX). Note that GPT-4 does worse on average than Llama3-70B not because it is a worse at context-specific prediction necessarily, but because we do not have access to its logits so that its scores are restricted to an integer scale and often result in ties.   New context-aware finetune Reviewer rZtt made a good point that "it remains unclear how applicable the proposed method is to other LMs". As part of our rebuttal we finetuned a reward model based on Gemma 2B that performed relatively well for its size on Reward Bench. This shows that the benefit of training on RPR is not restricted to the Mistral 7B model we used in our submission. We report the results in our response to Reviewer rZtt below, which we will add to the paper.   Added sample from RPR We realized after submission that our manuscript did not include any samples from the dataset itself. We have since added a sample from RPR to the main text to help with the exposition. We are happy to share it during the extended discussion period if the reviewers would find it helpful. Pdf: /pdf/1e38ca21990f3cf8cb78cd20fd103261c65b3d27.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Bootstrapping Top-down Information for Self-modulating Slot Attention	Accept (poster)	Summary: This paper proposes a method that improves the performance of object-centric learning by modulating Slot Attention with semantic and location information obtained based on the output slots and attention maps of Slot Attention. For a given output slot, the semantic information is chosen as the vector closest to this slot in a codebook that is learned from all output slots of the entire dataset via vector quantization, and the location information is chosen as the shifted attention map with a mean value of 1. The proposed method is implemented based on DINASOUR and is compared with DINASOUR and other methods on two synthetic and two real datasets. The proposed method outperforms the compared method in most cases when the size of codebook is chosen appropriately. Strengths: 1. Improving the performance of object-centric learning with top-down information is an important and interesting research direction. 2. The proposed method outperforms the compared methods when the size of codebook is chosen appropriately. Weaknesses: 1. My main concern about the paper is that the performance of the proposed method is very sensitive to the size of codebook. If the size of codebook is not chosen appropriately, the performance can be even worse than the original DINASOUR method. Moreover, how to choose the best size of codebook is not described clearly. 2. From my understanding, the core of the proposed method is a self-modulation module (along with a codebook) that is compatible with all the object-centric learning methods using the slot attention mechanism. However, the proposed method is only implemented based on DINASOUR. The quality of the paper could be significantly improved if the proposed method is implemented based on at least another object-centric learning method that is developed based on slot attention. 3. The proposed self-modulation includes both semantic and spatial modulations. However, the codebook used in the semantic modulation also contains spatial information. This design can be more elegant, and the codebook can be much more useful if it only contains semantic information. In this way, whether two objects belong to the same category can be determined automatically. 4. Some data in the experimental part are inconsistent. For example, the mBO^i of the proposed method on COCO is 33.0 in Table 2, 33.3 in Table 3, and 32.7 in Table 4. 5. The authors attribute both semantic modulation and spatial modulation as top-down modulations. I agree that the semantic modulation is a top-down modulation because the codebook contains knowledge learned from all images (not just the inferred image). However, I don’t think that the spatial modulation can be considered top-down because it just uses attention maps computed in a bottom-up manner (only based on the content of one image). Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How is the best size of codebook chosen? 2. Is it possible to learn a codebook that is independent of the locations of objects in the image? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitation in the conclusion part, and there is no potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Sensitivity to codebook size We acknowledge the performance dependence on the codebook size, as noted in our limitations section. However, we would like to emphasize that the optimal codebook size was determined automatically using the training set only (without the validation set). To choose the codebook size, we monitor the perplexity of the code usage distribution during training. Perplexity—the exponent of entropy—indicates how uniformly the codes are being used and typically increases as the codebook becomes larger. However, when the codebook size exceeds the number of semantic patterns in the data, some codes become neglected, causing low code utilization [D, 45, 26]. Hence, the ideal codebook size leads to the maximum perplexity. To find the optimal codebook size, we start with a size of 64 and monitor the perplexity of the codebook over training. After sufficient iteration (e.g., 250K), the codebook size is doubled repeatedly until the perplexity reaches a plateau. For example, on COCO, the perplexity when the codebook size is 256, 512, and 1024 are 176.9, 253.9, and 242.8, respectively, where 512 was chosen as the final size. Thereby, an adequate codebook size could be efficiently determined without using the validation set. ## Experiment on models other than DINOSAUR We appreciate your suggestion and have implemented our self-modulation technique with the original Slot Attention setting [28], which trains the encoder from scratch and uses an image reconstruction objective. The following table summarizes the results on the CLEVR6 dataset: \| \| FG-ARI \| mBO \| \|----------\|----------\|----------\| \| Original slot attention \| 98.7 \| 21.2 \| \| Original slot attention + SelfMod \| 88.7 \| 61.9 \| We observe a significant improvement in mBO, showing that our self-modulation technique is applicable to slot attention and provides complementary benefits. Although FG-ARI decreased, mBO is considered more robust and should be prioritized when evaluating model performance, as detailed in our response to reviewer 3ckb ("Necessity & effect of VQ: (3) Difference in FG-ARI and mBO in VQ Ablation"). We have also included qualitative results in Figure R3 of the attached PDF, which demonstrates that self-modulation markedly improves segmentation. These results indicate our method's effectiveness with different encoder configurations and training objectives, suggesting it does not heavily rely on pretrained features. We will add the results and discussion in the revision. ## Does the codebook also contain positional information? Our codebook primarily encodes semantic content, largely independent of the object position (Figure 2 of the main paper). We hypothesize that this is because semantic information dominates slot representations, with positional information occupying a small subspace. This is supported by $K$-means clustering of DINOSAUR slot representations, which automatically groups semantic categories using simple L2 distance (Figure R6 of the attached PDF). While our approach empirically extracts mainly semantic information, explicitly disentangling positional and semantic information remains an interesting future direction, with work like [3] offering promising leads. ## Inconsistent numbers between tables We apologize for the confusion. Our mBO$_i$ on COCO of 33.3 in Table 3 is a typo, and the correct number is 33.0. We will revise the paper to include the correct measurement. Despite this error, our method still outperforms the LSD [18] by 2.6%p, showing the effectiveness of our approach. Additionally, as mentioned in line 234 on page 6, the in-depth analysis and ablation studies (Tables 4, 5, and 6) were conducted with models trained for 200K iterations, rather than the full 500K iterations, due to computational constraints. This explains the differences between Tables 1-3 and Tables 4-6. We found that the models mostly converged at 200K iterations and thus was enough to reflect the performance trends of models trained for the full 500K iterations. To improve clarity, we will update the captions of Tables 4-6. We appreciate your attention to these details. ## Can spatial modulation be called “top-down”? We appreciate the reviewer’s comment on our top-down approach. We respectfully clarify our reasons for calling spatial modulation a top-down approach. A top-down approach is defined by the use of prior knowledge (contextual or task-relevant) to guide visual processing, while a bottom-up approach relies solely on sensory data. As the reviewer noted, our semantic modulation using vector quantization aligns with this definition, leveraging contextual knowledge learned from the dataset. Although the spatial information is derived from the same image, the spatial modulation provides rough object saliency maps to the subsequent slot attention, which is crucial task-relevant information for object discovery. It is task-relevant because it explicitly guides the model to discover objects within the provided saliency map, narrowing the search space. Such spatial cues are recognized as a form of top-down knowledge in human vision research [A, B, C], guiding attention to regions likely containing objects of interest. In conclusion, while our semantic and spatial modulations differ in their information sources, both embody top-down processing principles by providing crucial task-relevant guidance to the visual processing pipeline. --- [A] SUN: Top-down saliency using natural statistics, Visual cognition, 2009 [B] Components of visual orienting, Attention and performance X: Control of language processes, 1984 [C] Oculomotor strategies for the direction of gaze tested with a real-world activity, Vision research, 2003 [D] Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks, ICML, 2023 --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response. Since the codebook size is determined solely based on the training set in an unsupervised way and applying the proposed self-modulation to Slot Attention also improves the performance on the CLEVR6 dataset, my concerns about technical flaws disappear and the rating is increased. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your thoughtful reconsideration of our work. We are pleased that our clarifications regarding codebook size determination and the CLEVR6 experiment have addressed your concerns. We will improve the paper in future revision based on your feedback. Thank you for your valuable feedback and updated assessment.	Summary: Slot Attention is a popular component for object-centric learning methods. In this paper, the authors propose an extension of Slot Attention named "top-down pathway". After the last iteration of Slot Attention, the slots are mapped to a discrete, learnable codebook. Jointly with the final attention map, this information is used in a second iteration of Slot Attention to modulate the QKV-like attention mechanism. Previous models are consistently outperformed on established benchmarks. Moreover, a more detailed analysis of the learned codebook shows that distinct semantic concepts are represented by different codebook entries. Strengths: - The proposed extension of Slot Attention is clearly explained. The original Slot Attention is described at an adequate level of detail, which helps to make the paper accessible. - The model consistently outperforms DINOSAUR, from which all design decisions are derived. This demonstrates the effectiveness of the novel mechanism. - The analysis of the learned codebook is insightful and confirms the motivation of the proposed method. Weaknesses: I am not fully convinced by the term "top-down pathway". Only the recurrent Slot Attention module has been extended, the rest of the model is still bottom-up. From the initial description, I would have expected top-down modulation reaching to the encoder. Did you consider a variant of the model that continues with the slot embeddings from the first iteration instead of starting from scratch? If this performs similarly well, it might indiciate that the quantization due to the codebook is driving the improvement. Technical Quality: 3 Clarity: 3 Questions for Authors: - A recent approach to object-centric learning which is not based on Slot Attention is CutLER (Wang et al. 2023). How does the proposed method compare to CutLER? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are very briefly discussed. An additional point could be the dependence on the pretrained feature encoder, which might not work well in all scenarios. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Clarification on term “top-down pathway” / modulating encoder We appreciate the reviewer's comment on our use of the terminology, "top-down pathway." We respectfully explain that our terminology is appropriate for the following reasons: In our context, "top-down" refers to using higher-level task-relevant information to guide lower-level processing. Our method achieves this by: a) Extracting semantic information via vector quantization to enforce the focusing on certain features. b) Using spatial information from the previous slot attention to guide object localization. Both of the above mechanisms bootstrap top-down information beyond pure bottom-up processing, without requiring additional supervision. While our top-down pathway is confined to the slot attention, it still introduces a meaningful top-down influence on the features by modulating values in slot attention. We agree that extending the top-down modulation to the encoder is an interesting direction for future research. However, our current approach strikes a balance between effectiveness and efficiency, as modulating the encoder would significantly increase computational cost (the encoder accounts for over 71% of total compute). Our method achieves meaningful improvement without this additional overhead. For a detailed discussion of our method's efficiency, please refer to our response to Reviewer 3ckb (“Computation cost analysis”). ## Value modulation vs. slot initialization We appreciate the insightful suggestion. We did consider such a variant that initializes the slots with the quantized representations from the previous slot attention. However, this approach did not perform well in our experiments. Given that slot attention consists of 3 iterations, initializing the second slot attention with quantized representations from the first is equivalent to inserting VQ in the middle of 6 iterations. To evaluate this, we insert VQ into DINOSAUR's slot attention with 6 iterations, after the third iteration. Results on COCO with 200K training schedule are presented below, together with the results presented in the Table 5 of the main paper: \| \| FG-ARI \| mBO$_i$ \| \|----------\|----------\|----------\| \| DINOSAUR 6-iter \| 28.5 \| 28.2 \| \| Ours \| 37.3 \| 32.7 \| \| DINOSAUR 6-iter + VQ \| 17.4 \| 12.9 \| Introducing VQ to DINOSAUR with 6 iterations leads to a performance decrease of 11.1%p and 15.3%p in FG-ARI and mBO, respectively (row 1 vs. 3 of the table). The severe performance degradation is likely due to the disruption in slot attention's convergence process. As shown in [6], slot representations naturally converge over its recurrent iterations. Discretizing in mid-process introduces a sudden shift in the slot representation, which will harm the overall optimization and convergence. Our method avoids this issue by modulating inner value activations, which can modulate the recurrent slot update process without directly changing the slot representation. By doing so, we can exploit the top-down information without hindering convergence of the slot attention. ## Comparison to CutLER We appreciate the suggestion to compare with CutLER [A], a recent approach to unsupervised object segmentation. While CutLER's full pipeline includes self-training of Mask R-CNN, we focus on comparing our method to CutLER's MaskCut algorithm for pseudo-mask generation, as this is the most direct comparison to our approach. Our method’s predictions can also be used as pseudo labels for training Mask R-CNN. We reproduced MaskCut using the official CutLER repository and evaluated it on the COCO dataset: \| \| FG-ARI \| mBO$_i$ \| mIoU \| \|----------\|----------\|----------\|----------\| \| MaskCut (CutLER )\| 31.5 \| 28.9 \| 26.7 \| \| Ours \| 37.4 $\pm$ 0.0 \| 33.0 $\pm$ 0.3 \| 31.2 $\pm$ 0.3 \| Our method significantly outperforms MaskCut across all metrics, demonstrating its effectiveness for object discovery. We will add these results and discussion in the revision. ## Dependence on the pretrained feature encoder We appreciate the insightful suggestion. Regarding the inquiry, we conducted an experiment with the settings of the original slot attention [28], which trains the encoder from scratch (response to reviewer Yith, PJTV: “Experiment on models other than DINOSAUR”). Interestingly we observe substantial performance gain even in this setting, demonstrating that our method's improvements does not solely arise from using a pretrained feature encoder. For more detailed results and discussion, please refer to the corresponding response. We will discuss this experiment in the future revision as well. --- [A] Cut and Learn for Unsupervised Object Detection and Instance Segmentation, CVPR, 2023 --- Rebuttal Comment 1.1: Comment: Thank you very much for the detailed response. - I am still not entirely convinced about the tmerin "top-down pathway", however I think this not a major issue that speaks against accepting the paper. - The additional comparisons to MaskCut and other methods in the responses to other reviewers are very helpful and confirm the consistent improvements of the proposed model. - I agree with the other reviewers on the concerns regarding the codebook size and the unclear role of the vector quantization. (1) While the codebook size can be chosen without supervision, it seems that the required multiple training runs substantially increase the computational cost. (2) The impact of the codebook on the performance seems to be small and heavily decreases the performance when used alone, as shown by the additional ablation study. While I still think the analysis of the codebook is interesting, I am not sure to which extent it is necessary for the performance improvement. Overall, I still think the paper should be accepted, since the proposed method is interesting and consistently outperforms previous methods. Due to the remaining concerns I keep my rating for now. But I am looking forward to read the other reviewers comments on these points and happy to discuss further. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your detailed response and thoughtful consideration of our rebuttal. We will address your feedback as follows: - We will elaborate on the term "top-down pathway" in the paper to clarify its usage and context. - We're pleased you found the additional comparisons to MaskCut helpful. We will include this in the future revision. - Regarding codebook concerns: a) While determining optimal codebook size requires multiple training runs, perplexity is measured at half the full training schedule (250K), reducing computational burden. b) On the necessity of vector quantization, we kindly refer you to our response to reviewer 3ckb, specifically points (1), (3), and (4) under "Necessity & effect of VQ", which provide further insights into VQ's role in our method's performance. We truly appreciate your valuable feedback and will incorporate your suggestions to improve our paper's final version. Thank you for your support!	Summary: This paper proposes a modification to Slot Attention incorporating top-down information into the algorithm. After an iteration of Slot Attention, the slots are quantized into a learned codebook. The quantized slots and attention maps are then used in another iteration of Slot Attention, refining the representations. The algorithm is evaluated in the DINOSAUR setting on the MOVi, Pascal VOC, and COCO datasets, showing improved segmentation quality over vanilla DINOSAUR. Visualizations of the codebook show that meaningful semantic concepts are learned and several ablations are performed. Strengths: The proposed method is well-motivated and seems to show an improvement over previous methods. The use of top-down semantic information from a learned codebook is novel, from my understanding. Overall, the paper is well-written and the authors provide a good analysis of their method including ablations of the different design choices. Weaknesses: 1. The experiments are only performed in the DINOSAUR setting with pretrained, frozen ViT features as the input and reconstruction target. It would be very informative to also include experiments with the original image as the input and reconstruction target since that is also a common use case for Slot Attention. By only evaluating in the DINOSAUR setting, it is unclear how reliant the performance of the proposed model is on the pretrained ViT features. 2. In Figure 3 and 5.B, it seems the slots are sometimes reassigned before and after the self-modulation step? For example, the skater in Figure 3 is captured by the 4th slot before the modulation and the 5th slot after the modulation. The 5th slot before the modulation seems to capture part of the background, not any part of the skater. It is unclear to me why this would happen during the modulation update. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Since the slots also contain position information (”where” information), it seems possible that this gets captured in the learned codebook. Is this something the authors observed? 2. How cherry-picked is Figure 2? Are there cases where these codes do not correspond to the same semantic concept? 3. In Figure 2, it seems that sometimes, multiple objects with similar semantics (e.g. multiple zebras or signs) are being captured by one slot. Did you notice your proposed method does more semantic grouping instead of instance grouping compared with vanilla DINOSAUR? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Experiment on models other than DINOSAUR We appreciate your suggestion for additional experiments. We have implemented our self-modulation technique with the original slot attention setting [28], which includes training encoders from scratch and using an image reconstruction objective. The following table summarizes the results on the CLEVR6 dataset: \| \| FG-ARI \| mBO \| \|----------\|----------\|----------\| \| Original slot attention \| 98.7 \| 21.2 \| \| Original slot attention + SelfMod \| 88.7 \| 61.9 \| We observe a significant improvement in mBO, showing that our self-modulation technique is applicable to slot attention and provides complementary benefits. Although FG-ARI decreased, mBO is considered more robust and should be prioritized when evaluating model performance, as detailed in our response to reviewer 3ckb ("Necessity & effect of VQ: (3) Difference in FG-ARI and mBO in VQ Ablation"). We have also included qualitative results in Figure R3 of the attached PDF, which demonstrates that self-modulation markedly improves segmentation. These results indicate our method's effectiveness with different encoder configurations and training objectives, suggesting it does not heavily rely on pretrained features. We will add the results and discussion in the revision. ## Slots reassigned to other object after modulation Thank you for your acute observation. We sincerely apologize for the discrepancy in Figures 3 of the main paper and B.5 of the appendix, which resulted from inadvertently pairing slot attention maps from different samples: the slot attention maps after self-modulation are correct but attention maps before self-modulation were wrongly paired. We have provided the updated figures as Figure R1 and R5 in the attached PDF. As the reviewer correctly pointed out, slots are modulated based on the information captured during the first slot attention. In the corrected Figure 3 (Figure R1 of the attached PDF), row 3, the 5th slot, which initially captures the skater with less certainty, is modulated to capture the skater with higher confidence by leveraging the semantic and spatial information from the initial slot. We would like to reaffirm that our original observation—“ modulation dynamically refines the attention maps, depending on how well they have captured the scene”—remains valid. For instance, in the 2nd row, the 1st slot initially captures both the plates and the orange, but after modulation, it is associated exclusively with the orange, yielding the plates to the 7th slot which more accurately captures the plates during the first slot attention. Similarly, in the 4th sample, the 3rd slot, which initially associates with the straw in a trivial manner, is modulated to identify and locate the straw more accurately. On the other hand, the attention maps of the 1st row are refined at the boundaries only since the initial attention maps were well structured. Once again, we apologize for any confusion caused by this error. We appreciate your understanding and the opportunity to clarify our findings. ## Does the codebook also contain positional information? While slot representations can indeed contain positional information due to the reconstruction objective, our observations suggest that the codebook primarily encodes semantic content. Performing $K$-means clustering on slot representations from DINOSAUR, we observed that the clusters are primarily grouped by semantic categories rather than spatial positions, as shown in Figure R6 of the attached PDF. This indicates that semantic information dominates in slot representations. Figure 2 in the main paper further supports this, showing objects with similar semantics but different positions mapped to the same code. Thus, while present, positional information likely occupies a small subspace in slot representations, with semantic content being the primary factor captured by our codebook. ## Question about codebook visualization/semantics Figure 2 is not heavily cherry-picked. Most codes consistently represent a single semantic concept, with a few codes that portray supercategories such as animals and humans (people and their hands), as shown in Figure R4 of the attached PDF. Also, we did notice an interesting edge case where some codes capture only the top-left-most patch, as shown in Figure R2 of the attached PDF. We attribute this to the autoregressive decoding process. During decoding, the top-left patch must be reconstructed first without any context from previous patches. Consequently, in images with less visual complexity, some slots appear dedicated to capturing these top-left patches. However, these special codes are minimal (1 or 2 out of 512 codes) and have limited impact. We will include more codebook visualizations and a detailed discussion in the revision. ## Do modulated slots perform semantic grouping more than instance grouping? Thank you for the insightful observation. To quantitatively assess whether our method favors semantic grouping over instance grouping compared to DINOSAUR, we conducted an analysis using two metrics: Instance Precision: The precision score with instance-level GTs Semantic Recall: The recall score with semantic segmentation GTs If our model was biased toward semantic grouping, we would expect to see a lower instance precision (as semantic segmentation ground truth masks often have larger sizes) and a higher semantic recall compared to DINOSAUR. However, our analysis on the COCO dataset shows that our method improves both metrics, as shown in the table below. \| \| Instance Precision \| Semantic Recall \| \|----------\|----------\|----------\| \| DINOSAUR \| 41.1 \| 75.0 \| \| Ours \| 42.3 \| 75.3 \| These results suggest that the proposed top-down pathway enhances identifying both individual instances and semantic categories rather than favoring one. --- Rebuttal Comment 1.1: Title: Reply to Rebuttal Comment: Thank you for the rebuttal and running additional experiments. The updated figures for the slots before and after modulation make much more sense now. The additional results with vanilla Slot Attention on CLEVR are also promising, although I would encourage the authors to run additional seeds and/or make sure the hyperparameters are correct, as I've seen better results with vanilla slot attention on CLEVR (looking at the qualitative results it seems the background is being split). For what it's worth, I think it would still be a useful contribution if the proposed approach only improves the metrics on more complex datasets (since as a field, we should be generally pushing towards more complex datasets), but it is important to know the effect on simpler datasets as well so researchers are aware of the limitations. I would encourage the authors to include these results in the final version of the paper. After reading the rebuttal as well as the other reviews and rebuttals, I have decided to increase my score to 7. --- Reply to Comment 1.1.1: Comment: We appreciate your thorough review and time evaluating our rebuttal. We're pleased our updated figures and CLEVR experiments have resolved your concerns. We acknowledge your suggestion and will include CLEVR results from multiple seeds in the final paper. Thank you for your constructive feedback and recognizing the strength of our work.	Summary: This work proposes a novel unsupervised object-centric learning method. In particular, the proposed approach enhances the slot-attention mechanism by incorporating a "top-down pathway" that highlights features relevant to objects in the image. The method first employs a standard self-attention mechanism to extract slot vectors and the corresponding attention masks. These initial slot vectors and attention masks are then fed into another round of the slot-attention mechanism. In this round, they modulate the value maps of the cross-attention mechanism separately for each slot. This modulation is channel-wise for slot vectors and spatial-wise for attention masks, enabling the slot-attention mechanism to emphasize to features relevant to the slots discovered on the first slot-attention round. Furthermore, the method quantizes the slot vectors from the first slot-attention round using the VQ approach before utilizing them for modulation in the second round. To evaluate the proposed self-modulated-based slot-attention mechanism, the authors integrate it into the DINOSAUR framework and apply it to the COCO, VOC, MOVIE-E, and MOVIE-C datasets for the downstream task of object segmentation. The results demonstrate improvements over the vanilla slot-attention mechanism. Ablation studies are conducted on the COCO dataset. Strengths: STRENGTHS: - The idea of modulating the value maps for each slot, using slots found from a previous self-attention round, is interesting. This technique allows the subsequent self-attention round to focus more on features that are relevant to the objects discovered in the initial round. Essentially, it's an additional iterative improvement for the extracted slots, supplementing the existing iterative mechanism in the standard slot-attention. - The proposed method has been demonstrated to improve object segmentation results on all tested datasets. Furthermore, the authors utilized challenging datasets, such as COCO, to validate the effectiveness of their method, rather than resorting to simpler one, which is often among other object-centric works. - The paper has strong results and detailed ablation studies. Weaknesses: - (W1) It's unusual that the mIoU results in the MOVIE-C and MOVIE-E datasets surpass the mBO results. The mBO is computed by assigning each ground truth mask the predicted mask with the largest overlap, and then averaging the IoUs of the assigned mask pairs. On the other hand, the mIoU metric is more strict and employs Hungarian matching (rather than greedy matching) to assign predicted masks to the ground truth masks. Consequently, mBO results should be either greater than or equal to the mIoU results. Therefore, it's highly likely that there's a bug in the computation of the mBO or the mIoU metrics on the MOVIE-C and MOVIE-E datasets. Or a typo / mistake when copying the results to the paper maybe. - As a relevant side note for the upcoming comments, it's important to mention that the FG-ARI metric is generally viewed as unreliable for assessing unsupervised object-centric methods. This is because it only takes foreground pixels into account, which can provide a deceptive understanding of segmentation quality by disregarding the localization accuracy of predicted masks [A, B, C, 30, 42]. - (W2) The motivation behind the Vector Quantization (VQ) component of the method is unclear. In particular it's uncertain why quantizing the slot vectors is essential for self-modulation. In fact, Table 4 indicates that the method is sensitive to the codebook size (also acknowledged by the authors in the limitations section). The mBO performance for sub-optimal codebook sizes 128, 256, and 1024 (30.2%, 30.5%, and 29.8%, respectively) is either the same or worse than the reproduced DINOSAUR mBO performance (30.5% from Table 6). Conversely, removing VQ achieves 32.3%, which is only marginally worse than the 32.7% obtained with VQ and the optimal codebook size. Therefore, VQ unnecessarily complicates the proposed method. As mentioned earlier, FG-ARI is unreliable, and the higher differences in this metric (36.3% w/o VQ vs 37.7% w/ VQ) should be disregarded. - It's also somewhat concerning that Table 6 shows only a minimal impact on mBO performance when using channel-wise modulation (from 32.5% to 32.7%). However, this does not really increase the complexity of the method, as VQ does. - (W3) Lacks discussions on how the proposed method affects the training and test time - (W4) Table 3 is missing comparisons with related works, such as Rotating Features [D] and SPOT [C], which also demonstrate strong results. [A] Genesis: Generative scene inference and sampling with object-centric latent representations, ICLR 2020. [B] Unsupervised Layered Image Decomposition into Object Prototypes, ICCV 2021. [C] SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers, CVPR 2024 [D] Rotating features for object discovery. NeurIPs 2023. [30] Bridging the gap to real-world object-centric learning, ICLR 2023. [42] Slotdiffusion: Object-centric generative modeling with diffusion models, NeurIPs 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: - My primary concern regarding this work (mentioned as (W2) in the weaknesses section) is that the channel-wise and Vector Quantization (VQ) components unnecessarily complicate the method without demonstrating significant improvements in the ablation results. I would appreciate it if the authors could provide compelling arguments, such as additional experimental evidence on other datasets (e.g., VOC or MOVIE-E/C) or different settings (e.g., longer training), that show (if it is the case) the benefits of channel-wise modulation with VQ. - Please address the concerns (W1), (W3), and (W4) mentioned in the weaknesses section. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Validity of mIoU scores in MOVI-C/E Thank you for your astute feedback. We sincerely apologize for the mistake in the mIoU calculations on the MOVI datasets and any confusion it may have caused. As you correctly pointed out, the mBO should be greater than or equal to the mIoU. The reported mIoU is inflated due to omitting the first ground-truth mask. We have provided the corrected mIoUs in Table R1 of the attached PDF. Importantly, despite this error, our revised results still demonstrate that our method outperforms DINOSAUR, substantiating our original claim about the benefits of our top-down pathway. We greatly appreciate your diligence in identifying this issue and allowing us to correct it. ## Necessity & effect of VQ Thank you for your insightful comment regarding Vector Quantization (VQ). We explain the necessity of VQ by addressing the inquiries below: (1) Necessity of VQ for Self-Modulation VQ is fundamental to our method since it allows the model to capture recurring semantic patterns from rough object representations produced by the first slot attention. Similar to online clustering, VQ maps a continuous slot representation to the nearest discrete code. Throughout this process, codes in the codebook are updated to encode only the semantic information from the noisy object representations; this is consistent with clustering being used to capture high-level semantic knowledge from noisy inputs in various tasks [D, 44]. Therefore, incorporating VQ into the top-down pathway enables our model to focus on core semantic concepts of the dataset, which are bootstrapped from noisy representations without extra supervision. This leads to more refined object-centric representations. (2) Sensitivity to Codebook Size We acknowledge the performance dependence on the codebook size, as noted in our limitations section. This characteristic is indeed common in clustering-based methods. However, we would like to emphasize that the optimal codebook size was determined automatically using the training set only (without the validation set). To choose the codebook size, we monitor the perplexity of the code usage distribution during training. Perplexity—the exponent of entropy—indicates how uniformly the codes are being used and typically increases as the codebook becomes larger. However, when the codebook size exceeds the number of semantic patterns in the data, some codes become neglected, causing low code utilization [E, 45, 26]. Hence, the ideal codebook size leads to the maximum perplexity. To find the optimal codebook size, we start with a size of 64 and monitor the perplexity of the codebook over training. After sufficient iteration (e.g., 250K), the codebook size is doubled repeatedly until the perplexity reaches a plateau. For example, on COCO, the perplexity when the codebook size is 256, 512, and 1024 are 176.9, 253.9, and 242.8, respectively, where 512 was chosen as the final size. Thereby, an adequate codebook size could be efficiently determined without using the validation set. (3) Difference in FG-ARI and mBO in VQ Ablation While we agree that mBO is the more reliable metric, FG-ARI can provide complementary information when mBOs are similar. FG-ARI is considered deceptive because it does not penalize under-segmentations that include background pixels [A], sometimes resulting in misguidedly high FG-ARIs despite low mBOs. However, in Table 6 of the main paper, VQ improves both metrics. The increased mBO, which accounts for background pixels, suggests that the FG-ARI increase is not due to under-segmentation but due to improved foreground object segmentation. Thus, we respectfully claim that the increase in FG-ARI with VQ should be considered as valid evidence for the effectiveness of VQ, especially considering the slight improvement in mBO. (4) VQ Ablation with Full Training Schedule We further validate the effectiveness of VQ with an experiment under a full training schedule (500K iterations) on COCO as suggested. The results demonstrate consistent improvement in both FG-ARI and mBO, clearly showing that VQ plays a crucial role in our model: \| \| FG-ARI \| mBO$_i$ \| \|----------\|----------\|----------\| \| Ours wo/ VQ \| 36.3 $\pm$ 0.5 \| 32.4 $\pm$ 0.1 \| \| Ours \| 37.4 $\pm$ 0.0 \| 33.0 $\pm$ 0.3 \| In summary, while we acknowledge the sensitivity to codebook size, we believe the benefits of VQ outweigh this limitation. Additionally, an appropriate codebook size can be efficiently determined by monitoring perplexity during training. ## Computation cost analysis While our model requires one more forward pass for slot attention, the additional computation cost is negligible. Slot attention accounts for only 0.64% of the total FLOPs, compared to 71.26% for the encoder and 28.10% for the decoder. Thus, our model requires 47.62 GFLOPs while DINOSAUR needs 47.32 GFLOPs. Also, processing the entire COCO 2017 val split takes 71.4 seconds for our model, while 70.5 seconds in the DINOSAUR. ## Missing comparison with recent work Thank you for pointing out the missing related work. We will add comparisons with Rotating Features [C] and SPOT [B] in the revision. We would like to kindly remind the reviewer that SPOT was published at CVPR 2024, after our submission deadline. Moreover, SPOT's self-training and sequence permutation techniques are completely orthogonal to our work on top-down pathways, suggesting potential for future integration of these complementary approaches. --- [A] Genesis: Generative Scene Inference and Sampling with Object-centric Latent Representations, ICLR, 2020. [B] SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers, CVPR, 2024 [C] Rotating Features for Object Discovery, NeurIPS, 2023 [D] Deep Clustering for Unsupervised Learning of Visual Features, ECCV, 2018 [E] Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks, ICML, 2023 --- Rebuttal Comment 1.1: Comment: I thank the authors for the detailed responses to my comments. Their response addresses my main concerns regarding the necessity of VQ: the codebook size selection is automatic and uses the training split, and VQ offers bigger improvement for longer training. Therefore, and after reading the other reviews and the respective rebuttals, I am going to increase my score. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your careful review of our responses and your reconsideration. We're glad our explanations about codebook size selection and VQ benefits clarified your concerns. In our future revision, we will include discussions and experiments about VQ, which will significantly improve the paper's quality. Thank you for your valuable feedback!	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for the constructive and insightful comments. Our work introduces a novel top-down pathway for object-centric learning that consistently improves performance across multiple benchmarks. Reviewers highlighted several strengths of our work, including the clarity of the manuscript, proposed method’s motivation and novelty, consistent improvements over the baseline across challenging real-world datasets, and extensive analyses. In each response to the reviewer, we have carefully addressed every comment and question to the best of our ability, providing additional results, clarifications, and analyses where requested. Key points we elaborate on include the clarification for vector quantization, comparisons to additional baselines, and further insights into our method's behavior. We also provide additional results and visualizations in the attached PDF. Due to the word limit, we referenced papers cited in the main paper with numbers and newly cited ones with alphabetical letters. Thank you again, and we look forward to a constructive interaction in the following discussion period! Pdf: /pdf/1e0ca2318957ee7eafb66fbfd864d4ce14def0c0.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
DiTFastAttn: Attention Compression for Diffusion Transformer Models	Accept (poster)	Summary: In this paper, the authors introduce a novel post-training model compression method aimed at accelerating Diffusion Transformers (DiT) used for image and video generation tasks. They identify the spatial, temporal, and conditional redundancy in attention blocks and propose the corresponding method to tackle them. Strengths: + The proposed methods are interesting and effective. +The writing is easy to understand. Weaknesses: - The authors sample 5K images to evaluate the generation quality while the standard setting is sampling 50K images (following original DiT). - What is the adopted compression strategy in each layer and each timestep? It would be better to illustrate the searched compression plan in a figure. - Can the proposed method be combined with other acceleration methods \eg Deepcache? - Why do the authors only demonstrate the Inception Score in Figure 9? FID is the most common metric in image generation. - What are the FLOPs and Latency in Table 1 exactly? The authors should demonstrate the value rather than only the fraction since there is still enough space. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We truly appreciate the reviewer for the insightful and constructive comments. > W1. The authors sample 5K images to evaluate the generation quality while the standard setting is sampling 50K images (following original DiT). Thanks. Following your suggestion, we have increased the evaluation size of ImageNet to 50K and COCO to 30K. Tables can be found in the global rebuttal (Table 1 & 2). More discussions can be found in the global rebuttal Q1. > W2. What is the adopted compression strategy in each layer and each timestep? It would be better to illustrate the searched compression plan in a figure. Thanks for your suggestion. We generated figures for adopted compression strategy in different threshold settings, check Q2 in global rebuttal and Figure 1 in the rebuttal PDF. > W3. Can the proposed method be combined with other acceleration methods \eg Deepcache? We believed that Deepcache benefits from the hierarchical nature of U-Nets so it can not be directly applied on DiT model. However, we do think our method can be combined with other acceleration methods. For example, our method is orthogonal with quantization method, so we can apply both quantization and our method to the model to achieve further acceleration. > W4. Why do the authors only demonstrate the Inception Score in Figure 9? FID is the most common metric in image generation. Showing only one metric in a figure tends to be more clear. In a future version of our paper, we will include an ablation study plot to further demonstrate the FID results. > W5. What are the FLOPs and Latency in Table 1 exactly? The authors should demonstrate the value rather than only the fraction since there is still enough space. Thanks. We make a summary table in the global rebuttal Q1 (Table 1 & 2) to contain exactly FLOPS and Latency. Summary tables like that will be included in the future version of our paper. --- Rebuttal Comment 1.1: Comment: How much time do we need to search the compression strategy? --- Rebuttal 2: Title: Compression strategy search time Comment: Thanks for your comment. Search time varies for different models and parameter settings. We use a greedy method in compression plan search so increase threshold (more compression) will result in a shorter search time. The maximum search time for a model can be determined by setting the threshold to 0. One thing to note is that once the search is completed, the compression strategy will be locally cached, eliminating the need for a repeated search. The cached strategy will be utilized the next time our method is applied for image or video generation. For reference, here we post the search times for DiT and PixArt-Sigma 1K using the default settings. #### Table1. Search time for DiT-XL-2-512 \| Threshold \| Search Time \| \|-----------\|-------------\| \| 0 \| 04m39s \| \| 0.05 \| 04m08s \| \| 0.1 \| 03m49s \| \| 0.15 \| 03m14s \| #### Table2. Search time for PixArt-Sigma-XL-2-1024-MS \| Threshold \| Search Time \| \|-----------\|-------------\| \| 0 \| 22m02s \| \| 0.05 \| 20m12s \| \| 0.1 \| 17m50s \| \| 0.15 \| 15m49s \| --- Rebuttal 3: Comment: Thanks for your patience. Here we post the results for DiT after changing our experiment setting to align with DiT paper. From this table, we observe that when the threshold is increased from 0 to 0.10, the IS drops from 210 to 180, and the FID decreases from 3.16 to 3.09, then increases to 4.52. The behavior of the FID in different settings follows the pattern reported in the previous discussion[1]. We observe a higher reduction in FLOPs and Attention FLOPs in this setting. We believe the current results demonstrate the effectiveness of our method across different settings. \| Threshold \| IS@50K \| FID@50K \| macs \| attn_mac \| \|-----------\|--------\|---------\|------\|----------\| \| Raw setting \| 219.9721 \| 3.16 \| 262359 \| 33823 \| \| 0.025 \| 218.1955 \| 3.09 \| 236265 \| 18041 \| \| 0.050 \| 210.3559 \| 3.10 \| 218865 \| 12339 \| \| 0.075 \| 196.05 \| 3.54 \| 203420 \| 8777 \| \| 0.100 \| 180.34 \| 4.52 \| 195682 \| 7137 \| So far, we increased evaluation sample size, changed experiments settings and provided results that aligned with the original paper according to W1; provided compression plan and search time that asked in W2 and FLOPs number that asked in W4; explained the combination with other acceleration model for W3; revised our paper according to W5. We are happy to answer additional questions and provide more clarification if needed. We kindly ask you to reconsider the score since we nearly reached the end of the discussion period. [1] Jayasumana S, Ramalingam S, Veit A, et al. Rethinking fid: Towards a better evaluation metric for image generation[C]// CVPR. 2024: 9307-9315.	Summary: The authors observe computational redundancies across the three main dimensions of the generation process in DiTs -- space, sampling time, and conditional vs. unconditional forward passes due to CFG. Based on these observations, the authors introduce a set of approaches leveraging them to enable more efficient inference with negligible quality loss for pre-trained DiTs. Strengths: - This paper addresses an important avenue of research, as diffusion models are inherently very inefficient (as compared to, e.g., GANs) for inference, due to having to evaluate large models for a multitude of steps. While previous works such as DeepCache (Ma et al., 2023) or Cache Me If You Can (Wimbauer et al., 2023) have already investigated this problem in general, they do so specifically for diffusion U-Nets and many of these methods are not directly applicable to DiTs that do not share the hierarchical nature of U-Nets. Given the recent rise in popularity of DiT-based diffusion backbones for successful large-scale diffusion models (PixArt-alpha, Sora, Stable Diffusion 3, ...), providing methods that can successfully speed up the inference of these models is important. The authors present a method that is able to substantially reduce the cost incurred from evaluating attention, which makes up a large fraction of the FLOPS in high-resolution settings, in these diffusion transformers during inference while retaining most of the quality of the generated images, addressing this problem well. The authors investigate and identify a set of redundancies during the generation process in diffusion transformers that relate to evaluating the attention mechanism. These findings are then used to motivate the presented methods. - Going beyond naive approaches to exploit the presented redundancies, the authors introduce further, non-obvious tricks such as residual sharing to make them work well. - The main method is thoroughly evaluated on two image diffusion models where it shows promising results. Weaknesses: - While the method as a whole is evaluated reasonably well, the evaluation of the effect of parts of the method & design choices seems lacking: there are missing ablations regarding different variants (e.g., naive approaches) of leveraging specific redundancies, and especially the evaluation of the compression plan algorithm is lacking and its hyperparameters (threshold $\delta$, number of samples). - Open questions remain, especially regarding unorthodox design choices and whether the CFG redundancy is actually an attention redundancy as described or a general redundancy already thoroughly described in previous works. See the questions below. Technical Quality: 3 Clarity: 3 Questions for Authors: Questions regarding the main contributions of the paper - While their approach to only computing attention locally does seem to be effective, I am surprised by the fact that the authors chose to apply a 1D window attention on the flattened image token sequence instead of one of the many 2D-aware local attention mechanisms such as (Shifting) Window Attention. These should have been the obvious choice in this situation, and the authors fail to motivate their unorthodox choice. While the attention map patterns presented in Fig. 3a do look somewhat like the effective attention pattern given by 1D window attention, I would presume that the additional diagonal lines next to the main diagonal correspond to the same location, but offset by one row. Besides generally being a more intuitive choice, the aforementioned approaches should also correspond to this observed pattern significantly better. - Is it possible that the observed CFG redundancy is not an attention-specific aspect but a general one that is already well-described in previous literature? I would be very interested in a comparison with CFG not being applied at all for low noise timesteps, given the large range of prior works that found it beneficial to, substantially anneal the CFG scale for low noise timesteps (MDT, Gao et al., 2023; Analysis of Classifier-Free Guidance Weight Schedulers, Wang et al., 2024; and a wide range of non-published sources such as https://enzokro.dev/blog/posts/2022-11-28-sd-v2-schedules-1/, https://x.com/jeremyphoward/status/1584771100378288129, https://github.com/mcmonkeyprojects/sd-dynamic-thresholding), or even completely deactivate it (Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models, Karras et al., 2024). - What is the distribution of which acceleration method is used how often, and where in the generation process? Additional small questions - Can the authors provide an explanation regarding the intuition of FID improving when reducing the attention FLOPS in Fig. 5? - Subjectively, more compression results in a loss of contrast, both for generated images (albeit slightly), as well as videos. Do the authors have an explanation as to what might be causing this? - For their similarity analyses in Fig. 4, the authors only analyze the cosine similarity of attention outputs. Are the magnitudes similar as well? Wouldn't this be a requirement to enable successful caching? Presentation suggestions - Fig. 5 is hard to parse due to the multitude and different scales of the presented metrics. I'd suggest that the authors separate the metrics into separate stacked graphs, which should also render these graphs more accessible for people with color perception impairments. I would also suggest adding arrows to indicate whether lower or higher is better for each individual metric. Similar improvements to other figures could also help make the paper more accessible. - For a camera-ready version, I think the paper would benefit from additional uncurated qualitative examples Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations and societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer for the insightful and constructive comments. > W1. the evaluation of the effect of parts of the method & design choices seems lacking: there are missing ablations regarding different variants (e.g., naive approaches) of leveraging specific redundancies, and especially the evaluation of the compression plan algorithm is lacking and its hyperparameters (threshold $\delta$, number of samples). We show the effect of some naive approaches in Figure 9 in our paper, and we have added more images generated (like Figure 5a in the attached pdf) and compression plan (like Figure 5c in the attached pdf) by naive approaches. We conducted an ablation study on number of samples, and we found that adding number of samples had little effect on the compression plan. DiT n_samples ablation \| num of samples \| Threshold \| IS \| FID \| \|---------\|-----------\|-------------\|----------\| \| 8 \| 0.05 \| 206.43 \| 29.97 \| \| 8 \| 0.1 \| 200.30 \| 26.91 \| \| 8 \| 0.15 \| 179.13 \| 23.11 \| \| 16 \| 0.05 \| 205.76 \| 30.33 \| \| 16 \| 0.1 \| 197.16 \| 27.62 \| \| 16 \| 0.15 \| 180.68 \| 23.24 \| \| 32 \| 0.05 \| 205.63 \| 30.35 \| \| 32 \| 0.1 \| 202.73 \| 26.97 \| \| 32 \| 0.15 \| 180.22 \| 24.55 \| In global rebuttal Q5, we added more ablation studies. > W2. & Q2. Open questions remain, especially regarding unorthodox design choices and whether the CFG redundancy is actually an attention redundancy as described or a general redundancy already thoroughly described in previous works. Thanks for your comments. We believe that the issue of conditional guidance (CFG) redundancy is not limited to the attention mechanism alone. There has been some prior work exploring methods to reduce CFG redundancy. However, our approach differs from previous efforts, as we have examined the use of CFG sharing at a finer granularity. Specifically, we have assessed the influence of conditional and unconditional computations across different layers and different timesteps. Our work applies CFG sharing selectively, only on certain layers and time steps, in order to avoid significantly degrading the model's performance. In this way, our work acknowledges and builds upon existing research, while offering a novel perspective on addressing redundancy within the unique framework of diffusion models. Going forward, we plan to extend our method to other model components in the future. > Q1. While their approach to only computing attention locally does seem to be effective, I am surprised by the fact that the authors chose to apply a 1D window attention on the flattened image token sequence instead of one of the many 2D-aware local attention mechanisms such as (Shifting) Window Attention. These should have been the obvious choice in this situation, and the authors fail to motivate their unorthodox choice. Please see global rebuttal Q4. > Q3. What is the distribution of which acceleration method is used how often, and where in the generation process? Please check Q2 in global rebuttal and Figure 1 in the rebuttal pdf attached. > Q4. Can the authors provide an explanation regarding the intuition of FID improving when reducing the attention FLOPS in Fig. 5? We have thoroughly reviewed the evaluation code and confirmed that there are no issues on our end. We believe the discrepancy may be due to the inaccuracy of the FID metric, as indicated in previous research [1]. The FID does not fully capture the visual quality of the generated images. However, changes in the FID score can still be meaningful, as they reflect shifts in the distribution of the generated images. [1] Jayasumana S, Ramalingam S, Veit A, et al. Rethinking fid: Towards a better evaluation metric for image generation[C]// CVPR. 2024: 9307-9315. > Q5. Subjectively, more compression results in a loss of contrast, both for generated images (albeit slightly), as well as videos. Do the authors have an explanation as to what might be causing this? We evaluated the images generated using different methods and found that the high threshold of ASC was likely the cause. We included an example generated by different methods and at varying ASC thresholds in the rebuttal PDF (Figure 5). We suspect that excessive CFG sharing can lead to a loss of detail and less pronounced edges in the generated images. If the trade-off between contract quality and ASC is a significant concern, the user should consider disabling the ASC method. > Q6. For their similarity analyses in Fig. 4, the authors only analyze the cosine similarity of attention outputs. Are the magnitudes similar as well? Wouldn't this be a requirement to enable successful caching? Thanks. We plot the magnitude of attention output, the plot can be found in the rebuttal pdf (Figure 6). It is similar to cosine similarity. > Q7. Fig. 5 is hard to parse due to the multitude and different scales of the presented metrics. I'd suggest that the authors separate the metrics into separate stacked graphs, which should also render these graphs more accessible for people with color perception impairments. I would also suggest adding arrows to indicate whether lower or higher is better for each individual metric. Similar improvements to other figures could also help make the paper more accessible. Thanks for your advice. Please see Q3 in global rebuttal. > Q8. For a camera-ready version, I think the paper would benefit from additional uncurated qualitative examples Thanks. We plan to include more uncurated qualitative examples, ablation studies, plots, and tables in the future version of our paper, including the results presented in the rebuttal PDF. --- Rebuttal Comment 1.1: Comment: Thank you for the extensive response and for running the additional ablations in the short rebuttal timespan! Regarding your response to my CFG sharing comment, I'd like to clarify that I was primarily referring to the practice of omitting CFG for the final few steps altogether, which also seems to have a negligible effect in practice, and is substantially simpler. Regarding the Natten comparison, I'm really surprised that the Natten kernel is seemingly so inefficient despite the substantially smaller number of elements, and I thank the authors for providing that valuable context. I'd also suggest including that in a future revised version of the paper and potentially still running practical quality evaluations with Natten, as it should be substantially more FLOP-efficient and might catch up w.r.t. practical speed with innovations like flex attention. For now, the provided context is sufficient for me. --- Reply to Comment 1.1.1: Comment: Thanks for your comments and valuable suggestions. > Regarding your response to my CFG sharing comment, I'd like to clarify that I was primarily referring to the practice of omitting CFG for the final few steps altogether, which also seems to have a negligible effect in practice, and is substantially simpler. Thanks for your clarification. We conducted an ablation study on CFG dropping in the final steps. The results are shown in the following table. We observed that the method of CFG dropping in the final steps is effective in DiT. We also observed that our DiTFastAttn can work well with this method. We will add this to the appendix in our revised version of the paper, along with more evaluation results and generated images. \| Setting \| IS@5K \| FID@5K \| \|---------------------\|----------\|------\| \| Raw \| 208.3134 \| 31.65\| \| CFG Dropping final 10% steps \| 278.4405 \| 30.5 \| \| Threshold = 0.1 \| 200.2979 \| 26.91\| \| Threshold = 0.1 + CFG Dropping final 10% steps \| 262.5407 \| 26.77\| Note that the method only focuses on CFG dropping in the final steps, while our CFG sharing in our method reduces redundancy in other steps. Since the CFG dropping is manually set now, it is an interesting idea to automatically search for where to drop the CFG. We are considering researching this point in the future. > I'd also suggest including that in a future revised version of the paper and potentially still running practical quality evaluations with Natten, as it should be substantially more FLOP-efficient and might catch up w.r.t. practical speed with innovations like flex attention. For now, the provided context is sufficient for me. Thanks for your suggestion. We will include the Natten results in the revised version of our paper. So far, we have provided the results of 1D vs 2D local attention, n_samples, loss of contrast, magnitude, and the compression plan for Q1, 3, 5, 6, and 7. We have also provided a reasonable explanation for Q2, and an analysis of CFG sharing to address W2 and Q2. We are open to further discussion and happy to answer any additional questions. As we are nearing the end of the discussion period, we kindly ask you to reconsider the score.	Summary: The paper presents a post-training model compression method aimed at reducing the computational complexity of Diffusion Transformers. The authors identify three key redundancies in the attention computation. To address these, they propose three techniques. The proposed methods compress the model FLOPs and enable more efficient deployment in resource-constrained environments. The paper includes extensive experimental results demonstrating significant reductions in computational cost and latency while maintaining generation quality. Strengths: - The identification and analysis of spatial, temporal, and conditional redundancies in attention mechanisms are thorough and well-founded. - The authors conduct extensive experiments on multiple DiT models and various tasks, providing strong evidence of the effectiveness and generalizability of their methods. Weaknesses: - Limited Theoretical Analysis: I am in general agreement with the overall concept presented in the paper. However, the paper primarily focuses on empirical results, with limited theoretical analysis to support the proposed methods. A deeper theoretical foundation could help in understanding the generalizability and limitations of the techniques. - Complexity and Scalability: The paper does not extensively discuss the complexity of the compressed models. The figure 1 presents the efficacy under different resolutions. But it does not discuss the model complexity, running speed, and model size in these settings. - Insufficient Comparisons: The paper does not compare the related methods, such as flashattention, and KV cache. This mainly weakens the contribution. A more detailed comparison with existing attention acceleration and model compression techniques would be beneficial. - Vague Illustration：The authors present three key redundancy strategy. However, the paper lacks sufficient annotations and explanations and does not provide clear definitions for the three scenarios. - Application: The paper mainly focuses on attention reduction, but It is not very effective at low resolutions because the attention mechanism accounts for a small proportion of the computational cost. Technical Quality: 2 Clarity: 3 Questions for Authors: - How do the authors ensure the generalizability of their method across different datasets and tasks? A more rigorous theoretical analysis could help understand the limitations and potential extensions of the proposed techniques. - What is the model complexity, running speed, and model size under different compression settings? - Can the authors offer clearer definitions and detailed illustrations for each redundancy type (spatial, temporal, and conditional redundancy)? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors have acknowledged several limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer for the insightful and constructive comments. > W1 & Q1. How do the authors ensure the generalizability of their method across different datasets and tasks? A more rigorous theoretical analysis could help understand the limitations and potential extensions of the proposed techniques. About generalizability, our experiments demonstrate the effectiveness of DiTFastAttn across multiple datasets and tasks: - Image generation: We evaluated on class conditional generation (for DiT) and text conditional generation (for PixArt-Sigma). - Video generation: We successfully applied DiTFastAttn to OpenSora for video generation tasks. Our method shows improved compression ratios and quality preservation as resolution increases from 512x512 to 2048x2048, indicating good scalability. This diversity in datasets, tasks, and model architectures provides strong evidence for the generalizability of our approach. However, we acknowledge that further evaluation on additional domains could further strengthen this claim. For theoretical analysis, we thank the reviewer for this valuable suggestion. While our current work focuses on empirical results, we agree that a more rigorous theoretical analysis would be beneficial. - We have provided rigorous computational complexity analysis for each proposed technique (WA-RS, AST, ASC) and their combinations. Some directions we could explore in future work include: - Investigate how the error propagates through the denoising steps when using approximated attention outputs. - Quantify the information loss when using window attention compared to full attention. We will consider adding a discussion of these theoretical aspects in an extended version of the paper. > W2 & Q2. What is the model complexity, running speed, and model size under different compression settings? In Appendix A.4 of our paper, we discussed the complexity of our compression method. Additionally, in the global rebuttal, we have added summary tables (Table 1 and 2) to present the exact FLOPs and latency (running speed). The result shows our method is effective to reduce both the complexity and improve the running speed. The model size will not change. For model size and VRAM usage, when AST or WA-RS is applied, the hidden states from the previous timestep are stored. The shape of stored hidden states of a layer is [batchsize*2, num_of_head, seqlength, hidden_size]. For example, with the DiT XL/2 512 model, at a batchsize of 1, the VRAM usage of the original model is 11652MB. To store the FP16 hidden states, an additional 2016MB of VRAM is needed, which is 17.3% of the original model's VRAM usage. We will include a VRAM analysis in the limitations section. > W3. Insufficient Comparisons: The paper does not compare the related methods, such as flashattention, and KV cache. This mainly weakens the contribution. A more detailed comparison with existing attention acceleration and model compression techniques would be beneficial. Thanks. Our baseline (we denoted as raw setting in our method) has used flashattention in their attention computation, and our method can achieve up to 1.6x speedup compared to it. We'll add a note to the paper to avoid confusion. KV cache is a method used in large language models for auto-regressive generation. In diffusion model, DiT don't need to attend the tokens from previous timesteps. KV cache isn't applicable to diffusion models like DiT. So we cannot compare our method with KV cache. Note that our method is inspired by caching concepts but applied differently: - Caching activation outputs across denoising timesteps and conditional/unconditional branches. - WA-RS technique caches residuals to maintain long-range dependencies with efficient local attention. > W4 & Q3. Can the authors offer clearer definitions and detailed illustrations for each redundancy type (spatial, temporal, and conditional redundancy)? Although we have provided some definitions in our paper, we agree that providing more clarity on these concepts will strengthen our paper. Here we denote $X$ as input, $Y$ as output Window Attention with Residual Sharing (WA-RS): - Definition: The original paper Eq.1 and Eq.2. - Illustration: The original paper Figure.3. We will add a heat map visualization of the attention matrix, clearly showing the concentration of attention values along the diagonal. Attention Sharing across Timesteps (AST): - Definition: For step k, $Y_k=W_{o}O_k$ and $Y_{k+1} = \\{Y_k$ if $AST_k = 1$; $W_{o}O_{k+1}$ if $AST_k = 0\\}$ - Illustration: The original paper Figure.2 is clear Attention Sharing across CFG (ASC): - Definition: For step k, $Q_{k} = \\{W_{Q}X_{k,:c/2}$ if $ASC_k = 1$, $W_{Q}X_{k}$ if $ASC_k = 0\\}$ and so as $K_{k}$ and $V_{k}$. The output is $Y_{k} = \\{[W_{o}O_k, W_{o}O_k]$ if $ASC_k = 1$; $W_{o}O_{k}$ if $ASC_k = 0\\}$ - Illustration: The original paper Figure.2 is clear > W5. Application: The paper mainly focuses on attention reduction, but It is not very effective at low resolutions because the attention mechanism accounts for a small proportion of the computational cost. While it's true that our method provides greater benefits at higher resolutions, we believe this is actually a strength rather than a limitation: - Even at lower resolutions, our method still provides measurable improvements. For example, on the 512x512 DiT model, we achieve up to 31% FLOP reduction and 10% latency reduction for attention computation. - As transformer-based diffusion models continue to scale up in size and target higher resolutions, the relative importance of attention computation increases. Our method is thus well-positioned to provide even greater benefits for future large-scale models. --- Rebuttal Comment 1.1: Comment: Dear Reviewer e43K, Thanks so much again for the time and effort in our work. According to the comments and concerns, we provided more analysis and further discuss the related points. We also conduct a variety of ablation studies and experiments that we showed in the Global rebuttal. As the rebuttal period is about to close, may I know if our rebuttal addresses the concerns? If there are further concerns or questions, we are welcome to address them. Thanks again for taking the time to review our work and provide insightful comments. --- Rebuttal Comment 1.2: Comment: After reviewing the authors’ rebuttal and all subsequent responses, I find that most of my concerns have been addressed. Consequently, I have decided to revise my rating to a positive one. I recommend that the additional experiments and analyses discussed be incorporated into the revised version of the manuscript.	Summary: This paper proposes a combination of novel techniques to reduce the self-attention computation in diffusion transformers (DiTs), without requiring finetuning. This methods leverage locality in attention scores, similarity in timesteps, and similarity with classifier free guidance. Overall, the proposed method is capable of maintaining image quality while significantly reducing self-attention FLOPS. Strengths: The paper proposes three novel approaches with compelling empirical justification for each. The methods were then ablated and demonstrate good visual performance while reducing computational overhead during inference. While the authors acknowledge limitations of their kernel implementation, their results may have a strong impact in both downstream deployment and in explainable AI for generative image models. Weaknesses: While the visual results are compelling, the paper has limited quantitative evaluations, and may be in a regime with low statistical SNR requiring additional samples. Furthermore, the authors focus on improving FLOPS, however, they do not discuss VRAM usage which would be significant for consumer grade devices and generating longer videos. Finally, there are issues with figure 5 which significantly hinders readability. Technical Quality: 2 Clarity: 3 Questions for Authors: 1) You should cite Sora in the introduction and not OpenSora. 2) While the inclusion of residual caching is novel, how is your windowed attention method different from neighborhood attention (Natten)? This should be cited with local attention. 3) Including the 1-sigma boundaries (e.g. aggregated over 1k samples) in the MSE plots on Fig 3a would help understand how much variation there is for the attention layers. Additionally, is that DiT-XL or Pixart-Sigma? Identifying the model in the caption would be a good idea. 4) You explored the impact with CFG but have you looked at the impact with negative conditioning? Does sharing the conditional attention perform similarly or degrade? 5) There appears to be some interesting behavior in Figure 4, which may be interesting to explore from an explainability perspective. 6) For your compression plan search, did you consider using other metrics such as LPIPS or SSIM? Would they impact the chosen plan? 7) The use of 5k image samples for evaluation is unusual. Typically the evaluation is 50k for ImageNet and 30k for COCO, as a larger sample size will reduce statistical noise (FID can be very noisy below 15k samples). 8) When you evaluated FID, IS, CLIP, did you do so on the same 5k samples used for determining the plan? I would have used the ImageNet/COCO validation set for the plan search and then evaluated on the test set to ensure the two distributions were non-overlapping. 9) What dataset did you use for your OpenSora evaluation? 10) Did you consider looking at the impact on VRAM? If there is no impact, stating so in the limitations would be appropriate. 11) If increasing the step count improves the result in Fig 9, then this suggests that there is an optimal tradeoff between attention FLOPs and step count (iso-sampling FLOPS). Performing that analysis may strengthen adoption of your technique. 12) Reporting relative performance as percentages is difficult to read. It would be clearer to report them as ratios (e.g. 0.93 vs 93%) Finally, Figure 5 is unreadable and requires revision. - FID and IS should not be plotted on the same axis as they are very different in scale. Use a dual-axis plot if you want to combine them. - The points should be connected with a line like in Figure 8 and 9. - I would only plot CLIP and FID (not IS) for Pixart, which will make the plot easier to read. - Your CLIP scores appear too high, where they should be between 0.2 and 0.35. Please provide a revision to this figure or a table of the data plotted therein. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors adequately discuss limitations of their approach, although should include a note about VRAM. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We truly appreciate the reviewer for the insightful and constructive comments. > Q1. You should cite Sora in the introduction and not OpenSora. Thanks. We have added the citation of Sora in the introduction. > Q2. How is your windowed attention method different from neighborhood attention (Natten)? This should be cited with local attention. We adopted 1D window attention, which can be recognized as 1D neighborhood attention, and we have added references to Natten in the article. The evaluation shows our 1D window attention is more efficient (See global rebuttal Q4). > Q3. Including the 1-sigma boundaries (e.g. aggregated over 1k samples) in the MSE plots on Fig 3a would help understand how much variation there is for the attention layers. We have added the MSE change plot on 1k samples and included the 1-sigma boundaries. We have also added the model type in the figure title and caption. The results indicate that the variation is not significant. Please refer to Figure 4.b in the attached rebuttal PDF. > Q4. You explored the impact with CFG but have you looked at the impact with negative conditioning? Does sharing the conditional attention perform similarly or degrade? We investigated the impact of negative conditioning. The attentions with positive conditioning and negative conditioning have a high degree of similarity, and our method is also effective in reducing computation with negative conditioning. We provided an example in the attached PDF (Figure 4c, "low quality" as the negative prompt). We also added a plot to show that similarity across CFG with negative prompt behave similar as the one without negative prompt (Figure 4d). We will add more examples and analysis of negative conditioning in our paper. > Q5. There appears to be some interesting behavior in Figure 4, which may be interesting to explore from an explainability perspective. In Figure 4a, we observe that the middle timesteps exhibit less similarity, which may suggest that the structural transformation occurs primarily in the middle timesteps, while the other timesteps focus on noise removal and refinement. In Figure 4b, we note that the middle layers show more differences between the conditional and unconditional cases, indicating that the middle layers are responsible for the condition-specific processing. > Q6. For your compression plan search, did you consider using other metrics such as LPIPS or SSIM? Would they impact the chosen plan? Yes, we've considered LPIPS and SSIM for our comparison. LPIPS only works for image comparisons and requires the channel number to be set to 3, whereas we are comparing latent space (with 4 channels). LPIPS also needs to use a network as a feature extractor, which results in a significant inference overhead. We also tried the SSIM metric, and the results are included in Table 4 of the global rebuttal. The SSIM metric appears to achieve comparable results to our own metric. However, SSIM is more computationally complex to calculate. Therefore, we have decided to use our own metric for this task. We have included the compression plan in Figure 2 of the rebuttal PDF. Interestingly, the SSIM metric selects more lower-level layers for compression, while the existing metric chooses to compress the higher layers. We will research this in the future. > Q7. Evaluation sample number Please see global rebuttal Q1. Our conclusions remain unchanged at the larger evaluation size. > Q8. When you evaluated FID, IS, CLIP, did you do so on the same 5k samples used for determining the plan? I would have used the ImageNet/COCO validation set for the plan search and then evaluated on the test set to ensure the two distributions were non-overlapping. For the calibration process, we only use class labels (for DiT) and text prompts (for PixArt-Sigma) to generate a small set of images (8 images for DiT, 6 images for PixArt 1K, and 4 for PixArt 2K). No real images are used in the calibration process, so there is no risk of overlapping. > Q9. What dataset did you use for your OpenSora evaluation? Since the video generation process is too resource-intensive, and we found the evaluation metrics to be unstable when the sample size is small, we did not apply these metrics to the generated videos. Instead, we included some sample generated examples. These example videos were generated using the default prompts provided by OpenSora. > Q10. Did you consider looking at the impact on VRAM? If there is no impact, stating so in the limitations would be appropriate. When AST or WA-RS is applied, the hidden states from the previous timestep are stored. The shape of stored hidden states of a layer is [batchsize2, num_of_head, seqlength, hidden_size]. For example, with the DiT XL/2 512 model, at a batchsize of 1, the VRAM usage of the original model is 11652MB. To store the FP16 hidden states, an additional 2016MB of VRAM is needed, which is 17.3% of the original model's VRAM usage. We will include a VRAM analysis in the limitations section. > Q11. If increasing the step count improves the result in Fig 9, then this suggests that there is an optimal tradeoff between attention FLOPs and step count (iso-sampling FLOPS). In the rebuttal PDF, we have added a FLOPs-IS Pareto front plot (Figure 4.a). Interestingly, we observed that different compression ratios require different optimal settings (steps and DiTFastAttn threshold). We will include this plot and the related discussion in the appendix of the future version of the paper. > Q12. Reporting relative performance as percentages is difficult to read. It would be clearer to report them as ratios Thanks. We've changed the format of relative performance from percentage to ratio. > Fig 5 & CLIP issues Figure 5 is now changed accordingly (check global rebuttal and Figure 3 in the pdf). For CLIP, we use the package torchmetric to calculate it so the scale is different (0-100 vs 0-1). It is now changed to 0-1. --- Rebuttal Comment 1.1: Comment: I thank the authors for their effort with additional evaluations, the SSIM difference and compression plans are especially interesting. There are a few points that I would like clarification on. Q1.* There appears to be an issue with your evaluation metrics. The FID scores are too high (DiT-XL/2 should be around 2.2 with CFG, and Pixart-sigma around 9). Similarly, your IS scores for DiT-XL are too high and too low for Pixart. Were you able to reproduce the metrics from the respective papers without your compression method? Q2. > LPIPS only works for image comparisons and requires the channel number to be set to 3... This is partially true, where it would be possible to retrain a vgg model as per [1] in the latent space. However, it would also be feasible to simply decode the latent images and apply LPIPS in RGB space (this is also a differentiable operation). Q3. > For the calibration process, we only use class labels (for DiT) and text prompts (for PixArt-Sigma) to generate a small set of images (8 images for DiT, 6 images for PixArt 1K, and 4 for PixArt 2K). No real images are used in the calibration process, so there is no risk of overlapping. Is this a sufficient number of images to establish a robust compression plan? If the goal is to find a method for downstream inference performance, I would want to have at least 1k samples (maybe more). Q4. Given the early focus on full self-attention in the compression plans, would some downscaled method like HiDiffusion [2] be applicable? Q5. > ...we have added a FLOPs-IS Pareto front plot... Interpreting this plot is difficult, perhaps adding marker shapes for each compression ratio would make it clearer. However, if I understand correctly, it suggests that using no compression at 20 steps performs similarly to using moderate compression at 30 steps? If so, can you give an example for where your approach would be preferable to simply reducing the sampling steps? [1] Zhang, R., et.al. (2018), "The Unreasonable Effectiveness of Deep Features as a Perceptual Metric" [2] Zhang, S., et.al. (2023), "HiDiffusion: Unlocking Higher-Resolution Creativity and Efficiency in Pretrained Diffusion Models" --- Rebuttal 2: Comment: Thanks for your additional questions. Here is our reply: > Q1. Evaluation metrics The experiment settings are different in the following ways: - 1. The number of timesteps. To compare DiT with ADM and LDM, DiT uses a relatively high number of diffusion timesteps (250). For most image generation applications, 20-50 steps would be reasonable. Other methods, such as SD-XL and PixArt series, also evaluate on 20-50 steps. That is why we use 20-50 steps in our paper. - 2. Evaluation software. We use pytorch-fid & torchmetric to calculate FID and IS scores, while DiT uses the TensorFlow evaluation suite from the ADM paper. This resulted in some differences in the FID and IS scores. - 3. We use a cfg_scale of 4, which is the default setting for both DiT official code and diffusers DiT pipeline. In the DiT paper, they use a cfg_scale of 1.5 to match the settings used for ADM and LDM. We have modified the code to reproduce the results. However, it is not feasible to resample the 50K images under the timestep=250 setting during the rebuttal period (this would take approximately 55 GPU hours). We will provide a partial set of the results soon. For PixArt-Sigma, they used a curated set of 30,000 images instead of COCO to calculate FID. Since the dataset is not publicly released, we cannot reproduce those results. We will clearly explain the experiment settings in the appendix. > Q2. LPIPS implementation Thank you for the suggestion. We have implemented LPIPS as an additional evaluation metric (by decoding the hidden states into RGB space). However, we found that LPIPS has problems for our use case: - 1. Insensitivity to Value Changes: We observed that LPIPS is quite insensitive to value changes in our experiments. Only small value changes were observed when switching between different methods, and LPIPS always suggested using sharing across timesteps when the threshold was set to 0.005 or smaller. - 2. Computational Overhead: Using LPIPS as a metric takes significantly more time. Therefore, we were unable to provide the LPIPS results within the rebuttal period due to the computational requirements. > Q3. Number of calibration images We conducted an ablation study on the number of samples generated during the plan search, and found that increasing the number of samples had little effect on the resulting compression plan. \| n samples \| threshold \| IS \| FID \| \|---------\|-----------\|-------\|---------------\| \| 8 \| 0.05 \| 206.4319 \| 29.97 \| \| 8 \| 0.1 \| 200.2979 \| 26.91 \| \| 8 \| 0.15 \| 179.1331 \| 23.114 \| \| 16 \| 0.05 \| 205.7642 \| 30.33 \| \| 16 \| 0.1 \| 197.1631 \| 27.62 \| \| 16 \| 0.15 \| 180.6786 \| 23.24 \| \| 32 \| 0.05 \| 205.626 \| 30.35 \| \| 32 \| 0.1 \| 202.7274 \| 26.97 \| \| 32 \| 0.15 \| 180.2231 \| 24.55 \| From the results, the IS and FID did not change significantly as we varied the number of samples. Nevertheless, users can easily set 'n' to a larger number if desired, as it is a configurable parameter in our code. >Q4. downscaled method like HiDiffusion be applicable? We have carefully reviewed the HiDiffusion approach, which is designed for UNet-based diffusion models. We have considered how this method may be applicable to our framework: - 1. HiDiffusion proposed using local attention to replace global attention in the top layers of the UNet, as global attention in the upper blocks is computationally dominant. However, in DiT, the amount of self-attention computation is equal across each layer. Therefore, we do not need to focus solely on specific layers, and instead use a search-based method to automatically identify which layers can be replaced. 2. HiDiffusion use the Modified Shifted Window Attention, where the window area is set differently across diffusion timesteps. This approach may be applicable to our method as well. However, as mentioned in the Global rebuttal (Q4), the inference speed of these attention mechanisms could be slower. It is also an interesting topic to research. >Q5. Pareto front question Thanks for your suggestion. We have followed your advice and added marker shapes. Here is our observation from the plot: - The point representing compression at 30 steps (yellow marker at (2.7 TFLOPs, 208 IS)) is superior to using no compression at 20 steps (blue marker at (2.9 TFLOPs, 207 IS)). This is because the compressed 30-step model has lower computational cost (TFLOPs) while achieving a higher Inception Score (IS). - For a high FLOPs budget (> 3.5T), it is better to use moderate compression with 50 steps. - For a medium FLOPs budget (2-3.5T), it is preferable to use slight compression with 20 steps or moderate compression with 30 steps. - For a low FLOPs budget (< 2T), the optimal choice would be to use high compression with either 20 or 30 steps. --- Rebuttal 3: Comment: Thank you for your patience. Here we present the results for DiT after changing our experimental settings to align with the DiT paper. \| Threshold \| IS@50K \| FID@50K \| macs \| attn_mac \| \|-----------\|--------\|---------\|------\|----------\| \| Raw \| 219.9721 \| 3.16 \| 262359 \| 33823 \| \| 0.025 \| 218.1955 \| 3.09 \| 236265 \| 18041 \| \| 0.050 \| 210.3559 \| 3.10 \| 218865 \| 12339 \| \| 0.075 \| 196.05 \| 3.54 \| 203420 \| 8777 \| \| 0.100 \| 180.34 \| 4.52 \| 195682 \| 7137 \| From this table, we observe that when the threshold is increased from 0 to 0.10, the IS drops from 210 to 180, and the FID decreases from 3.16 to 3.09, then increases to 4.52. The behavior of the FID in different settings follows the pattern reported in the previous discussion[1]. We observe a higher reduction in FLOPs and Attention FLOPs in this setting. We believe the current results demonstrate the effectiveness of our method across different settings. So far, we provide explanations, revision and examples according to each concerns. We are open to more discussion and will to provide more clarification if needed. As we nearly reached the end of the discussion period, we kindly ask you to reconsider the score. [1] Jayasumana S, Ramalingam S, Veit A, et al. Rethinking fid: Towards a better evaluation metric for image generation[C]// CVPR. 2024: 9307-9315. --- Rebuttal Comment 3.1: Comment: Thank you for the follow-up clarification. Q1. The updated FID and IS scores are closer to what is expected. While I understand your desire to evaluate under more practical settings, maintaining a consistent evaluation protocol is necessary to compare with other works. This requirement for comparison is much more sensitive to image count and cfg scale rather than sampler, which is why many recent works deviate in sampler algorithm and step count (computational practicality). From your updated evaluations, I interpret the results as significant degradation beyond a threshold of 0.05, below which the speedup is not significant. Q2. The results you described with LPIPS are interesting, and should be included in the revision for completeness. If anything, they will at the very least indicate that a less computational method achieves similar or better results for less overhead. Q3. Your results exhibit a significant improvement using more samples for a threshold above 0.1. When interpreting FID, a shift of 1-2 points can be significant, although this may be reduced with the updated conditions from Q1. I remain concerned with the low number of images used in calibration, which I believe may serve to highlight issues in the generality of FID and IS as standard evaluation metrics. While I understand that more samples are computationally expensive, the cost largely becomes irrelevant if it need only be performed once per model. --- I thank the authors for their effort in addressing my questions and those of the other reviewers. Additionally, I commend the authors on their detailed investigations, which make this an interesting technical work. As such, I am inclined to raise my score. However, I believe the variable quality as a function of calibration sample count brings into question the efficacy of the results. Furthermore, the results presented bring into question the appropriateness of using FID and IS as primary evaluation metrics rather than as a sanity check against significant degradation. Given the aforementioned pros and cons, I raise my score from 4 to 5. --- Rebuttal 4: Title: Clarification on updated results Comment: We appreciate your valuable feedback. In the revised version, we will include the LPIPS results and additional evaluation metrics, as well as more uncurated qualitative examples, as you suggested. Regarding your interpretation of the updated evaluations (Q1), there seems to be some misunderstanding. Our findings indicate that even at a threshold of 0.025, the method was able to significantly reduce the attention FLOPs by 46%, while maintaining a satisfactory generation effect. At a threshold of 0.05, the attention FLOPs were reduced by 63% without significant degradation. The results show that as the number of steps increases, the threshold required to achieve the same compression ratio decreases. For example, with 250 steps, a threshold of 0.025 can achieve the same compression ratio as a threshold of 0.125 with 20 steps. This means that for larger timesteps, a smaller threshold is needed to reach the same compression ratio. Therefore, we believe our results demonstrate that in the experiment setting of 250 steps, a small threshold can indeed achieve good compression performance. We will adjust the threshold values and include more results in the range of 0 to 0.05 to better illustrate this point. We kindly ask you to reconsider the score if we have addressed all of your concern. --- Rebuttal Comment 4.1: Comment: Thank you for the additional context. I remind the authors that FLOPs is not a good metric for inference speedup, especially considering that many inference workloads are memory bound. While FLOPs can serve as a heuristic, the bigger picture is more complicated. Instead, overall inference latency is a more appropriate metric of which a significant deviation is not observed for thresholds below 0.1. From the other results, it is my belief that your upper performance bound is limited by the number of calibration samples, which results in plans that are unable to effectively generalize across the standard evaluation metrics. The authors should focus on that aspect: reducing degradation for higher thresholds to achieve significant latency reductions, rather than more results with thresholds below 0.05. --- Reply to Comment 4.1.1: Comment: Thank you for your reminder. We have tested the latency to the table for your reference. We observed a 30% reduction in DiT-XL/2 attention latency at a threshold of 0.05 (timesteps=250, 512x512 generation), showing that our method not only reduces FLOPs but also results in a significant speedup in the attention calculation part. Regarding the overall inference latency, your observation is correct that the latency reduction fraction is lower than the attention latency reduction fraction. However, the proportion of attention computation in the overall computation would increase along with resolution (as shown in Figure 1 of our paper). Our method can significantly reduce the overall latency on the generation of high-resolution images. For 2K image generation, attention computation contributes more than 70% of the total computation time, so a 30% reduction in attention latency will result in more than a 21% reduction in overall latency. Regarding the number of calibration samples, we agree with your idea and will increase the number to further reduce degradation. However, it will take more computation cost. This is a trade-off between calibration time and performance. We will carefully evaluate the number of samples in the future. Thanks again for your remind and advice. If we have addressed all of your concern, we kindly ask you to reconsider the score.	Rebuttal 1: Rebuttal: Thanks all the reviewers for the time and effort taken to provide valuable insights and comments on our work. Here we provide experiment results and answers for some common questions and comments: > 1. The use of 5k image samples for evaluation is not enough. We have increased the evaluation size of ImageNet to 50K samples and COCO to 30K samples. The table below showing the results of DiT (Table 1) and PixArt-Sigma 1K (Table 2). Plots of the ImageNet@50K and COCO@30K results can be found in the rebuttal PDF (Figure 3). The results are more stable and demonstrate the same trend of change as the previous 5K results. These updated results will be included in the future version of the paper. #### Table1. DiT ImageNet 50K Results \| Threshold \| MACs (GFLOPs) \| Attn_MAC (GFLOPs) \| Latency \| Attn_Latency \| IS \| FID \| \|-----------\|------:\|----------:\|---------:\|--------:\|---:\|----:\| \| 0 \| 20989 \| 2706 \| 2.841s \| 0.890s \| 400.64 \| 23.99 \| \| 0.025 \| 20623 \| 2383 \| 2.863s \| 0.918s \| 402.24 \| 23.86 \| \| 0.05 \| 20105 \| 2088 \| 2.854s \| 0.914s \| 400.28 \| 22.95 \| \| 0.075 \| 19560 \| 1832 \| 2.892s \| 0.909s \| 401.37 \| 21.75 \| \| 0.1 \| 19032 \| 1598 \| 2.782s \| 0.828s \| 385.48 \| 19.74 \| \| 0.125 \| 18432 \| 1361 \| 2.769s \| 0.799s \| 330.35 \| 18.09 \| \| 0.15 \| 17796 \| 1209 \| 2.658s \| 0.710s \| 328.21 \| 15.10 \| #### Table2. PixArt 1K COCO 30K Results \| Threshold \| MACs(GFLOPs) \| Attn_MAC (GFLOPs)\| Latency\| Attn_Latency \| IS \| FID \| CLIP \| \|----\|----:\|----:\|-----:\|----:\|---:\|---:\|-----:\| \| 0 \| 132693 \| 46464 \| 2.437s \| 0.937s \| 24.25\| 55.70\| 0.31377 \| \| 0.025 \| 129478 \| 43662 \| 2.459s \| 0.958s \| 24.28\| 55.67\| 0.31378 \| \| 0.05 \| 123701 \| 38385 \| 2.508s \| 0.993s \| 24.23\| 55.58\| 0.31371 \| \| 0.075 \| 117661 \| 33041 \| 2.472s \| 0.962s \| 24.18\| 55.28\| 0.31365 \| \| 0.1 \| 113351 \| 29427 \| 2.417s \| 0.922s \| 23.98\| 55.11\| 0.31342 \| \| 0.125 \| 108101 \| 25330 \| 2.372s \| 0.868s \| 23.74\| 53.90\| 0.31342 \| \| 0.15 \| 103403 \| 22131 \| 2.300s \| 0.798s \| 23.40\| 51.68\| 0.31314 \| > 2. What is the adopted compression strategy? We generated plots for the adopted compression strategies under different threshold settings. Please refer to Figure 1 in the rebuttal PDF. The results show that all the strategies (WA-RS, ASC, WA-RS+ASC, and AST) are well-distributed, demonstrating that our strategy search algorithm is effective. An interesting observation is that the different strategies tend to be concentrated in specific layers and timesteps. > 3. Figure 5 is unreadable. We have optimized the display of Figure 5. We show the result of each metrics seperately. Arrows are added to indicate whether lower or higher is better for each individual metric. We show 4 plots in the rebuttal pdf (Figure 3) as an example due to limited spaces. > 4. What kind of local attention is used? Why use 1d local attention instead of 2d local attention? We use sliding window attention for WA-RS, which is the same as the sliding window attention used in Longformer and the 1D neighborhood attention. Our method can also be generalized to 2D local attention. However, we observe that 2D local attention is not efficient enough. This is because the GPU memory access of the K and V tensors is not sequential, resulting in extra data gathering overhead. We evaluate the inference latency of adopting Natten 2D attention and our sliding window attention. We use a kernel size of 127 for the sliding window, and a kernel size of 5 for the Natten 2D attention. The latency results are shown in Table 3 below. Based on these results, we chose to use 1D local attention. #### Table3. Latency of different kernel \| Kernel \| Attention Latency \| \|-----------------\|--------------------\| \| Window Attention \| 0.828s \| \| Natten 2D Attention \| 0.926s \| > 5. More ablations and examples are needed. In the rebuttal PDF, 1. we conducted additional ablation studies on different metrics used in the search (Figure 2). We tested different SSIM compression schemes and found that when SSIM is chosen as the metric, to ensure the quality of the images generated, the threshold should be set at a small value of about 1/10 of the existing metric. 2. We investigated different the impact of the negative prompt (Figure 4c) and found that negative prompt remain effective with our method. 3. We checked the effect of CFG sharing on contrast (Figure 5). Loss of contrast can happen when using a large threshold for ASC. 4. We provided examples of the Pareto front of number of steps vs. TFLOPS (Figure 4a), which can served as guidance for user to choose appropriate compression setting. 5. We provided the MSE plot of 1K samples with 1-sigma in DiT (Figure 4b) to show that window part have a higher MSE and larger variation across samples. 6. We plotted the Magnitude of Attention Outputs Across Step and CFG Dimensions in DiT (Figure 6) to show magnitude of attention output show a similar result with cosine similarity. #### Table4. SSIM vs Current Metric \| Threshold \| Metric \| FID@5K \| IS@5K \| MACs (GFLOPs)\| Attn_MAC (GFLOPs)\| Latency \| Attn_Latency \| \|---\|----\|------\|------\|----\|------\|------\|-----\| \| 0.005 \| SSIM \| 28.77 \| 201.92 \| 19324 \| 1770 \| 2.851s \| 0.862s \| \| 0.01 \| SSIM \| 26.71 \| 200.92 \| 18611 \| 1492 \| 2.794s \| 0.825s \| \| 0.015 \| SSIM \| 25.44 \| 192.42 \| 18042 \| 1286 \| 2.764s \| 0.774s \| \| 0.05 \| Current \| 30.33 \| 205.76 \| 20102 \| 2064 \| 2.957s \| 0.950s \| \| 0.1 \| Current \| 27.62 \| 197.16 \| 18976 \| 1579 \| 2.834s \| 0.851s \| \| 0.15 \| Current \| 23.24 \| 180.68 \| 18028 \| 1229 \| 2.713s \| 0.756s \| Pdf: /pdf/9bbb8153a9c7d198b90cc34b26e7a0cbaf8f0de1.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Gaussian Process Bandits for Top-k Recommendations	Accept (poster)	Summary: This paper addresses the problem of top-k recommendation with bandit feedback, where the goal is to recommend a list of k items and receive feedback only on the overall quality of the list, not individual items. This captures the notion that user satisfaction depends on the overall quality of the recommendations, not just the single most relevant item. To tackle this, the authors propose a novel algorithm combining Gaussian Process Upper Confidence Bound (GP-UCB) with a specifically designed weighted Kendall kernel. The paper proposes faster inference exploiting features of the Kendall kernel, improving computational complexity from $O(T^4)$ to $O(T^2)$. An upper bound on cumulative regret is derived, showing sublinear regret. Finally, synthetic experiments show that the proposed algorithm can outperform a few baselines. Disclosure: I have reviewed a previous version of this paper. Strengths: Strengths: * Extends the literature on top-k recommendations with a GP-UCB algorithm * Well presented Weaknesses: Weaknesses: * The approach does not seem scalable to real-world problems * The theoretical contribution could be further clarified Technical Quality: 3 Clarity: 3 Questions for Authors: * The theoretical analysis builds on the analysis of Krause & Ong (2011), adapted to the Kendall kernel. Can the authors please clarify the technical contribution (novelty) compared to Krause & Ong (2011). * The approach does not seem practical. The empirical evaluation is done with a small number of items (n<=50) and short horizon (T<500). In real-world applications the number of items is often much larger and the number of rounds can also be large. * The acquisition function is not applied to all possible top-k rankings, rather a local search approach is used. The regret analysis does not account for the approximation due to the local search. The implementation also uses an approximate posterior rather than an exact posterior. This creates a discrepancy between the implementation of the algorithm and the theoretical analysis. The authors are somewhat upfront about this. * L304 (Evaluation for large arm space): First, I wouldn’t say that n=50 is large. In real-world applications n is in the millions (e.g., music, books) or billions (e.g., videos). Second, when showing the regret in Figrue 3 you should actually show all iterations (1-500), not just the ones that include a model update (every 5 steps). * A few comments from the previous review that are still not fixed: * You still have an overloaded notation $\sigma$ for both a ranking (L88) and the sigmoid function (L273). In the previous version you had a three-way overload of $\sigma$ and now it’s only two, but this is easy to fix. * Another example from the rebuttal: “CG is the conjugate gradient algorithm. We will clarify this in the future.” In the current version this part is moved to the appendix (L707), but still appears as “CG”, without clarification. * L283: “We set $\lambda = 0.25$ to emphasize relevance over diversity.” This value actually gives a 0.25 weight to ndcg and a 0.75 weight to diversity so the weight for diversity is higher. I also mentioned this in my review to the previous version, did I not understand this correctly and this is why you ignored the comment? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: No societal issues. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer N1XB, We are thankful for your continued positive outlook on this work and encouraged to find recognition of our contributions. Thanks for identifying the typos; we will fix them. Next, we respond to your questions and weaknesses below. > clarify the technical contribution (novelty) compared to Krause & Ong (2011). We utilize Theorem 1 from Krause & Ong (2011) (presented as Proposition 1), which relates the regret bound to the maximum mutual information, generalizing the contextual GP-UCB bandits from Srinivas et al. (2010). Building on Proposition 1, we focus on bounding the maximum mutual information by leveraging Weinstein–Aronszajn identity and the feature representation of top-k rankings kernels introduced in Section 3 (as given in Proposition 2). This result in Proposition 2 is not and yields significantly tighter regret bounds, as shown in Table 3, following your previous review suggestion. For further details, please see Section B.4. > The approach does not seem practical, and the empirical evaluation is done with a small number of items (n<=50) and a short horizon (T<500). The empirical assessment was designed to demonstrate the effectiveness of our approach in terms of regret minimization, focusing on a large arm space setting (e.g., 10+ billion arms for n=50 and k=3). The chosen parameters of n and T were intended to provide a conclusive comparison of the regret minimization effectiveness against other baseline algorithms rather than to showcase scalability to "practical" real-world scales. Regarding the application in practical scenarios, we recognize that real-world systems often deal with billions of items, far exceeding the 50 items considered in our study. However, practical recommendation systems typically employ a multi-staged architecture involving rankers and re-rankers that progressively filter and narrow down the item pool from billions to often fewer than 50 items (see [1] and [2]). Bandit algorithms, or Bayesian optimization techniques, are usually applied at the final stage of these systems, suggesting that our approach could be more applicable to practical settings than it might seem. Our theoretical analysis indicates that the proposed algorithm is expected to yield further improvements for larger horizons, i.e., for $T>500$, the approach shall scale much better. It's important to note that the primary aim of this work is not to deliver a ready-to-deploy industrial-scale recommendation bandit algorithm but rather to advance research in this area under a more generalized setting with fewer assumptions. Nevertheless, we are open to refining our approach based on your feedback to better align with practical demands and real-world applicability. * [1]. LinkedIn blog on building large-scale system https://www.linkedin.com/blog/engineering/recommendations/building-a-large-scale-recommendation-system-people-you-may-know * [2]. IJCAI tutorial on Bayesian Optimization for Balancing Metrics in Recommender Systems https://ijcai20.org/t03/ > The acquisition function is not applied to all possible top-k rankings, and the regret analysis does not account for the approximation In principle, the local search can explore all possible top-k rankings. Still, we acknowledge that it may not always yield a global optimizer. Addressing this approximation challenge is notably complex. Most literature, including the foundational GP-UCB paper by Srinivas et al. (2010), which received the ICML-2020 Test of Time Award, assumes exact optimization of the acquisition function and proves regret bounds under this assumption. The gap between theory and practical optimization remains underexplored. The only related study we are aware of is by Jungtaek Kim and collaborators in their paper "On Local Optimizers of Acquisition Functions in Bayesian Optimization," which investigates this issue in continuous input spaces and presents challenging results under several assumptions about the behavior of the reward function, http://mlg.postech.ac.kr/~jtkim/papers/ecmlpkdd_2020.pdf. > Figure 3 should show all iterations (i.e., in the batch mode) Thanks for raising this point. The added PDF provides the regret computation for all iterations. It yields higher variance and slightly more regret while keeping the conclusions and observations consistent, as we anticipated and overserved in our experiments earlier. We value your detailed feedback and welcome further engagement to strengthen this work. --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications.	Summary: The paper introduces a contextual bandit algorithm for top-k recommendations, leveraging Gaussian processes with a Kendall kernel to model the reward function. The proposed method utilizes full-bandit feedback without assumptions such as semi-bandit feedback or cascade browsing. Theoretical analysis demonstrates a sublinear regret with the number of rounds and arms, and empirical results from simulations show superior performance compared to other baselines. The authors also improve the computational efficiency and memory requirements of the algorithm. Strengths: 1. The paper is well-organized and well-written, with clear explanations and detailed experimental results. 2. The proposed algorithm achieves state-of-the-art performance in top-k recommendation tasks compared with various baselines. The theoretical analysis provides a solid foundation for the algorithm's performance guarantees. 3. The improvements in computational efficiency and memory requirements are novel, and technical, making the algorithm more practical for real-world applications. Weaknesses: 1. The paper claims that existing bandit algorithms impose strict assumptions about feedback models, e.g., semi-bandit feedback or cascade browsing. I agree with this statement. However, the paper addresses this problem by directly using Gaussian processes to model the reward function. Isn't this also a restrictive assumption? The authors should clarify this point. 2. The paper lacks a detailed discussion on the choice of the Kendall kernel. The authors should provide more insights into why this kernel was chosen and its advantages over other kernels. 3. There are existing works about top-k combinatorial bandits with full-bandit feedback, such as Rejwan and Mansour (2020) [1]. However, the paper does not compare its method with these existing works. The authors should compare their method with these existing works to demonstrate the difference and novelty of their approach. [1] Rejwan, Idan, and Yishay Mansour. "Top-$ k $ combinatorial bandits with full-bandit feedback." Algorithmic Learning Theory. PMLR, 2020. Technical Quality: 2 Clarity: 3 Questions for Authors: Questions: 1. Why did you choose the Kendall kernel for the Gaussian process? What are the advantages of this kernel over other kernels? 2. How does your algorithm's performance compare to existing works that use full-bandit feedback, such as Rejwan and Mansour (2020)? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Limitations are adequately discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer DrUM, We appreciate your positive feedback on multiple aspects, including writing, results, theoretical analysis, and novel improvements in computational efficiency and memory requirements. We respond to your questions below. > Are Gaussian processes restrictive for modeling rewards? The ability of GPs to accurately model a specific reward function depends on the expressivity of the RKHS associated with the utilized kernel. Informally, as long as there exists a vector 𝑤, s.t, the norm $\|\|𝑤^T𝜙(𝜋)−𝑓\|\|$ is small, where 𝑓 is the true reward function and \phi(\pi) is the feature space, the proposed algorithm should be effective. This statement can be extended to contextual setup as well. It’s worth noting that Assumption 2 outlines the same principle, similar to prior works such as Krause & Ong (2011). Following your suggestion, we assessed the accuracy of the RKHS of the proposed kernel in approximating reward signals through linear regression in the feature space. We trained on 10,000 arm configurations and tested on another 10,000, repeating this six times to measure MSE performance for the nDCG reward as detailed in the paper. The MSE for the NDCG was 0.816 +/- 0.017 for WK, 0.157 +/- 0.009 for CK, and 0.069 +/- 0.010 for WCK. > Could you provide a Detailed discussion on the choice of the Kendall kernel? Why did you choose it for the Gaussian process? What are its advantages over other kernels? Section 2.1 discusses the choice of the Kendall kernel, emphasizing its suitability for Gaussian processes due to being positive definite, right-invariant, and appropriate for rankings. It stands out as one of the few kernels that meet these requirements, alongside the Mellow kernel, built on top of the Kendall kernel. We opted for the Kendall kernel for its simplicity and potential ease of adaptation to top-k rankings. Considering Mellow Kernels would be an exciting avenue for future work. Section 2.2 elaborates on the appropriateness of the Kendal kernel, with examples provided in Table 2. > Compared to existing works that use full-bandit feedback, such as Rejwan and Mansour (2020)? Rejwan and Mansour (2020) present a study on full-bandit feedback for combinatorial bandits. They focus on selecting a subset of top-k items without concern for their arrangement, which is a different problem setting than what this work considers. We appreciate your valuable feedback and look forward to learning from it to enhance this work. --- Rebuttal Comment 1.1: Comment: Thanks for your responses and I will keep my rating. I also recommend adding a 'Related Work' section to help readers better understand the distinctions between your work and other related research. --- Reply to Comment 1.1.1: Title: Thanks for acknowledging the comment. Comment: Thank you, Reviewer DrUM, for your prompt response. We genuinely appreciate it. We’ll consider adding a related work section in the appendix due to space constraints if it’s not feasible in the main section. At your convenience, we would greatly appreciate it if you could clarify whether your concerns have been addressed or not. While your clarification is very important to us and is a crucial step in the review process, we understand if you’re unable to reply. Thanks again!	Summary: The authors propose a new bandit algorithm for top-k recommendations that utilizes a Gaussian Process to model the reward of the exponentially large set of arms, and provide regret bounds which improve upon the naive approach of modeling each arm independently. The authors also present a new kernel and show that the associated kernel matrix displays a particular structure that can be exploited to efficiently compute GP predictions. Finally, experimental validation show the improved regret with respect to baseline algorithms and other similar kernels. Strengths: - The paper is very well presented and easy to read, and the proofs are also very clear. The authors provide meaningful insight into the strengths of their new kernel through small examples and the results of the experiments. - The computational costs of a Gaussian Process using the new kernel are extensively addressed, and their approach is significantly more efficient than a standard procedure. Weaknesses: - The authors claim the new algorithm to be their primary contribution, though it is a simple adaptation of a contextual GP-UCB to top-k recommendations, and a lot of attention is dedicated instead to the new kernel and its computational costs. - Despite the strengths of the new kernel, it is not particularly novel as it is a straightforward combination of existing kernels. - Figures in the Experiments section should be more clear about the confidence ranges (the confidence level in Figure 2 is not mentioned, and the ranges appear to be missing entirely in Figures 3 and 5). I found a few typos and confusing passages, but overall, they didn't affect the reading experience. I will report them here as a feedback: line 99: typo ww(...) -> w(...) lines 254-256: this sentence sounds like the new terms involving n in the bounds are n^{k/2-1} and n^{k-1}, maybe the intended meaning was to highlight the improvement with respect to the cited results? lines 281-283: \lambda = 0.25 emphasizes diversity over relevance line 751: typo shoes -> shows line 821: typo very -> every (also possibly missing a subject after "observed"?) Technical Quality: 3 Clarity: 4 Questions for Authors: 1) The number of local searches is a substantial factor in the overall time complexity of the proposed algorithm, and the number of local searches seems to be particularly high for the large arm space scenarios. Since optimizing the computational costs of the algorithm is a big part of this work, I would like to ask the authors whether they have considered and whether it is possible and/or practically useful to further exploit the structure of the kernel to avoid recomputing the predictor entirely when moving between neighbors. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer 9Xg2, We appreciate your positive feedback on the readability and clarity of the proofs. We respond to your comments below. Thanks for pointing out typos and flipped values of \lambda parameter, i.e., when we wrote \lambda = 0.25, we meant to write \lambda = 0.75. > new algorithm to be their primary contribution We agree that our primary contribution can be viewed as an efficient implementation of the GP-UCB for the ranking kernels studied—this is a significant and essential part of our work. At the same time, although we do not alter the schematic of the GP-UCB, our contributions extend beyond the efficient implementation of the kernels, novel kernel for top-k rankings, exploring the full-bandit feedback settings, empirically showing optimal arm selection for the contextual acquisition function over the top-k domain, and providing its regret analysis. > more clear about the confidence ranges Thank you for your critical observation. The confidence level for Figure 2 is 95%. In Figure 5, confidence intervals were omitted due to poor visibility, while Figure 3 exhibits confidence behavior similar to that of Figure 2. We have included these figures in the attached pdf for your reference. > on the possibility of efficient local search exploiting the kernel structure Thank you for your insightful question. The number of local searches significantly influences our algorithm, which grows linearly with the number of items in our implementation due to the neighbors considered by the local search. There might be a strategy to improve this further to reduce the number of neighbors. However, it does not seem trivial to create smaller and more relevant neighborhoods for the local search by leveraging the relationship between the UCB acquisition function and the feature representation of the arm, i.e., $\phi(\pi)^T{w}_t + \beta \sqrt{\phi(\pi)^T M_t \phi(\pi)}$, where ${w}_t$ and $M_t$ are fixed matrices at the t-th step. We are grateful for the valuable feedback on our work. We welcome further engagement and look forward to learning from your perspective. --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications. Thanks for your responses and I will keep my rating.	Summary: This paper considers the slate recommendation problem where k items are simultaneously recommended to a user at the same time (in a banner or "slate"). In order to solve this problem the authors adapt existing Gaussian process methods to the top-k setting by modifying Kendall kernels to the top-k setting. They also introduce some computational tricks that allow for the algorithm to be quadratic in the embedding size. They also provide a regret analysis and empirical studies to defend their algorithm. Strengths: The paper is extremely readable and well presented, I particularly like the boxes highlighting key claims and the examples e.g. Table 2 and Figure 1. The mathematical notation is clean and consistently used, and the authors are clearly in command of the literature they are building on. This is particularly evident when the authors are able to draw a clean distinction between existing work and their contributions. The experiments are well designed, cleanly implemented, well executed and presented immaculately. Weaknesses: I would like to provide some commentary on this paper and how it might relate to production recommender systems without in any way wanting to diminish the excellent work done in this paper. NeurIPS is a primarily academic research venue and this paper is very well executed and it is very pleasing to see a recommender systems paper of such quality. I would however be very surprised if this method or a method derived from it was applied in a real system for a number of reasons. First, I like the authors use of Figure 1 in order to highlight limitations of the cascade model (the even simpler position based model has similar problems), there have however been models proposed for this case including the Probabilistic Rank and Reward model - https://arxiv.org/pdf/2208.06263 . The "full bandit framework" while useful does discard the useful preference information about what was clicked (as well as saying the banner was clicked somewhere), this information is usually quite valuable. The covariance functions of the model (equation 6) is very well explained and the idea is interesting. A further useful note is that using this multiplied form the reward is correlated if both the context is similar _and_ the actions (top-k) are similar, due to multiplication if either is approximately zero the covariance is zero. The covariance in the action space is rather limited however as it has no notion for some actions being similar. For example if you are going to recommend action movies then you might fill the top-k with totally different action movies yet still have a correlated reward. The model in this paper has some similarities, including the factorized covariance matrix https://arxiv.org/abs/2008.12504 Finally, this paper learns a unique reward for every top-k ordering. This results in a combinatorial explosion both in learning and in delivering recommendations. While the assumption is interesting it does not allow for finding the best recommendations performing an argsort often approximated using a fast maximum inner product search (e.g. LSH or HNSW), it seems difficult to relax these assumptions in real world systems. Technical Quality: 4 Clarity: 4 Questions for Authors: Do you agree with the limitations discussed above? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer tLdn, We appreciate your positive feedback on the readability and design of our experiments. We address your comments in the order they were presented. > On the cascade model, utilize click information. We acknowledge that exploiting click information can be valuable. Still, it can also be misleading when the cascade model or item presentation order is ambiguous, e.g., in the scenarios given in the paper. Our approach addresses scenarios where the cascade model may not be appropriate. We certainly agree that click information is valuable, and it’s worth exploring a middle ground between full-bandit feedback and more restricted models like the cascade model. > On covariance functions and its shortcomings. Thank you for highlighting the limitations of the proposed product kernel, which requires both context and top-k rankings to be similar to utilize data from previous rounds. This limitation can be mitigated using other kernels, such as the additive kernel (similar to Krause et al. [14]). The contributions to accelerate the bandit algorithm remain applicable even with the additive kernel. We will clarify this further in the final draft. Exploring even broader classes of kernels for ranking/context spaces that admit efficient algorithms could be an exciting direction for future work. > Finding recommendations with inner product search We agree that the proposed approach does not allow for straightforward inner product search and requires an optimization algorithm due to the more generic setting, i.e., non-composability of rewards over items. This is also true for the GP bandit algorithm given by Wang et al. [32], despite having a much more relaxed scenario of the semi-bandit feedback. Thank you for the insightful review!	Rebuttal 1: Rebuttal: Dear Program Chairs, Area Chairs, and Reviewers, We appreciate the constructive feedback on our work. It is encouraging that all reviewers affirm this work's soundness, presentation, and contributions. Below, we address the questions and concerns raised by each reviewer. We provide additional figures as reviewers 9Xg2 and DrUM requested in the attached single-page pdf. Thanks again! Pdf: /pdf/6f0aeba2ebafebc4353f589046b81863ae4f3cf6.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Exploring Low-Dimensional Subspace in Diffusion Models for Controllable Image Editing	Accept (poster)	Summary: Even though there have been many research papers on conditional Diffusion Models, sampling-based disentangling method remains a challenge. The authors empirically/theoretically show that PMP is locally linear and the singular vectors of the gradient of PMP are in low-rank space. Based this, the authors propose LOCO Edit that can benefit homogeneity, composability and linearity in semantic controls. Empirical studies are provided to prove the effectiveness of the method. Strengths: - The proposed idea is theoretically interesting. - Introduction is well organized. Weaknesses: 1. Literature review of the recent related works 2. Lack of baselines and comparisons with other methods (Please add comparisons with other local/global editing papers.) 3. Uninterpretable semantic changes. For example, in Fig. 5, color, light angle, tower architecture are not controllable. Is it red? Blue? What is the meaning of color? It would be more interesting if the authors can add some interpretable ways to control the semantics (e.g., using label information). 4. Almost everything this paper describes is based on PMP while the actual results are not from PMP but from DDIM. 5. Lack of ablation studies; null space projection 6. The metrics used in Fig. 2 are not intuitive and descriptions are not clear. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. (Fig. 2) - Can the results be generalized? It seems like only a few x_T’s are used, and only one prompt is used (as mentioned in Appendix D.1) - fig2(a) The authors mentioned in L79 that “rank (A) denotes the numerical rank of A”. However, in Eq. 10 and Fig. 2, it is shown as rank ratio. First of all, what is the rank ratio? and how Eq. 10 can be interpreted? Please elaborate them. - fig2(b) To my understanding, $f$ and $l$ (in L146) seem to be a sort of the denoised images. Why not directly computing their distance? It would be more natural measure for the statement $f \sim l$ in L142. 2. (L80) the x_0 sampled from p_data might need to be the input to get x_t, not the output of the posterior distribution. 3. (L37) What are the various other unexplored aspects? 4. (Fig. 4) What are the undirected t2i edit? If text is not given, why is it called as t2i edit? 5. (L159-165) Are the three properties shown to be true in the experiments? For example, what about composing multiple singular vectors? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: 1. (L157) and (Alg. 1) The authors mentioned that the proposed method benefits the one-step editing while the actual algorithm needs to sample by DDIM twice (DDIM inv and DDIM). 2. (L189-L190) I think there would have been many papers regarding local editing based on Diffusion Models such as Blended diffusion [1]. [1] Blended Diffusion: Text-driven Editing of Natural Images, CVPR’22 Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive feedback that helps improve the quality of our work. Below, we address the reviewer's major concerns and clarify potential misunderstandings in some parts of our work. And we will incooperate those valuable points into our final version. >Q1: >Literature review of related works. (L189-190) I think ... such as Blended diffusion. A1: 1. First, a detailed discussion of related works is provided in Appendix B for space limitations. There, we thoroughly review the existing literature on (1) semantic latent spaces of diffusion models, (2) global image editing in unconditional diffusion models, and (3) local/global image editing in conditional diffusion models. 2. Several existing methods address local editing in conditional diffusion models, such as CLIP-guided diffusion in BlendedDiffusion and classifier-free guidance diffusion in SEGA [22]. However, local editing remains challenging for purely unconditional diffusion models. Thanks the reviewer. We will clarify this point in lines 189-190 and include BlendedDiffusion in related works. >Q2: Lack of baselines and comparisons... A2: Thanks to the suggestions, we conduct comparisons with BlendedDiffusion and NoiseCLR [1], with details in Q&A 1 of the global response. >Q3: Lack of ablation studies; null space projection A3: We showed the role of nullspace projection in Fig. 8 in Appendix, and refer the summary of ablation studies to Q&A 2 in the global response. >Q4: Uninterpretable semantic change… A4: These are great points, but we want to give further clarification: 1. Unsupervised vs supervised edit. Unsupervised edit finds edit directions without any label or text prompt such as Pullback [25] and NoiseCLR, while supervised edit utilizes labels or text supervision such as BlendedDiffusion and SEGA [22]. Both are important research directions. Our method mostly focuses on unsupervised edit (Fig. 1 and 5), but can be extended to supervised edit with a target text prompt as shown in Fig. 4(b). 2. Interpretation of semantic changes. Even in unsupervised edit, the semantic change directions can be interpreted only after being identified. For example, in Fig. 5, these directions indeed have semantic meanings: colors are changed from white to red, tower architecture are changed from simple to complex. We will make the descriptions more concrete. >Q5: Misunderstandings on our paper. >1. The actual results are not from PMP but from DDIM >2. (L80) The $x_0$ sampled from $p_{data}$ might need to be the input to get $x_t$, not the output of the posterior distribution >3. (L157) The proposed method is not one-step editing... >4. (L159-165) Are the three properties... >5. What are the undirected t2i edit... A5: We want to clarify these misunderstandings: 1. PMP and DDIM. We use PMP at time step $t$ to find the direction $v_p$ and edit $x_t$ as $x_t + \lambda v_p$. In contrast, DDIM-Inv is used to get $x_t$ from $x_0$, and DDIM is to generate the edited image by denoising $x_t + \lambda v_p$. We use both PMP and DDIM, whereas PMP takes the most important role in finding the edit direction $v_p$. 2. Clarification on L80. $\mathbb{E}_{x_0 \sim p\_{data}(x)}[x_0\|x_t]$ is the expectation of $x_0$ given the observed $x_t$ and the prior distribution $p\_{data}(x)$. Its output is indeed not $x_0$. 3. One-step editing. In [25], “one step” means only changing $x_t$ at one timestep $t$ for image editing. In comparison, as mentioned by reviewer 9wnQ, NoiseCLR requires edits across multiple timesteps. 4. Experimental verification of three properties. These properties have been demonstrated in Fig. 1 (b,c,d) and discussed in L 64 - 69. Particularly, in Fig. 1$c$, the disentangled editing directions are composable: linear combination of two editing directions result in changes of two attributes simultaneously. 5. Undirected and text-directed T2I edit. Suppose T2I diffusion models generate an image based on text prompt $c_o$. - In Undirected T2I edit, image is edited without additional editing prompt for the target edit direction. - In Directed T2I edit, an extra editing prompt $c_e$ is provided. So it is called “text-directed T2I edit”. >Q6: For Fig. 2: >1. Whether the results be generalized to more data samples and prompts >2. Define rank ratio and interpret Eq. 10 >3. Why use the norm ratio A6: Thanks for raising these points. 1. We test 15 more text prompts. As shown in Fig. F of the global response PDF. Results demonstrate the generalizability of more prompts and initial noises. 2. The rank ratio is defined in Line 144, which is the ratio between the numerical rank and the ambient dimension $d$. Eq.10 can be interpreted as finding the smallest $r$ such that: compared to the top-$r$ largest singular values of the given Jacobian, the remaining singular values are smaller than a threshold and can be neglected. 3. The norm ratio $\|\|l - f\|\|_2/\|\|l\|_2$ measures relative distance to better reflect linearlity. If $l$ and $f$ are apart from each other but $\|\|l\|\|_2$ and $\|\|f\|\|_2$ are too small, the absolute norm differences $\|\|l - f\|\|_2$ may be too small to reflect violations of linearity, but the norm ratio can. >Q7: (L37) What are the various other unexplored aspects? A7: Specifically, these aspects include but not limited to (1) whether diffusion models have semantic spaces; (2) what features lie in the semantic spaces; (3) theoretically justify those semantic features; (4) utilize those features unsupervised local edit in unconditional diffusion models. In this work, we looked into all these aspects and proposed LOCO edit that tackles many of the questions raised. We will make the discussion clearer. [1] Yusuf et al., "Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models." CVPR 2024. --- Rebuttal Comment 1.1: Comment: Thank you for clarifying my concerns during the rebuttal. Overall, most of my concerns have been resolved. I will raise my score to 5. One of the remaining issues (which could be important) is: if the semantic direction means white to red, what happens if input image does not have anything related to white and red, e.g., black and blue? --- Rebuttal 2: Title: looking forward to your reply Comment: Dear Reviewer hJrH, We have worked diligently to resolve your concerns. As the rebuttal period comes to an end, please do not hesitate to contact us if you have any last-minute questions or require further clarification. Best regards, The Authors --- Rebuttal 3: Comment: Dear reviewer, Thanks for the positive response and we appreciate your acknowledgment of our responses in addressing the existing problems. For the proposed precise point: (1) We have tested transferring edit direction from “white to red” to flowers in other colors including blue, pink, and orange. We saw the colors change to darker blue, darker pink, and darker orange. Considering the effects of transferring, the edit direction can be described as “darker color”, which is semantically meaningful for flowers in general. (2) To further discuss the point, when transferring “enlarging eyes” to images of a person wearing sunglasses, only the details of the sunglasses are changed which is hard to notice. This is because “enlarging eyes” is not semantically meaningful for people wearing sunglasses. (3) Such edge cases are interesting, but since we consider practical editing scenarios where the edit directions are semantically meaningful, we do not present them in the paper. Besides, we can not show additional figures, but we will add those in our revision to make the discussion more complete. Thanks for the question and hope we have further addressed your concerns. We are happy to answer any other questions. Best Regards, Authors	Summary: This paper proposes an diffusion image editing framework called LOw-rank COntrollable edit (LOCO Edit) based on two observations: 1) the learned posterior mean predictor (PMP) is locally linear during the middle timesteps during denoising, and 2) the singular vectors of the PMP's Jacobian lie in low-dimensional semantic subspaces. These observations are backed both empirically and theoretically. To conduct LOCO edit, an image is inverted with DDIM inversion to an intermediate timestep. SVD is conducted on the Jacobian of the PMP w.r.t. to this noisy image in order to discover semantic directions. Furthermore, a mask can be applied along with a null space projection in order to achieve more disentangled and localized edits. A variant of LOCO edit is also proposed to enable text direction. A variety of experiments are conducted for the unconditional LOCO edit on various domains using human evaluation and CLIP score and edit discovery time, achieving superior results against other methods. Qualitative results are shown for various diffusion model variants (e.g., Stable Diffusion, Latent Consistency Model) and domains, demonstrating linearity and the ability to modulate strength, the ability to compose with other directions, as well as the ability to transfer the edits to other images. Strengths: * The paper tackles an important problem in diffusion models: disentangled and continuous editing. * This is done in a principled way, based on two important observations of the PMP and its Jacobian across different timesteps. Furthermore, there are both empirical experiments as well as theoretical analysis validating these. * These edits can be composed, transferred, and modulated in a continuous manner. Additionally, the time required for discovering these directions is less than other methods, such as Asyrp. * The qualitative and quantitative results demonstrate strong editing capabilities. * A variety of ablations are provided in the paper and the appendix, such as the effect of nullspace projection, timestep used for editing, etc. Weaknesses: * The qualitative results for editing are not quite the best I have seen, and they are compared against relatively weak baselines (e.g. Asyrp). However, I find that this is okay since the approach is quite principled. Many impressive diffusion based image editing papers have been released in the past year [1,2]. I don't think it is necessary to compare against them, but including them in the paper in the related works section would keep this paper up-to-date with the current state of image editing. * Since the linearity and low-rankness properties are more apparent in the middle time-steps, applying more global edits (e.g. changing the shape of something while keeping appearance) would be difficult since more coarse level features are constructed in the earlier timesteps of denoising. I see this effect in Fig. 7 in the appendix, as other parts of the image are edited as well. [1] @inproceedings{dalva2024noiseclr, title={Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models}, author={Dalva, Yusuf and Yanardag, Pinar}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={24209--24218}, year={2024} } [2] @article{gandikota2023concept, title={Concept sliders: Lora adaptors for precise control in diffusion models}, author={Gandikota, Rohit and Materzynska, Joanna and Zhou, Tingrui and Torralba, Antonio and Bau, David}, journal={arXiv preprint arXiv:2311.12092}, year={2023} } Technical Quality: 4 Clarity: 3 Questions for Authors: * I am quite surprised that an edit at a single timestep has enough effect to edit the image. For instance, [1,2] require over multiple timesteps. Is there an intuition for why this edit at a single timestep has enough causal effect? * For more open-domain diffusion models like Stable Diffusion, can the edit direction be applied to another subject? For instance, in Figure 1, you transfer human opening eyes to different images. Would this transfer to other domains, such as an animal? * A visualization of all the edit directions discovered via SVD would be quite interesting. Are they all semantically meaningful? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: A more substantive discussions of limitations is missing. Although Appendix Section H discusses future directions, there is little mention of limitations of the current method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for constructive feedback and interesting questions. During rebuttal, we address the reviewer’s concerns on the experiments and other questions as follows. > Q1: The qualitative resultes are not quite the best... up-to-date with the current state of image editing. A1: Thank the reviewer for pointing out the interesting works NoiseCLR and Concept sliders. We have looked into them and will cite and add discussions on NoiseCLR and Concept sliders in the revised manuscript. Besides, based on the reviewer's suggestion, we conducted more qualitative and quantitative baseline comparisons with NoiseCLR. Moreover, we extend both qualitative and quantitative results per the reviewer's suggestion. We refer detailed results and discussion to Q&A 1 of the global response. >Q2: Since the linearity and low-rankness properties are more apparent in the middle time-steps, applying more global edits... would be difficult since more coarse level features are constructed in the earlier timesteps of denoising...are edited as well. A2: Thanks for raising this point here, and we would like to discuss further. 1. For coarse features that are controlled in early time steps close to random noise, LOCO is not guaranteed to disentangle these features in the high-rank space. This is closely related to our method is principled in the low-rank subspace of diffusion models in the middle time steps, and hence we mainly focus on unsupervised precise local edit. 2. It would be interesting to non-trivially model the space at those high-rank timepoints, maybe taken inspiration from works such as NoiseCLR. Our current focus is to study the low-rank subspaces in diffusion models and leave the understanding of semantics in high-rank spaces to be explored in future studies. > Q3: I am quite surprised that an edit at a single timestep has enough effect to edit the image. For instance, [1,2] require over multiple timesteps. Is there an intuition for why this edit at a single timestep has enough causal effect? A3: Thanks for the interesting question and we try to provide possible intuitions. 1. From the perspective of method principles, LOCO initiates from the local linearity of PMP, and tries to find directions that lead to the largest changes in the posterior mean. Under the assumption of local linearity within the low-rank semantic subspace, such identified direction $v$ can achieve one step edit with appropriate edit strength. Following the intuition above, the one-step edit ability is experimentally verified. 2. In contrast, NoiseCLR optimizes a conditional variable representing specific features to be used in classifier-free guidance, and Concept sliders use LoRA to finetune a Slider that steers the generation to be more aligned with using conditional variable $c+$ instead of $c-$. Such indirect optimization and finetuning based on conditional variables may benefit from multiple timesteps to achieve faster convergence in optimization and better performance in editing. > Q4: For more open-domain diffusion models like Stable Diffusion, can the edit direction be applied to another subject? For instance, in Figure 1, you transfer human opening eyes to different images. Would this transfer to other domains, such as an animal? A4: Thanks for the interesting question, and we would like to discuss it further. 1. The transferability has a high success rate for unconditional diffusion, which matches our theoretical analysis for unconditional diffusion models. However, it’s currently hard to transfer edit directions for more open-domain T2I diffusion models as we have experimented. 2. It is potentially because the feature space of these diffusion models is more complicated correlating with various text prompts. Such feature space does not align with the assumptions in the theoretical analysis for unconditional diffusion models. 3. The modeling of feature space in T2I diffusion models is still an open and interesting question in the area. The exploration of transferability in these models requires further studies that may involve conditional text variables, noisy images, and various other non-trivial aspects. > Q5: A visualization of all the edit directions discovered via SVD would be quite interesting. Are they all semantically meaningful? A5: Thanks for proposing the interesting question. We have attached the identified editing directions for different regions of interest in Figure C of the global response PDF. These editing directions are selected from CelebA, FFHQ, and AFHQ. 1. We observe that the editing directions demonstrate semantics correspondences to the region of interest. We also notice for these datasets, the position of objects is biased in the center of images, which benefits the transferability of editing directions. 2. However, from our observation, the transferability is robust to gender difference, facial feature shape difference, and moderate position difference, as presented in Figure 1(b) of the main paper and Figure B of the global response PDF. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response. I have read through the global rebuttal and PDF, as well as other responses. Overall, I am satisfied with the responses. I suggest that the authors include the limitations of global editing and the difficulty of transferring edits in conditional diffusion models. After going through this paper again and revisiting other papers, I think there are some similarities to [1], such as linear properties of intermediate noise space, and the one-step editing. I suggest that the authors include this paper in their work and discuss it. I keep my overall score as is. [1] "Boundary Guided Learning Free Semantic Control with Diffusion Models." Zhu et al. 2023. --- Rebuttal 2: Comment: Thank you for all the valuable suggestions and for recognizing the value of our work! We will expand the discussion on global editing and transferability in conditional diffusion models and will discuss BoundaryDiffusion in the paper. We have carefully read through BoundaryDiffusion, which is an interesting supervised one-step global edit method. BoundaryDiffusion uses linear SVMs to classify image latent by hyperplanes using label annotations, and about 100 images are required for each class. The edit intuitively moves the image latent in both $\epsilon_t$ and $h_t$ spaces to the negative or positive side of the hyperplane, so that the corresponding semantic is reduced or enhanced. Although the linear and one-step edit properties are similar, our method is (a) localized, (b) requires no label annotation, (c) can find transferable directions using a single image, (d) requires edits only in $x_t$ space, and (e) has a theoretical basis to support the approach. Thanks for the constructive feedback, we will make sure all of them are reflected in our final manuscript.	Summary: The paper examines the use of low-dimensional subspaces in diffusion models for precise and disentangled image editing. The paper observes that the Posterior Mean Predictor (PMP) in diffusion models shows local linearity across various noise levels, and the singular vectors of its Jacobian exist in low-dimensional semantic subspaces. Building on these observations, they introduce LOw-rank COntrollable image editing (LOCO Edit), a technique that enables efficient, training-free, and precise localized image manipulation by utilizing the Jacobian's low-rank nature. This approach has a strong theoretical basis and has been shown to be effective across different architectures and datasets. Strengths: 1. The paper introduces an approach to image editing within diffusion models by identifying and exploiting low-dimensional subspaces. The idea of using local linearity and low-rank properties of the Jacobian in diffusion models for controllable editing sounds novel. 2. The methodology looks sound, supported by robust empirical evidence demonstrating local linearity and low-rankness of the PMP's Jacobian. 3. The proposed method tackles a key challenge in diffusion models: achieving precise and disentangled image editing without additional training. This has wide-ranging implications for various image generation and manipulation applications. 4. Extensive empirical evaluations are presented, showcasing the method's effectiveness across diverse network architectures (UNet and Transformers) and datasets (e.g., CIFAR-10, CelebA, ImageNet). The results show performance improvements and validate the approach's generalizability. 5. The paper is well-organized and clearly written. Explanations are thorough, and the visuals (e.g., figures) effectively convey the main concepts and outcomes. Weaknesses: 1. The experimental evidence appears highly limited. The quantitative and qualitative comparison results (confined to Table 1 and Fig. 6 with just one example in the main paper) seem inadequate. The study would be strengthened by a more comprehensive comparison with additional state-of-the-art image editing methods, beyond just [24] and [25]. Highlighting specific advantages and potential weaknesses in relation to these methods would provide a more balanced perspective. 2. The assumptions regarding low-rank Gaussian distributions used for theoretical validation warrant more rigorous examination. Expanding the discussion to include potential limitations or circumstances where these assumptions may not hold would add depth and nuance to the analysis. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The paper should clarify that the Posterior Mean Predictor (PMP) is a network function within diffusion models, and detail how it is computed. Specifically, it should describe how PMP is derived and utilized in the context of local linearity and low-rankness. 2. Elaborate on what P_{\omega} and P_{\omega^C} represent in Fig.3 and lines 193-195, and clarify the decomposition of x_{0,t}^{\hat} into ROI and null space using masking techniques. 3. Explain how performing the Jacobian over different masked images helps in disentangling features for precise local editing. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: In addition to the listed weakness above, the sensitivity of the method to various parameters (e.g., noise levels, perturbation strengths) could be explored in greater detail. Understanding the robustness of the approach under different conditions would be valuable for practical applications. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for constructive feedback. During rebuttal, we address the reviewer’s concerns on the experiments and other questions as follows. >Q1. Limited experimental evidence and quantitative and qualitative comparisons. A1: Thanks to the reviewer's suggestions, we added more results, and conducted more qualitative and quantitative comparisons with more baselines to better evaluate our method. We refer detailed results and discussion to Q&A 1 of the global response. >Q2. Detailed ablation study on sensitivity to various parameters... A2: Thanks to the reviewer's suggestions, we conducted more detailed ablation studies and gave a more comprehensive summary. We refer detailed results and discussion to Q&A 2 of the global response. > Q3: Validation of using the low-rank Gaussian distributions. A3: In [1], the authors conducted extensive empirical experiments and found that for diffusion models trained on the FFHQ and AFHQ datasets, the learned score can be approximated by the linear score of a Gaussian, particularly at high noise levels. Furthermore, [2] confirms the low-dimensional nature of various image datasets, such as CIFAR-10 and ImageNet, and calculates their intrinsic dimensions. Building on these findings, we incorporate the concept of Gaussian distribution and low-dimensional properties to study the low-rank Gaussian case. This approach enables our model to effectively capture the core structure of image data while remaining tractable for theoretical analysis, thus serving as a practical foundation for theoretical studies on real-world image datasets. We will integrate these discussions into the final version to enhance motivation. > Q4. Clarification on Posterior Mean Predictor (PMP). A4: 1. Definition & computation of PMP and reference for it’s derivation: Indeed, PMP is a network function in diffusion models, which takes an input pair ($x_t$, $t$), and outputs the predicted posterior mean $\hat{x}\_{0,t}$. Specifically, the computation of PMP in diffusion models is defined in Equation (2), where $\epsilon_{\theta}$ is the learned unet denoiser in the diffusion model. We refer how PMP is derived to the derivation of Equation (12) in the DDIM paper [3]. 2. How local linearity and low-rankness in PMP is utilized: (1) PMP’s local linearity is the key assumption for finding edit direction via SVD of PMP’s Jacobian, since directions found via SVD are meaningful only if PMP is linear; (2) PMP’s local linearity also leads to the linear and composable properties of the LOCO edit method; (3) PMP’s low-rankness ensures us to use a low-rank estimation to find the nullspace, achieving efficient and effective nullspace projection. We will make the above points clear in the final version. > Q5: Clarification on finding the precise and disentangled edit direction. > - Elaborate on what $P_{\Omega}$... using masking techniques. > - Explain how performing the Jacobian... helps... for precise local editing A5: We would like to clarify concepts, process, and intuition in finding such local edit direction. We will also revise the writing to make the points clearer. 1. Definition of ROI, $P_{\Omega}$ and $P_{\Omega^C}$: Here, $\Omega$ is an index set that covers the region of interest (ROI) and $\Omega^C$ covers any other region in the image outside ROI. Based upon this, $P_{\Omega}$ and $P_{\Omega^C}$ denote the projections onto $\Omega$ and $\Omega^C$, respectively. Intuitively speaking, $P_{\Omega}(I)$ crops the content of an image $I$ within the mask, and $P_{\Omega^C}(I)$ crops the contents of $I$ outside of the mask. 2. Decomposition of $\hat{x}\_{0,t}$ and calculation of Jacobians: Based on the above definition, we decompose PMP’s output $\hat{x}\_{0,t}$ into $\tilde{x}\_{0,t}$ and $\bar{x}\_{0,t}$ as visualized in Figure 3, where $\tilde{x}\_{0,t} = P_{\Omega}(\hat{x}\_{0,t})$ and $\bar{x}\_{0,t} = P_{\Omega^C}(\hat{x}\_{0,t})$. We further define $\tilde{J}\_{\theta,t} = \partial \tilde{x}\_{0,t} /\partial x_t$ and $\bar{J}\_{\theta,t} = \partial \bar{x}\_{0,t} /\partial x_t$. 3. Finding image editing directions and nullspace via SVD of Jacobians: Let $\tilde{J}\_{\theta,t} = \tilde{U}\tilde{S}\tilde{V}^T$, and $\bar{J}\_{\theta,t} = \bar{U}\bar{S}\bar{V}^T$ be the compact SVD. Intuitively, $span(\tilde{V}) = range(\tilde{J}\_{\theta,t}^T)$ is the subspace containing change directions of $x_t$ that leads to change within the $\tilde{x}\_{0,t}$. Besides, $span(\bar{V}) = range(\bar{J}\_{\theta,t}^T)$ is the subspace containing change directions of $x_t$ that leads to edit in $\bar{x}\_{0,t}$. Moreover, $nullspace(\bar{J}\_{\theta,t})$ is the subspace leads to no edit in $\bar{x}\_{0,t}$. 4. Nullspace projection for more precise and disentangled image editing: For a direction $v \in range(\tilde{J}\_{\theta,t}^T)$, nullspace projection means projecting $v$ onto $nullspace(\bar{J}\_{\theta,t})$. In practice, this nullspace projection can be calculated as $v_p = (I - \bar{V}\bar{V}^T) v$. Such nullspace projection can further eliminate the effect of $v_p$ to change within $\bar{x}\_{0,t}$, and disentangle the change to be only within $\tilde{x}\_{0,t}$. Then by denoising $x_t + \lambda v_p$, the new image is more precisely edited within the ROI $\Omega$. [1] Binxu et al., "The Hidden Linear Structure in Score-Based Models and its Application." 2023. [2] Phillip et al., "The intrinsic dimension of images and its impact on learning." ICLR 2021. [3] Jiaming et al., Denoising diffusion implicit models. 2020. --- Rebuttal Comment 1.1: Comment: I appreciate the authors' efforts in addressing concerns through their rebuttal. While some issues have been clarified, I still have reservations about the depth of empirical evidence presented in the original paper: 1. Limited comparative analysis: The paper only includes two comparison methods ([24, 25]), which may not provide a comprehensive benchmark. 2. Small sample sizes: The evaluation relies on relatively few samples (90 cases for transferability testing, 100 examples for other metrics), potentially limiting the robustness of the findings. 3. Absence of standard metrics: Apart from using the CLIP score, the study doesn't utilize other widely accepted measures such as Fréchet Inception Distance and Inception Score, which could have strengthened the evaluation. 4. Insufficient ablation studies: The paper lacks thorough ablation studies to demonstrate the contribution of individual components. While the rebuttal has partially addressed these concerns, the original paper's limitations in empirical validation remain significant. Consequently, I am inclined to either maintain my initial score or consider a slight downward adjustment. --- Rebuttal 2: Comment: Thanks for the questions, we would like to further make clarifications on these points. 1. We have extended the comparisons with two additional baselines BlendedDiffusion [2] and NoiseCLR [1] for comparison. See global rebuttal Q&A 1 and pdf Tab. A and Fig. E. 2. (1) The human evaluation dataset is randomly selected have high diversity to cover various semantic directions; (2) Asyrp [3] uses only 40 data samples for human evaluation on CelebA (Appendix K.1), Pullback [4] only shows qualitative results; (3) BlendedDiffusion and NoiseCLR do not mention the size of evaluation dataset; (4) A further detail is, we used 400 samples for other added metrics LPIPS and SSIM. Hence, we think the evaluation dataset size is robust enough to provide fair comparisons. 3. The mentioned FID and IS are good metrics measuring whether generated image distributions are similar to the training images, but not standard metrics for image editing methods. (1) FID and IS will be increased if additional features are edited in comparison to the original one; (2) Representative baslines including NoiseCLR, BlendedDiffusion, Asyrp, and Pullback use neither of the FID and IS; (3) As in global rebuttal Q&A 1 and pdf Tab. A, we have conducted a comprehensive comparison using 7 metrics and 4 attributes to show the superiority of our method. This indeed supports that we have conducted comprehensive empirical evidence in addition to the solid theoretical basis which is lacking in the previous papers. 4. For more ablation studies, as summarized in Q2&A2 for the global response, we do more ablation studies on Noise levels (i.e., time steps), Perturbation strength, Nullspace projection, and ranks. We have shown representative examples, and more results will be included in the revision. We appreciate the reviewers' prompt responses and make further clarification and references on how we conduct comprehensive experiments to show the superiority of our method. We kindly request the reviewer evaluate our work fairly based on the rebuttal and our further clarifications. [1] Yusuf et al., Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. CVPR 2024. [2] Omri et al., Blended diffusion for text-driven editing of natural images. CVPR 2022. [3] Mingi et al. Diffusion models already have a semantic latent space. ICLR 2023. [4] Yong et al. Understanding the latent space of diffusion models through the lens of riemannian geometry. NeurIPS 2023.	Summary: The paper presents a method for steering the generation in a diffusion model, without any further training. The proposed method is based on two insights about the Posterior Mean Predictor (PMP) and its Jacobian -- that is, the former being locally linear and the latter having singular vectors lying on low-dimensional subspaces. Some theoretical results are provided towards the justification of the above. Local linearity of the PMP allows for a single-step, training-free method for local editing of regions of interest, while the low-rank nature allows for the effective identification of semantic directions using subspace power methods. Some qualitative and quantitative experimental results are provided. Strengths: The paper is well-written and sound. The provided theoretical results wrt local linearity of the PMP, along with the low-rankness of its Jacobian are interesting (however not extremely surprising) and might be useful to the research community. Weaknesses: The weakest aspect of the paper concern the reported experimental results. Both qualitative and quantitative results are extremely limited, whilst comparisons with existing work are also not convincing. Besides, the significance of the empirical results is not convincing. For instance, in Fig. 6, when the proposed method (LOCO) is compared to Asyrp, it's not clear at all why localized editing with LOCO is better than Asyrp. Similarly, in Table 1, the Transfer CLIP Score (an important metric in the context of the task studied in the paper) of Asyrp is significantly better that the proposed method's. The authors do discuss this, but the fact that the proposed framework is learning-free, in contrast to Asyrp that learns how to perform editing, does not justify the significant difference wrt to CLIP Score. Finally, as the authors acknowledge in the Appendix H ("Future Direction", where limitations are also discussed), the provided theoretical framework concerns mainly the undirected image editing part; how text-directed image editing behaves is not addressed by the proposed method. This is a certain limitation, that weakens the generality of the proposed method, yet it doesn't reduce the importance of the theoretical results provided. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss adequately the limitations of the proposed method in the Appendix H (Future Direction). A separate "Limitations" section is missing, but the authors discuss the limitations of the paper honestly and comprehensively. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer’s constructive feedback, and in the following we address the reviewer’s concerns one-by-one. >Q1: Not enough qualitative and quantitative results and comparisons with existing work. Thanks to the reviewer's suggestions, we conduct more qualitative and quantitative comparisons with more baselines, and add more results to better evaluate our method. We refer detailed results and discussion to Q&A 1 of the global response. >Q2. Not convincing empirical results. >1. Fig. 6 is not convincing to show better localized editing with LOCO than Asyrp >2. Why Transfer CLIP Score of Asyrp is better than LOCO, and whether it’s an issue A2: We thank the reviewer for raising these constructive points which improve our work. 1. More convincing visualizations. - To demonstrate the effectiveness of our method, we selected examples in Fig. 6 where Asyrp achieves its best performance, yet our method still performs even better. For instance, in Fig. 6, to edit lips, Asyrp noticeably alters undesired regions such as face color, hair shape, and face shape more than our method. - There are many other instances where Asyrp's editing changes the image much more significantly than ours. We add more random examples in Figure E of the global PDF to visualize the local edit ability of our method in comparison of others. 2. Discussion on the Transfer CLIP Score in Table 1\. Very good question for a deeper discussion. - The higher transfer CLIP score of Asyrp is because it is directly supervised by the CLIP score to learn each concept and predict editing directions. In contrast, our method is unsupervised and learning-free, and has the best CLIP score among other unsupervised methods. - Moreover, the transfer CLIP score is biased by large changes in Asyrp: For successful edit in both Asyrp and LOCO, Asyrp tends to have larger global changes resulting in a much higher transfer CLIP score, sometimes leading to corruption in other parts though the desired region is edited. Examples are shown in Figure E of the global PDF. - Additionally, the transfer CLIP score relies on CLIP's intrinsic failures to capture detailed semantics as discovered in [1], which may lead to edit failures of Asyrp. Failure examples that are potentially related are shown in Figure E (row 5-8) of the global PDF, where Asyrp fails to edit darker eyebrows for all random examples. - Therefore, to compensate for the biases of the transfer CLIP score, we also measure the local edit success rate, where our method outperformed all others. Thanks to the reviewer for raising the point, we also add LPIPS [2] and SSIM [3] as guardians of dramatic changes and use the less biased local edit success rate as the major evaluation metric for local edit ability. In the revision, we will include those discussions in the paper to clarify our findings. >Q3: The provided theoretical framework concerns mainly the undirected image editing part; how text-directed image editing behaves is not addressed by the proposed method. This is a certain limitation, that weakens the generality of the proposed method, yet it doesn't reduce the importance of the theoretical results provided. A3: We thank the reviewer for raising the points, but we want to clarify this further. 1. In Figure 4, we showed that our method can be extended to text-directed image editing. This demonstrates the generality of our edit approach. 2. Second, the empirical observation of low-rankness and local linearity is generalizable to text-to-image diffusion models as shown in Figure 2 and Figure 9. 3. Although our theoretical study part is for unconditional diffusion models, to the best of our knowledge, our work is the first theoretically ground method compared to all previous diffusion model based editing methods. It's an advantage compared to other image edit methods. 4. Besides, for studying text-directed diffusion models, the challenges lie in the more complicated feature space caused by a mixture of image and text distributions. Understanding their feature space is yet an unsolved problem in the area. Hope our work can provide some inspiration for future exploration. [1] Shengbang et al., Eyes wide shut? exploring the visual shortcomings of multimodal llms. CVPR 2024 [2] Richard et al. The unreasonable effectiveness of deep features as a perceptual metric. CVPR 2018. [3] Zhou et al., Image Quality Assessment: From Error Visibility to Structural Similarity. 2004	Rebuttal 1: Rebuttal: We thank all reviewers for carefully reviewing our work with constructive and positive feedback. Most reviewers find our empirical observation “interesting”, “well-validated”, “important” (eLFT, h6gb, 9wnQ), our theoretical analysis “useful”, “strong”, (h6gb, pgJR), our edit method “novel”, “quite principled”, “well-motivated”, “insightful” (eLFT, pgJR, 9wnQ, hJrH), and our presentation “well-written”, “sound”, “well-organized” and “clear” (h6gb, pgJR, hJrH). Summary of our results. Our work introduced a simple yet effective method for local edit in diffusion models by exploring (i) the linearity of the posterior mean predictor (PMP), and (ii) low-rankness of its Jacobian. The advantages of our method can be highlighted as follows: - One-step, local editing. To our knowledge, this is the first work on image editing using unconditional diffusion models that allow for one-step, localized editing. In contrast, most works require editing on multiple (all) timesteps and only perform global edit. - An intuitive and theoretically grounded approach. Our method is highly interpretable, leveraging the benign properties of PMP. The identified properties are well supported by empirical observation (Fig, 2 in paper) and theoretical justifications in Section 4. In contrast, most previous methods are heuristic and lack theoretical justifications. Addressing reviewers’ major concerns. We appreciate the reviewer's comments on our limited experiments. During the revision, we addressed reviewers’ concerns with more comprehensive results, comparisons, and ablation studies as follows. They are presented in Fig. A-F and Tab. A of the global response PDF, and will be included in the revised paper, with code released. >Q1: More evaluation of our method. A1: 1. More qualitative results: We add more qualitative results on different datasets, shown in Fig. A of the attached PDF. 2. More quantitative metrics: The quantitative comparison is extended in Tab. A of the attached PDF using the additional metrics: - LPIPS [4] and SSIM [5] to measure the consistency between edited and original images. - Learning time, Transfer Edit Time to measure the time each method requires to find the editing direction and apply it to edit an image. - #Images for Learning to measure the number of images used to find directions. - One-step edit, No Additional Supervision, Theoretically Grounded, and Localized Edit are different properties for each editing method. 3. More qualitative comparisons: We also extend the qualitative comparison in Fig. E to showcase our method's strong local editing capability. 4. Detailed comparison with more baselines: We compare our work with two other studies: NoiseCLR [1] and BlendedDiffusion [2], together with the previous baselines. We discuss key observations as follows. - Local edit ability: Tab. A shows LOCO achieves the best Local Edit Success Rate. For LPIPS and SSIM, our method performs better than global edit methods but worse than BlendedDiffusion. However, BlendedDiffusion sometimes fails the edit within the masks (as visualized in Fig. F, rows 1, 3, 4, and 5). We discuss the potential causes from CLIP bias in the last point below. Other methods like NoiseCLR find semantic direction more globally, such as style and race, leading to worse performance in Local Edit Success Rate, LPIPS, and SSIM for localized edits. - Efficiency and transferability. First, LOCO requires less learning time than most other methods, and the learning needs only a single time step and a single image. Moreover, LOCO is highly transferable, having the highest Transfer Success Rate in Tab. A. In contrast, BlendedDiffusion can't transfer and requires optimization for each image. While NoiseCLR excels at open-domain transfer, it performs worse than LOCO in closed-domain transfer (e.g., on the CelebA dataset). Other methods exhibit even weaker transferability. - Theoretically grounded and supervision-free. LOCO is theoretically grounded. Besides, it is supervision-free, thus integrating no biases from other modules such as CLIP. [3] shows CLIP sometimes can't capture detailed semantics such as color. We observe failures to capture detailed semantics in methods that utilize CLIP guidance such as BlendedDiffusion and Asyrp in Fig. E. >Q2: More ablation studies. A2: 1. Noise levels (i.e., time steps). We test different noise levels and show results in Fig. 7. Results for more noise levels are in Fig. D with key observations: (a) Edit at large noise level (i.e., large time step) perform coarse changes while small noise level performs finer edit; (b) LOCO applies to a large range of noise levels ([0.2T, 0.7T]) for precise edit. 2. Perturbation strength. The linearity concerning edit strengths is visualized in Fig. 1d, and detailed ablation results of perturbation strength at 0.6T are in Fig. D as an example, with key observations: LOCO applies to a generally wide range of perturbation strengths ([-15, 15]) to achieve localized edit. 3. Nullspace projection and ranks. Ablabtion study on nullspace projection is in Fig. 8, with key observations: (a) the local edit ability with no nullspace projection is weaker than that with nullspace projection; (b) when conducting nullspace projection, an effective low-rank estimation with $r=5$ can achieve good local edit results. [1] Yusuf et al., Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. CVPR 2024. [2] Omri et al., Blended diffusion for text-driven editing of natural images. CVPR 2022. [3] Shengbang et al., Eyes wide shut? exploring the visual shortcomings of multimodal llms. CVPR 2024 [4] Richard et al. The unreasonable effectiveness of deep features as a perceptual metric. CVPR 2018. [5] Zhou et al., Image Quality Assessment: From Error Visibility to Structural Similarity. 2004 Pdf: /pdf/a9b1c7fe01c8f866394b17be0fa73bede756d2f0.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper made an interesting observation on local linearity of the diffusion model's denoiser. Based on the observation, the author introduces a novel method of performing one-step closed-forrm operation to achieve semantic image editing. Empirical and numerical results are given to demonstrate the effectiveness of the method Strengths: 1. The observations in terms of local linearity of the denoiser is interesting, and is well-validated across different architecture and datasets. 2. The presented method is novel and well-motivated, and simple. 3. The discovery of homogeneity and transferability of the editing direction is insightful. 4. The empirical and numerical result demonstrate consistent image editing. Weaknesses: 1. The computation time of the method is still a bit long (taking around 70s if I understand correctly). Technical Quality: 3 Clarity: 3 Questions for Authors: 1. I wonder whether the homogeneity of the editing can be observe on samples that are very different: e.g., 1) computing editing direction on an image with face at the left, and transfer to an image with face on the right; 2) computing editing direction on an image with realistic style, and transfer to an image with cartoon style. 2. In addition, since current models such as SD3 are using flow-matching objective to train the model, I wonder if such observed phenomenon would be more prominent on these models? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the limitations have been adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer’s constructive feedback, and in the following we address the reviewer’s concerns one-by-one. >Q1: The computation time of the method is still a bit long (taking around 70s if I understand correctly) A1: 1. Compared with the global edit method Pullback [25], the additional cost is introduced from finding local edit direction, yet the addition cost is not significant. 2. Besides, we add additional comparisons with a local edit method BlendedDiffusion [1] in Table A of the global response PDF, where our method is more efficient in learning time. 3. Moreover, our method can transfer the local edit direction to other images with a high successful rate and no additional time cost in learning. On the contrary, other local/global edit methods cannot transfer the edit direction (BlendedDiffusion) or have weaker transferability as summarized in Table A. >Q2: Whether the homogeneity of the editing can be observed on samples that are very different: e.g., >- computing editing direction on an image with a face at the left, and transfer to an image with a face on the right; >- computing editing direction on an image with realistic style, and transfer to an image with cartoon style. A2: This is a good question. We present more results and extend the discussions. 1. - For unconditional diffusion models on FFHQ and CelebA, we can transfer edit directions to faces on the different side, as presented in Figure B of the global response PDF. But we do notice that these datasets tend to have the object in the center, which benefits the transferability. - We have further attached the visualization of editing directions in Figure C, and the edit direction has semantic correspondences to the target edit region. From our observation, the transferability is robust to gender differences, facial feature shape differences, and moderate position differences in FFHQ and CelebA. 2. - To achieve goal (2), conditional diffusion models are required. We have previously explored such transferability in stable diffusion models, and the transfer success rate is low. By our analysis, the homogeneity needs the images lie in the same subspace, but the manifold of images with different styles in conditional diffusion models may violate this. - Understanding the feature space of conditional diffusion models is a challenging and unsolved problem in the area, since the feature space is related to text prompts and noisy images. We are interested in exploring the question, and hope our theoretical analysis in unconditional diffusion models can provide some inspiration for future works. >Q3: In addition, since current models such as SD3 are using the flow-matching objective to train the model, I wonder if such observed phenomenon would be more prominent on these models? A3: 1. This is an interesting question and we have explored it further. Due to time constraints, we study the low-rankness and local linearity of SiT [2], a simpler diffusion model using the flow-matching objective at training. As presented in Figure F of the global response PDF, we do see generalized local linearity, and a similar low-rank trend in SiT, though the rank is not exactly as low as other diffusion models. It would be interesting to study the differences in low-rank subspace between diffusion models trained under different objectives, and test other diffusion models trained using flow-matching objectives such as Stable Diffusion 3. [1] Omri et al., Blended diffusion for text-driven editing of natural images. CVPR 2022. [2] Nanye et al., "Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers." 2024.	null	null	null	null	null	null
Motion Graph Unleashed: A Novel Approach to Video Prediction	Accept (poster)	Summary: The paper proposes the motion graph method in predicting the video frames by exploring the spatio-temporal relations among frames from the limited data. Strengths: It is significant to propose methods of few-shot prediction techniques in video inputs. Weaknesses: Please provide more details about the setup of the datasets such as KITTI and Cityscapes that are not originally used for testing video prediction. The results on Cityscapes and KITTI datasets based on LPIPS, SSIM, PSNR are needed to be explained in detail in the evaluation methods either in the main text or Appendix. Technical Quality: 3 Clarity: 4 Questions for Authors: . Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The authors address the limitations briefly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question 1: More details about the setup of the datasets such as KITTI and Cityscapes that are not originally used for testing video prediction. Answer 1: We follow the experiment setup from previous works which originated from Wu, Yue, et al. "Future video synthesis with object motion prediction". Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. We use the data preprocessing code from this repo, https://github.com/hzwer/CVPR2023-DMVFN/blob/main/utils/, to generate the train/val data. --- Question 2: The results on Cityscapes and KITTI datasets based on LPIPS, SSIM, PSNR are needed to be explained in detail in the evaluation methods either in the main text or Appendix. Answer 2: Thank you for the suggestion. We will add the calculation details in the appendix. --- Rebuttal Comment 1.1: Comment: I have read the rebuttal of the authors. I have seen the authors' main point on generating their literature from the previous work using KITTI and Cityscapes datasets for video prediction. I appreciate the effort put into this and hope that further researchers can benefit from that literature in the future and the related works multiply based on video prediction. As I tend to accept this work now, I hope to see the evaluation methods of LPIPS, SSIM, and PSNR on Cityscapes and KITTI datasets in the appendix in your revised version of the manuscript. --- Reply to Comment 1.1.1: Comment: Thank you for the valuable suggestion. We recognize that few existing studies report PSNR and SSIM metrics for evaluations on the KITTI and Cityscapes datasets. To support future research in the video prediction community, we will add a new table and discussion to Appendix A.3, detailing these metrics for our method. Additionally, we will compile and compare available results from other methods. We also plan to release the result images from these two datasets to enable more convenient quantitative and qualitative comparisons for future researchers.	Summary: The paper introduces a graph-based methods to predict video frames through motion prediction. The proposed motion graph captures complex spatial-temporal relationships by transforming video frame patches into interconnected graph nodes. This method improves performance and reduces computational costs compared to SOTA methods on UCF Sports, KITTI and Cityscapes. Strengths: 1. The paper introduces a novel motion graph representation, combining graph theory with video prediction to overcome existing limitations. 2. It has comprehensive evaluations across multiple datasets and achieve superior performance than prior methods. 3. The manuscript is well organized and has detailed experimental results. Weaknesses: 1. Slower inference speed compared to optimized methods like DMVFN. 2. Struggles with predicting abrupt or unpredictable motions. 3. Unclear if the motion graph captures hierarchical / semantic feature of motions. 4. Does not address the stochastic nature of human videos. 5. Limited diversity in evaluation datasets, focusing mainly on sports and driving scenarios. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Do the authors see a future where such work might be useful for improving video generative models? 2. Can the authors explain what the weight components are for? It seems that the dynamic vectors can represent motion even without the weighted component. If it does represent something meaningful, can the authors show what they represent after training? For example, might it correspond to the change in depth? 3. How does the motion graph approach compare to recent transformer-based video prediction methods in terms of performance and efficiency? 4. How sensitive is the method to the choices made in graph construction (e.g., number of neighbors, edge types)? 5. How does the motion graph approach scale with longer video sequences? Is there a practical limit to the number of frames it can handle efficiently? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors briefly mention limitations regarding inference speed and handling sudden motions. However, they could provide a more comprehensive discussion of potential limitations, such as scalability to longer videos or generalization to diverse video types. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question 1: Do the authors see a future where such work might be useful for improving video generative models? Answer 1: Yes. This work uses video prediction task as an example to validate the efficiency and effectiveness of motion graph as a comprehensive video motion representation. In the future, we will explore motion graph's potential as an efficient motion representation tool, developing advanced motion graph-based systems for long-term video generation. --- Question 2: Can the authors explain what the weight components are for? It seems that the dynamic vectors can represent motion even without the weighted component. If it does represent something meaningful, can the authors show what they represent after training? For example, might it correspond to the change in depth? Answer 2: After training, the graph construction process uses weights from the cosine similarity scores between pairs of image patches to determine their connectivity. Similarly, the weights output by the motion upsampler function as confidence scores for each predicted motion vector. This weighting strategy offers several advantages: a) Graph Construction Robustness: In the motion graph, each image patch ideally connects to $k$ patches in the subsequent frame. If a patch is occluded in the next frame, selecting the top $k$ connections could lead to random, inaccurate connections. The weighting strategy helps identify and minimize these erroneous connections. b) Selective Information Aggregation: During graph learning, this approach allows the model to focus on high-confidence correspondences and reduce the influence of ambiguous connections, enhancing the overall learning effectiveness. c) Enhanced Frame Synthesis: During the synthesis of future frames, multiple source pixels may correspond to a single pixel in the future frame. The weighted approach allows for strategic aggregation of these source pixels, improving the quality of the resulting image. --- Question 3: How does the motion graph approach compare to recent transformer-based video prediction methods in terms of efficiency and performance? Answer 3: We appreciate the reviewer initiating this discussion. Both the motion graph approach and transformer-based video prediction methods (e.g. MaskViT) share the fundamental goal of predicting unknown relational information by modeling existing temporal correspondences between adjacent frames. However, significant differences in efficiency, performance, and adaptability set them apart: a) Efficiency: Transformer-based methods create a complete graph for all feature patches between two adjacent frames, employing complex models that require a large number of parameters. This often leads to high computational demands, making these methods less suitable for high-resolution video processing. In contrast, the motion graph approach constructs a sparser graph, with edge weights determined by feature similarities from a CNN-based image encoder. This results in a more lightweight model. Furthermore, motion graphs leverage Graph Convolutional Networks (GCNs) composed of stacked linear layers, enabling efficient learning from the graph structure. b) Performance: The efficiency differences make direct performance comparisons with transformer-based methods challenging. Currently, few transformer-based approaches are tested on the high-resolution datasets discussed in this manuscript, leaving performance comparisons inconclusive. c) Adaptability to Different Video Resolutions: A major drawback of transformer-based methods is their lack of flexibility with varying input resolutions during testing. They struggle to process videos of different resolutions effectively. Conversely, our motion graph approach employs a CNN-based image encoder and motion upsampler, offering greater adaptability to various video resolutions. This adaptability enhances the suitability of the motion graph method for real-world applications. --- Question 4: How sensitive is the method to the choices made in graph construction (e.g., number of neighbors, edge types)? Answer 4: Our ablation study (detailed in Table 7) reveals a monotonic performance increase as the value of $k$ increases, although the rate of gain diminishes with larger $k$ values. This suggests that performance is more sensitive to changes in $k$ when $k$ is small, as it more closely resembles an optical flow-based method. However, as $k$ increases and motion information approaches saturation, the impact of $k$ lessens. Additionally, Table 12 in Appendix A.4 demonstrates that including both spatial and backward edges enhances model performance. --- Question 5: How does the motion graph approach scale with longer video sequences? Is there a practical limit to the number of frames it can handle efficiently? Answer 5: Longer video sequence can add more nodes and edges to the motion graph, which will result in an increase in the parameter number and computational overhead of motion graph learning. However, in terms of graph construction and feature extraction, the impact of video sequences length will be less, since we extract each image frame individually through an image encoder. Lastly, the practical limit to the number of frames a motion graph approach can handle efficiently is typically dependent on the hardware constraint. The available GPU memory significantly influences how many frames can be effectively managed.	Summary: This paper proposes a motion-based method for video prediction. They design a new motion representation named motion graph that transforms patches of video frames into interconnected graph nodes. The proposed video prediction pipeline, empowered by the motion graph, exhibits substantial performance improvements and cost reductions. Experiments on UCF Sports, KITTI, and Cityscapes are conducted. Strengths: + Motion-based video prediction is quite an interesting topic and benefits the research community with an affordable training cost. + The evaluation of both efficiency and effectiveness is conducted. + The paper is well-structured. Weaknesses: - Motivation and Occlusion / Out-of-View cases. As the authors claimed in lines 33-37 and Figure 1, existing methods struggle to effectively handle occlusion cases. However, the paper does not provide a corresponding evaluation to demonstrate the effectiveness of the proposed motion graph in complex situations involving occlusion or out-of-view scenarios. This is a significant omission, as these challenges are prevalent in real-world video prediction tasks. Furthermore, based on my understanding of the work, the proposed motion graph may not adequately address these problems. Specifically, occluded and out-of-view object/background do not appear to be considered in the graph, as the connections and features captured in the motion graph are derived from the visible parts of the targets. This raises concerns about the method's ability to maintain accurate predictions when dealing with occlusions or objects moving out of the camera's field of view. The authors should provide a thorough discussion that includes both the theoretical principles and empirical evidence. - Video prediction evaluation setting. While the proposed motion graph method shows promising results in predicting a few frames (up to 10), the paper lacks an evaluation of its performance in predicting a larger number of frames. Other video prediction works [A, B, C] have evaluated their methods on longer prediction horizons, providing a more comprehensive assessment of their models' robustness and accuracy. [A] ExtDM: Distribution Extrapolation Diffusion Model for Video Prediction, CVPR24 [B] Efficient Video Prediction via Sparsely Conditioned Flow Matching, ICCV23 [C] MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation, NeurIPS22 - Comparison with MMVP. The proposed motion graph method appears to be quite similar to the motion matrix used in MMVP, which also provides a dense connection among frames. Although I appreciate the discussion provided in the related work, the authors do not sufficiently clarify the differences and advantages of the proposed method compared to MMVP. This lack of distinction makes it difficult to understand the unique contributions and potential improvements offered by the motion graph. - Comparison with other motion-cues-guided video prediction methods. The paper lacks a comparison with other motion-cues-guided video prediction methods, such as MotionRNN, which also leverages a vector-based motion representation. The authors should explain how the feature warping mechanism in MotionRNN compares to the feature processing in the motion graph. What are the potential benefits or drawbacks of each approach? Are there specific scenarios where the motion graph's approach to feature handling provides a clear advantage? - Lack of a component-wise ablation. The reviewer noticed that there is no component-wise ablation study for motion feature learning, upsampling, decoding, image wrapping, and interaction modules. It makes it difficult for the audience to identify which parts of the methodology are effective and contribute most to the overall performance. The authors should conduct a thorough ablation study that isolates and evaluates the impact of each component of their proposed method. - Complexity of the proposed method and presentation. Actually, the proposed method is somehow non-trivial, making this work hard to follow. Accurate presentation can make the work easier to understand. Technical Quality: 4 Clarity: 2 Questions for Authors: See weakness Confidence: 4 Soundness: 4 Presentation: 2 Contribution: 4 Limitations: The authors indicate limitations in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1. Motivation and occlusion / out-of-view cases. A1: As a non-generative model, our proposed video prediction system may face challenges with occlusions that require the generation of unseen objects. However, it excels in scenarios with scenarios involving occlusions of known objects, as showcased in the third column of Figure 6, where our method outperforms existing SOTA methods. Notably, in scenes with partial obstructions—such as the white wall behind the cyclist—our system adeptly employs multiple motion vectors per pixel to reconstruct occluded areas. This feature also supports precise management of object expansion due to perspective projection, as exemplified by the green car in Figure 6's first column. Moreover, our approach is adept at managing objects exiting the camera’s view. By explicitly modeling the motion of image patches, the motion graph predicts when features are about to leave the scene. Thus, any image patches projected to move out of view are not included in the final prediction. This is demonstrated in the 1st, 3rd, and 4th columns of Fig 6, where objects like a cyclist and the front of a blue truck are shown as moving out of frame. Our method successfully and uniquely captures and represents these movements, unlike how the other methods evaluated. --- Q2. Video prediction evaluation setting. A2: Thank you for initiating this discussion. Our research on video prediction has identified key differences between systems designed for short-term and long-term predictions. Short-term systems typically use fewer frames to predict the immediate next few frames, while long-term systems are tasked with forecasting an extensive sequence of future frames. The design logic and objectives of these systems, thus, vary significantly. For a more detailed comparison, we have outlined these distinctions in the table below: \|Task type\|Output length\|Resolution\|Dataset\|Recent work\|Objectives\| \|---\|---\|---\|---\|---\|---\| \|Short-term\|Short\|Up to 4K\|UCF 101, UCF Sports, KITTI, Cityscapes, SJTU4k, Vimeo, DAVIS\|SIMVP, MMVP, STRPM, STIP, DMVFN\|High resolution videos, pixel-level accuracy, real-time application\| \|Long-term\|Long ($\gg 10$)\|Up to $256\times 256$\|KTH, moving MNIST, BAIR, cropped Cityscapes\|MVCD, MaskViT,VIDM,ExtDM\|Conditional video generation, semantic-level accuracy\| In this study, we emphasize our method's advanced motion modeling and significant reduction in computational costs, essential for short-term prediction of high-resolution videos. Our system is tailored for short-term video prediction, with evaluations conducted along this line of work. We plan to further exploit the motion graph's potential as an efficient motion representation tool and develop advanced, motion graph-based systems for long-term video generation in our next work. --- Q3: Comparison with MMVP. A3: We highlight in the manuscript three major differences to MMVP. \|Method\|Motion representation\|Prediction module architecture\|Motion & Appearance Composition method\| \|---\|---\|---\|---\| \|MMVP\|Motion Matrix ($H\times W\times H\times W$)\|3D Convolution based network\|Downsampled-feature-level matrix multiplication\| \|Motion Graph (Ours)\|Motion Graph ($H\times W\times k\times 3$) \|Graph convolution network\|Pixel-level image forward warping\| The advantages of our method are evidenced by its performance in various tests: a) Our sparser motion representation allows the system to process videos with higher resolutions using fewer computational resources. For instance, Table 6 indicates that GPU consumption for our method is 52% of that required by MMVP. b) Our graph-based motion prediction module aggregates relational information from motion embeddings more efficiently, resulting in a lighter and more effective model. As shown in Tables 4 and 6, our method requires only 21% of the model size of MMVP while improving the LPIPS metric by 22%. c) Pixel-level image warping maintains critical appearance information from existing frames, significantly boosting detail recovery. Figure 5 demonstrates this by comparing the clarity of details like the horse’s facial features, the rider’s face, and the athlete’s feet, with our method outperforming others in detail retention. --- Q4: Comparison with other motion-cues-guided video prediction methods. A4: Although MotionRNN also utilizes motion vectors, there are several fundamental differences: a) MotionRNN employs local $3\times3$ motion filters, which are limited in handling large-scale motion; b) MotionRNN does not account for the interaction of motion across different spatial and temporal regions, a feature that our motion graph learning process emphasizes; c) MotionRNN performs feature-level warping and our approach conducts pixel-level warping, providing finer detail. These distinctions contribute to the lesser robustness of MotionRNN in handling high-resolution videos with complex motion scenarios, as detailed in Tab 3 of the manuscript. We have reviewed various motion-based video prediction methods in Sec 2. Moving forward, we plan to conduct a further literature review to expand and update this section accordingly. --- Q5: Lack of component-wise ablation study. A5: The components of motion upsampling, motion decoding, and image warping modules are each essential to the functionality of the system. So we have opted not to isolate each component for ablation studies. Instead, we focus the ablation analysis on the composition of motion features: please see details in Table 8 of the manuscript. Appendix A.4 also provides additional ablation studies, with Table 12 specifically exploring the effects of various interaction modules. --- Q6. Presentation of this complex method. A6: Thank you for this comment. To more effectively present our proposed method, for it to be more comprehensible and accessible, in the revised manuscript we worked on the clarity of descriptions and incorporated additional visual explanations. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer Mw8x Comment: Thanks for the rebuttal. Most of my concerns are addressed. I'd like to see the discussions about occlusions, long-term VP, and other related methods in the revised version. --- Reply to Comment 1.1.1: Comment: We deeply appreciate your suggestions and will incorporate a discussion on occlusion, long-term video prediction, and comparisons with other methods into the revised manuscript.	Summary: The authors propose a Motion Graph for predicting future video pixels. To achieve this, they introduce three modules: 1. Motion Graph Node Construction; 2. Edge Construction; 3. Graph Interaction Module; 4. Video Prediction Model. The first three modules encode patches and their interactions in a frame, while the fourth module decodes future frame pixels. Experiments on UCF Sport, KITTI, and CityScape demonstrate the method's high efficiency and construction quality. Strengths: 1. The author propose to predict future pixels from motion graph perspective, to achieve this, they propose motion graph constuction and graph decoding modules. 2. The method demonstrate high computation efficieny and good reconstruction quality on few video benchmarks. Weaknesses: 1. It's better to present an overview figure for the framework. It's not easy to grasp the overall structure by reading about the four separate modules. 2. Figures 2-4 have fonts that are too small, making them hard to read. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Which backbone are used to extract features during motion graph construction. Does the method robust to different backbones? 2. Does the prediction quality decrease with respect to resolutions? is it possible to analyze prediction quality with large resoultion settings. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors discuss and partially address the limitations of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question 1: Which backbone are used to extract features during motion graph construction. Does the method robust to different backbones? Answer 1: ResNet was used as the backbone of our image encoder to extract features during motion graph construction. In the manuscript submission, Figure 7 of Appendix A.2 illustrates the network architecture of the image encoder. Our experiments demonstrate the robustness of our design across different image encoder architectures. We selected a comparatively lightweight model to ensure both model compactness and system computational efficiency. --- Question 2: Does the prediction quality decrease with respect to resolutions? is it possible to analyze prediction quality with large resolution settings. Answer 2: To evaluate the model's performance on high-resolution images, we tested our method on the SJTU4K dataset, which has a resolution of $2160\times3840$. Due to the absence of publicly available data splits for direct comparison with existing state-of-the-art (SOTA) methods, we created our own training and validation splits, allocating 80% of the sequences to training and 20% to validation. \| Dataset \| Resolution \| LPIPS $\downarrow$ \| MS_SSIM $\uparrow$ \| \|:---:\|:---:\|:---:\|:---:\| \| KITTI \| $256\times 832$ \| 9.50 \| 87.70 \| \| Cityscapes \| $512\times 1024$ \| 4.13 \| 94.85 \| \| SJTU4K \| $2160\times 3840$ \| 8.22 \| 90.70 \| While the model's performance on the SJTU4K dataset surpasses its performance on KITTI, which has a much lower resolution, it does not match the performance on Cityscapes, which also has a smaller resolution than SJTU4K. Our findings suggest that factors such as motion complexity, frame rate, and scene complexity have a greater influence on prediction performance than frame resolution alone.	Rebuttal 1: Rebuttal: We are grateful for the thoughtful suggestions and comments from the reviewers. In response, we have implemented several enhancements to the manuscript, including, but not limited to: a) Enhancing the visual presentation throughout, such as adding an overview figure; b) Adjusting the text font size of the figures for better readability; c) Expanding the literature review to include more comprehensive discussions; d) Clarifying the descriptions of our methods for improved understanding; e) Providing additional details about the evaluation metrics and the development of the testing datasets. Please see below for our point-by-point response to the reviewers' feedback. Thank you for your time.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
NoMAD-Attention: Efficient LLM Inference on CPUs Through Multiply-add-free Attention	Accept (poster)	Summary: The paper presents NoMAD-Attention, using SIMD registers in CPU, to speedup LLM inference in CPU. Strengths: (1) Important problem with interesting solution. (2) The idea is straightforward, and the figure is very clear. (3) Good system speedup and maintaining the original performance of attention. Weaknesses: The ML benchmarks seem a bit weak (e.g. perplexity, and PIQA-like easy benchmark). Can you evaluate on harder tasks like MT-Bench? Technical Quality: 3 Clarity: 3 Questions for Authors: Please see weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors include the limitation that this only works on CPUs with SIMD, but the reviewer thinks this is not a significant limitation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate the reviewer's careful review and valuable suggestions. We address their comments as follows. [W1] Evaluations on harder tasks. To further assess NoMAD-Attention, we conducted additional evaluations on the more challenging MMLU, GPQA, and MGSM (English) benchmarks. Our results demonstrate that NoMAD-Attention with $d_{sub}=1$ effectively maintains model quality across these diverse tasks. \| \| MMLU (STEM) \| MMLU (Social Sciences) \| MMLU (Humanities) \| MMLU (Other) \| GPQA \| MGSM \| \|-------------------------------------\|-------------\|------------------------\|-------------------\|--------------\|-------\|------\| \| LLaMA-7b (Attention) \| 26.39 \| 29.57 \| 29.73 \| 33.15 \| 20.98 \| 4.8 \| \| LLaMA-7b (NoMAD-Attention $d_{sub}=1$) \| 27.31 \| 29.12 \| 29.44 \| 32.8 \| 23.66 \| 4.0 \| \| LLaMA-7b (NoMAD-Attention $d_{sub}=2$) \| 25.25 \| 24.99 \| 26.82 \| 30.09 \| 20.08 \| 2.0 \| \| LLaMA-7b (NoMAD-Attention $d_{sub}=4$) \| 25.21 \| 22.98 \| 24.85 \| 26.94 \| 25.67 \| 1.6 \| \| LLaMA-13b (Attention) \| 34.13 \| 44.39 \| 40.60 \| 46.48 \| 28.35 \| 7.2 \| \| LLaMA-13b (NoMAD-Attention $d_{sub}=1$) \| 33.43 \| 43.71 \| 39.57 \| 45.83 \| 28.35 \| 7.6 \| \| LLaMA-13b (NoMAD-Attention $d_{sub}=2$) \| 30.70 \| 37.96 \| 34.39 \| 40.20 \| 27.01 \| 6.8 \| \| LLaMA-13b (NoMAD-Attention $d_{sub}=4$) \| 25.88 \| 25.58 \| 27.46 \| 27.55 \| 25.67 \| 1.2 \| --- Rebuttal Comment 1.1: Comment: Thank you! This is good. I raised the score to 6. Please consider accept this paper.	Summary: This paper proposes using SIMD instructions on CPUs to speed up Transformers by removing multiply-add instructions. The paper replaces the attention operation a lookup-table based alternative. The paper motivates the application well for Transformer inference on CPUs. Strengths: Strengths: * Important problem (efficiency) on an understudied platform (CPUs) * Shows speedup compared to standard attention * The quality studies are strong, with evaluation on upstream and downstream benchmarks. There is some degradation in quality, but it can be adjusted based on hyperparameters. I have also reviewed this paper in a previous conference (ICML) and I see that the paper is much improved from that version. I recommend accept based on the strength of the motivation and quality of the evaluation. Weaknesses: Evaluation on additional and more recent open models (LLaMa-2, LLaMa-3, Gemma, QWEN) would make the paper more convincing. Two open questions from the current evaluation are how well the method generalizes to other pretrained models, and whether changes in the attention algorithm from more recent models change it. These are not necessary for the paper, but are natural follow-ups for the work and would improve the paper's impact. Technical Quality: 4 Clarity: 4 Questions for Authors: How well does the method generalize to other open models and to changes in the attention algorithm? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for their support of our paper and the thoughtful suggestions. We address the reviewer's concerns as follows. [W1] Evaluation on additional and more recent open models. - We conducted additional experiments on the LLaMA-3-8b model with NoMAD-Attention across a range of downstream tasks. Our results demonstrate that NoMAD-Attention with $d_{sub}=1$ effectively maintains model quality. \| \| SciQ \| Arc-E \| Arc-C \| Hellaswag \| WinoGrande \| PIQA \| \|--------------------------------------\|------\|-------\|-------\|-----------\|------------\|-------\| \| LLaMA-3-8b (Attention) \| 96.4 \| 80.09 \| 50.51 \| 60.18 \| 72.77 \| 79.71 \| \| LLaMA-3-8b (NoMAD-Attention d_sub=1) \| 96.1 \| 80.05 \| 49.49 \| 59.86 \| 73.16 \| 79.65 \| \| LLaMA-3-8b (NoMAD-Attention d_sub=2) \| 94.8 \| 78.32 \| 46.59 \| 57.52 \| 70.17 \| 78.35 \| \| LLaMA-3-8b (NoMAD-Attention d_sub=4) \| 86.3 \| 70.08 \| 37.54 \| 47.38 \| 57.38 \| 76.61 \| [W2, Q1] How well does the method generalize to other open models and to changes in the attention algorithm? NoMAD-Attention can generalize to other pretrained models and attention variants such as grouped-query attention (GQA) [1] and Attention with Linear Biases (ALiBi) [2]. Given that GQA employs shared key heads across multiple query heads, we can adapt NoMAD-Attention to reuse the same key codes for performing in-register lookups. The ALiBi method adds a linear bias term to the query-key dot products, a process fully compatible with the NoMAD-Attention approach. References [1] Ainslie, Joshua, et al. "GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints." Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. [2] Press, Ofir, Noah Smith, and Mike Lewis. "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation." International Conference on Learning Representations. --- Rebuttal Comment 1.1: Title: Response Comment: Thank you for the rebuttal - I will be keeping my high score. Good work!	Summary: This paper utilizes product quantization (PQ) to replace dot product operations in the matrix multiplications involved in the attention mechanism of transformers with memory lookup operations, showcased on language models. To my knowledge, this technique has been first introduced by Blalok et al (reference [5] in the paper). This paper goes further into the systems aspect of deploying PQ on GPUs by specifically utilizing SIMD registers to store the codebook used in PQ. This promises to significantly improve the memory access overhead even compared to alternative implementations that keep the PQ codebook in the L1 cache. However, this poses very stringent codebook size constraints that are addressed in the paper. PQ is done completely post-training unlike prior work, but uses a calibration dataset. Strengths: - This paper is addressing an important topic: LLM optimization on commodity CPUs. - The use of SIMD registers is innovative and results in significant speedup. - PQ has been demonstrated a number of times in the past couple of years, but never with strong results on LLMs. Weaknesses: - d_sub is very confusing to me. What is the length or your codeword? if d_sub = 1, is the size of the codeword = 1? how is that product quantization? How does it result in a speedup? - the proposed product quantization has been proposed to replace MAD in matrix multiplications in reference [5]. Line 148 in this paper is overclaiming its contribution suggesting that this submitted paper has proposed this method. - Can you please clarify the additional memory overhead introduced by the codebook? - "double quantization" of PQ has been introduced in prior work and is not cited: Abouelhamayed et al: PQA: Exploring the Potential of Product Quantization in DNN Hardware Acceleration. Technical Quality: 3 Clarity: 3 Questions for Authors: see above Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for the reviewers' careful review and insightful comments. We have addressed their feedback in detail below. [W1] d_sub is very confusing to me. What is the length or your codeword? if d_sub = 1, is the size of the codeword = 1? how is that product quantization? How does it result in a speedup? - We are sorry for the confusion. To clarify, $d_{sub}$ represents the dimensionality of each sub-quantizer, which is equivalent to the dimensionality of each cluster centroid or the number of dimensions encoded by each product-quantized key code. Notably, the special case where $d_{sub}=1$ can be considered a generalization of multi-dimensional product quantization. - $d_{sub}=1$ results in a speedup primarily due to two reasons: 1. Leveraging Lower-Latency, Higher-Throughput Instructions. Unlike the vanilla multiply-add attention, which relies on batched multiplication and addition SIMD instructions (e.g., `_mm256_mul_ps` and `_mm256_add_ps` in AVX2), NoMAD-Attention utilizes the SIMD lookup instruction (`_mm256_shuffle_epi8`). This latter instruction operates on more elements at once (32 elements versus 8) and exhibits lower latency (1 cycle vs. 4 cycles on most architectures) [1], contributing significantly to the efficiency gains. 2. Minimized Data Movement. The product-quantized key cache employed in NoMAD-Attention effectively reduces the volume of data transferred between RAM and registers, hence speeding up computations. [W2] Line 148 in this paper is overclaiming its contribution. - We apologize for this oversight. We will revise line 148 in the final paper to the following: "Building upon previous work [5], we employ product quantization to approximate dot products within the attention mechanism." [W3] Can you please clarify the additional memory overhead introduced by the codebook? --- Sure! - The memory overhead introduced by the codebook can be computed as $l \times h \times d \times 16 \times 4$ bytes, where $l$ represents the number of layers, $h$ the number of attention key heads, $d$ the dimensionality in an attention head, $16$ the number of centroids, and $4$ the number of bytes for storing each centroid parameter. - Therefore, the codebook memory overhead for LLaMA-7b/LLaMA-2-7b with $d_{sub} \in \\{1,2,4\\}$ is 8.4MB, and the memory overhead for LLaMA-13b/LLaMA-2-13b with $d_{sub} \in \\{1,2,4\\}$ is 13.1MB, which is a relatively modest memory footprint compared to the overall model size. - We thank the reviewer for highlighting this point, and we will incorporate a detailed discussion of codebook memory overhead into the final paper. [W4] "double quantization" of PQ has been introduced in prior work and is not cited. - We thank the reviewer for bringing PQA to our attention. We will incorporate a citation to this relevant work in the final version of our paper. References [1] https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html --- Rebuttal Comment 1.1: Title: Thank you for the response Comment: The clarifications proposed by the authors should improve the paper writeup. I will maintain my positive-leaning score for this paper..	Summary: This paper proposes an algorithm to compute vector inner product efficiently on CPU by exploiting the fast access speed of in-register memory for fast self-attention computation on CPU for model inference. Instead of use multiply and add to compute the inner product between query and key vectors, the authors propose to break down the inner product between two vectors to the summation of multiple inner product between sub-vectors. For each sub-vector product, given a query vector, the algorithm enumerates all possible dot-product results with all possible key sub-vectors in a lookup table during preprocessing. The key sub-vectors are quantized so that the the lookup table is small enough to fit into the in-register memory. The authors further optimize the efficiency via optimizing key cache memory layout. Strengths: 1. The algorithm uses lookup operation to avoid multiplication and allows calculating multiple dimensions at once when the dimension of sub-vectors is larger than 1. 2. The algorithm is able to utilize an extremely small memory (128 bits) to accelerate the calculation. Weaknesses: 1. It would be better to also give some data about the latency of GPU decoding so that we know how far CPU is behind the GPU in Figure 2. 2. When the $d_{sub} > 1$, as shown in Table 1, the performance of these models got a severe hit, so it seems that the algorithm only works well on enumerating all possible scalar products for $d_{sub} = 1$. 3. Even when $d_{sub} = 1$, there are still latency improvement on Figure 2, but I think the number of operations should be the same as the vanilla multiply-add attention, except that each multiply is replaced with a lookup. I am wondering where is the efficiency improvement. 4. For $d_{sub} = 1$, the algorithm is quite relevant to quantized matmul for self-attention computation, I am wondering how the method performs compared to int8 or lower bit width quantization in terms of model accuracy and efficiency. Technical Quality: 3 Clarity: 2 Questions for Authors: see weakness section. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors discussed limitations and societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewers' careful consideration of our paper and their valuable feedback. We have addressed each of their comments below. [W1] It would be better to also give some data about the latency of GPU decoding so that we know how far CPU is behind the GPU in Figure 2. - We thank the reviewer for their valuable suggestion. We will incorporate the latency of GPU decoding into the final version of the paper to provide a more comprehensive evaluation. [W2 & W3] Values of $d_{sub}=1$. - As demonstrated in Table 1, a $d_{sub}$ value of 2 still yields reasonable model performance, as measured by perplexity and accuracy. - The $d_{sub}=1$ configuration already achieves significant speedups compared to the original model, with a 1.78x improvement observed at a context length of 16K (Figure 2). - The efficiency improvement achieved when $d_{sub}=1$ is primarily due to two factors: 1. Leveraging Lower-Latency, Higher-Throughput Instructions. Unlike the vanilla multiply-add attention, which relies on batched multiplication and addition SIMD instructions (e.g., `_mm256_mul_ps` and `_mm256_add_ps` in AVX2), NoMAD-Attention utilizes the SIMD lookup instruction (`_mm256_shuffle_epi8`). This latter instruction operates on more elements at once (32 elements versus 8) and exhibits lower latency (1 cycle vs. 4 cycles on most architectures) [1], contributing significantly to the efficiency gains. 2. Minimized Data Movement. The product-quantized key cache employed in NoMAD-Attention effectively reduces the volume of data transferred between RAM and registers, hence speeding up computations. [W4] How does the method perform compared to int8 or lower bit width quantization in terms of model accuracy and efficiency? --- Here is a comparison. - We present additional experiments comparing NoMAD-Attention with INT8 and INT4 key cache quantization (q8_0 and q4_0 in llama.cpp). We report the accuracy on various benchmarks and the decoding latency at a context length of 16K tokens for LLaMA-7b. - NoMAD-Attention ($d_{sub}$=1) demonstrates better model quality than INT4 quantized key cache and significantly higher speedup than INT4 and INT8 quantized key cache. \| \| SciQ \| Arc-E \| Arc-C \| Hellaswag \| WinoGrande \| PIQA \| Avg \| Decoding Latency (16K) \| Speedup \| \|------------------------\|------\|-------\|-------\|-----------\|------------\|-------\|--------\|---------------------------------------\|---------\| \| Original Attention \| 94.6 \| 75.21 \| 41.89 \| 56.93 \| 70.09 \| 78.67 \| 69.565 \| 572.497 \| - \| \| INT8 Quantized Key Cache \| 94.7 \| 75.29 \| 42.15 \| 57.00 \| 70.09 \| 78.63 \| 69.643 \| 562.377 \| 1.018x \| \| INT4 Quantized Key Cache \| 93.6 \| 74.33 \| 41.04 \| 55.34 \| 67.96 \| 77.69 \| 68.327 \| 540.319 \| 1.060x \| \| NoMAD (d_sub=1) \| 94.9 \| 75.34 \| 41.81 \| 56.57 \| 70.56 \| 78.56 \| 69.623 \| 391.098 \| 1.464x \| References [1] https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewers' careful evaluation of our paper and their valuable feedback. In the following section, we address common concerns raised by multiple reviewers. We are happy to provide further clarification during the discussion period. 1. Regarding $d_{sub}$ NoMAD-Attention achieves speedup when $d_{sub}=1$ primarily due to two factors: 1. Leveraging Lower-Latency, Higher-Throughput Instructions. Unlike the vanilla multiply-add attention, which relies on batched multiplication and addition SIMD instructions (e.g., `_mm256_mul_ps` and `_mm256_add_ps` in AVX2), NoMAD-Attention utilizes the SIMD lookup instruction (`_mm256_shuffle_epi8`). This latter instruction operates on more elements at once (32 elements versus 8) and exhibits lower latency (1 cycle vs. 4 cycles on most architectures) [1], contributing significantly to the efficiency gains. 2. Minimized Data Movement. The product-quantized key cache employed in NoMAD-Attention effectively reduces the volume of data transferred between RAM and registers, hence speeding up computations. References [1] https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D	Accept (poster)	Summary: The paper attempts to address the Janus Problem in SDS-based text-to-3D methods. It first analyzes the cause of the Janus Problem in SDS-based approaches, identifying that discrete view encoding and shared priors in 2D lifting are the primary causes. To address this, it proposes the LCGen method, which guides text-to-3D generation to obtain different priors with varying certainty from different perspectives, thereby ensuring view consistency. Experiments show that the LCGen method can be seamlessly integrated into various SDS-based text-to-3D methods, effectively mitigating the Janus Problem without significant side effects. Strengths: 1. This work attempts to address the challenge of the Janus Problem by tackling its root causes and identifying the key factors that contribute to this issue, making a significant contribution to solving this critical problem. 2. The proposed method addresses the issue from a novel perspective, and experimental results have confirmed its effectiveness. Weaknesses: 1. The examples presented in this work are relatively homogeneous. Have you tried other types of examples? (For instance, a finely designed object) 2. The limitations section and the appendix in the paper show some limitations and failure cases. If this method is supposed to fundamentally address the Janus problem, what are the reasons for these failure cases and limitations? I did not see much detailed discussion about this in the paper. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: n/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your effort in reviewing our work! Our responses are as follows: ## 1. Diversified Examples In Rebuttal File, we demonstrate how our method alleviates the Janus Problem in various other types of examples. In Fig. C and D, we tested on "a sunflower" and "a piano" to show the effectiveness of our method when the object does not contain a "head". It can be observed that the original Prolificdreamer generates sunflowers and keyboards of piano on both the front and back, resulting in spatial inconsistency. After applying our method, it correctly renders the front and back images of the sunflower and piano, resolving the spatial inconsistency issue. In Fig. F, we carefully designed a prompt to show the effect of our method when a large amount of descriptive language is added: "A sleek, silver-gray dolphin leaping gracefully out of the crystal-clear ocean, its body glistening in the sunlight as it arcs through the air with joyful exuberance." It can be observed that the original Prolificdreamer generates two dolphin heads, while our LCGen method correctly generates the image. In Fig. G, we also carefully designed a prompt: "A Matte painting of a Christmas cake made of cheese surrounded by a moat made of ice cream". It can be observed that in the original prolificdreamer, the front and back of Santa Claus on the cake both have faces. However, using our method, the correct samples are generated. ## 2. Limitation Discussion For the failure cases, our analysis is as follows: 1. When objects are in very unusual poses, our method has a certain probability of failing. For example, "an upside-down lion." In this case, using the original certainty to separate different views fails because the bottom might be the highest certainty view. For such samples, we need to redesign the view-based guidance function to accommodate the generation of these special samples. 2. Our method may fail in multi-object generation scenarios. This is due to the limitations of current text-to-3D methods in multi-object generation tasks. As shown in Fig. H in Rebuttal File, when the text is "three corgis facing different directions," baseline method Prolificdreamer fails to generate three corgis correctly, with an overlap between different objects. On this basis, the current best view-consistent generation methods, such as DreamControl and our method, cannot directly address the Janus Problem. To solve this issue, the primary task is to solve the multi-object generation problem in text-to-3D, which is a significant research area. For example, a current mainstream idea is to generate individual assets for each object and then stitch them together[1]. Our method still has the potential to adapt to such methods. 3. Due to the lack of 3D prior knowledge, like other SDS-based methods, our method can only model 3D representations that appear more realistic from various perspectives, but cannot ensure that these 3D representations adhere to the physical laws of the real world. For example, during the generation of octopus tentacles, since there is no difference in tentacles from different views and the model does not know how many tentacles should be generated, it may produce 3D representations that do not conform to objective reality. To address this issue, we need to endow the model with the ability to understand the 3D world. One possible approach is to collect massive 3D data and design appropriate representation forms to establish a pre-trained 3D generation model (e.g., 3D diffusion). Given the enormous data requirements and training difficulty, this requires the collective effort of the entire AIGC community. Thank you again for your efforts and suggestions! We will include detailed discussions on the above points in the Main Paper and Appendix. ## Reference [1] Zhou, Xiaoyu, et al. "Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting." arXiv preprint arXiv:2402.07207 (2024).	Summary: This paper presents a simple and effective method to address the Janus Problem for Score Distillation Sampling (SDS)-based text-to-3D methods. This paper argues view consistency is related to that the 3D model tends to learn content with higher certainty from each perspective, and using different priors with different certainty will help with more consistent generation. Specifically, they assume that the certainty of one denoising step follows a Gaussian distribution and design additional loss to constrain the distribution so that different viewpoints have different distributions. Experiments illustrate that their proposed method alleviates the Janus Problem. Strengths: 1. The proposed method is simple and effective. Compared to previous work, this method does not require additional data or models, and can be well integrated into the computation of the original SDS loss. 2. This paper is well-motivated by a detailed analysis of the Janus Problem and the underlying reason for decoupling the relationship between viewpoints and distributions. 3. The empirical experimental results show that it can effectively alleviate the Janus Problem. Weaknesses: 1. One major concern is that this paper lacks enough comparison with previous work. Although in Figure 6, the authors mention several advantages (on the resources, universality, and training efficacy), comparing against the empirical performance quantitatively is still useful. The model may not outperform other models that have more requisites, but it is still important to know where we stand. 2. In addition, for the comparison with DreamControl, the main difference is that this model does not need an additional fine-tuning stage. Because it does not affect the inference time consumption (which is in general more important when we consider the efficacy issue), the contribution will be a bit limited if the proposed method does not outperform DreamControl. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can you provide a quantitative performance comparison with the previous work on the Janus Problem? 2. Is there a quantitative analysis of the advantages of training time cost compared to DreamControl? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author mentioned the limitation with regard to the modeling of multiple objects or complex objects. Appendix H also includes a failure example of the multiple object issue. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your effort in reviewing our work! Our responses are as follows: ## 1. Quantitative Comparison with Other Methods Dealing with the Janus Problem We have provided the quantitative comparison results with other methods in the table in Rebuttal File (as well as in the table below). It can be observed that our method can comprehensively surpass or partially match the performance of other methods. Specifically, our method shows the best suppression of the Janus Problem when embedded in dreamfusion; when embedded in prolificdreamer, it achieves the best overall performance, strongly suppressing the Janus Problem while maintaining the highest CLIP score. Notably, our method can achieve good performance without requiring additional 3D priors or multi-stage fine-tuning, consuming minimal extra computational power (see quantitative comparison in the next response) and can be directly applied to SDS-based text-to-3D methods. \| Method \| No Additional Prior \| Single Stage \| No Fine-tuning \| No Object-specificity \| JR(%)↓ \| CS(%)↑ \| \|------------------\|---------------------\|--------------\|----------------\|-----------------------\|--------\|--------\| \| Zero-1-to-3 \| ✘ \| ✘ \| ✘ \| ✔ \| 23.33 \| 22.94 \| \| MVDream \| ✘ \| ✘ \| ✘ \| ✔ \| 20.00 \| 26.31 \| \| Prep-Neg \| ✔ \| ✔ \| ✔ \| ✘ \| 26.67 \| 26.23 \| \| D-SDS \| ✘ \| ✔ \| ✔ \| ✔ \| 23.33 \| 24.82 \| \| DreamControl \| ✔ \| ✘ \| ✘ \| ✔ \|20.00 \| 28.03 \| \| Dreamfusion \| \| \| \| \| \| \| \| Origin \| ✔ \| ✔ \| ✔ \| ✔ \| 56.67 \| 22.73 \| \| LCGen (Ours) \| ✔ \| ✔ \| ✔ \| ✔ \| 16.67 \| 22.95 \| \| Magic3D \| \| \| \| \| \| \| \| Origin \| ✔ \| ✔ \| ✔ \| ✔ \| 46.67 \| 23.77 \| \| LCGen (Ours) \| ✔ \| ✔ \| ✔ \| ✔ \| 23.33 \| 23.61 \| \| ProlificDreamer \| \| \| \| \| \| \| \| Origin \| ✔ \| ✔ \| ✔ \| ✔ \| 63.33 \| 26.23 \| \| LCGen (Ours) \| ✔ \| ✔ \| ✔ \| ✔ \| 20.00 \| 28.94 \| ## 2. Quantitative Comparison of Computation Cost with DreamControl We conducted a quantitative comparison of computation costs with Dreamcontrol. Notably, in the text-to-3D process, Dreamcontrol consists of two stages: Stage 1 - 3D Self-Prior Generation and Stage 2 - Control-Based Score Distribution. However, only the code for the latter has been released in their official codebase, while the former remains on their to-do list. Using Stage 2 alone cannot directly perform text-to-3D generation when given a text prompt. According to the official instructions, we need to provide an obj file or a threestudio checkpoint as a condition during the inference stage. Therefore, in the quantitative computation process for text-to-3D, we need to consider the costs of both stages (baseline+stage 2 or stage 1+stage 2). We used prolificdreamer as the baseline model for both methods, since it is one of the best text-to-3D methods currently. We conducted our experiments on a single NVIDIA RTX A6000 GPU with the max steps set to 10,000. It is worth noting that the following data were all calculated during the text-to-3D generation (i.e., generating the corresponding 3D representation from the given text, not include pre-training time), which can be considered as the inference time in other methods, since the trainable params are the 3d representation we need. The quantitative comparison of computational overhead in text-to-3D task for one sample is as follows: \| Method \| Trainable params in text-to-3D ↓ \| Total estimated model params size ↓ \| GPU memory usage ↓ \| Text-to-3D Generation Runtime ↓ \| \|------------------\|---------------------\|--------------\|----------------\|-----------------------\| Prolificdreamer (Baseline) \| 15.1 M \| 60.384 M \| 28366 M \| 1h30min38s \| DreamControl (Stage2) \| 17.6 M \| 70.422 M \| 35554 M \| 1h55min21s\| DreamControl (Baseline + Stage2) \| 32.7 M \| 130.806 M \| 35554 M \| > 3h25min \| LCGen (Ours) \| 15.1 M \| 60.384 M \| 31458 M \| 1h35min54s \| It can be observed that compared to DreamControl, our method: 1) Our method does not require two-stage processing, and the model parameter count, and runtime are reduced by more than 50% compared to DreamControl, GPU memory usage is also lower than DreamControl. 2) When comparing DreamControl (Stage 2) and LCGen (Ours), our method also performs better than DreamControl in all computational overhead metrics. 3) According to the Table in response 1, the performance of our method is not inferior to DreamControl across all metrics. Thank you again for your efforts and suggestions! We will include detailed discussions on the above points in the Main Paper and Appendix. --- Rebuttal Comment 1.1: Title: Response to Author's response Comment: I read other reviews and the authors' responses. It seems the authors addressed my issues. Other reviews now also have positive scores.	Summary: This paper presents a method to tackle the issue of the Janus Problem in text-to-3D content generation method. A method named LCGen has been proposed that focuses on low certainty regions to generate view-consistent generation. Strengths: Some causes of the Janus Problem have been analysed visually. We then introduce LCGen method is proposed to guide text-to-3D generation toward spatial consistency by establishing varied certainty priors across viewpoints. Proposed method works without adding extra data requirements, and excessive computational overhead. Weaknesses: Proposed method encourage single head, but there could be scenarios where multiple heads needs to be generated e.g. three corgis facing different directions. What about those examples where there is no concept of head. How this method will behave in those situations e.g a street lamp, tree etc. It seems that the proposed low uncertainty approach would lead to low quality samples where fine-grain details are important. There is no comparison with the existing approaches mitigating Janus problem. A table in the main paper with metrics CS and Janus Rate is a MUST. I acknowledge Table in the appendix G, where only functional level comparison is given. Quantitative comparison is missing. Technical Quality: 1 Clarity: 2 Questions for Authors: Please see the limitations and weakness sections, answer those questions. Why is there a blue shade in the generated corgi in your method in fig 7 and 10 ?. Is this some kind of artefact due to low uncertainty generation ? Confidence: 5 Soundness: 1 Presentation: 2 Contribution: 2 Limitations: It seems the method will work on single object generations. It is also not clear how good this method is compared to the existing method quantitatively. Examples used for visual validation are also limited corgi and pig, It is not clear which 30 prompts are selected from the library for the experiments. No clarity on the dataset point of view. A clear selection of data samples MUST be added for reproducibility and validation of the proposed method. Many generated samples in fig 10 are not even of pig. Only one failure case is shown, while there could be many other scenarios, a detailed analysis of failure cases is also missing. It is important to demonstrate the limitations within which method will work. While the evaluation is mostly based on the visuals or derived metrics from visuals. The main paper and appendix both doesn’t have sufficient diverse visual results to validate the proposed method Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your efforts in reviewing our paper! Below is our response. For Figure and Table, please see Rebuttal File. ## 1. Diverse Examples The main purpose of our method is to help address Janus Problem appearing on a single object in text-to-3D baselines. In the *limitations section of the Main Paper* (lines 291-294), we mention, "Our method performs well in generating individual objects but has limitations with complex multi-object scenes." We will provide more detailed explanations in the limitations section and include the following examples: 1. Examples without concept of heads. In Rebuttal File Fig.A and B, we show the results of 'a street lamp' and 'a tree' as you mentioned. It can be seen that both the original Prolificdreamer and our method achieve spatial consistency generation. Considering that the above two examples do not have obvious differences between different views, we conducted experiments on 'a sunflower' and 'a piano' as shown in Fig.C and D in Rebuttal File. It can be seen that the original Prolificdreamer produced spatial inconsistencies, showing multiple frontal images of the flower and multiple keyboards in single piano from different views. Our method successfully alleviated this issue, generating a sunflower and piano with correct front and back images. This indicates that our method is effective not only for objects with heads but also for spatial consistency modeling of other objects. 2. Multiple objects with multiple heads.** We will supplement the discussion details in the limitations section of the Main Paper. Multi-object generation is another important and challenging field in text-to-3D tasks. For SDS-based methods, the current most important issue in multi-object generation is not the Janus problem but how to handle the relationships between different objects. In *Fig.H in Rebuttal File*, we show the results of 'three corgis facing different directions.' It can be seen that the original Prolificdreamer cannot correctly generate 3 corgis, and there is some degree of sticking between the objects. Both our method and the current state-of-the-art multi-view consistent generation method DreamControl cannot correctly handle this example. Once the multi-view generation issue is resolved, our work will have more exploration possibilities. For example, a current mainstream idea is to generate individual assets for each object and then stitch them together[1]. Our method still has the potential to adapt to such methods. ## 2. Generation Quality 1) Regarding the blue shading you mentioned, this is an normal process handled by prolificdreamer during the generation. As shown in Fig.E in Rebuttal File*, the original Prolificdreamer can also produce blue shading when changing seed, and our method can generate results without blue shading. 2) As shown in the Main Paper, our method does not suffer from a loss in generation quality compared to baseline methods and can improve the CLIP-Score. Low certainty generation helps the model find an optimization direction that better aligns with specific views, while the fine-grain detail of the generated images is determined by the generation model rather than certainty. Additionally, by aiding in view consistency modeling, the final generation results exhibit spatial consistency and better overall quality. 3) We also compared some fine-designed prompts, as shown in Fig.F and G in Rebuttal File*. Our method ensures high-quality generation even when dealing with finely designed objects. ## 3. Quantitative Comparison Please see the Table in Rebuttal File. It presents a quantitative comparison with other methods dealing with the Janus Problem. It can be seen that our method, while having a series of advantages, surpasses existing methods in various metrics: it achieves the best overall performance when integrated with Prolificdreamer, strongly suppresses the Janus problem, and maintains the highest CLIP-score; when integrated with Dreamfusion, it has the best suppression effect on the Janus problem. ## 4. Prompt Library Due to constraint of 6000 max length, please see Prompt Library in General Rebuttal to all reviewers*. We present 30 prompts selected from the Dreamfusion prompt library. Since our focus is on the Janus Problem, we have specifically chosen prompts that can evaluate method's effectiveness. ## 5. The Pig Generated by Stable Diffusion All the samples in Fig.10 in Main Paper* are generated by Stable Diffusion 2.1 base (not our method) to demonstrate that Stable Diffusion 2.1 base has view biases during generation. This supports our analysis and thus introduces our method. In some samples, the pig is not generated correctly, reflecting the limitations of the Stable Diffusion pre-trained model. We generated all the images at once on Stable Diffusion 2.1 base and evaluated their views without any selecting to ensure the accuracy of the presented results. ## 6. Failure Cases Due to constraint of 6000 max length, please see Limitation Discussion in respone 2 for reviewer wiP5. Thanks very much. Thank you again for your efforts and suggestions! We will include detailed discussions on the above points in the Main Paper and Appendix. ## Reference [1] Zhou, Xiaoyu, et al. "Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting." arXiv preprint arXiv:2402.07207 (2024).	null	null	Rebuttal 1: Rebuttal: Thank you to all the reviewers for their hard work! We are very honored that our work has been recognized for: 1) significant contribution to addressing key challenges, 2) being well-motivated by a detailed analysis of root causes, 3) simplicity and effectiveness, and 4) good experimental results. ## Contribution Restatement Here, we would like to restate our core contributions: 1) We modeled and analyzed the root causes of the Janus Problem in SDS-based text-to-3D, and designed the LCGen (Low Certainty Generation) method to alleviate the Janus Problem in single-object generation. 2) Our method can be directly embedded into existing SDS-based text-to-3D methods, effectively alleviating the Janus Problem without compromising generation quality and with minimal computational cost. ## Main Questions The reviewers' questions mainly focused on two aspects: 1) Quantitative comparison with other methods dealing with the Janus Problem. We conducted a quantitative comparison with the methods mentioned in the paper, and the specific results can be found in the table in the rebuttal file. It can be observed that our method outperforms others: it achieves the best overall performance when embedded in prolificdreamer, significantly suppressing the Janus Problem while maintaining the highest CLIP score; our method with dreamfusion achieves the best suppression of the Janus Problem. Additionally, our method does not require the introduction of prior knowledge, multi-stage training, or fine-tuning, and achieves good results with minimal additional computational cost (see response to Reviewer FWvW). 2) Performance in diversified examples. Our method aims to alleviate the single-object Janus Problem in SDS-based text-to-3D. We also conducted experiments on various other examples, including objects without heads, fine-designed text, and multiple objects, as detailed in the figures in the rebuttal file. The baseline method is ProlificDreamer, since it is one of the best text-to-3D methods currently. Our method can achieve spatially consistent results when dealing with objects without heads and fine-designed text. For the Janus Problem of multiple objects, as mentioned in the limitations section of the Main Paper, this is constrained by the text-to-3D multi-object generation capability, which will be a direction for our future research. ## Prompt Library "a bald eagle carved out of wood", "a beagle in a detective's outfit", "a beautiful rainbow fish", "a bichon frise wearing academic regalia", "a cat with a mullet", "a ceramic lion", "a chihuahua wearing a tutu", "a chimpanzee holding a peeled banana", "a chimpanzee looking through a telescope", "a confused beagle sitting at a desk working on homework", "a corgi taking a selfie", "a cute steampunk elephant", "a DSLR photo of a baby dragon drinking boba", "a DSLR photo of a cat wearing a bee costume", "a DSLR photo of a corgi puppy", "a DSLR photo of a dog made out of salad", "a DSLR photo of a frog wearing a sweater", "a DSLR photo of a humanoid robot using a laptop", "a DSLR photo of a lion reading the newspaper", "a DSLR photo of a mouse playing the tuba", "a DSLR photo of a pig playing a drum set", "a DSLR photo of a robot dinosaur", "a fox playing the cello", "a highland cow", "a lionfish", "a pig wearing a backpack", "a red panda", "a tiger playing the violin", "a zoomed out DSLR photo of a baby dragon", "a zoomed out DSLR photo of a monkey riding a bike" We sincerely thank all the reviewers once again. For the detailed responses to each reviewer's questions, please refer to the individual replies. Pdf: /pdf/d89c8059d1845d7df276f505869d834e272f8158.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Continual Learning with Global Alignment	Accept (poster)	Summary: This paper tackles continual learning by addressing interference between tasks. For the interference between different tasks, the authors propose a method called ‘global alignment’ to align the data representations using task-specific compositions of pre-trained token representations. Then the authors conduct extensive experiments to verify the effectiveness of the proposed method. Strengths: 1. This paper is well-written and easy to understand. 2. The motivation is clear. From the perspective of cross-task interference, this paper provides some simple but effective methods to avoid the interference of the data representation and classifier. 3. The proposed methods are easy to follow and experiments verify their effectiveness. Weaknesses: 1. The analysis of interference in Section 3.2 does not consider the activation function between the two layers of the network. Multiplying the two linear weight matrices is equivalent to using only one weight matrix, so the analysis of this linear case is different from that of the non-linear case and I’m curious about the inference in the non-linear case. 2. About wiring with neighbor attention: a) The author needs to explain the rationality of neighborhood tokens and the impact of the size of K on the model. b) This method requires matching the most similar K tokens for each token given a sample. This process may require a large amount of computation. c) The author needs to further explain the difference between this method and controlled-LoRA. Both these two methods add a learnable term to the original token representations but adopt different generation strategies. Specifically, this method adds a term composed of neighborhood tokens, while controlled-LoRA uses low-rank matrices. So what are the advantages of this method? 3. About Controlled-LoRA: a) Computational burden and parameter size: As the number of tasks increases, linearly increasing model parameters is not in line with the spirit of CL, and the increased parameter size should be limited even if LoRA only adds small parameters compared to the large model. b) Why does Controlled-LoRA perform worse than other methods? In the case of task-IL, given the task ID for each test sample, we can directly use the best LoRA model. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to Weaknesses. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and feedback. We address your concerns as follows. ___ W1. The analysis of interference in Section 3.2 does not consider the activation function between the two layers of the network. * We consider the ReLU activation which is widely used in neural networks. After ReLU, for each layer we have the output ${\bf{h}}^{(l)} = {\bf{D}}^{(l)}{\bf{W}}^{(l)}{\bf{h}}^{(l-1)}$. ${\bf{D}}^{(l)}$ is a diagonal matrix with the $i$-th diagonal value as 1 if the $i$-th element in ${\bf{W}}^{(l)}{\bf{h}}^{(l-1)}$ is positive, and otherwise 0. Then we have $\frac{\partial {\bf{h}}^{(l)}}{\partial {\bf{W}}^{(l)}} = {\bf{D}}^{(l)} \otimes ({\bf{h}}^{(l-1)})^T$ and $\frac{\partial {\bf{h}}^{(l)}}{\partial {\bf{h}}^{(l-1)}} = {\bf{D}}^{(l)}{\bf{W}}^{(l)}$ where $\otimes$ is the Kronecker product. We can calculate the gradient of each ${\bf{W}}^{(l)}$ based on the chain rule. * However, adding the non-linearity will complicate our analysis and also make the results hard to understand. Since it's typical to remove activation analysis for a first-pass analysis, and we do not believe the nonlinearity significantly changes the observations, we did not include it in the paper. ___ W2.a. The rationality of neighborhood tokens and the impact of neighbor size k. * Why using neighborhood tokens: Based on Eq. 6, our model wires the pre-trained representations of input tokens to learn a task, and thus its task-learning ability depends on the range of input tokens. However, the task information may not be limited to the information of input tokens, but also their neighbor tokens. For example, in a text entailment task, give a sentence pair: {s1: the boy is crying; s2: he’s happy about the view.} with the label ‘contradiction’. The pre-trained representations of ‘crying’ and 'happy’ may not be negatively correlated, which makes the model hard to learn their contradiction. However, ‘crying’ has a neighbor token 'sad', and pre-trained representations of 'sad' and 'happy' are likely to have negative correlations. Therefore, introducing the information of 'sad' may make the model easier to learn the task, and thus enhance its task learning capacity. * The impact of the size of K: Please see the general response 3. ___ W2.b. The computation cost of Wire-Neigh. Please see the general response 2. ___ W2.c. Comparison between Wire-Neigh and Controlled-LoRA. Please see the general response 4. ___ W3. Computational burden and parameter size of Controlled-LoRA. Why does Controlled-LoRA perform worse than other methods? * a. Controlled-LoRA does not linearly increase model parameters with tasks. Controlled-LoRA adds low-rank query and value matrices as LoRA does. These low-rank matrices are shared across all tasks. We agree with your point that it is not desirable to increase model parameters when tasks grow. And that’s why we use shared parameters across tasks, and focus on alignment methods to reduce destructive interference and forgetting. * b. Controlled-LoRA still has forgetting because we do not progressively store parameters for each task. And it performs worse than wiring models because by learning both query and value matrices, it has weaker alignment effects. This can be compensated by applying PF (probing and then fine-tuning). After adding PF it outperforms wiring models in Table 1. --- Rebuttal Comment 1.1: Title: Thank you for the rebuttal Comment: Thanks for answering my questions. After going through all the other reviews and the given rebuttal, I've increased my score to 6. W1. I understand the difficulty of the non-linearity analysis and thanks for the authors' explanation. W2. The comparison between Wire-Fixed, Wire-Neigh, and C-LoRA is clear, and I suggest adding this part to the main text for better understanding. W3. I'm sorry for the misunderstanding about C-LoRA and my concerns have been addressed. --- Reply to Comment 1.1.1: Title: Thank you for the response Comment: We thank the reviewer for the thoughtful response. ___ W1 and W3. Thank you for your understanding. W2. Thank you for the suggestion. We will add this part to the main paper.	Summary: In Continual Learning (CL), the interference caused by the constant modification of the representation is the leading cause of catastrophic forgetting. Motivated by this and the idea that gradients in opposite directions are one cause of this interference, the authors proposed new ways of adding knowledge to a model. The authors motivate the proposal with a toy model that exemplifies where and why the interference occurs, concluding that it comes from the discrepancy between the representations of the different tasks and the relationship between the class vectors. These conclusions lead the authors to propose three proposals that slightly modify the transformer architecture, specifically adding learnable weights in the self-attention layers to generate a task-specific attention representation. The authors also propose using the probing and then fine-tuning approach presented in the past to help initialize class vectors. The results are presented in task and class incremental learning benchmarks in sequences of text datasets. Strengths: - The authors present a motivation from which they exemplify the problem in a simple way and where they want to aim their solution. This motivation may help the reader understand the context and problem to be solved. - The paper introduces a method that alters the architecture of a transformer model. While the paper does not explicitly state this, the provided motivation helps to understand why certain parts of the self-attention layer (the k matrix) are modified. - The authors must describe why only the K matrix is modified and not the others. Weaknesses: - What is the actual contribution of applying wiring or C-LoRA? The results show that applying FP can benefit the methods much more than adding the proposals presented. - What are the implications of applying FP to the Adaptation or CL methods? The performance gain from the proposed methods is relatively low (compared to L2P), and this could further decrease with the application of FP. This requires a deeper exploration of the proposed methods and the alternatives. - The writing is unclear and often unnecessarily complex, making reading difficult for someone unfamiliar with the subject. The notation used is only sometimes in line with the literature, which can confuse readers. In addition, there are problems explaining some terms and easily avoidable problems. For example: - Line 46, FP, is mentioned but needs to be correctly defined. However, the document defines the abbreviation FP multiple times after that when only one should be enough. - Line 182, there is an extra “192”. Technical Quality: 2 Clarity: 1 Questions for Authors: - My main concern is the causes of interference that motivate this work. The author hypothesised that a gradient in the same direction can alleviate forgetting between tasks, a concept also adopted by GEM, despite its susceptibility to forgetting. Even with a gradient in the same direction, if it induces significant weight changes, the representations could be compromised, leading to catastrophic forgetting. - How much does having a model pre-trained in tasks similar to those used in CL affect the proposed methods? The gradient will likely be low for these tasks. - Text benchmarks need minor modifications to the model weights due to the similar distribution between the pre-trained model and the new tasks. In cases where higher modifications are needed, can the scaling factor (s) help reduce this effect? Did you explore different values of the scaling factors? - Can this method work on image benchmarks? Images can have a more complex distribution than text, making avoiding interference between tasks difficult. - Is FP applied to CL or adaptation methods in Table 1? - Another technique used in CL is freezing class vectors in the classifier so that only the classes present in the batch during training are modified. This is especially useful when there is no memory. Can this help further reduce interference in the proposed methods? - Fig 2.b, what method of global alignment was used? - Three methods to mitigate interference are presented. Can these methods be complemented by each other? - How many neighbours are used in Wire-Neigh? Did you perform experiments to find the optimal number of neighbours? - Did you train only the classifier for the Task Incremental Learning problem? This method should not forget, and the pre-trained model can have good prior knowledge. - For the CIL experiments, the proposed methods have very similar results to previous methods when applying FT in Yahoo. This somewhat contradicts the claim that the proposed methods mitigate forgetting. Do you have any intuition about why FT+PF works so well? - In Fig 4.b., is the performance with or without FP? Confidence: 4 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: As the authors mention, this method heavily relies on a pre-trained model with a similar distribution to the incoming tasks. This assumption is only sometimes true and can be more complex in scenarios with a more complex input data distribution. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and feedback. We address your concerns as follows. ___ W1. What is the actual contribution of applying wiring or C-LoRA? * As shown in Eq. 1, the interference depends on two factors: (1). Correlation between representations. The wiring and C-LoRA models are designed to address this, by learning aligned data representations with the reference of pre-trained token representations. (2). Correlation between class vectors. Probing and then fine-tuning (PF) are used to address this. Both alignment models and PF are our proposals to address interference in CL. * We respectfully disagree with the point that ‘applying PF can benefit the methods much more than adding the proposals presented’. Without PF, wiring models can already achieve superior performance (Table 1). PF is especially helpful when the alignment between representations is not strong, as discussed in paper l327 -l330. * About the comparison to L2P, almost all of our models, with and without PF, achieve better performance than L2P in Table 1. If we misunderstood your point, please let us know and we will be happy to discuss it. ___ Q1.a. Gradient direction and significant weight changes. * The interference problem which focuses on the direction of gradients is fundamental in CL, as studied by many previous works like GEM you mentioned. The weight change is another important problem. We believe these two do not contradict each other. * The weight changes depend on both the learning rate and the gradients. In our experiments, models with similar structures (e.g. models with pre-trained encoder frozen) are tested with similar learning rates for a fair comparison. ___ Q1.b. Effect of having a model pre-trained in tasks similar to those used in CL. * Pre-trained and downstream CL tasks: In pre-training, models are pre-trained in a self-supervised manner (masked language modeling), which does not have implications for any supervised task used in CL. * Gradient will likely be low: In downstream tasks for CL, the task data usually have different distributions from that in pre-training (e.g., detecting text entailment of sentence pairs), and the classifier for the task has to be learned from random initializations. Both of these will make the gradient not low (i.e., the loss is not small) at the beginning of tuning. * Use of pre-trained model in our work: Our work does not hypothesize the similarity between pre-training and fine-tuning tasks. On the other hand, we use pre-trained token representations as a basis/reference, and the models have to learn task-specific attention to them (with the randomly initialized key matrix). A better pre-trained model can help if it learns true correlations between pre-trained token representations, which will provide better pre-trained token representations for better alignment effects in our models. ___ Q1.c. In cases where higher modifications are needed, can the scaling factor(s) help reduce this effect? We assume the ‘higher modifications are needed’ means we need more plasticity to learn every single task. In this case, we can increase the scaling factor in Controlled-LoRA or expand the range of neighbor tokens in Wire-Neigh. The scaling factor controls the balance between the alignment effect and the models’ tuning capacity. If the datasets are hard and we need more plasticity, then increasing the scaling factor will help. However, when the scaling values become too large, the alignment effect will be reduced and the model may have more risk of forgetting as well. For comparison, please refer to the results of C-LoRA (with the scaling factor 0.1) and LoRA (with the scaling factor 1) in Table 1. Although LoRA achieves better single-task performance than C-LoRA, it also forgets more in our CL experiments. ___ Q2. Is FP applied to CL or adaptation methods in Table 1? * No. Since PF is proposed based on our analysis in Eq. 1, applying it to CL is also a part of our proposed models. So we did not apply it to other baselines. * Modify only the classes in the batch: Our standard training process has already used the mentioned strategy. In our training, for each task, only the class vectors of classes presented in that task are trained. For each batch, data are randomly sampled over all classes in the task. Besides this, our proposed model can achieve additional improvements in CL by the global alignment design and PF. ___ Q3. Fig 2.b, what method of global alignment was used? Sorry for the confusion. Fig 2.b is plotted under the wiring model with neighbor attention. ___ Q6. Did you train only the classifier for the Task-IL? Yes. We have provided this result in the paper, row for ‘Probing’ (above adaptation models) in Table 1. The pre-trained model has good prior knowledge in DB, but the prior knowledge is not sufficient for Yahoo and News Series as there are large performance gaps between MTL and Probing. And all of our methods outperform the probing performance in Table 1. ___ Q7. Why FT+PF works so well? * As shown in Eq. 1, both correlations between data representations and class vectors will affect the interference. And PF can reduce the interference caused by class vectors. This may be the reason that using FT + PF works well, as discussed in l323 - l330. * Why Alignment models + PF have similar results to FT + PF: for CIL on DB, alignment models outperform FT + PF. For CIL on Yahoo, models may need more separation ability in each task's learning. Since FT has a strong separation ability for each task, it may have potential in CIL with PF. Similarly, C-LoRA has more single-task capacity than wiring models and also has alignment effects, so it achieves the best results in CIL on Yahoo. * Fig 4.b.: The performance is without PF, since wiring and C-LoRA themselves perform well in Fig 4.a. ___ For questions Q4 and Q5, please refer to the general response 4, 3. --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed answers. W1. Yes, I am sorry about my error. I associate IDBR's results with L2P, and because IDBR uses memory, I understand that their comparison is not completely fair. For Q1.a, I was referring to the gradient in the model. I understand that the gradient in the classifier can be high, and it should be high to learn the downstream task. However, some techniques have been proposed to mitigate the classifier's forgetting. For example, fix the columns of those classes not present in the batch (to mitigate the interference at the gradient level). Q1.c. Yes, this is exactly what I was referring to. If the task distribution is very different, the strategy must modify the model weights, increasing the scaling factor. I understand that this can increase forgetting, but it is interesting to understand how much the scaling factor can move to learn a very out-of-distribution downstream task but mitigate forgetting. This is the main challenge in CL when using pre-trained models. Q.6. Thanks. I didn't see it. Would it be a good idea to add something in the first column? Due to this, I will increase my score to a 5. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the thoughtful response and detailed explanation. ___ W1. Thank you for your understanding. ___ Q1.a. We agree that fixing the columns of classes not present in the batch can mitigate interference in the classifier. Under this setting, updating the columns of class vectors from new tasks will not influence the class vectors learned in the previous tasks (i.e., zero gradient for class vectors of previous tasks). As stated in the rebuttal Q2, we have already used this technique in training all our models, including baselines. However, this technique can not fully mitigate interference in the encoder. Since the encoder is shared across all tasks, the value of new class vectors will influence the gradients on the encoder (Eq. 1), in both the magnitude and direction. Since the loss is usually non-zero at the beginning of task learning, the gradients of the encoder are unlikely to be zero. This will cause interference based on the encoder’s gradients, on which we focus in this paper. If we have a good pre-trained model that can solve the target task by only tuning the classifier, then the gradients on the encoder are zero (i.e., the loss is zero) after probing. However, that is not always the case as shown in the probing results of Table 1. ___ Q1.c. Thanks for your suggestions. We show the Task-IL accuracies of different scaling factors $s$ for controlled-LoRA (C-LoRA) and Wire-Neigh in the tables below. We test on the News-Series sequence, where the tasks are from nature language inference (NLI) datasets that have different data distributions from pre-training. The probing performance on News Series also has a large gap to MTL, which indicates that models need more plasticity to solve the tasks. \| Model \| s = 0 (Probing) \| s = 0.1 \|s = 0.4\|s = 0.7\|s = 1.0\| \|-\|-\|-\|-\|-\|-\| \| C-LoRA \| 74.81 \| 74.83 \|72.99\|71.02\|69.59\| \| C-LoRA + PF \| 74.81 \| 78.59\|77.41\|76.83\|76.81\| * For C-Lora, when $s$ increases, the model's global alignment effect decreases while the plasticity increases. The results suggest: (1) C-LoRA’s CL performances decreases when $s > 0.1$. This may be because the encoder loses the global alignment effect, which also misleads the learning of the classifier and increases the interference. (2) After applying PF, C-LoRA’s accuracy first increases and then slightly decreases, overall consistently outperforms probing. This may be because PF reduces the interference caused by the class vectors, and the model can fully utilize its global alignment ability when increasing plasticity. However, when the scaling factor goes too large, the loss of alignment ability will lead to more forgetting even with PF. \| Model \| s = 0 (Wire-Fixed) \| s = 0.1 \|s = 0.3\|s = 0.5\| \|-\|-\|-\|-\|-\| \| Wire-Neigh \| 76.28\|77.20 \|75.19\|72.46\| * For Wire-Neigh, we can improve its plasticity by expanding the neighborhood (general response 3), or increasing the scaling factor $s$ that represents the interpolation between the pre-trained token representations and their neighbor representations. The observation is similar to C-LoRA: when $s$ goes up, the model’s accuracy first increases and then decreases because of the trade-offs between global alignment and plasticity. ___ Q6. Yes, we will mark the probing as the classifier-only learning baseline in the first column.	Summary: This paper addresses the problem of Task-Incremental-Learning (TIL) with pre-trained transformer in the context of NLP. The author extended their experiments in the Class-Incremental Learning (CIL) scenario. The authors identify potential forgetting causes as (1) negative correlation between data representations and (2) negative correlation between class prototypes. After theoretically justifying their claim, the authors propose to align the model representations with either a) learning new key matrices for pre-trained queries and values b) learning new key matrices for with additional token chosen as the nearest neighbors of the considered token c) learning new queries and values from the original ones with a low-rank adaptation strategy. The author additionally align prototypes (class vectors) leveraging a probing and fine-tuning strategy. This paper shows superior performances on various dataset in CIL and briefly discuss the effect of various components. Strengths: - clear and understandable paper - the equations are well derived and clear - the proposed approach has interesting theoretical justifications - The obtained performances on the TIL settings are compelling Weaknesses: 1. While I believe the evaluation to be sufficient, it would improve the paper to include more recent prompt-learning techniques such as CODA [1] which showed stronger performances than L2P. I believe it would equally be interesting to see this approach applied to Computer Vision problems, even though I understand this paper focuses on NLP. 2. Although the proposed approach is theoretically justified, it would be interesting to quantify such alignment through experiments, by computing data representations covariance/correlation between transformed class vectors; for each alignment strategy. 3. I appreciate the effort to include results in CIL scenarios. However, I think more discussions as to how the proposed approach performs poorly compared to ER-ACE could be introduced. 4. What is the link between the findings of this paper and previous work on orthogonal subspaces in continual learning [2, 3]? In these works, amongst others, it seems that the correlation between hidden representation should be zero, and not positive. I think such work should be included in related work section. 5. A discussion regarding the extra computation induced by the alignment strategies and the PF would be welcomed as well, as it seems to increase it considerably. 6. The code is unfortunately not accessible to the reviewers. If the authors can address most of the above points I would happily increase my score. Typos and presentation - LoRA l44 and PF l.46 are not defined. Please either define the acronyms or cite the corresponding paper. - l.114 $h_{\tau}$ should be h_{i} - I do not like the use of RHS l.163 and 164. - l. 182: "grounded192"? [1] Smith, James Seale, et al. "Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [2] Wang, Xiao, et al. "Orthogonal subspace learning for language model continual learning." Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10658–10671 [3] Chaudhry, Arslan, et al. "Continual learning in low-rank orthogonal subspaces." Advances in Neural Information Processing Systems 33 (2020): 9900-9911. Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Limitations have been partially addressed in the main draft. I believe a discussion regarding the poor performances on CIL scenarios should be included, as well as a discussion on the potential computation overhead. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and feedback. We address your concerns as follows. ___ W1. Comparison to more recent prompt-learning techniques and the application to CV tasks. * Comparison to CODA: we show CODA's average accuracy on Task-IL below. Since CODA's hyperparameters are set for CV tasks, the recommended settings may not be optimal for our NLP evaluation. This may be the reason that CODA does not outperform L2P in our case. However, after finding the optimal NLP hyperparameters CODA may perform better than L2P as shown on CV tasks. We leave this into our future work. And we will add more recent prompt-learning techniques to our related works. \| Model \| DB \| Yahoo \|News Series \|-\|-\|-\|-\| CODA\|98.67\|88.05\|75.07 L2P\|99.63\|90.82\|73.99 * Application to CV tasks: please see the general response 5. ___ W2. Quantify alignment through experiments. * Representations' correlation between transformed class vectors: By evaluating our Task-IL models in the Class-IL evaluation, the Class-IL results show the correlation between learned representations and class vectors. If data representations and class vectors are not correlated well across tasks, then models may fail to assign data representations to corresponding classes (across all tasks). Our models have improvements in Class-IL as well, which suggests they have learned representations correlated to all class vectors. * Representations' correlation in the pre-trained space: To further evaluate the correlation between data representations, we decode representations to the token space using the pre-trained decoder. If data representations are well correlated and guided by the pre-trained representations, they should be decoded to tokens that are related to tasks, as shown in paper Table 2. We quantify the model's alignment ability using E-SNLI [1] data, where the data's task-related tokens are highlighted by human annotators. We calculate the Recall@20 of task-related tokens decoded after training on News Series and SNLI (single task). The results are shown below. \| Model \| SNLI \|News Series \|-\|-\|-\| FT\|6.80\|7.74 C-LoRA\|14.53\|23.13 Wire-Fixed\|37.01\|27.80 Wire-Neigh\|36.24\|32.32 Results suggest that wiring models have more alignment ability than C-LoRA, in both in-task (SNLI) and CL evaluations on similar NLI tasks (News Series). And Wire-Neigh has a better alignment ability in CL evaluations. ___ W3. How the proposed approach performs poorly compared to ER-ACE. * ER-ACE replays previous tasks’ data at each training step. This means it is learned to explicitly distinguish classes from different tasks, which is effective for class-IL. However, our methods are only trained in a task-IL manner without experience replay, which means (1). They have no information about classes in previous tasks; (2). They do not have access to previous data so they are not explicitly trained to distinguish classes from different tasks. * We believe it is unfair to directly compare our models to ER-ACE since ER-ACE uses more information for class-IL during training. However, even under this setting, our models perform close to ER-ACE on DB, and achieve good class-IL performance on Yahoo compared to fine-tuning. Their ability to separate representations from different tasks/classes is a result of our global alignment design. ___ W4. What is the link between the findings of this paper and previous work on orthogonal subspaces in continual learning? * Thanks for pointing out these related works. As you mentioned, they focus on getting zero interference between gradients, which minimizes the interference but also reduces the chances of the gradients being in the same direction and enhancing different tasks’ learning. * Our work does not restrict the gradients to be orthogonal. On the other hand, we guide the gradients by the pre-trained representations. If there are two tasks that have positive knowledge transfer between them, models may focus on tokens that have positively correlated pre-trained representations, which make the inner product between their gradients positive and enhance the two tasks’ learning (if the class vectors are not negatively correlated). ___ W5. Extra computation induced by the alignment strategies and the PF. Please see the general response 2. ___ W6. The code is unfortunately not accessible to the reviewers. We are cleaning the code and will send the link to AC soon. Typos and presentations. Thanks for pointing out the typos and the presentation suggestions. We will correct and modify the paper accordingly. ___ [1] Camburu et al. e-SNLI: Natural Language Inference with Natural Language Explanations. NIPS 2018. --- Rebuttal Comment 1.1: Title: Thank you for the rebuttal Comment: I thank the authors for taking the time to answer my question and improve my understanding of the manuscript. W1 I believe it would make it a fairer comparison to include CODA as a SoTA prompt-learning based method after hyper-parameter, since it outperforms L2P on vision tasks, at least for CIL scenarios. I acknowledge that results might differ for the TIL scenario. I still appreciate the effort and I understand that the limited time for rebuttal is not necessarily sufficient to conduct an extensive hyper-parameter search. W2 Thank you for the extra experiments, which are convincing. Including them in the main draft could also clarify the impact of the proposed loss on the model alignment. W3 Thank you for the clarification. I agree that direct comparison to replay-based methods might be unfair. W4 Thank you for the clarification. I would advise the authors to include such discussion in the related work. Time consumption Thank you for the additional information. Such details could be included in the paper or the appendix, as computation is a major focus focus in Continual Learning. Overall, I do not have major concern and I will increase my score to 7. --- Reply to Comment 1.1.1: Title: Thank you for the response Comment: We thank the reviewer for the thoughtful response and helpful suggestions. ___ W1 We agree with that. We are working on the hyperparameter searching on CODA and will include it for comparison in the later version of the paper. W2 Thank you for the suggestions. We will include them in the main draft. W3 Thank you for your understanding. W4 and Time consumption Thank you for the suggestions. We will include the discussions in the paper.	Summary: This paper studies the cause of cross-task interference in class-incremental learning of transformer-based language models. The authors disentangle the cause into the correlation (i) between data representations and (ii) between class vectors in the linear classifier. To tackle (i), the authors propose three ways to construct data representations at each layer by learning an attention over the pretrained token representations. To tackle (ii), the authors propose to only train the classifier for a new task (to obtain a good initialization) before jointly training both the classifier and the encoder. The authors perform experiments with the pretrained BERT-base model and various text datasets. Training is in the task-incremental setup (where task labels are provided and the model predicts over in-task classes); the model is evaluated on both task-incremental and class-incremental (where task labels are not provided and the model predicts among all classes) setups. The authors found that their methods, "alignment models", outperform existing adaptation and continual models. Strengths: 1. This paper is well-written and I did not find major technical flaws. 2. I enjoyed reading the motivation of the paper in Sec. 3 where the authors examine the causes of cross-task interference. 3. I find that the initialization of class vectors may influence cross-task interference interesting, although I have some related questions. Weaknesses: 1. While the main goal of the paper is reducing cross-task interference, the main results (Table 1) is using the task-incremental setup, where task labels are known during prediction. The class-incremental learning results are only in Fig. 4, comparing the proposed method only with LoRA and ERACE. 2. I find it a bit hard to infer where the performance improvement comes from from the results. I wonder if it is possible to do some fine-grained ablation analysis that verifies that the proposed method indeed helps by reducing overlap in data representations and in class vectors' features for different tasks. For example, one could measure the accuracy on the task level or look at the confusion matrix, which may reveal some information about cross-task confusions. 3. Is $\Delta W$ shared across tasks? If so, I don't understand why the proposed method does not forget. The authors hypothesize that this is potentially due to referencing pretrained representations. However, since the [CLS] representations are constructed involving $\Delta W$, I'd expect the model to lose some ability to generate good representations for past tasks. Could the authors elaborate on this point? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the authors explain the choice of training in TIL while evaluating in CIL, as opposed to training and evaluating in the same setup? 2. (Eq. 2) Since weights are usually initialized from a distribution centered around 0, shouldn't the first term be quite small? 3. (Sec 4.3) Does probing perform softmax over only the new classes? If so, how does it help two class vectors for different tasks to focus on different features, since the loss does not require differentiating between the two classes? 4. (Sec. 4.2) In the wiring models, what is the intuition behind only replacing the key matrix but not the query and value matrices? Minor: 5. (Eq. 1) It seems that interference is eliminated once any of the three components is 0. Is it possible to fix the class vectors to be the canonical basis which are guaranteed to be perpendicular, and only learn the data representations? 6. (Sec. 4.2) In Wire-Neigh, do you need to find the $K$ nearest neighbors for each token? If so, how expensive it is in practice? 7. (L#219) Could you elaborate on what "grounded" means here? 8. Missing references: [1, 2] are two prompting-based CL methods that are shown to perform better than L2P. 9. (L#41-44) I recommend changing the indices to, e.g., "(i), (ii), (iii)", as a different "(2)" is referred to in the next paragraph. 10. (L#53) "by" -> "after" 11. (L#182) "grounded192" -> "grounded" [1] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning. Wang et al. ECCV 2022. [2] CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning. Smith et al. CVPR 2023. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitations clearly in Sec. 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and feedback. We address your concerns as follows. ___ W1. While the main goal of the paper is reducing cross-task interference, the main results are using the Task-IL setup. We focus on Task-IL because: * Cross-task interference is a fundamental problem in Task-IL, which focuses on the gradients of shared parameters (i.e., the encoder) across all tasks. If destructive interference happens, the encoder may be updated to generate drifted representations for previous tasks, and the previous classifier may not be able to distinguish them (i.e., forget previous knowledge). Previous works study interference over the Task-IL setting as well, as cited in the paper. * Task-IL focuses on the interference in the shared parameters, while Class-IL further requires distinguishing between classes from different tasks. Although our paper shows they are inherently related, the distinction between cross-task classes is not our original goal. We do not compare with more Class-IL methods because they are explicitly designed to distinguish cross-task classes during training like ERACE, while we only use Task-IL training without information of previous tasks’ classes. However, our method can be combined with those methods to achieve stronger performance in Class-IL. ___ W2. Ablation analysis of proposed models in reducing overlap in data representations and class vectors for different tasks. * In paper Fig. 2, we show a T-SNE plot of data representations on DB sequence for our model and the fine-tuning model. Our model generates non-overlapping representations after the first and last task. However, fine-tuning representations overlap after the first task, and further mix up after the last task. * By evaluating our Task-IL models in the Class-IL evaluation, the Class-IL results show the separation of learned representations and class vectors. If data representations and class vectors are not separable enough across tasks, then models may fail to assign data representations to corresponding classes (across all tasks). Our models have improvements in Class-IL as well, which suggests they reduce overlapping in representations and class vectors. ___ W3. Is $\Delta W$ shared across tasks? Yes, $\Delta W$ is shared across all tasks. The proposed model still has forgetting as shown in experiments. However, the forgetting is significantly reduced by using our global alignment methods, which results in less interference during training and thus less forgetting. ___ Q1. The choice of training in TIL while evaluating in CIL. * For Class_IL, without replay or pre-separating class vectors, training with the loss on in-task classes can reduce forgetting compared to the loss over all classes [1]. The reason is that, if we use the loss on all seen tasks' classes but do not have data from all those classes (since no replay), it will cause data imbalance on classes during training. This may distort previously learned representations and cause forgetting. * As mentioned in W2, this setting can also evaluate the separation of data representations and class vectors learned in Task-IL. ___ Q2. Influence of the first term in Eq. 2 when the weights are usually initialized to be centered around 0。 We write Eq. 2 with the learning rate $\alpha$ as: ${\bf{v}}\_{y_i}^T{\bf{v}}\_{y_j} = {\bf{v}}\_{y_i}^T{\bf{v}}\_{{y_j},0} - \alpha{\bf{v}}\_{y_i}^T\sum\nolimits_{t}\nabla_{{\bf{v}}\_{y_j}} \mathcal{L}({\bf{h}}\_{j, t}, y_{j})$. * Whether the first term is small depends on the class vector learned from the previous task, i.e., ${\bf v}\_{y_i}$. * The influence of the first term also depends on the scale of the second term. Since models are fine-tuned with relatively small learning rates like $\alpha$ = \{1e-3, 1e-4, 1e-5\}, the second term is also 'small'. So the first term can still have a large impact even if the class vectors are initialized centered around 0. ___ Q3. How does probing help two class vectors for different tasks to focus on different features? In the probing stage, we fix the encoder learned from previous tasks and only train the classifier for the new task. Therefore, the encoder generates the new task’s representations based on the previous task’s knowledge. And the class vectors of new and previous tasks can have different focuses on such knowledge, based on different objectives. ___ Q4. Replacing the key but not the query and value matrices. Please see the general response 1. ___ Q5. Interference is eliminated once any of the three components is 0. Is it possible to fix the class vectors to be the canonical basis? * Our goal is to avoid destructive interference when learning across tasks, which encourages both zero and positive interference at adequate times. If we fix the class vectors to be perpendicular to each other, there will be no positive interference that can help positive knowledge transfer. * Also, the perpendicular class vectors may not be good class vectors that prevent feature distortions when tuning a pre-trained model. This will cause the loss of pre-trained knowledge during learning and make models lose generalization on OOD data, which can also make models perform inferior in CL. ___ Q6. Computation cost for Wire-Neigh. Please see the general response 2. ___ Q7. 'Grounded' in L219. “Grounded” means that the correlations between data representations are decided by the correlations between corresponding pre-trained representations $G_i$, $G_j$, and the attention on these representations (learned in tasks). For example, if the two tasks pay attention to tokens that have orthogonal pre-trained token representations ($G_iG^T_j = 0$), e.g. when the two tasks’ information is irrelevant, then the alignment models will generate orthogonal data representations as well. ___ [1] Masana et.al. Class-incremental learning: survey and performance evaluation on image classification. TPAMI, 2022. --- Rebuttal Comment 1.1: Title: Thanks for rebuttal Comment: I thank the authors for answers my questions. My concerns are resolved and I've increased my score to a 6. Regarding W3, I wonder if the reason that a shared $\Delta W$ shows low forgetting could also be due to that it does not contain a lot of parameters. This recent paper [1] shows that PEFT methods enjoy low forgetting when you don't have too many tunable parameters. [1] Thede et al. Reflecting on the State of Rehearsal-free Continual Learning with Pretrained Models. CoLLAs 2024. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for the thoughtful response and helpful reference. ___ Yes, we agree that tuning fewer parameters may help to mitigate forgetting. In the paper Table 1, we observe popular PEFT methods (Adapter, LoRA, Prefix-Tuning) have overall less forgetting than pure fine-tuning (FT). In this paper, we provide an angle of interference that connects PEFT models’ superior performance to a global alignment effect. According to [1], token representations in different PEFT models can be viewed as combinations of the pre-trained token representations and some modification vectors. When the models have limited parameters, the modifications may also be limited (e.g., prefix tuning with a limited number of prompts). This makes the token representations strongly guided by the pre-trained token representations, which has the global alignment effect to reduce interference and mitigate forgetting. However, if models have too few parameters, they may lose the plasticity to learn hard tasks and perform inferiorly in CL. Therefore, we also study ways to reduce forgetting when increasing model parameters/adaptations for plasticity. ___ [1] He et al, Towards a Unified View of Parameter-Efficient Transfer Learning, ICLR 22	Rebuttal 1: Rebuttal: We thank all reviewers for their thoughtful reviews and feedback. We address common questions as follows. ___ 1. The intuition behind only replacing the key matrix but not the query and value matrices in wiring models. * Why not replace value matrices: when replacing the value matrices with $\Delta {\bf W}_v$, the data representation ${\bf{h}}^{(l)}$ turns out to be: $({\bf{h}}^{(l)})^T = ({\bf{b}}^{(l)})^T{\bf{A}}^{(l)}{\bf{G}}^{(l-1)}({\bf W}\_v^{(l)} + \Delta {\bf W}^{(l)}_v)$. Compared to our goal in Eq. 6, this will reduce the guidance of pre-trained token representations, which does not fit our wiring goal. * Why not replace query matrices: since the wiring models only relearn the attention for [CLS] token, for all tasks they only query from [CLS] (other tokens' queries are pre-trained). However, the keys are from all tokens in the input and are different when tokens from different tasks have different distributions. Intuitively, we learn the key matrices applied over all input tokens instead of the query matrices applied on the same [CLS] across different tasks. ___ 2. Extra computation induced by the alignment strategies and the PF. * Computation induced by searching for neighbors in Wire-Neigh: We find $k$ neighbors for each token based on their embeddings at the embedding layer. The time complexity is $O(vn)$ where $v$ is the size of vocabulary and $n$ is the size of input tokens. Then the neighbor representations are updated for each layer, with a complexity of $O(nk^2)$. In practice, since the embedding layer is fixed in our model, for each data instance we only need to find their neighbors once and then store the neighbor indices for iterative training (i.e., for several training epochs). The neighbor selection can also be accelerated by reducing the search space of neighbor tokens, for example, only searching neighbors from frequently used tokens instead of the whole vocabulary. * Computation induced by PF: in the probing stage, we fix the encoder and only train the classifier. This takes less than 40% training time (including LM forward) and 30% GPU memory compared to full fine-tuning. ___ 3. Impact of hyperparameter K in Wire-Neigh. For computation efficiency, we fix the number of neighbors as $k=5$, and randomly select the neighbors from top-$K$ nearest neighbors to control the range of tokens' neighborhood. We show the Task-IL accuracies with different $K$ below: \| Wire-Neigh \| DB \| Yahoo \|News Series \|-\|-\|-\|-\| \| $K=5$ \| 99.86 \| 91.16 \|76.90\| \| $K=20$ \| 99.86\| 90.98 \|77.10\| \| $K=50$ \| 99.86\| 91.16 \|77.20\| \| $K=100$ \| 99.87 \| 91.13\|76.58\| For relatively simple sequences DB and Yahoo, Wire-Neigh under different $K$ has stable performance. However, for hard sequence News Series, when $K$ increases the model has more neighbor information (more capacity) to solve the task, which first improves its CL performance. However, when $K$ is too large ($K=100$), the neighbor information may become noisy, which makes the CL performance drop. ___ 4. Difference between Wire-Fixed, Wire-Neigh and C-LoRA. We illustrate the difference between alignment models based on Eq. 6, where aligned data representations ${\bf{h}}^{(l)}$ are expected to have the form $({\bf{h}}^{(l)})^T = ({\bf{b}}^{(l)})^T{\bf{A}}^{(l)}{\bf{G}}^{(l-1)}{\bf W}\_v^{(l)}$. * Wire-Fixed: keep the pre-trained representations ${\bf{G}}^{(l-1)}{\bf W}\_v^{(l)}$ fixed, only learn the attention $ ({\bf{b}}^{(l)})^T{\bf{A}}^{(l)}$ by the self-attention mechanism with new key matrices, denoted as $SA(\Delta {\bf{W}}^{(l)}_k)$. This is only applied to [CLS] tokens. Then $({\bf{h}}^{(l)})^T = SA(\Delta {\bf{W}}^{(l)}_k){\bf{G}}^{(l-1)}{\bf W}\_v^{(l)}$. Wire-Fixed is strongly guided by pre-trained token representations but may have limited capacity in solving hard tasks. * Wire-Neigh: apply the same wiring strategy as Wire-Fixed but add neighbor representations ${\bf{G}}\_{neigh}^{(l-1)}$ for better capacity. Then $({\bf{h}}^{(l)})^T = SA(\Delta {\bf{W}}^{(l)}_k)[{\bf{G}}^{(l-1)}; {\bf{G}}\_{neigh}^{(l-1)}]{\bf W}\_v^{(l)}$. The advantage of Wire-Neigh is it keeps the guidance of pre-trained token representations but increases the model capacity by expanding the neighborhood. * C-LoRA: adapt both query and value matrices with $\Delta {\bf{W}}^{(l)}_q, \Delta {\bf{W}}^{(l)}_v$ but with a small scaling factor $s = 0.1$. This adaptation is applied to all tokens. Then $({\bf{h}}^{(l)})^T = SA({\bf{W}}^{(l)}_q + s\Delta {\bf{W}}^{(l)}_q){\bf{H}}^{(l-1)}({\bf W}\_v^{(l)}+s\Delta{\bf W}^{(l)}\_v)$, where ${\bf{H}}^{(l-1)}$ is the adapted token representation of the previous layer. Compared to wiring models, C-LoRA modifies both the attention and the value matrices applied to all tokens. It can still enjoy the guidance of pre-trained representations with the small scaling factor $s$, but the guidance may be weaker than wiring models. Meanwhile, it has more task-learning capacity. C-LoRA does not use neighbor information. ___ 5. The application of the methods to computer vision (CV) problems. * We believe our main contributions on alignment models and utilizing PF to reduce interference are general to CV tasks. First, our analysis of interference in Section 3 is general to both NLP and CV. Second, there are effective pre-trained models in CV as well, which can be used for alignment purposes. * We think the keys to applying our model to CV tasks are: (1) how effective the pre-trained model is in providing self-supervised token representations for alignment; (2) how to properly set scaling ratios or select neighborhoods to balance the alignment effect and the task capacity on CV tasks. We will study our models' application to CV in our future works. ___ 6. Typos and definitions in the writing. Thanks for pointing out the typos and the presentation suggestions. We will correct and modify the paper accordingly.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
When are dynamical systems learned from time series data statistically accurate?	Accept (poster)	Summary: The present manuscript concerns the use of neural networks to fit physical dynamical systems of generic kind (including and focusing on chaotic maps). It is shown that adding the information of the Jacobian of the dynamical map in the supervised learning process leads to better performances. Strengths: Originality: the work builds on previous ideas and problems. The original part (as claimed by the authors) concerns adding the information on the Jacobian of the dynamical map to the supervised learning process. The author show that this significantly improves the statistical properties of the points generated by the fitted dynamical systems. Quality: The paper is very well written, well structured and clear. Section 4 is very clear. Significance: I believe that the problem tackled in this work is a relevant one and the results show an interesting perspective on it. Weaknesses: I believe that the most important weakness of the manuscript relies on the limitations of the results. The paper clearly shows that having information on the Jacobian of a dynamical map is essential for reproducing the statistical properties of the generated orbits. However such information is typically lacking. Technical Quality: 3 Clarity: 4 Questions for Authors: 1) I believe that it must be carfully stated that it is assumed that the map F does not depend on time. This is an important assumption which may be relaxed. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The main and severe limitation is that the improvement training method needs the knowledge of the Jacobian of the ground truth dynamical system. It is unclear how this can be estimated from data contain just points in the orbits of the maps and what are realistic settings where this is known. I believe that this is a severe limitation of the work. This does not spoil the result in the sense that the manuscript clearly shows that the information coming from the Jacobian is essential, but on the other hand it seems to me that this information is hard to get. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Results require full Jacobian but it can be approximated with available derivative information in practice You are absolutely right in that the Jacobian is computationally hard to calculate (through AD or finite difference) in high dimensions; thank you and Reviewer vAp3 for pointing this out! This in turn would slow the training process and might make it prohibitive for learning high-dimensional chaotic systems. We emphasize that our primary contribution is toward the foundational question of understanding generalization in the context of learning chaotic dynamics. Our main result deconstructs the effect of the ergodic properties of the true system on the generalization ability of the neural network. To isolate this effect, we consider only a minimal illustrative setting of regression using the common MSE loss. When some Jacobian information (local information about the short-term dynamics and its linear perturbation structure) is added to the loss, surprisingly we find that the physical measure (global statistics) is learned. Our results are aimed at explaining this observation via the mechanism of shadowing. As a minimal setting to verify our results in practice without any confounders, we consider systems in which the full Jacobian can be calculated and used in training. Having said that however, our results lend themselves to many practical training strategies where the Jacobian information is only approximated. For instance, in scenarios where adjoints or other surrogate models for the Jacobian are available, which are common in many scientific applications in the geosciences and aerospace turbulent flows, these can be used as proxies for the full Jacobian matrix. As we mention in response to Reviewer vAp3, who raised the same concern, we are interested in analyzing whether random Jacobian matrix-vector products can yield informative generalization bounds such as Theorem 2. Since the novelty and focus of our present paper is in the theoretical connection between using Jacobian information and statistical accuracy and why this obviates the need for performing generative modeling, we defer such analyses to future work. ### Focus on autonomous systems In this paper, we focus on autonomous systems -- wherein the flow map itself does not depend on time -- which occupy a large class of chaotic systems encountered in nature. As you correctly point out, the analysis and even the background, starting from the existence of a unique physical measure, will completely change when we consider nonautonomous systems such as chaotic systems with control terms and random dynamical systems with chaotic orbits. For these systems, there is a lot of exciting recent development in the theoretical dynamical systems literature (see e.g. this survey article: https://link.springer.com/article/10.1007/s10955-016-1639-0) proving the existence of random physical measures. The next step would be connect the theory of such measures with statistical learning theory. In the present work, we show a path for such a connection in deterministic chaotic systems, deferring random chaotic systems (which are increasingly recognized as being useful models for stochastically parameterized climate systems) to future work. We will add this remark to the appendix in the next revision. Thank you very much for bringing this to our attention!	Summary: The paper addresses the problem that classical ERM training of dynamical systems models often fails to capture invariant measures of the observed dynamics, even when test errors are low. The authors use ergodic theory to explain this failure from a theoretical viewpoint. They further demonstrate that incorporating Jacobian information during training enables much better reconstructions of invariant properties and validate their approach on various neural flow operators trained on several common benchmark chaotic systems. Strengths: - I think the authors study an interesting problem, i.e. when classical ERM for dynamical models can still lead to a reconstruction of invariant properties of the observed dynamical system, as classical ERM (i.e. here simply “one-step-ahead” predictions) comes with benefits of decreased training time and less training difficulties - The authors use dynamical systems and ergodic theory to accompany and explain the practical behavior of several neural flow operators (N-ODEs, RNNs, MLPs, etc.) when training them using ERM on observed dynamical systems - The paper is generally well written Weaknesses: - l. 53 - 56: Indeed there are also connections of training recurrent architectures to model chaotic DS and exploding loss gradients in gradient descent based training [1]. Moreover, recently methods for dealing with these training instabilities have been introduced and proven useful [1, 2]. - There is also a recent paper which discuss (out-of-domain) generalization in dynamical systems reconstruction [3] which might make sense to add to related work, as it also introduces measures to assess generalization with a focus on evaluating invariant measures of the underlying dynamics and addresses the problem that classical ERM framework is not enough to assess generalization in dynamical systems reconstruction - While it is good to see that the Jacobian matching loss improves the reconstruction of invariant measures of the underlying dynamical system, it’s application in practice is fairly limited as Jacobians have to be estimated from data if the ground truth vector field is not known. - I see that the authors want to study the case where training is only perform using “one-step-ahead” predictions (i.e. no unrolling of dynamics), however, I think considering the case of unrolling dynamics as a comparison to the Jacobian matching loss would be helpful as 1) unrolling dynamics is often enough to get a good reconstruction of the invariant measures of the observed attractor and 2) it is easy to apply in practice (no need for extra knowledge of the Jacobians). I.e. I think the more interesting question is whether knowledge of the Jacobians outperforms unrolling of dynamics during training significantly - I think section 4, i.e. dynamic generative models, could be written a tad clearer, i.e. more explicit in terms of what exactly is now assumed to be stochastic in the specific VAE architecture and which part is deterministic. Maybe a Figure demonstrating the architecture setup in favor of moving e.g. proof of Theorem 1 into the Appx. - To me the paper lacks a clear connection of the theoretical findings to the practical real-world setting, i.e. how these insights influence how we should train models to reconstruct real-world dynamics I really like the idea and insights of this paper, but given the Weaknesses above, I do think it lacks actual practical relevance and the novelty is also very much limited to the (seemingly impractical) theoretical findings. Miscellaneous comments: - Figure 2 readability could be improved by removing the legend from within the plots and moving it to the outside (e.g. right, or between the plots), as the legend is the same for both plots - Hyperlink of Table 3 jumps to section 3 - typo l. 192: “shdaowing” [1] [Mikhaeil et al. (NeurIPS, 2022), On the difficulty of learning chaotic dynamics with RNNs](https://proceedings.neurips.cc/paper_files/paper/2022/hash/495e55f361708bedbab5d81f92048dcd-Abstract-Conference.html) [2] [Hess et al. (ICML, 2023), Generalized teacher forcing for learning chaotic dynamics.](https://proceedings.mlr.press/v202/hess23a.html) [3] [Göring et al. (ICML, 2024), Out-of-domain generalization in dynamical systems reconstruction.](https://arxiv.org/abs/2402.18377) Technical Quality: 3 Clarity: 2 Questions for Authors: - how exactly are the empirical measures $\mu$ and $\mu_{NN}$ and their $W_1$ distance computed (for Table 1 e.g.)? - Could the authors include comparisons of the Jacobian matching loss to the more common training method of unrolling the dynamics for at least some time $t$? - How can one draw practical consequences from the theoretical findings? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors do mention limitations of their approach (e.g. that the Jacobian matching loss is hard to implement in practice when ground truth is not known), but do not address how their findings might be used in real-world data settings. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Related work on invariant measures Thank you for sharing these references -- we have added them to Related Work, which has spilled into the Appendix! Since the focus is on deriving dynamics-aware generalization bounds, we consider a minimal regression setup for one-step dynamics, which does not require specialized training methods such as the one introduced in [1]. In future work, it will be interesting to derive bounds similar to Theorem 2 for the sparse teacher forcing in [1] and other RNN/ESN-based training methods, as you point out. Many thanks for reference [3] that we have been remiss to overlook. Our setting is that of unique, ergodic, invariant physical measures in this paper, which is less complicated than multiple ergodic basins considered in [3]. It is reassuring that generalization with MSE loss does not imply statistical accuracy for the non-ergodic case considered in Sec 3.2 of [3]. Another interesting connection would be to specialize notions of generalization (Definition 3.2 specifically) introduced in [3] to the unique SRB/Gibbs measure case. While this notion or the W1 distance can be used to derive ERM problems, the sample complexity of these problems may be large. From the theoretical standpoint, we will extend our shadowing-based results to the non-ergodic case by considering weaker notions of finite-time shadowing in every connected component of the attractor. Different from [3], we demonstrate that even when we study ergodic systems, and sample from the entire state space, the traditional MSE does not yield the correct notion for generalization. ### Practicality of Jacobian loss and implications of theory This is indeed an important point, thank you and Reviewer PvwR for pointing this out! The computation of the Jacobian matrix estimated via automatic differentiation or via finite differences scales quadratically with the problem dimension. However, as illustrated in our numerical results, the sample complexity is indeed much smaller than for directly learning the physical measure via a generative model. Moreover, since the training strategy is an elementary regression on any simple architecture (like an MLP), it can be competitive when compared to more sophisticated training algorithms with recurrent architectures or transformers. In practice, whether to choose simple regression or to resort to transformers or recurrent architectures with innovative training stabilization strategies will depend on the dimension and the dynamical complexity (attractor dimension, number of positive Lyapunov exponents, correlation decay time etc). The computational complexity of complicated models such as transformers during inference time is quadratic in the dimension as well, making an elementary architecture appealing as a surrogate model. Furthermore, our result that adding Jacobian information implies statistical accuracy invites the development of practical supervised learning methods in high dimensions. Even though the theoretical results are derived for $C^1$ generalization, we are currently working on incorporating random Jacobian-vector products into the loss function. Such matrix vector products are easily tractable in several scientific applications where adjoint/tangent equation codes are available (see e.g geophysical turbulence models). One subject of our future work is to evaluate the efficacy of black box Jacobian-vector products in learning the physical measure. ### Unrolling dynamics We have run experiments on the Lorenz '63 system with an unrolled loss function: $\ell_{\rm u}(x) := (1/k) \sum_{t \leq k} \\|v_t(x)\\|^2, $ where $v_t(x) = F^t_{\rm nn}(x) - F^t(x)$ are vector fields on learned orbits. With $k = 10$ timesteps of unrolling time, we see that the [attractor is reproduced well](https://ibb.co/YBKhPjT), but [atypical orbits are still produced](https://ibb.co/dr9bPr6) for random initializations. The learned LEs are still incorrect: $[ 0.905, -0.03, -5.6]$ (the true LEs are :$[0.85, 0, -14.5].$) When the unrolling time $k$ is increased, we still obtain the same results, with even the positive LE being overestimated. Overall, unrolling seems to learn more accurate representations than the MSE model but worse than the JAC model. We also observe that the unrolling time $,k,$ needs to be fine-tuned as a hyperparameter to achieve good generalization; a small perturbation in $k$ can lead to training instabilities. We have these results to Appendix C of our revision. To understand these results, for short times, $t,$ (when compared to the Lyapunov time, $1/\lambda_\mathrm{max}),$ $v_t$ are in tangent spaces along the NN-orbit, $\{F^t_\mathrm{nn}(x).\}$ Now, we note that $v_t$ is the pushforward through the Jacobian/linearized dynamics of $v_{t-1},$ which yields the recursive relationship, $v_t(x) = dF(F^t(x)) v_{t-1}(x) + v_0(F_{\rm nn}^t(x)).$ Thus, the unrolled loss does contain Jacobian information implicitly although it does not enforce the learned trajectory to be close to the true trajectory in $C^1$-distance. This serves to explain, via Theorem 2, why it works better than the MSE (even in some practical climate emulators, e.g., FourcastNet). Thank you very much for raising this important point, which supports the central argument of this paper: adding Jacobian information leads to statistical accuracy. ### W1 distance We use Python Optimal Transport package, which uses the Sinkhorn algorithm. For our 1D distributions, the optimal transport map is analytically the target inverse CDF $\circ$ CDF of source, which is estimated via projected gradient descent in the package. ### Edits Many thanks for a careful reading and suggesting many improvements (implemented in the revision, including adding limitations to sec 7) to the paper! We have rewritten ``Dynamic generative models'': here, the dynamics $\Phi^t_\phi$ is stochastic, while the encoder and decoder are deterministic. This is needed for greater expressivity of the conditional measures. --- Rebuttal Comment 1.1: Title: Re: Rebuttal Comment: > Related work on invariant measures Happy to hear that the authors add the references to the related work and thanks for the additional clarification and connection to their own work! > Practicality of Jacobian loss and implications of theory Indeed, the complexity of computing the model Jacobian using AD/Finite differences is one part, however, did the authors also consider the added complexity by including this Jacobian in the loss function, which means one has to a backward-over-forward AD problem (-> second order differentiation)? Does runtime/complexity still only scale with problem dimensionality squared? My main concern with the practicality is the estimation of the Jacobians from time series data, where no ground truth knowledge is available (I appreciate the geophysics examples and I see that in these cases the Jacobian matching loss might improve things). I really miss any experiments on trying to make this loss more applicable in real-world scenarios (e.g. estimating Jacobians from time delay embeddings of real-world data -> do crude approximations already help learning dynamical invariants?). > Unrolling dynamics Thanks for conducting the additional experiments! That zero/negative LEs are still underesimated when unrolling dynamics is indeed a problem I encountered in practice, too. Hence it is great to see that knowledge of the Jacobians can eliminate this problem. However, I checked the paper again and found that the authors report a ground-truth max. LE for the Lorenz63 system with standard settings of ~0.86 (cf. Table 2). However, to my knowledge the Lyapunov spectrum is given by ~[0.905, 0, -14.57]. I checked this with minimal code using Julia (see below). Do the authors mind to share how they estimated the ground-truth Lyapunov spectrum or whether they used specific literature values? ``` using DynamicalSystems, ChaosTools # Lorenz63 ds = Systems.lorenz(ρ=28.0, σ=10.0, β=8 / 3) # Lyapunov spectrum lyap = lyapunovspectrum(ds, 100_000, Ttr=1000) # ~ [0.905, 6e-6, -14.57] ``` > W1 distance Thanks for the clarification! > Edits I see, thanks! --- Reply to Comment 1.1.1: Title: Jacobian and LEs Comment: Thank you very much for reading through the rebuttal! We really appreciate your interest and excellent questions! > Does runtime/complexity still only scale with problem dimensionality squared? This is a very keen observation -- thank you. We meant that computing the Jacobian (using Finite Difference or AD) has quadratic complexity, but you are absolutely right that training time increases also because of differentiating the model at the training points. During training the Jacobian (with respect to the input state) is differentiated with respect to the parameters of the neural network. The complexity with respect to the dimension of the problem still remains quadratic but the training complexity increases just like a physics-informed neural training; the factor of increase therefore also depends on the complexity of the network. Inference time is the same as a vanilla neural ODE however since only the training is modified. Practically we found that KS system (128-dimensional system) was still training in about 10 hours on a single RTX30xx GPU. > estimation of the Jacobians from time series data, where no ground truth knowledge is available This is exactly the scenario we had in mind for the rebuttal. We can estimate the Jacobian from timeseries data using finite difference as long as the data sample the system frequently (there are classical results known for the impossibility of system identification for longer observation frequency and very noisy observations of chaotic systems). As you correctly point out, there is an error made by finite difference (which is on order of epsilon^2, where epsilon is the step size norm). However, even though our theoretical results only use exact Jacobians, the results with unrolling dynamics, e.g., where only indirect Jacobian information is available, suggest that even inexact Jacobians can help us learn the dynamical invariants better. For instance, even Lyapunov vectors and exponents can be estimated with high accuracy using just finite difference estimates of the Jacobian (this is easy to try by replacing the AD with inexact Jacobians in the code snippet below). But as you suggest, we should test this on real world high-dimensional examples (even ones where the attractor is only known through delay embedding), which we defer to a future work. Our present paper emphasizes on understanding why Jacobian information leads to statistical accuracy. Our main contribution is using shadowing theory to provide an explanation and suggest that there is hope of learning chaotic systems with elementary supervised learning methods, and that these can still learn the invariant measure. About the Lyapunov exponents, small differences (statistical error/noise) may arise depending on the integration time (len(traj_gpu) in the code below), time integration method (RK4 in the code below), and subsequently how we estimate Jacobians. Here is the code snippet we wrote for computing 3 LEs: ``` def lyap_exps(dyn_sys_info, traj, iters): model, dim, time_step = dyn_sys_info LE = torch.zeros(dim).to(device) traj_gpu = traj.to(device) f = lambda x: rk4(x, model, time_step) Jac = torch.vmap(torch.func.jacrev(f))(traj_gpu) Q = torch.rand(dim,dim).to(device) eye_cuda = torch.eye(dim).to(device) for i in range(iters): if i > 0 and i % 1000 == 0: print("Iteration: ", i, ", LE[0]: ", LE[0].detach().cpu().numpy()/i/time_step) Q = torch.matmul(Jac[i], Q) Q, R = torch.linalg.qr(Q) LE += torch.log(abs(torch.diag(R))) return LE/iters/time_step ``` Thank you very much for your questions and please feel free to let us know if anything is unclear! Thank you again for your time and help!	Summary: This paper extends generalization results to models trained on dynamical data, especially Neural ODEs. The paper shows and then attempts to explain why Neural ODE trained without a Jacobian matching term fail to capture physical behaviour even when they have low generalization error. Under a generalization assumption, the paper shows that the Jacobian-matching loss can lead to statistically accurate models. The paper shows a number of experiments that show the behaviour stipulated by the theoretical results. Strengths: The paper is technical but well written and ideas are clearly explained. The mathematical formulation is quite clean and targets an important problem of generalization in ODE models for dynamical data. The explanation of generalization for Neural ODEs is I think novel. Weaknesses: I think the main weakness is assumption 1. There should be some explanation of the difference between C1 and strong C1 generalization. Furthermore, there should at least be some justification of why this should hold and when we could not expect it to hold. Another possible problem I see is this: The paper explains why minimizing loss 3 (with the jacobian) implies statistical accuracy. I don’t think that it properly explains why minimizing loss 2 does not imply statistical accuracy. I would understand why a standard MLP might not match derivatives. But I have a harder time understanding why a neural ODE with low generalization error isn’t able to do that. Some more intuition about this would be useful. Technical Quality: 3 Clarity: 3 Questions for Authors: Related to the above. In lines 264-265 the paper states that C0 generalization is insufficient for learning shadowing orbits. Is it an obvious fact? Further explanation would be useful. Also useful would be a comparison of a standard MLP minimizing loss 2 and a neural ODE minimizing the same loss. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper does not have a section on limitations. I think a discussion of limitations would add to the value of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### When Assumption 1 is expected to hold A necessary condition for Assumption 1 ($\mathcal{C}^1$ strong generalization) is that the optimization problem with the Jacobian loss is solved "well" -- resulting in low generalization errors. But, as you have carefully observed, this is insufficient to claim this notion of strong generalization. The Jacobian loss must be small along orbits of the learned NN (NN with small Jacobian generalization error). Note that these orbits are $(\epsilon_0, \epsilon_1)$ orbits of the true system. When these orbits are on or near the support of $\mu$ (the physical measure of the true system), then, the Jacobian loss is small at those points, and $\mathcal{C}^1$ strong generalization holds. Thus, a sufficient condition is for the true dynamics to be such that small $\mathcal{C}^1$ perturbations of it lead to small perturbations of the physical measure. Such a condition is called smooth linear response in the parlance of dynamical systems and has been extensively studied in the theoretical literature (e.g., see the review by Baladi here: https://arxiv.org/abs/1408.2937). For uniformly hyperbolic systems, which are considered in this paper, and many chaotic systems observed across physics, such a smooth linear response holds (https://iopscience.iop.org/article/10.1088/0951-7715/22/4/009/meta). But, of course, this is only a sufficient condition, and in practice, Assumption 1 may be satisfied more easily. For instance, when we train by sampling points randomly in a box around the attractor, we automatically sample from points near the support of $\mu$ (the attractor). Thus, in input space regions where orbits of $F_{\rm nn}$ live, the Jacobian loss might evaluate to a small value. That is, whether or not linear response holds, we achieve $\mathcal{C}^1$ strong generalization by choosing training points randomly distributed (according to any density) around the attractor and minimizing the Jacobian loss. We thank you for this excellent question, which has led to an important clarification in the paper. Due to space constraints in the paper, we have added more details in the appendix and only briefly make a clarification after Assumption 1 in the main text. ### Why minimizing MSE loss does not yield statistical accuracy This is an astute observation, thank you. Indeed the fundamental contribution of the paper is to suggest shadowing as a mechanism for learning the physical measure. Shadowing does not hold when we only learn well with respect to the $\mathcal{C}^0$ loss! This is because in order to shadow a chaotic orbit, informally, we need to predict the next step as well as local directions of infinitesimal linear perturbation (linear/Jacobian structure) induced by the next step dynamics. This intuition is implicit in the proof of the shadowing lemma (see e.g., Chapter 18 of Katok and Hasselblatt), and therefore extends to the proof of our high-probability version of the shadowing lemma that leads to our main result. Just learning the short-term dynamics without learning the local directions of growth/decay of infinitesimal perturbations cannot lead to learning shadowing orbits. Since we prove that shadowing is the underlying mechanism for learning the invariant measure, the MSE loss is insufficient. We hope this clarifies our result -- thank you for the great question! Due to space constraints, we have not added this explanation, but we will in the next revision right after Theorem 2. ### Neural ODEs are neural parameterizations of the vector field Neural ODEs (trained with MSE) learn the vector fields, but we need to learn the derivatives of the vector fields (with respect to the state vector) for statistical accuracy. Our results indeed already compare Neural ODEs where the vector field is parameterized by an MLP, MLP with Fourier Layers and ResNet blocks. Our observation is the same: without the Jacobian information in the loss function, even parameterizing the vector field, like a Neural ODE does, is not sufficient to learn physical measure. Intuitively, one can argue that learning more and more derivatives of a function (like a vector field of a flow or a discrete-time dynamical system, $F$) would lead to more accurate learning the function and hence also give rise to statistical accuracy. But shadowing (which only requires first derivatives) says that higher-order derivative matching is not necessary. ### Limitations We add the discussion on limitations in Section 7 of the revision. Mainly, we illustrate our theoretical results using the full Jacobian, but this does not yield a practical scheme for training. Secondly, our results are derived for mathematically ideal chaotic systems trained with a vanilla regression setup. Training interventions that enforce learning invariant measures are not considered in our analysis. In this regard, please see our responses to Reviewer vAp3 and sRDx. --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications. From the other reviews, I see there are some concerns regarding practicality of obtaining the Jacobian. However, I think the explanation of generalization also has value. I maintain my accept recommendation. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you very much -- we especially appreciate your questions! The purpose of this paper, as you indeed correctly point out, is dynamics-aware generalization bounds. We show that learning the Jacobian leads to statistical accuracy and explain why by deriving a generalization bound via shadowing theory. Computing the Jacobian is actually not impractical compared to more sophisticated training approaches using recurrent architectures or generative modeling of the SRB measure (physical measure in the paper). We appreciate your time and effort! Please let us know if we can provide any further clarification!	Summary: The authors focus on analyzing why MSE loss fails to capture the physical behavior of dynamical systems. Narrowing their analysis to invariant ergodic systems, they provide theoretical insights on when generalization implies statistical accuracy. They propose that for models to be statistically accurate, they must reproduce dynamical invariants, which is achieved by ensuring that the learned model closely follows the true system's orbits. Specifically, the authors provide theoretical justification for why Jacobian matching loss can better capture statistical properties than MSE loss. Empirically, they verify their analysis on the Lorenz 63 system using different architectures, including MLP, Neural ODE, and FNO. Strengths: The authors provide a thorough theoretical analysis and propose some useful notions to characterize models' ability to reproduce the dynamics statistical measure. Overall, the paper is well-written and easy to follow. Weaknesses: 1. Jacobian loss (Eqn.3) considered in the paper is a special case of the Sobolev norm in [1]. The authors should consider extending their analysis to the Sobolev norm. 2. Recent related works have also discussed the problem of MSE and shown multiple ways to improve upon MSE loss, e.g., the Wasserstein loss used in [2,3] and the theoretical analysis provided in [4]. Although some of these works might focus more on the empirical side, my understanding is that this problem has raised more attention than the authors' claim, and some comparison to the existing work is necessary. [1] Learning Dissipative Dynamics in Chaotic Systems (https://arxiv.org/pdf/2106.06898) [2] DySLIM: Dynamics Stable Learning by Invariant Measure for Chaotic Systems (https://arxiv.org/abs/2402.04467) [3] Training neural operators to preserve invariant measures of chaotic attractors (https://arxiv.org/abs/2306.01187) [4] On the difficulty of learning chaotic dynamics with RNNs (https://arxiv.org/pdf/2110.07238) Technical Quality: 2 Clarity: 3 Questions for Authors: Q1: Figure 1 is unclear, especially when the y-axis is not aligned in the 4&5 columns, and the message shown in Figure 1 seems to contradict the authors' theoretical claims, as the probability distribution of the modeled dynamics is still mismatched when the dynamics of the attract seems to be well learned. Q2: It's confusing when the authors interchangeably used notations $h$ and $F$ for the dynamics map. Is there a particular reason for switching between these notations, and could you elaborate on why? Q3: Line 183: It's not clear why the authors state that implementing Wasserstein distance is difficult when the map $h$ is chaotic. Q4: As Lyapunov exponents are calculated using local dynamics, could you show a comparison using some long-term evaluation metrics? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Implications for training with Sobolev norm It is indeed interesting to consider errors in the learned dynamics as distributions in a Sobolev space, $W^{k,p}$, as you point out. You are absolutely correct that our MSE loss is a special case for $k = 0, p =2$ and the Jacobian loss for $k= 1, p =2.$ Here, we only consider classical functions (as opposed to distributions) for the true map $F$; the neural network, $F_{\rm nn},$ also does not have singularities since we use a smoothed out ReLU as activation. Hence, the weak derivatives above are indeed just classical derivatives. Our results (Theorem 2) means the following for the Sobolev loss: we do not gain more statistical accuracy by minimizing errors with larger $k,$ that is, by considering higher-order derivatives. When $p=2,$ we do not gain by including in the error or loss term higher order Fourier coefficients. In this paper, we prove that a sufficient condition for statistical accuracy is the prevalence of shadowing and its typicality (that is, distribution of shadowing orbits according to $\mu$ with high probability). Our high-probability version of shadowing only requires $C^1$ strong generalization: that is, matching derivatives only up to first order. Therefore, when a shadowing based mechanism for generating the physical measure holds in a supervised learning problem, our results imply that matching higher-order derivatives are not necessary. We thank you for making this important suggestion! Due to space constraints, we have not added this to the revision but plan to add a more detailed analysis to the next revision. ### Related work These are valuable additions to our related work, we are greatly indebted to you and Reviewer vAp3 for showing us these references! We have added all of the references in the Related Work and continued the section into an appendix due to lack of space. Regarding the first cited reference, it is very interesting to derive generalization bounds for the dissipativity-enforcing loss proposed by the authors. Without enforcing dissipativity, in our work, we find that incorporating the Jacobian learns an attractor of zero volume (the attractor dimension is related to the LEs and these are correctly obtained) in dissipative systems. We do not consider training interventions as done in [2] and [4] to stabilize the training of chaotic dynamics over longer time horizons. We leave performing theoretical analysis of the statistical accuracy of the empirical approach to train RNNs in [4] for future research. Instead, our focus here is on obtaining provable guarantees for learning of the physical measure; hence, we only consider here the minimal training setting of regression over short time horizons, wherein such stabilization interventions are not needed. The reference [3] is indeed very close in motivation to our work in that invariant measures of chaotic systems are learned by neural network models. However, the approach taken in [3] is markedly different since optimal transport is performed on key statistics or such summary statistics are approximated through contrastive learning. We argue that supervised learning with Jacobian information is less complicated and more tractable in high dimensions even when compared to an efficient OT algorithm (Sinkhorn -- on discrete measures -- is used in the reference). Furthermore, we obtain theoretical guarantees for the error in the learned invariant measure (e.g. in Wasserstein distance), while it would be interesting to derive such guarantees for the contrastive learning approach taken in [3]. ### Clarification of Figure 1 The y-axis of the 4th and 5th columns of Figure 1 were the empirical PDFs of a random orbit generated by the MSE and Jacobian models respectively. The different scale of the plots show that the empirical distribution of the MSE orbit is incorrect. This exemplifies the thesis of the paper: adding Jacobian information in the regression problem leads to learning the physical measure. To avoid confusion, we have now [combined the 4th and 5th plots](https://app.gemoo.com/share/image-annotation/679582047652405248?codeId=vzaQe6gZzGngO&origin=imageurlgenerator&card=679582046868070400) into one figure, revised the caption and added a Gist (please see the link). Thank you! ### Clarifying notation of map We apologize for the confusion: the h is used as an argument for writing the generalization error as a function of the model $h$. The learned chaotic map is denoted throughout as $F_{\rm nn}$ and the true map by $F$. ### Solving OT problems with ergodic measures of chaotic maps Thank you for the careful observation! Suppose we are trying to minimize the Wasserstein distance, whose dual form we can lower bound, for some $g \in {\rm Lip}^1(M),$ as $$ W^1(\mu, \mu_{\rm nn}) \geq \lim_{t\to\infty} \dfrac{1}{t} \sum_{n\leq t} \|g(F^n(x)) - g(F^n_{\rm nn}(x))\|, $$ for Leb a.e. x. In practice, e.g., when solving OT problems with the Sinkhorn algorithm, we replace the continuous measure with a discrete measure, which in the above case reduces to replacing the ergodic average ($t\to \infty$) with an average over an orbit of finite length. Then, the estimate of the above error is noisy when $F$ and $F_{\rm nn}$ are chaotic, and indeed the variance grows exponentially with $t$ and then saturates (due to both attractors being bounded). ### LEs are statistical measures LEs are indeed long-term evaluation metrics (now added in Section 2). For a vector field $E_i$ in the $i$th Oseledets subspace, we can rewrite the definition of the $i$th LE as an ergodic average: $$\lambda_i := \lim_{t \to \infty} \dfrac{1}{t} \sum_{n =1}^t \log \\|dF (F^n(x)) E_i(F^n(x))\\|.$$ We also show empirical distributions of various quantities (like the components of the state vector) estimated over the long orbits directly in Figure 1 and Table 1. Thank you for this suggestion! --- Rebuttal 2: Title: Response to your rebuttal Comment: Thank you for your response! Your response regarding the Sobolev norm was valuable, and your further clarifications cleared up my concerns. I appreciate your discussion regarding the related works, which helped me better assess the status of your work. I will increase my score to accept. --- Rebuttal Comment 2.1: Title: Thank you! Comment: We really appreciate your time and for re-assessing our work based on our rebuttal! Thank you!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Gradient-based Discrete Sampling with Automatic Cyclical Scheduling	Accept (poster)	Summary: The paper presents a novel gradient-based algorithm to sample from complex multimodal discrete distributions based on differentiable energy functions. Overall, the method is based on locally balanced proposals, previously introduced, and instantiates it with parametrized functions and a cyclical schedule for the "learning rate" that promotes the alternation of modes discovery and modes refinement. The authors also provide an algorithm to automatically tune the introduced parameters based on an input acceptance rate and initial and final balancing parameters. The paper features a theoretical analysis that includes concrete convergence rates under (somewhat restrictive) assumptions and finishes with some experiments in learning and sampling from RBMs and EBMs and finally in text infilling with masked language models. Strengths: - to the best of my (limited) knowledge in this area, the algorithm presented in the paper seem novel and potentially quite impactful (provided the author release an "easy-to-use" implementation) - the illustrative example sets the stage nicely for the need of further development in the field of gradient-based sampling from discrete distribution and provides a very compelling visualization of the efficacy of the method. - the non-asymptotic bounds may offer concrete guarantees when assumptions are met. Weaknesses: - I think too much of the paper is in the appendix, and the information presented in the paper is not sufficient to fully follow the logical flow (see also points below). In my opinion, some details regarding the development of the method and the theory could be moved to the appendix to make more space for both preliminaries (like how are these methods used in practice) and developing more intuition. - the algorithm is fairly complex and the paper fails in provide enough intuition for some parts of its functioning. Like the authors claim that "it is fairly easy to choose initial and final balancing factors", but do not elaborate why (in the main paper) - the only "real-world" experiment is only sketched in the main paper, not providing enough information to appreciate the task. What is the precise difficulty here? Since roberta is a masked language model I believe it is possible to derive "pseudo-distributions" like in [1] - the theoretical analysis seems to require strong assumptions that may be violated in compelling real-world use cases (like LLMs) [1] Hennigen, Lucas Torroba, and Yoon Kim. "Deriving language models from masked language models." arXiv preprint arXiv:2305.15501 (2023). Technical Quality: 3 Clarity: 2 Questions for Authors: - Can you please discuss in which cases the locally concave hypothesis holds for realistic models such as LLMs? - Even if this is probably more of a sanity check, I'd like to see how the method behave on unimodal, or mildly multimodal distributions - Can you please describe some other applications of the method to real-world problems (like text infilling) Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Partially discussed, see questions Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your supportive comments. We include our responses to your points below: # Q1: Insufficient Content in Main Body regarding Tuning Algorithm In section 4.4, we first provide an intuition for our algorithm under “Main idea” and then present our algorithm by separating it into three separate components: the estimation of $\alpha_\text{max}$, the estimation of $\alpha_\text{min}$, and constructing the schedule of $\beta_i$. Section A of the appendix provides the detailed algorithm that shows how each sub-routine works. We will move some content from the Appendix to the main body and try our best to further improve its clarity. # Q2: Choosing $\beta_\text{max}, \beta_\text{min}$ In section 4.3 where we introduce the cyclical balancing parameter schedule in lines 139-146, we have discussed that the selection of these values are dependent on the theoretical results from [1], [2], which demonstrate that when the step-size $\alpha \to 0$, the optimal $\beta_\text{min} = .5$; but when $\alpha \to \infty$, the optimal $\beta_\text{max} = 1$. We will move the explanation of this from the Appendix to the main body. # Q3: Difficulty of Infilling Task The difficulty of the text infilling task comes from the fact that the probability distribution is over a very large sample space due to the size of the vocabulary and the number of positions to be filled, as discussed in [3]. Furthermore, the paper you mention [4] only considers how to find the distribution of two masked tokens, whereas we mask up to 50% of the sentence. This greatly increases the complexity of the task as it exponentially increases the sample space of potential combinations. # Q4: Locally Concave Hypothesis in Practice The locally log-concave assumption is common in both sampling and optimization literature [4], [5]. It holds on several practical discrete sampling tasks, such as Ising Models with negative definite weight matrix $W$ and Poisson distributions. The locally log-concave assumption does not hold with complex models such as LLMs, since theoretical results of sampling on such models are in general difficult to obtain. Our work provides the first non-asymptotic convergence bound for gradient-based discrete samplers. We leave analysis on non log-concave distributions for future work. Besides, we provide extensive empirical work demonstrating that our sampler converges on models where this assumption does not hold, such as deep EBMs and LLMs. # Q5: Unimodal, mildly multimodal distributions We demonstrate the performance of ACS on a unimodal distribution, similar to the experiment from Figure 1. We put the links for the visual results below. Target Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/single_mode_init_dist.pdf \ DMALA Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/single_mode_est_dist_dmala.pdf \ ACS Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/single_mode_est_dist_acs.pdf Below we provide the results for DMALA and ACS on a mildly multimodal distribution, where the majority of the mass is placed on one mode. Target Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_target_dist.pdf \ DMALA Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_dmala_est_dist.pdf \ ACS Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_acs_est_dist.pdf We provide quantitative results below, where we compare the KL Divergence between the estimated and true distribution. \| Distribution \| DMALA \| ACS \| \| :-\| :-: \| :-: \| \| Slightly Multimodal \| $0.7011$ \| $0.1250$ \| \| Unimodal \| $0.0089$ \| $0.0032$ \| In both cases, we see ACS achieves a lower KL divergence than DMALA, thus demonstrating that our proposed method is capable of capturing both a unimodal and slightly multimodal distribution. # Q6: Other applications of method In addition to sampling from language models, we also demonstrate that our proposed sampler can be used to train deep discrete energy based models more efficiently than previous discrete samplers. Within the domain of language modeling, there are many additional applications beyond text infilling. In work such as [7], [8], language models are framed as Energy Based Models and then sampled from in order to perform controlled generation tasks. These tasks include abductive reasoning and counterfactual story rewriting [8]; sentiment guided generation and detoxification as in [7]; and keyword generation as in [7], [8]. Given the success of applying ACS and discrete sampling to the task of text infilling, a natural step would be to investigate the application of ACS to controlled language generation. Furthermore, controlled generation is remarkably similar to language model alignment as described in [9]. One interesting direction would be to apply the discrete Energy Based Model framework to this task as a decoding time algorithm, similar to [10]. [1] Informed proposals for local MCMC in discrete spaces. 2017.\ [2] Any-scale Balanced Samplers for Discrete Space. ICLR 2022.\ [3]. A Langevin-like Sampler for Discrete Distributions. ICML 2022.\ [4] Deriving language models from masked language models. arXiv Preprint 2023.\ [5] Optimization methods for large-scale machine learning. SIAM Review 2018.\ [6] Theoretical guarantees for approximate sampling from smooth and log-concave densities. Journal of the Royal Statistical Society Series B: Statistical Methodology 2017.\ [7] Gradient-Based Constrained Sampling from Language Models. EMNLP 2022.\ [8] COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamic. NeurIPS 2022.\ [9] Reward-augmented decoding: Efficient controlled text generation with a unidirectional reward model. EMNLP 2023.\ [10] Args: Alignment as reward-guided search. ICLR 2024. --- Rebuttal Comment 1.1: Title: Thanks Comment: Dear authors. Thank you very much for your rebuttal. I appreciate the additional experiments clarifications on the infilling task. I keep my opinion that this work is a valid addition to the conference.	Summary: The paper proposes a solution to the challenge of sampling from high-dimensional discrete spaces, where conventional discrete samplers often get trapped in local modes. To address this, the authors introduce a discrete Langevin sampler with automatic cyclical scheduling. This method comprises three components: a cyclical step size schedule, a cyclical balancing schedule, and an automatic hyperparameter tuning scheme. The authors provide theoretical guarantees for non-asymptotic convergence and inference, and extensive experiments demonstrate the method's superiority in sampling complex multimodal discrete distributions. Strengths: The paper is well-motivated, and the proposed automatic cyclical scheduling method is presented clearly, making it accessible to readers. The theoretical results, which offer non-asymptotic convergence, support the method's robustness. Additionally, the empirical study is solid, with extensive experiments demonstrating the method's superiority in sampling from high-dimensional spaces. Weaknesses: - My primary concern lies in the complexity of the proposed methods. The automatic schedule tuning scheme appears to be quite time-consuming, particularly the grid search required for the balancing parameters \beta_i, demanding significant computational resources. - Another concern pertains to the theoretical assumptions underlying the analysis. The non-asymptotic convergence of the proposed samplers relies on the strong convexity of the negative energy function, an assumption that may not hold in practical deep EBM scenarios. Despite this potentially restrictive condition, the analysis provides valuable insights, and empirically, the proposed method demonstrates effective performance, as supported by extensive studies. Technical Quality: 3 Clarity: 3 Questions for Authors: - In equations 8 and 9, how do you estimate the acceptance rate A? Is it estimated by averaging across the training batch? If so, that implies the complexity would increase to ns times compared to the original DMALA samplers at each step, where n is the number of grids for parameters \beta_i and s is the number of sampling steps per cycle. This would be highly time-consuming. - Could you provide a comparison of the running times between the proposed ACS sampler and the DMALA samplers? - Could you elaborate on the rationale behind setting the target acceptance rate $\rho^$ to 0.5 in your experiments? What are the implications of setting it to 1 or 0.234 instead? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Despite the strong assumptions in the theoretical analysis and the perceived complexity of the algorithm, this paper presents a substantial contribution to the field by offering a well-explained, theoretically sound, and empirically validated approach to enhancing discrete sampling in high-dimensional spaces. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your supportive and valuable comments. We will address the issues you raise below. # Q1: Complexity of Tuning Algorithm In total, the automatic tuning algorithm takes 500 sampling steps as a budget at maximum, which is much smaller when compared to the 5,000 sampling steps we use for EBM sampling and the 10,000 steps we use for RBM sampling. Furthermore, the cost of a tuning step is almost the same as a standard sampling step — as shown in Algorithm 4 and Algorithm 5, most of the additional steps for the estimation of $\alpha_\text{min}, \alpha_\text{max}$, and the $\beta_i$ schedule are arithmetic operations that take constant time or averaging over acceptance rates that are already computed within a normal sampling step. Therefore, the tuning algorithm is neither time-consuming nor costly compared to the main sampling costs. # Q2: Estimation of Acceptance rate The acceptance rate for a pair of $\alpha, \beta$ is calculated by averaging over the current batch for one iteration. As most MCMC algorithms run multiple chains at the same time by sampling in batches, we found that averaging over the batch for a single time step provides useful acceptance rates to use during the tuning process. For a given time step, we take the average acceptance probability as calculated by Equation 6 across the batch. Thus each tested pair of $\alpha, \beta$ requires only one step to estimate the acceptance rate. While this does mean that the complexity is $O(n * s)$, where $n$ is the number of potential $\beta_i$ and $s$ is the number of steps per cycle, we found that the number of potential $\beta_i$ does not have to be very large. In practice, we found that testing 5 different $\beta_i$ for each step produces good schedules. The largest cycle length that we use in our experiments is 20 steps — this means that this step takes 100 sampling steps total, which is small compared to the total number of sampling steps of 10,000 in the RBM sampling experiment and 5,000 in the EBM sampling experiment. # Q3: Run Time Comparison We found that using a budget of 500 sampling steps enabled the discovery of good hyper-parameter schedules. As discussed in our response to Q2, a tuning step has essentially the same cost as a sampling step. Thus the total overhead amounts to 5% of the total budget of the RBM sampling task and 10% of the total budget of the EBM task. Beyond this overhead, the run times of DMALA and ACS are the same, as both calculate the proposal function and the acceptance rate in the exact same manner. We use the RBM sampling experiment to further compare the run times between DMALA and ACS. We we make the results available at the following link: https://anonymous.4open.science/r/neurips_rebuttal-B010/rbm_log_mmds_dmala_acs_comp_time.pdf As visible in the results, even if we restrict ACS (including the tuning phase) and DMALA to the same total time budget of 20 seconds, we see that ACS is able to achieve a log-MMDs superior to that of DMALA. Thus even considering the overhead, our proposed sampler outperforms DMALA within a fixed time budget. # Q4: Rationale behind setting target acceptance rate We base our target acceptance rate of .5 on the study done in [1], where the authors demonstrate that the ideal acceptance rate for locally balanced samplers is close to .5. Besides, .5 is also the commonly used target acceptance rate for gradient-based samplers. Therefore, we use .5 in the paper and find that it works well in practice. If we set the target acceptance rate to be close to 1, we will end up with step-sizes that are too small, resulting in insufficient exploration of the distribution. If we set the target acceptance rate to be around .234, then the sampler will end up rejecting most of the proposed movement, decreasing the efficiency of the sampler. An acceptance rate of .5 avoids either scenario, allowing for efficient and thorough characterization of the target distribution. [1]. Optimal Scaling for Locally Balanced Proposals in Discrete Spaces. NeurIPS 2022. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. It has solved my concern about the complexity. I keep my opinion that this work is good to be in, and I highly recommend including the discussion of complexity in the camera ready.	Summary: The paper introduces a novel method for sampling from multimodal discrete distributions, which presents an innovative approach to address the challenge of local modes trapping in gradient-based discrete sampling, together with non-asymptotic convergence guarantee and empirical validation of the proposed method. Strengths: 1. The proposed method seems novel to address the challenge of sampling from multimodal discrete distributions. 2. The hyperparameter tuning algorithm seems useful for practical use. Weaknesses: 1. In the experiments, there seems to be no error bars in Figure 1 and Table 1. 2. The quality of the samples seems worse than DMALA in Table 2 and Figure 12. Is it possible that the proposed method might sacrifice sample quality to achieve higher diversity? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The paper mentions the proof is not consistent with the specific tuning algorithm used in the experiments. Could you elaborate on the reasons? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful comments. We answer the questions below. # Q1: Error Bars Fig 1, Table 1 Figure 1 corresponds to density estimation, where error bars would not make sense. If you are referring to Figure 3, it should be noted that the shaded area represents the range within 1 standard error of the average performance across 11 seeds for RBM sampling. For the other missing error bars, we include the updated results in the summary response, which show that our claims still hold. # Q2: Sample quality The proposed ACS does not sacrifice sample quality to achieve higher diversity. It should be noted that Figure 12 does not show that ACS decreases quality — the generated sentences are reasonably fluent and comparable to those generated by DMALA. Furthermore, while our method does result in higher perplexity, which corresponds to lower likelihood under the model generations, it should be noted that we include the CoLA scores to measure the grammatical quality of the generations. As indicated by these scores, our method does not sacrifice grammatical correctness for diversity. While perplexity is a popular means of evaluating language generations, it is important to recognize that perplexity is based on the likelihood of the generation under the language model. This metric is biased towards frequent patterns and does not account for diverse modes. Therefore, it does not completely align with the goal of MCMC, which is to accurately characterize the target distribution. Minimizing the average perplexity of the sample corresponds to maximizing the average likelihood of the generations, which can be at odds with the goal of accurately capturing the target distribution. We illustrate this in the following experiment, where we construct a synthetic dataset of 25 modes where the top-left mode is weighted far more than all the others. Below are the anonymized links to the images for this experiment: Target Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_target_dist.pdf \ DMALA Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_dmala_est_dist.pdf \ ACS Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_acs_est_dist.pdf \| Method \| Average Energy \| KL Divergence \| \| :---------------- \| :-----------: \| :-----------: \| \| DMALA \| $-2.66 \pm 1.68$ \| $0.70$ \| \| ACS \| $-3.39\pm 1.63$ \| $0.13$ \| The average energy for the samples from DMALA is higher than that of ACS as DMALA ends up being trapped by the top left mode. This indicates that the samples generated by DMALA are more likely than the samples generated by ACS. However, the visualizations of the estimated distributions show that DMALA ignores the majority of the low-likelihood modes, whereas ACS is able to explore all of them. This is supported by the measured KL divergence between the estimated distribution and the target distribution. This demonstrates the disconnect between maximizing the average likelihood of the generated samples and accurately capturing the target distribution. # Q3: Consistency between Theory and Algorithm In the Conclusion and Limitations section, we mention that the proof is not based on a specific tuning algorithm. This means that our theoretical analysis does not consider the effect of the tuning algorithm. If we were to take into account the hyperparameters (alpha and beta) provided by the tuning algorithm, it might be possible to make the results more tailored to practical performance. The reason for this limitation is that conducting such a theoretical analysis is a challenging problem, as it requires analyzing an inhomogeneous Markov chain. Techniques to analyze such chains in discrete spaces, which may be applied here, are not known to us. In fact, we conjecture that any improvement on the current bound provided by us, in terms of the entire schedule, shall involve developing fundamental theory for inhomogeneous Markov chains in discrete spaces. [1]. Annealed Importance Sampling. Technical Report 1998.\ [2]. Oops I Took A Gradient: Scalable Sampling for Discrete Distributions. ICML 2021.\ [3]. A Langevin-like Sampler for Discrete Distributions. ICML 2022.\ [4]. Path auxiliary proposal for MCMC in discrete space. ICLR 2022. --- Rebuttal Comment 1.1: Comment: Thank you for the response and clarifications. I will keep my score.	Summary: This paper proposes a new discrete sampling method called ACS that addresses a common problem for existing gradient-based approach where they are susceptible to becoming trapped in local modes. ACS combines local-balancing proposals with a cyclic step size to balance local exploitation and global exploration; it is in essence an extension of cyclic stochastic-gradient MCMC to discrete distributions. To ensure proposals are still balanced with a varying step size, ACS uses a cyclic balancing schedule along with an automatic tuning scheme to easily adapt the schedules. Non-asymptotic convergence guarantees are provided. Results demonstrate ACS to outperform prior approaches for sampling from energy based models, training RBMs, and text-infilling. Strengths: - Using a cyclic step size schedule is a well motivated and effective approach for incorporating global considerations into the original local self-balancing MCMC approach presented in https://arxiv.org/pdf/2109.03867. - ACS is accompanied by an automated tuning scheme to make it easy to configure the two cyclic schedules. - Strong empirical results on EBM tasks. - ACS has non-asymptotic convergence guarantees. Weaknesses: - Experimental results for RBM have discrepancies wrt results reported in previous papers. In particular the ranges differ for average energy from other papers and the curves for log MMD for ACS show unexpected curvature. Please respond to the questions in the section below to clarify. - Inadequate discussion of text-infilling results. It is not clear why higher perplexity and diversity is good for ACS since the goal is to be able to efficiently sample from a target discrete distribution. - Error bars missing for results in Figure 3 and Table 1. Typos & formatting: - Inline citation format should show author name - Figure 3 caption: cpnvergence -> convergence Technical Quality: 3 Clarity: 3 Questions for Authors: - Why is average energy in Figure 3 negative? Also, the results for ACS on dynamic_mnist and omniglot are unexpected with better performance on fewer iterations before converging. - Why is the scale for average energy in Figure 3 different from that reported in Figure 4 of the [AB paper](https://openreview.net/pdf?id=lEkl0jdSb7B)? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are adequately discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. We answer your questions below. # Q1: Missing Error Bars, Figure 3 and Table 1 Thank you for pointing this out. For the log MMD curves for the RBM experiments in Figure 3, it should be noted that the filled in area corresponds to values within one standard error of the mean across 11 different random seeds. However, it is correct that our curves for the mixing time for deep energy based models in Figure 3 and the metrics from Table 1 do not have standard error bars. We have provided the results with error bars in our summary response, which show that our claims still hold. # Q2: Negative Average Energy The average energy is negative due to the definition of the distribution we use. We define the target distribution as $\pi(\theta) = \frac{\exp U(\theta)}{Z}$, where $Z$ is the partition function. The energy is $U(\theta) = log (\pi (\theta)) - log Z$. Because a log probability is negative and the partition function is the sum of exponentials, which are positive, it should be the case that the energy function is negative. In contrast, [5] use the definition of the distribution from [4] for the experiments with deep energy based models as they train these models using the sampling algorithm proposed in [4]. They define the distribution as $\pi(\theta) = \frac{\exp (-U(\theta))}{Z}$. Thus, their energy is positive. # Q3: Better performance with fewer iterations before converging Here, it should be noted that higher energy of the batch does not correspond to closeness to the target distribution. For a more in depth explanation on the difference between higher average energy and closeness to target distribution, we include a toy example in our response to Q5. The energy decrease can be seen as the direct result of ACS being able to escape from modes quicker than other samplers — ACS is able to find very likely samples quickly, but then explores the different modes of the distribution, causing the energy to decrease. Because the ground truth distribution in deep EBMs is unknown, we evaluate the sampling performance using average energy following the experiment in [5]. It tells us how quickly different samplers are able to reach likely modes, which gives insight into the speed of the various samplers. To evaluate the convergence to the target distribution, we include extensive experiments on RBMs, where it is possible to get an estimate as to how close the sampler is to the ground truth distribution via the maximum mean discrepancy to the block gibbs sampler, which takes advantage of the known architecture of an RBM. This enables us to get insight as to how accurate the samplers are. # Q4: Average Scale of energy different than AB This is due to the training of the EBMs — because the EBM represents an unnormalized probability distribution, different EBMs tend to have different scales of energy, depending on the sampler used within the contrastive learning routine. [5] uses Path Auxiliary MCMC from [4], whereas we use gibbs with gradient for the models in our EBM experiment. # Q5: Higher perplexity and diversity in text infilling task While we include perplexity as it is a popular means of measuring language quality, this metric faces significant limitations as discussed in the literature (e.g.[6,7]). It fails to capture logical or grammatical coherence, bias towards frequent patterns, does not align with human evaluation, and cannot capture diversity. Because of this, we further include CoLA (measure the grammatical quality) and Self-Bleu (measure diversity) to comprehensively evaluate the generated sentences. The results show that ACS generations achieve better CoLA and Self-Bleu scores. Furthermore, perplexity is especially limited when trying to measure the accuracy of a sampler in regards to a target distribution. As perplexity is based on the likelihood of the generation under the language model, it does not account for diverse modes. To illustrate this point, we provide the following toy example to show the goals of obtaining the most likely generations and capturing the language model distribution are orthogonal. Similar to Figure 1, we construct a synthetic distribution of 25 modes where 1 mode is weighted heavier than the others. We provide visualizations along with a quantitative comparison of the estimated and target distributions below. Target Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_target_dist.pdf \ DMALA Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_dmala_est_dist.pdf \ ACS Estimated Distribution: https://anonymous.4open.science/r/neurips_rebuttal-B010/slightly_multimodal_acs_est_dist.pdf \| Method \| Average Energy \| KL Divergence \| \| :---------------- \| :-----------: \| :-----------: \| \| DMALA \| $-2.66 \pm 1.68$ \| $0.70$ \| \| ACS \| $-3.39\pm 1.63$ \| $0.13$ \| During the sampling process, DMALA becomes stuck at the high likelihood mode, preventing the sampler from exploring the rest of the sample space. In contrast, ACS is able to visit all the modes in the distribution. This can be seen through a visual inspection of the density maps for DMALA and ACS. While DMALA ends up generating samples with higher energy, ACS estimates a distribution that is far closer to the ground truth as measured by the KL divergence. This demonstrates that generating more likely samples does not correspond to accuracy in terms of convergence to the target distribution. [1]. Annealed Importance Sampling. Technical Report 1998.\ [2]. Oops I Took A Gradient: Scalable Sampling for Discrete Distributions. ICML 2021.\ [3]. A Langevin-like Sampler for Discrete Distributions. ICML 2022.\ [4]. Path auxiliary proposal for MCMC in discrete space. ICLR 2022.\ [5]. Any-scale Balanced Samplers for Discrete Space. ICLR 2023.\ [6]. Lower Perplexity is Not Always Human-Like. ACL 2021\ [7]. Language model evaluation beyond perplexity. ACL 2021 --- Rebuttal Comment 1.1: Title: Post author response Comment: Thank you for responding to my questions. I have changed my score from 4 to 5.	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their constructive reviews. Multiple reviewers have pointed out that our results for sampling from Energy Based Models in Figure 3 and the Annealed Importance Sampling results in Table 1 do not have error bars. Below, we have provided the updated results for Table 1 and Figure 3, both of which confirm that our observations and conclusions remain valid. Updated Figure 3 EBM Mixing Results: https://anonymous.4open.science/r/neurips_rebuttal-B010/ebm_sample_avg_iter.pdf Updated Table 1: \| Dataset \| DMALA \| ACS \| \| :---------------- \| :---------: \| :---------: \| \| Static MNIST \| $-80.031 \pm 0.038$ \| $-79.905 \pm 0.057$ \| \| Dynamic MNIST \| $-80.120 \pm 0.036$ \| $-79.634 \pm 0.024$ \| \| Omniglot \| $-99.243 \pm 2.101$ \| $-91.487 \pm 0.128$ \| \| Caltech \| $-98.001 \pm 0.371$ \| $-89.262 \pm 0.290$ \| In the updated Table 1, we ran the AIS evaluation for the models trained with various samplers over 8 random seeds, and we show the mean as well as the standard error for the log likelihood over the test set. The results indicate that our proposed sampler is capable of training the models of better quality given the same computational budget. In the updated Figure 3, we show the area within one standard error of the mean across 11 different random seeds for each time step. Our proposed ACS has consistent performance across the datasets in terms of being able to mix in quickly. On Static MNIST, Dynamic MNIST, and Omniglot, we observe that our proposed ACS is able to consistently find high energy modes far quicker than the other methods. ACS is then able to find the less likely modes due to its ability to escape from local modes. On Caltech Silhouettes, ACS converges faster at the beginning and maintains competitive performance against the baselines. Finally, we would like to emphasize the key contributions of this work. There has been much work regarding the use of gradient-based discrete samplers, but not enough investigation as to how these gradient-based samplers are affected by the multi-modal nature of high dimensional discrete distributions. We demonstrate that this is a current limitation of discrete gradient-based samplers, and introduce a method that is capable of avoiding this pitfall while retaining the benefits associated with gradient information. Our work introduces cyclical step size and balancing parameter schedules with theoretical guarantees. Furthermore, our method can be automatically configured by a novel tuning algorithm with minimal overhead. We demonstrate on a synthetic highly multimodal distribution and a range of datasets that our sampler can achieve superior performance over existing methods. Given both the theoretical contributions and experimental results for our method, we believe this is a valuable contribution to the field of discrete MCMC sampling.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Regularized Q-Learning	Accept (poster)	Summary: This paper provides Q-learning convergence (asymptotic) in linear architectures with regularization power. Their algorithm is tested on mountain car example. Strengths: Q-learning convergence in linear architectures is an important problem in RL. Weaknesses: - The analysis follows ODE style analysis from Borkar and Meyn. So it is asymptotic. However, non-asymptotic guarantees (rate of convergence) can be provided when assuming non-zero stationary distribution (like Assumption 2.1). For e.g. see Chen et al, 22. There are research on extension to QL like TD-Learning convergence in linear architectures. So I think with the current tools one can give such finite time guarantees with some more effort. This work relies on old works such as Gosavi [2006] and Melo et al., [2008]. Although, I would be curious to know where the current hurdles are. - A recent work titled "Regularized Q-Learning with Linear Function Approximation" (https://arxiv.org/pdf/2401.15196) are providing non-asymptotic results for a similar problem. I have not looked at the details, but since this appeared in arxiv in Jan 2024, please do include what the contributions are of this work compared to this one. - Section 3.2 can be further improved w.r.t writing. I believe Lemma 3.3, 3.4, 3.5 are provided as independent results to satisfy Eq.(11) (which is crucial for existence and uniqueness of RPBE solution). But I am curious to know why 3 different Lemmas are provided. - How is eq.(15) constructed for $\eta$? I understand Lemma 3.3 helps for (S1). But I am not sure how (S2) came about. - It is mentioned at line 228 that $\eta>2$ is enough for Lemma 3.3. But $\eta<1$ for Lemma 3.4. I really hope Lemma 3.3 and 3.4 are not required to be satisfied simultaneously for the current results to hold. - Where is Lemma 3.1 used? We are in linear architecture setting. So Lemma 3.1 (b) inequality which involves $S\times A$ must be avoided. Linear architectures are helpful only when one can replace rate of convergences involving $S\times A$ can be replaced by feature-size $h\ll SA$. My score reflects this review provided here. Technical Quality: 1 Clarity: 1 Questions for Authors: na Confidence: 3 Soundness: 1 Presentation: 1 Contribution: 3 Limitations: na Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1 The analysis follows ODE style analysis from Borkar and Meyn. So it is asymptotic. However, non-asymptotic guarantees (rate of convergence) can be provided when assuming non-zero stationary distribution (like Assumption 2.1). For e.g. see Chen et al, 22. There are research on extension to QL like TD-Learning convergence in linear architectures. So I think with the current tools one can give such finite time guarantees with some more effort. This work relies on old works such as Gosavi [2006] and Melo et al., [2008]. Although, I would be curious to know where the current hurdles are A1 Thank you for the valuable comments. As the reviewer mentioned, we believe finite-time analysis of RegQ is possible. However, our primal goal is to develop a new Q-learning algorithm with linear function approximation that converges under relaxed or different scenarios. We consider proving the asymptotic convergence of Q-learning with linear function approximation as an initial step and the finite-time bound can be also proved following the spirit of related literature. We appreciate the provision of interesting research topics, which remain for future exploration. According to the reviewer's comment, the related discussions will be added in the revision. Moreover, we would like to note that our work does not rely on the work of Melo et al.,[2008]. We also note that Melo et al.,[2008] requires strong assumptions on the behavior policy to be met, whereas we do not requires such assumptions. Furthermore, we provided a comparison with existing algorithms in G1 of the global response, which clarifies current hurdles and the novelty of our approach. Following the reviewer's comment, we will add the associated discussions in the revised manuscript. Q2 A recent work titled "Regularized Q-Learning with Linear Function Approximation" are providing non-asymptotic results for a similar problem. I have not looked at the details, but since this appeared in arxiv in Jan 2024, please do include what the contributions are of this work compared to this one. A2 Thank you for the insightful comment. The regularization considered in ``Regularized Q-Learning with Linear Function Approximation'' requires to be bounded, e.g., an entropy regularizer. The $l$-2 type regularization which we consider in our work does not fall into this category, and the extension is non-trivial. Following the reviewer's recommendation, we will incorporate the discussions on the comparisons in the revised manuscript. Q3 Section 3.2 can be further improved w.r.t writing. I believe Lemma 3.3, 3.4, 3.5 are provided as independent results to satisfy Eq.(11) (which is crucial for existence and uniqueness of RPBE solution). But I am curious to know why 3 different Lemmas are provided. A3 Thank you for helping us to improve the clarity of our manuscript. We provided three different lemmas because each addresses different scenarios for (11) to hold. Lemma 3.3 covers when $\eta$ is larger than certain threshold, Lemma 3.4 considers the case when $\eta$ is nearby the origin, and Lemma 3.5 applies for for all $\eta$. Following the reviewer's comment, we will improve the clarity in the revised manuscript. Q4 How is eq.(15) constructed for $\eta$? I understand Lemma 3.3 helps for (S1). But I am not sure how (S2) came about. A4 We thank the reviewer for the insightful comments. The conditions in (S1) and (S2) is used to guarantee $(A_{\pi_{X\theta_k}} + \eta I)$ to have strictly negatively dominant diagonal or to be a negative definite matrix, respectively. In particular (S2) comes from the item 2 in Lemma 2.5, which guarantees the asymptotic stability of a switched system. The proof for the derivation of (S2) is given in Lemma A.6. in the Appendix of the manuscript. Note that the conditions (S1) and (S2) do not necessarily imply each others, which are discussed in Appendix A.15 of the manuscript. We will add the above discussion in the revised manuscript. Q5 It is mentioned at line 228 that $\eta>2$ is enough for Lemma 3.3. But $\eta<1$ for Lemma 3.4. I really hope Lemma 3.3 and 3.4 are not required to be satisfied simultaneously for the current results to hold. A5 Thank you for the valuable comments. Lemma 3.3 and Lemma 3.4 are only sufficient conditions but not necessary conditions for (11) to hold. Therefore, the two statements do not contradict each other. Moreover, we would like to note that Lemma~3.4 does not require the condition $\eta<1$. It only assumes the scenario $\|\| \gamma \Gamma \|\|_\infty < 1$. Following the reviewer's comment, we will clarify these points in the revised manuscript. Q6 Where is Lemma 3.1 used? We are in linear architecture setting. So Lemma 3.1 (b) inequality which involves $S\times A$ must be avoided. Linear architectures are helpful only when one can replace rate of convergences involving $S\times A$ can be replaced by feature-size $h<<SA$ A6 We thank the reviewer for the insightful comments. Lemma 3.1 is provided for a theoretical understanding on how $\Gamma_{\eta}$ behaves as a function of $\eta$. The inequality in Lemma 3.1 (b) demonstrates that $\Gamma_{\eta}$ remains bounded for all $\eta$ and does not diverge. We do not directly use the upper bound in Lemma 3.1 for our analysis, and it is not related to the rate of convergence. We will clarify this in the revised manuscript. --- Rebuttal Comment 1.1: Comment: Thank you for reflecting on the reviews. I will keep my score as the updates for Lemma 3.1-3.5 require some non-trivial updates that need to be evaluated further, which will be out of scope in this rebuttal period. Good luck. --- Rebuttal 2: Comment: We thank the reviewer for the response in the discussion period. However, we kindly disagree that Lemma 3.1-3.5 needs further evaluation or poses logical errors. 1) Lemma 3.1 : As mentioned in Q6, the lemma is only used to show that $\Gamma_{\eta}$ is bounded for any $\eta$, and we do not use the upper bound which is dependent on $\|S\|\|A\|$ in other result or proof. 2) Lemma 3.3-3.5 : Regarding Q5 in the initial rebuttal, the two lemmas do not contradict each other because they are only sufficient conditions for (11) to hold. They are totally independent results. Each lemmas cover different scenarios, and hence there are no logical errors between them. Furthermore, Lemma 3.3 covers the most practical scenario. If we scale the feature matrix such that $\max(\|\|X\|\|\_{\infty},\|\|X\|\|\_{\infty})<1$, then choosing $\eta>2$ satisfies the condition (11) by Lemma 3.3 in the manuscript. Scaling the values of feature matrix is a commonly employed technique in the both theoretical literature or in practice. We again thank the reviewer for the engagement in the discussion, and kindly request the re-evaluation of our manuscript.	Summary: This paper introduces a novel approach, RegQ, which is a framework for dealing with linear approximation of Q-function. Compared to the instability of the traditional Q-learning with function approximator, which is known as the deadly triad, RegQ addresses this problem by regularization term, making the algorithm more stable. Also the theoretical approach ensures convergence of linear function approximation. Strengths: This paper gives a novel approach with theoretical rigor. The most noteworthy strength of this paper is that it tackles a practical problem: the instability of Q-learning with linear function approximator, compared to other theoretical paper that gives impractical solutions. Although the theory is highly technical, this paper gives a good logical explanation, also Figure 1 helps readers to understand the proposed projection operator. Weaknesses: Most of the concerns arose from the lack of experiment scope, prior work comparison, and implementation. Authors claim about the strengths of the RegQ algorithms, but lots of claims are not confirmed by experiments. Also, the claims are given with the comparison of the prior work, but experimental results are not sufficient to validate those claims. It might be helpful to gain confidence about the novelty of the paper, Also, it is hard to get a precise understanding of the algorithm since there are lack of implementation details. Technical Quality: 3 Clarity: 3 Questions for Authors: In equation (13), what delta_k means? Also, in line 290, why m_k+1 is an i.i.d. noise? Is it an assumption, or can it be proved? In Figure2, the y-axis means the max norm of Q-values. I’m wondering about the value of y axis. At the initial stage of the episode, all of the algorithms have already shown 1e-18, and 1e-20 values, and what the authors claim is RegQ converges when others fail. My question is this: why initial point or other algorithms( which is shown 1e-18 error) is not considered as converged? Also, since values are too small, how could authors prevent the contamination by the floating point error? It will be easy to understand if the authors give more explanation about experiments. In Figure 1c, I can’t find the details about the phases: ‘Blowing up phase’, and ‘Shrinking phase’. I read the author’s error analysis, but there’s no explanation about the phase. Is it just a intuitive naming on the convergence behavior? If not, could you elaborate more about it in the theoretical context? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1. Most of the concerns arose from the lack of experiment scope, prior work comparison, and implementation. Authors claim about the strengths of the RegQ algorithms, but lots of claims are not confirmed by experiments. Also, the claims are given with the comparison of the prior work, but experimental results are not sufficient to validate those claims. It might be helpful to gain confidence about the novelty of the paper. Also, it is hard to get a precise understanding of the algorithm since there are lack of implementation details. A1. Thank you for the constructive comments. We would like to first note that due to the page limits, all the experiments are included in the Appendix B.. Similarly, the implementation details are given in the Appendix B. Moreover, we want to clarify that our main focus is on the theoretical analysis of proposed RegQ algorithm and regularized projected Bellman equation. We proposed a convergent algorithm under linear function approximation and mild assumptions and provided a thorough analysis on the conditions for the convergence. The comparison with prior works are mainly in theoretical sense, e.g., weaker assumption on the behavior policy or relaxing additional assumptions used in prior works. Moreover, the analysis on the properties of regularized projected Bellman equation is new in the literature. The experiments are provided to verify the convergence of the proposed algorithm and provide further insights on its behavior, e.g., convergence rate. Our claim is that RegQ can show faster convergence rate than other algorithms under certain scenarios, which is verified in Section 6. An intuitive reason why our algorithms shows faster convergence is that the baseline algorithms ( CQL, targetQ, and GreedyGQ ) are basically a two-time scale algorithm whereas our proposed algorithm uses a single-time scale step-size. Moreover, regarding the implementation detail, a pseudo-code is given in Appendix A.16. Following the reviewer's comments, we have added the discussion in the revised manuscript. Q2. In equation (13), what $\delta_k$ means? Also, in line 290, why $m_{k+1}$ is an i.i.d. noise? Is it an assumption, or can it be proved? A2. We thank the reviewer for the constructive comments. For the first question, $\delta_k$ is the TD error, which is defined below in the equation (2) in the manuscript. $m_{k+1}$ is not an i.i.d. noise but the Martingale difference sequence, which includes the i.i.d. noise scenario as a special case, and which is shown in Lemma A.14 in the Appendix. We will correct the typo in the revision. We would like to note that the Martingale difference sequence scenario is an assumption, and it cannot cover the case that the transitions are sampled from a single trajectory (or Markovian sampling scenario). In [1], the authors demonstrated an extension of the Borkar and Meyn Theorem to the Markovian sample cases. Since our proof relies on Borkar and Meyn Theorem, our result can also be extended to establish convergence under Markovian sample cases. We have incorporated this result into the revised manuscript. Q3. In Figure2, the y-axis means the max norm of Q-values. I’m wondering about the value of y axis. At the initial stage of the episode, all of the algorithms have already shown 1e-18, and 1e-20 values, and what the authors claim is RegQ converges when others fail. My question is this: why initial point or other algorithms( which is shown 1e-18 error) is not considered as converged? Also, since values are too small, how could authors prevent the contamination by the floating point error? It will be easy to understand if the authors give more explanation about experiments. A3. Thank you for the valuable comments. We agree with the reviewer. For clarification on the error bound, we have provided the first 50 steps in Figure 1a and 2a in the pdf file attached to the global response. The confusion arose because in the original manuscript, we have plotted the x-axis by episode, which consists of number of updates. In the corrected plot, we plotted the x-axis with number of updates. The initial point is not the convergence point because the error is larger than one, as can be seen in the corrected plot. According to the reviewer's comment, the figures have been replaced with the new ones in the revision, and more details of the figures have been newly added. Moreover, we would like to clarify that our claim does not mean that other algorithms fail while RegQ converges. Our claim is that under certain situations, the convergence rate of RegQ can be faster than that of other algorithms. An intuitive reason why our algorithms shows faster convergence is that the baseline algorithms ( CQL, targetQ, and GreedyGQ ) are basically a two-time scale algorithm whereas our proposed algorithm uses a single-time step-size. Following the reviewer's comment, we will update our manuscript. Q4. In Figure 1c, I can’t find the details about the phases: ‘Blowing up phase’, and ‘Shrinking phase’. I read the author’s error analysis, but there’s no explanation about the phase. Is it just a intuitive naming on the convergence behavior? If not, could you elaborate more about it in the theoretical context? A4. Thank you helping us to improve the clarity of our manuscript. Regarding the terms shrinking and blowing up phases, we have provided a response G2 in the global rebuttal. To clarify this further, the related discussions have been newly added in the revision. References [1] Liu, Shuze, Shuhang Chen, and Shangtong Zhang. "The ODE Method for Stochastic Approximation and Reinforcement Learning with Markovian Noise." arXiv preprint arXiv:2401.07844 (2024). --- Rebuttal 2: Comment: We appreciate the reviewer's time and effort in reviewing our manuscript. If there are any additional concerns, please let us know since the discussion period is coming to its end. Otherwise, we kindly request a re-evaluation of our manuscript.	Summary: This paper proposes a new Q-learning variant with linear function approximation called RegQ, and proves that its ODE form converges (even when associating it with linear function approximations). Q-learning is famously known to be affected by the 'deadly triad': it tends to diverge in practice when combined to off-policyness, function approximation and bootsrapping. Analyzing formally the convergence of Q-learning combined with simple linear function approximations helps understanding more precisely the mechanisms at play in the deadly triad, and paves the way toward RL with more solid foundations. Authors show empirically that RegQ is faster than two related algorithms with guaranteed convergence. Besides, RegQ relies on a single time-scale, while the other baseline algorithms use two time-scales. Strengths: - One of the most appealling properties of the framework proposed in this paper is its simplicity: just regularizing projected Bellman equations allows to obtain convergence proofs without relying on several artificial assumptions. Weaknesses: - In the related work, the paper lists several existing works proving the convergence of Q-learning under linear function approximation with some theoretical assumptions. The assumptions made in the present paper are weaker than most of the ones of existing works, which include restrictions on the Markov chain types, dependency between behavior and target policy, other guarantees than convergence, etc. However, the closest work is by Lee and He (2019), and the present paper does not explicitely state in what way the assumptions are now weaker than in Lee and He. It states that Lee and He's assumption on the behavior and feature matrix seems too stringent to check in practice, but nothing more precise. When Lee and He improved the sufficient condition of Melo et al. (2008), they showed that their new condition was strictly weaker than the previous one, and I believe that a similar analysis should be made. More structure in the comparison would make the paper look less like an incremental modification of Lee and He (2019). - The presented work follows the direction proposed in Lee and He (2019) by reducing the convergence analysis to that of a switching system, establishing simpler upper and lower bound systems and applying the Borkar and Meyn theorem to obtain a proof of the asymptotic convergence. There are of course différences in the approach, but they are in the details. - It can be regretted that the most interesting and novel parts of the paper are in Appendix, which shows that the conference format. might not be the best fit for it. Technical Quality: 3 Clarity: 3 Questions for Authors: - Although there are empirical evaluations, including one on Mountain Car, more ambitious empirical tests with regularized variants of Q-learning-based deep RL algorithms would be interesting. - Could the proposed algorithm fit the framework of regularized MDPs introduced in ["A Theory of Regularized Markov Decision Processes", Geist et al., 2019]? If yes, could results in ["A Theory of Regularized Markov Decision Processes", Geist et al., 2019] be directly applied to the proposed method, including for instance the analysis of the changes on the optimal policy due to regularization (error bound)? Including a discussion on this existing framework in the paper seems relevant. Typos: l9: 'has known to diverge' -> 'was known to diverge' l64: "guarantees convergence" => "guarantee convergence" Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1 In the related work, the paper lists several existing works proving the convergence of Q-learning under linear function approximation with some theoretical assumptions. The assumptions made in the present paper are weaker than most of the ones of existing works, which include restrictions on the Markov chain types, dependency between behavior and target policy, other guarantees than convergence, etc. However, the closest work is by Lee and He (2019), and the present paper does not explicitely state in what way the assumptions are now weaker than in Lee and He. It states that Lee and He's assumption on the behavior and feature matrix seems too stringent to check in practice, but nothing more precise. When Lee and He improved the sufficient condition of Melo et al. (2008), they showed that their new condition was strictly weaker than the previous one, and I believe that a similar analysis should be made. More structure in the comparison would make the paper look less like an incremental modification of Lee and He (2019). A1. We thank the reviewer for the valuable insights. Lee and He (2919) considered the following condition: $$\phi^\top_i D + \phi^\top_i \gamma DP \Pi_{\pi}\sum_{j\in \{1,2,\dots,n\}}\phi_j<0 ,\quad \pi \in \Theta_{\Phi},$$ where $\Theta_{\Phi}:= \{ \pi\in\Theta : \pi(s)=arg\max_{a\in\mathcal{A}} (\Phi\theta)(s,a) ,\forall s \in \mathcal{S},\theta\in\mathbb{R}^n \}$, $\Theta$ is the set of greedy policies, and \begin{align} \phi_i= \begin{bmatrix} \phi_i(1,1) & \phi_i(2,1) & \cdots & \phi_i(\|\mathcal{S}\|,\|\mathcal{A}\|)]^{\top} \end{bmatrix}. \end{align} The above condition is strict because it needs to hold for all the policy $\pi\in\Theta_{\Phi}$. In contrast, we do not require such condition. We only require the regularization coefficient $\eta$ to be larger than certain value, which can be easily met by scaling the feature matrix. In particular, the condition on $\eta$ can be met by scaling the norm of feature matrix to be smaller than one, and $\eta$ only needs to be larger than two. This follows from Lemma 3.3 in the manuscript. Moreover, we note that feature scaling is widely used in practice. Following the reviewer's comments, we will incorporate the discussion in the revised manuscript. Q2 The presented work follows the direction proposed in Lee and He (2019) by reducing the convergence analysis to that of a switching system, establishing simpler upper and lower bound systems and applying the Borkar and Meyn theorem to obtain a proof of the asymptotic convergence. There are of course differences in the approach, but they are in the details. A2 We appreciate the reviewer's constructive feedback. While our approach aligns with the principles outlined by Lee and He (2019), the following points are new in the literature: 1) We established the theoretical conditions for regularization term to ensure convergence of the algorithm; Moreover, as mentioned in A1, simply adding a regularization term allows to weaken the assumptions used in prior works including Lee and He (2019); 2) We characterized the existence and uniqueness of the regularized projected Bellman equation depending on $\eta$; 3) A tight error bound between the solution regularized projected Bellman equation and true solution, $Q^$ is provided. Each of the above points has not been studied previously, and our work provided a thorough analysis for the above points . Following the reviewer's comments, we will clarify this in the revised manuscript. Q3* It can be regretted that the most interesting and novel parts of the paper are in Appendix, which shows that the conference format. might not be the best fit for it. A3 We thank the reviewer for the valuable comments. As the reviewer mentioned, the proofs are deferred to the Appendix due to the space limit. However, we highlight that our contribution lies is the analysis of the properties of RPBE and the development of RegQ. We have provided thorough analysis of RPBE in Section 3, and have further elaborated on this, including additional details on Figure 1c of the manuscript in G1 of the global rebuttal. Additionally, as answered in A1 and A2, we will clarify the points regarding the analysis of RegQ. Following the reviewer's comments, we have incorporate these clarifications in the revised manuscript. Q4 Although there are empirical evaluations, including one on Mountain Car, more ambitious empirical tests with regularized variants of Q-learning-based deep RL algorithms would be interesting. A4 We thank the reviewer for the insightful comments. In the revised manuscript, we will consider more ambitious empirical evaluations on the variants of Q-learning-based deep RL algorithms. Q5 Could the proposed algorithm fit the framework of regularized MDPs introduced in ["A Theory of Regularized Markov Decision Processes", Geist et al., 2019]? If yes, could results in ["A Theory of Regularized Markov Decision Processes", Geist et al., 2019] be directly applied to the proposed method, including for instance the analysis of the changes on the optimal policy due to regularization (error bound)? Including a discussion on this existing framework in the paper seems relevant. A5 Thank you for the insightful comments. The domain of the regularization term used in Geist et al., 2019 is restricted to a probability simplex, for example an entropy regularization. The work focuses on the regularization over the policy space instead of the Q-function, and also focuses on the policy iteration algorithms instead of reinforcement learning or Q-learning. The $l$-2 type regularization we consider does not fall into this category, and the extension is non-trivial. We will incorporate the discussion in the revised manuscript. --- Rebuttal Comment 1.1: Comment: Thank you for these answers. As I still consider the work an incremental modification of Lee and He 2019, I keep my recommendation of borderline acceptance. --- Rebuttal 2: Comment: Q1. As I still consider the work an incremental modification of Lee and He 2019 A1 We thank the reviewer for the valuable feedback. We want to note that our contribution is not only proving the convergence of RegQ but also lies in thorough investigation of theoretical properties of regularized projected Bellman equation (RPBE), which has been explained in A2 in the initial response. We have provided thorough theoretical investigation of existence, uniqueness, and quality of the solution of RPBE. This is a unique contribution of our work and has not been presented in Lee and He 2019. Following the reviewer's comment, we will add the discussion in the revised manuscript. We thank the reviewer for the engagement in the discussion period, and providing constructive comments to improve the quality of the paper.	Summary: The paper introduces a new regularized Q-learning algorithm "RegQ" suitable for linear function approximation, which essentially adds an $\ell^2$ regularization term to the TD error in semi-gradient Q-learning. The authors prove that this addition ensures convergence of the algorithm and analyze the error with respect to the unregularized solution. Strengths: The paper centers on an important issue in reinforcement learning: the deadly triad, which is the failure of off-policy TD algorithms when combined with function approximation. This issue has been addressed in practical (deep) RL by expensive methods such as target networks. This paper proposes a simpler solution (specifically for the case of linear function approximation): regularization of the TD error. The paper is very clearly written and provides extensive sections on the background and related work. Their assumptions seem reasonable, and their analysis is rigorous. The overall contribution is highly significant. Weaknesses: 1. The biggest weakness of this paper is the limited experiment section. I understand that this is a theoretical work, and of course don't expect any large-scale experiments. However, the two experiments they do list are not properly explained. (What are these baseline algorithms?) The takeaway from the experiments is that their method has a faster convergence rate, but it is not explained why. The convergence rate is not mentioned elsewhere in the paper, where the focus lies on proving convergence, where other methods do not converge! It would be great to show an environment where RegQ converges while the baseline methods do not, or where the error of RegQ's approximate solution $\theta_\eta^\star$ is smaller than the baselines'. 2. Please comment on Assumption 2.2 (orthogonality of columns of $X$). Could it be relaxed? In high-dimensional ($\|\mathcal S\|\|\mathcal A\|$) spaces, $h \ll \|\mathcal S\|\|\mathcal A\|$ _random_ vectors are nearly orthogonal with high probability. Could this be used to show that the result will hold with high probability when using random features? 3. Figure 1c is not properly explained. What are the "shrinking" and "blowing up" phases? Why does the vector $x$, located on the _unit circle_ have norm $0$? 4. For Lemma 3.5, it is assumed that $X^\top D X = aI$. I understand that this is just an example of when a solution to the RPBE exists, but it should be clarified that this assumption is very unrealistic (earlier you wrote that $h \ll \|\mathcal S\|\|\mathcal A\|$, which is a contradiction). 5. In line 245, you state that, as $\theta^\star = \theta_\eta^\star$ if $\eta = 0$, it holds that $\theta^\star \to \theta_\eta^\star$ if $\eta \to 0$. You are implicitly assuming that $\theta_\eta^\star$ is a continuous function of $\eta$ (at $\eta = 0$), which you should at least mention, if not prove. 6. (minor) In the introduction, you do not mention at all that you also analyze the error $\theta_\eta^\star - \theta^\star$, which to me is a very important part of your work, and would not be out of place in the "summary of main contributions" at the end of your introduction. 7. (minor) As you talk about the deadly triad and how it has been addressed practically in deep RL, you might also want to cite "Deep reinforcement learning and the deadly triad" by van Hasselt et al. (2018). 8. (typo) In line 215, you swap (9) and (4), changing the meaning in a significant way. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Where does equation (6) come from? To solve an equation like $A\theta = b$ iteratively, I would construct a loss function $L(\theta) = \frac{1}{2}\|\|A\theta - b\|\|^2$ and do gradient descent: $\theta_{k+1} = \theta_k - \alpha_k\nabla L(\theta_k)$, where $\nabla L(\theta_k) = A^\top(A\theta - b) + E$, with $E$ containing additional terms if $\nabla A \neq 0$. Equation (6) looks similar at first but is in fact quite different. Could you explain how you arrived at equation (6)? 2. How should equation (15) be interpreted? How "large" does $\eta$ actually have to be to ensure convergence? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors adequately address the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful feedback, and the time and effort for reviewing our paper. Following the reviewer's comments, we have added the related discussion in the revised manuscript: Q1 The biggest weakness of this paper is the limited experiment ... . The two experiments they list are not properly explained. (What are these baseline algorithms?) The takeaway from the experiments is that their method has a faster convergence rate, but it is not explained why. ... It would be great to show an environment where RegQ converges while the baseline methods do not, or where the error of RegQ's approximate solution is smaller than the baselines. A1. Thank you for the insightful comments. The baseline algorithms are Q-learning variants that converges under linear function approximation, which are explained in Appendix Section D. The algorithms are Coupled Q-learning (CQL), GreedyGQ, and algorithms with target-network update (targetQ). As the reviewer suggested, we have newly provided an example, where RegQ converges while one of the baseline algorithm, CQL, does not converge in the PDF file attached in the global response. This is because CQL requires the norm of the feature matrix to be smaller than one. Meanwhile, we could not verify practical examples showing divergence of GreedyGQ or targetQ while RegQ diverges or vice-versa. This is because the Q-learning variants are developed to guarantee convergence to some solutions. However, we summarized the difference on the theoretical points in G1 in the global response. Lastly, the experiment results show that RegQ has a faster convergence rate under some scenarios. An intuitive reason for this is because the baseline algorithms are basically a two-time scale algorithms whereas RegQ uses a single-time step-size. A detailed exploration of convergence rates is a promising direction for future research. Following the reviewer's comments, we have added the related discussions in the revised manuscript. Q2 Please comment on Assumption 2.2 (orthogonality of columns of $\Phi$). Could it be relaxed? A2 Thank you for the insightful comments on the our paper. The assumption is required for the construction of comparison systems in the switched system analysis, and it seems that relaxing this condition in the current phase is non-trivial. However, the orthogonality assumption is not very restrictive and can be met easily in practice. For instance, we can use orthogonal Fourier basis functions as feature functions. Additionally, it may be possible to develop more advanced techniques in the future that can relax this assumption through coordinate transformations. This could be an interesting agenda for future research. Moreover, as the reviewer suggested, initialization of random vectors can guarantee such condition with high probability. We believe the suggested arguments can be used to justify the assumption. Q3 In Figure 1c, what are the "shrinking" and "blowing up" phases? Why does the vector $x$, located on the unit circle have norm $0$? A3 Thank you for the careful investigation of our manuscript. This a typo and the correct expression is $\|\|x\|\|_{\infty}=1$. We will correct this in the revised manuscript. Regarding the terms shrinking and blowing up phases, we have provided a response G2 in the global rebuttal. Q4 For Lemma 3.5, it is assumed that $X^\top D X = aI$. .. just an example ... but it should be clarified that this assumption is very unrealistic (earlier you wrote that $h\ll \|{\cal S}\|\|{\cal A}\|$, which is a contradiction). A4 Thank you for pointing out the issue. Following the reviewer's comments, we will clarify this point in the manuscript that such cases are unrealistic and differ from the case of $h\ll \|{\cal S}\|\|{\cal A}\|$. Q5 In line 245, you state that, as $\theta^=\theta_{\eta}^$ if $\eta=0$, it holds that $\theta^\to \theta^_{\eta}$ if $\eta\to 0$. You are implicitly assuming that is a continuous function of $\eta$ at 0. A5 We have provided a proof in G3 in the global response. We would like to note that before guaranteeing the continuity of $\theta^_{\eta}$, it should at least exist in a neighborhood of $\eta = 0$. Therefore, to guarantee the existence of $\theta^_{\eta}$ around $\eta = 0$, we have added the condition $\gamma\|\|\Gamma\|\|_\infty<1$. Then, under the existence, we can prove the continuity of $\theta^_{\eta}$ at $\eta=0$. Moreover, this result can be easily extended to the case that the solution $\theta^_{\eta}$ exists for a positive $\eta > 0$. Q6 Where does equation (6) come from? A6 Equation (6) is not a gradient of any objective function. It iteratively solves an equation $A\theta=b$ by $\theta_{k+1}\leftarrow \theta_k + \alpha(A\theta_{k}-b)$., which is a widely used algorithm called a Richardson iteration [1]. It simply updates by the difference of the left and right term of the equation. Q7 How should equation (15) be interpreted? How "large" does $\eta$ actually have to be to ensure convergence? A7 In equation (15), the terms (S1) and (S2) correspond to the condition for the stability of the switched system in each item of Lemma 2.5 to be satisfied, respectively. As noted in Appendix A.15, each condition covers different scenarios. However, the bound can be always chosen to be small because it only needs to be larger than the minimum of two quantity, (S1) and (S2). As show in Lemma 3.3, $\eta>2$ only needs to be met if we use feature scaling, which is widely used in practice. However, we note that there are several cases when (S2) can be smaller than (S1), as noted in the Appendix A.15. We thank again the reviewer for the constructive comments. Following the reviewer's comments, we have incorporated the above discussions in the revised manuscript. References [1] Kelley, Carl T. Iterative methods for linear and nonlinear equations. Society for Industrial and Applied Mathematics, 1995. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response. I will keep my score as is.	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewers’ constructive comments for our manuscript. The comments are valuable for improving the quality of our paper and provide important guidance for our research. In the following, we address the concerns commonly raised by the reviewers. G1 Reviewers yA4E and x1RW raised concerns regarding the comparison with existing works that guarantee the convergence of Q-learning with linear function approximation under mild conditions. The algorithms compared include Coupled Q-learning (CQL), GreedyGQ, and algorithms with target-network updates (targetQ). 1) GreedyGQ algorithm is guaranteed to converge to some solution. However, the solution is not a solution of the projected Bellman equation, of which its quality is difficult to quantify. However, we provide a tight bound for our solution in the error bound in Lemma 3.7. Moreover, it is a two-time scale algorithm which is known to be slower than a single-time scale algorithm like ours. 2) CQL is also a two-time scale algorithm whereas our algorithm is a single-time scale algorithm. As noted previously, two-time scale algorithms are often known to be slower than single-time scale algorithm. This is demonstrated experimentally in Section 6 in the manuscript. Moreover, as can be in the Figure 1c in the PDF file attached in the global response, the algorithm is sensitive to the scaling of the feature matrix, and the solution is not a solution of the projected Bellman equation, of which its quality is difficult to quantify. 3) To guarantee the convergence of targetQ (which uses target network update), it requires a projection or truncation method. This causes additional complexity in its implementation. The resulting solution lacks interpretability, as it may lie on the boundary of the projection or truncation ball. Lastly, the target network update can slow down the convergence rate, which can be verified in Section 6 in the manuscript. G2 Reviewers yA4E and cMMt raised concerns regarding the terminology of shrinking and blowing phase in Figure 1(c). We consider the scenario that $\|\| \gamma\Gamma \|\|\_{\infty} < 1$ holds, which implies that $\|\| \gamma \Gamma_0 \|\|\_\infty < 1$ because $\Gamma = \Gamma_\eta$ when $\eta = 0$. The figure implies that as $\eta\to\infty$, $\gamma \Gamma_\eta$ can potentially move outside of the unit ball, and this phase is indicated with the term blowing up phase. However, since $\lim\_{\eta\to\infty } {\|\| \gamma \Gamma _\eta \|\|\_\infty } = 0$, we know that $\gamma \Gamma\_\eta$ will eventually converge to the origin and move inside the unit ball. This behavior is indicated by the shrinking phase in the figure. - Additional proofs G3 (Response to Q5 of Reviewers yA4E ) Lemma ( Continuity of $\theta^_{\eta}$ in terms of $\eta$ ) Let $\eta_0$ be a non-negative real valued constant. Suppose $\gamma\|\|\Gamma_{\eta_0}\|\|\_{\infty}<1$. Then, $\theta^_{\eta}$ is continuous at $\eta_0$. Proof) Note that $\Gamma_{\eta}$ is continuous function of $\eta$, and we have, \begin{align} \Gamma_{\eta_0+\eta} = \Gamma_{\eta_0} + O(\eta), \end{align} where $O(\cdot)$ stands for the big O notation. Therefore, \begin{align} \|\| X\theta^_{\eta_0+\eta}-X\theta_{\eta_0}^\|\|\_{\infty} =& \|\| \Gamma_{\eta_0+\eta}\mathcal{T}X\theta^_{\eta_0+\eta}-\Gamma_{\eta_0} \mathcal{T}X\theta^_{\eta_0} \|\|\_{\infty}\\\\ \leq & \|\|\Gamma_{\eta_0} {\cal T} X (\theta^_{\eta_0+\eta}-\theta^_{\eta_0}) \|\|\_{\infty}+O(\eta)\\\\ \leq & \gamma \|\|\Gamma_{\eta_0} \|\|\_{\infty} \|\|X (\theta^_{\eta_0+\eta}-\theta_{\eta_0}^)\|\|\_{\infty}+O(\eta). \end{align} The first equality follows from the definition of $\theta^_{\eta_0+\eta}$ and $\theta^_{\eta_0}$. The second inequality follows from triangle inequality. The last inequality follows from the contraction property of the Bellman operator. Therefore, we have \begin{align} \|\|\theta^_{\eta_0+\eta}-\theta^_{\eta_0}\|\|\_{\infty} \leq C \|\|X\theta^_{\eta_0+\eta}-X\theta^_{\eta_0}\|\|_{\infty} \leq O(\eta), \end{align} where the first inequality holds because $X$ is full-column rank matrix, and $C$ is a universal constant. This completes the proof. Q.E.D We sincerely appreciate the reviewer's feedback. In response to the comments, we have incorporated the relevant discussions into the revised version. Pdf: /pdf/9f5c289912fe8c65225478f582ad612f5fb156fe.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: Q-learning is a popular RL algorithm. With function approximation, though, it is known that this algorithm can diverge. This issue is attributed to the `deadly triad': off-policy learning, bootstrapping, and function approximation. This work addresses this issue in the context of linear function approximation. Specifically, this work proposes a novel called Regularized Q-learning, in which a suitable regularization term is added to the standard update rule. The key result (Theorem 5.2) is that this modified algorithm almost surely converges. The proof is based on the switching systems. Strengths: S1. The paper introduces a new Q-learning algorithm called Regularized Q-learning (RegQ), which ensures convergence under linear function approximation. This addresses the known instability issue in traditional Q-learning with function approximation. S2. The paper uses the switched system theory to derive RegQ's convergence. Weaknesses: W1. The present work studies only the case of Q-learning with linear function approximation under a fixed behavior policy. This approach is extremely restrictive and practically not very useful since the quality of the resulting policy critically depends on the choice of the behavior policy. Specifically, as stated by Melo, Meyn, and Ribeiro (2008), for the approximate Q-learning algorithm to discover the optimal policy, the behavior policy would need to be close to the optimal policy itself, which is not feasible. For other choices of behavior policy, the policy estimated by the algorithm could be significantly different from the optimal policy. While Lemma 3.7 provides some guarantees, it is unclear how this result relates the greedy policy, with respect to $X \theta_{\eta}^*,$ to the optimal policy. This is why $\epsilon$-greedy exploration is commonly used in practice. However, the current work does not address this important case. Under $\epsilon$-greedy exploration, several recent studies [1] -- [3] have shown that Q-learning with function approximation suffers from various significant issues beyond instability. Notably, this algorithm can converge to non-locally optimal policies, sometimes even the worst, and exhibit policy oscillation. It remains unclear if the regularization term proposed in the present work would effectively address these issues with $\epsilon$-greedy exploration. References: [1] Patterson, A., Neumann, S., White, M. and White, A., 2023. Empirical design in reinforcement learning. arXiv preprint arXiv:2304.01315. [2] Young, K. and Sutton, R.S., 2020. Understanding the pathologies of approximate policy evaluation when combined with greedification in reinforcement learning. arXiv preprint arXiv:2010.15268. [3] Gopalan, A. and Thoppe, G., 2022. Demystifying Approximate Value-based RL with $\epsilon $-greedy Exploration: A Differential Inclusion View. arXiv preprint arXiv:2205.13617. Technical Quality: 3 Clarity: 3 Questions for Authors: L1. The current work only studies the setting where the $(s_k, a_k, r_{k + 1}, s_{k + 1})_{k \geq 0}$ sequence is sampled in an IID fashion in each iteration. Do you think the results carry over to the scenario with Markovian samples? L2. Are there any realistic examples of the feature matrix $X$ for which the condition in (11) is guaranteed? Minor issues: M1. Line 90: Shouldn't you emphasize what s_0's distribution is? M2. Line 96: Which Markov chain is assumed to be time-homogeneous? M3. Line 176: "... true action value may not lie..." Do you mean the "optimal" value function may not lie in the subspace...? M4. Line 181: `In this case, there are more chances...' Do you have any evidence for this statement? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes, the authors have discussed the limitations of their work. However, there are more serious issues with the work which I have highlighted. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1 The work studies only the case of a fixed behavior policy. This approach is extremely restrictive and practically not very useful .... Specifically, as stated by Melo, Meyn, and Ribeiro (2008), ..., the behavior policy would need to be close to the optimal policy, ...., the policy estimated by the algorithm could be significantly different from the optimal policy. While Lemma 3.7 provides some guarantees, it is unclear how this result relates ... to the optimal policy. .... , the current work does not address $\epsilon$-greedy exploration. Under $\epsilon$-greedy exploration, several recent studies ... have shown that Q-learning ... suffers from various issues beyond instability. .... It remains unclear if the regularization term proposed in the present work would effectively address these issues with $\epsilon$-greedy exploration. A1 We thank the reviewer for the insightful comments. In the following, we clarify the concerns raised by the reviewer: 1) We would like to first note that our assumption of a fixed behavior policy is standard in the literature [4,5]. It is important to note that even under such fixed behavior policy assumption, the convergence of Q-learning with linear function approximation has not been fully explored yet. We proposed a simple and convergent algorithm compared to previous works that guarantee convergence under linear function approximation. 2) We would like to emphasize that our work does not rely on the assumption used in Melo, Meyn, and Ribeiro (2008). Specifically, we do not require the behavior policy to be close to the optimal policy. In fact, we only require the state-action probability induced by the fixed behavior policy to be non-zero, which is a standard assumption in the literature [4,5]. 3) Lemma 3.7 provides a tight bound on the error of the estimated solution and the true solution, $Q^$. If the error bound with $Q^$ is small enough, then the estimated policy will be close to the optimal greedy policy. As the reviewer suggested, an interesting avenue for future research would be to investigate these properties under an epsilon-greedy behavior policy. Following the reviewer's comments, we have added the discussion in the revised manuscript. Q2 The current work only studies the setting where the state-action sequence is sampled in an IID fashion in each iteration. Do you think the results carry over to the scenario with Markovian samples? A2 Thank you for the valuable comments. We believe our result can be also extended to the Markovian sample cases. In [6], the authors demonstrated an extension of the Borkar and Meyn Theorem to the Markovian sample cases. Since our proof relies on Borkar and Meyn Theorem, our result can also be extended to establish convergence under Markovian samples. According to the reviewer's comment, we have incorporated this result into the revised manuscript. Q3 Are there any realistic examples of the feature matrix for which the condition in (11) is guaranteed? A3 We thank the reviewer for the constructive comments. If we scale the feature matrix such that $\max(\|\|X\|\|,\|\|X^{\top}\|\|)<1$, then choosing $\eta>2$ satisfies the condition (11) by Lemma 3.3 in the manuscript. Scaling the values of feature matrix is a commonly employed technique in the both theoretical literature or in practice. Consequently, the condition in (11) can be easily met in practice. Following the reviewer's comments, we have added the related discussion in the revised manuscript. Q4 l 90 : Shouldn't you emphasize what $s_0$'s distribution is? A4 Thank you for pointing out important insights on the paper. In our analysis, we assume that the state-action distribution to satisfy $d(s_0,a_0)>0$ for all $(s_0,a_0)\in{\cal S}\times{\cal A}$. Moreover, if we assume an ergodic Markov chain, the initial distribution does not matter because for any initial distribution, it will converge to stationary distribution of the Markov chain. We have added the discussion in the main manuscript. Q5 l 96 : Which Markov chain is assumed to be time-homogeneous? A5 Thank you for the insightful comments. A Markov chain is said to be time homogeneous if the transition probability does not change over time. We will clarify this in the revision. Q6 l 176 : "... true action value may not lie..." Do you mean the "optimal" value function may not lie in the subspace...? A6 Thank you for the valuable comments. As the reviewer mentioned, it means that optimal action-value function, $Q^$, may not lie in the subspace. We will clarify this in the revised manuscript. Q7* l 181: `In this case, there are more chances...' Do you have any evidence for this statement? A7 Thank you for carefully investigating our manuscript. This is an intuitive result based on the following facts: The Bellman equation $X \theta^* = {\cal T} X \theta^$ hardly admits a solution because the right-hand side in general does not lie in the column space of $X$ due to the Bellman operator $\cal T$. This situation can be partially alleviated by projecting the right-hand side ${\cal T} X \theta^$ onto the column space of $X$ using the projection operator $\Gamma$: $X \theta^* = \Gamma {\cal T} X \theta^$. According the reviewer's comment, the corresponding discussions will be modified in the revised manuscript in order to further clarify this intuition. References* [4] Sutton, Richard S., et al. "Fast gradient-descent methods for temporal-difference learning with linear function approximation." Proceedings of the 26th annual international conference on machine learning. 2009. [5] Lee, Donghwan, and Niao He. "A unified switching system perspective and ODE analysis of Q-learning algorithms." arXiv preprint arXiv:1912.02270 (2019). [6] Liu, Shuze, Shuhang Chen, and Shangtong Zhang. "The ODE Method for Stochastic Approximation and Reinforcement Learning with Markovian Noise." arXiv preprint arXiv:2401.07844 (2024). --- Rebuttal 2: Comment: We appreciate the reviewer's time and effort in reviewing our manuscript. If there are any additional concerns, please let us know since the discussion period is coming to its end. Otherwise, we kindly request a re-evaluation of our manuscript. --- Rebuttal Comment 2.1: Comment: I can never imagine a scenario where a fixed behavior policy will ever be used. However, the bounds given in Lemma 3.7 seem useful and could perhaps lead to some insights that could be exploited in the future for designing more effective algorithms. Hence, I have decided to change my score from 3 to 5. --- Reply to Comment 2.1.1: Comment: We thank the reviewer for the engagement in the discussion period. As the reviewer mentioned, fixed behavior policy is not a practical assumption. However, considering that this is a standard assumption in the literature in proving convergence of Q-learning or TD-learning, we believe our assumption aligns with the existing literature. As the reviewer mentioned, it would be an important direction to further explore on this topic. Following the reviewer's comments, we will incorporate this into the revised manuscript. We thank the reviewer for the time and effort in reviewing our paper.	null	null	null	null	null	null
VFIMamba: Video Frame Interpolation with State Space Models	Accept (poster)	Summary: Based on the popular S6 model's advantages of linear computational complexity and data-independent modelling capability, this paper applies it to VFI. Specifically, this paper proposes a token rearrangement strategy to learn the information of adjust frames, in addition to introducing a curriculum learning strategy to dynamically learn various motion magnitudes between adjust frames through joint training of vimeo90K and X-TRAIN. The model achieves the highest performance on existing commonly used VFI datasets. Strengths: 1. This paper is the first to adapt the S6 model to the VFI. 2. Experimental results show that the proposed VFIMamba achieve performance while using competitive FLOPs. Weaknesses: 1. It is recommended that the authors can demonstrate the accuracy of the interpolation results more intuitively by visualizing the error maps. 2. The experiments on table 4 are to demonstrate the computational validity of S6 model, and it is necessary to provide flops for the different models. 3. In limitations, the author says that the VFIMamba has faster speed, I would like to know the comparison of the runtime in table 2. 4. Lack of citations essential to the field, such as VFIT, TTVFI, ABME, EDSC, etc. Technical Quality: 3 Clarity: 3 Questions for Authors: I would like to see more analysis on the model effectiveness, such as the runtimes in Table 2. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors discuss methodological limitations in the supplementary material. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for your recognition and feedback on our work. We would like to respond to your concerns as follows: Q.1 Visualization of error maps R.1 Thank you for your suggestion. Visualizing the error maps can indeed provide a more intuitive demonstration of the accuracy of the interpolation results. We have provided an illustrative example in Figure 11 of the global response PDF, where VFIMamba shows clear advantages over other methods. We will add more comparative examples in the final version. Q.2 FLOPs of Table 4 R.2 Thank you for your suggestion. The FLOPs of the different models in Table 4 are reported in the table below. We will include this in the final version. \| Model \| FLOPs(T) \| \| --- \| --- \| \| w/o S6 \| 0.23 \| \| Convolution \| 0.27 \| \| Local Attention \| 0.23 \| \| Full Attention \| 0.59 \| \| S6 \| 0.24 \| Q.3 Runtime comparison R.3 Thank you for your suggestion. We have provided the runtime comparison in Table 2 of the global response PDF, and we welcome you to check that. We would also like to kindly remind you that in the limitations, we stated that VFIMamba is faster compared to the attention methods like SGM, with a runtime of 311ms vs. 942ms for 1024x1024 inputs. Our primary goal was to achieve high-performance video frame interpolation while maintain efficient processing, rather than solely pursuing the fastest runtime. We have discussed in detail how to further improve the runtime of our method in the response to Reviewer Bwpp's Q.1. Q.4 Additional citations R.4 Thank you for the reminder and these works are indeed important for the VFI task. We will add these citations in the final version. --- Rebuttal Comment 1.1: Comment: Thanks to the author's reply, I raise my score to Accept. Looking forward to the author's future work to further address on accelerating SSM. --- Reply to Comment 1.1.1: Comment: We will continue to make effort to enhance the efficiency of our model in the future. Thank you so much for your kind recognition of our work.	Summary: This paper introduces a novel video frame interpolation (VFI) method called VFIMamba. VFIMamba is the first method that combines the State-Space Model Mamba with VFI architectures and therefore, it has the advantage of a linearly growing complexity w.r.t. the resolution while maintaining the ability to utilize global receptive fields similar to vision transformers. In order to apply the idea to VFI, architectural modifications have been proposed to handle 2 frames as input. Further, a novel curriculum learning strategy is used to increase the models performance across a large range of motions. The method has been evaluated on multiple datasets and w.r.t. to various other methods, achieving state-of-the-art PSNR values, especially improving the performance for high-resolution frames (2K and 4K). Strengths: - Combination of an emerging alternative to transformers (MAMBA) with VFI methods that allow for higher resolution frame interpolation due to linear complexity growth. The paper is interesting to read. - Discussion and evaluation of different sequence arrangements. - Introduction of a relatively simple curriculum learning strategy for VFI with experiments confirming that this strategy is beneficial for VFI methods in general when having to deal with large motions. - Reaches new state-of-the-art performance - Exhaustive ablation studies proving the effectiveness for each of their introduced modules - Supplement contains a video with qualitative comparisons, although only a few short sequences and given the coarse time steps, it is difficult to clearly judge the temporal consistency. Weaknesses: - It is unclear how frames at arbitrary time steps are computed to perform 8x interpolation in table 3. It should also be directly clear from the caption of table 3, that 8x interpolation is evaluated (these details are only in the appendix). - In general, a lot of important information is in the appendix. It is helpful to add at least a reference from the main paper to the appendix that there is more information, e.g., such as the evaluation and the experiment on generlalization of curriculum learning. - Unclear how the FLOPs requirement scales with resolution. A plot/table showing the FLOPs compared to the input resolution for VFIMamba and other methods would be interesting to get a better feeling for the scaling w.r.t. frame resolution and a comparison of memory footprints might be interesting. - Ideally, it should be mentioned in table 2 and 3 on which dataset the other methods have been trained on, especially for methods where the original paper proposed multiple versions such as XVFI. - Some recent methods especially for high-resolution data are missing in Table 3 such as [A] and [B]. Especially [A] has only been included in Tab. 2 although their focus is also on high-resolution datasets and code for X-Test is directly available: https://github.com/feinanshan/M2M_VFI/blob/main/Test/bench_xtest.py [A] Hu et al. Many-to-many Splatting for Efficient Video Frame Interpolation, 2022 [B] Nottebaum et al., Efficient Feature Extraction for High-resolution Video Frame Interpolation, 2022. It would be nice to have longer sequences, and playing them a bit more fluently to get a feeling for the temporal and visual consistency. Minor: - Some colors difficult to see in overlay in Figure 1. - L. 126, Check sentence “where contains” Technical Quality: 4 Clarity: 3 Questions for Authors: - Table 3: There is a discrepancy between many of the reported number in their respective original papers and in this paper on X-TEST for 2k and 4k, e.g., EMA-VFI-S reported 30.89dB instead of 29.91dB and BiFormer reported 31.32dB instead 31.18. Based on the experience of the reviewer it is possible to reproduce these numbers using the correct evaluation protocol. - How is 8x interpolation performed? - Unclear why there is no FLOPs measurement and no evaluation of SNU-FILM for SoftSplat given that the source code + trained models are publicly available (https://github.com/sniklaus/softmax-splatting) Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Discussion of limitations is only in the appendix. It would be better to discuss them already in the main paper. Additionaly, has the method similar limitations as MAMBA (the base model for this entire work). Therefore, the compute requirements are still relatively high even though they do not have to compute an attention matrix anymore. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive and constructive suggestions. We have the following responses to your concerns: Q.1 How is 8x interpolation performed? R.1 As mentioned in line 497, we followed the testing procedure of FILM [1] and used an iterative approach for frame interpolation. Specifically, we first generated an intermediate frame based on the input two frames, and then, using a divide-and-conquer strategy, we further divided the first frame and the generated intermediate frame, as well as the generated intermediate frame and the last frame, to iteratively generate the remaining frames. Thank you for the reminder, and we will further emphasize the specific testing procedure in the main text. Q.2 A lot of important information is in the appendix R.2 We sincerely appreciate your suggestion and will incorporate the mentioned content into the main text in the final version, with more references to the appendix. Q.3 How the FLOPs/memory scales with resolution R.3 We have included the comparison of FLOPs and memory usage of different methods at various resolutions in the global response PDF. As shown in Figure 10, our method has a clear advantage in FLOPs and memory usage at high resolutions. Q.4 It should be mentioned in Table 2 and 3 on which dataset the other methods have been trained on R.4 Thank you for your suggestion. We have updated the information in Tables 2 and 3 in the global response PDF to include the specific training datasets for each method. We will also add this information in the final version. Q.5 Some recent methods especially for high-resolution data are missing in Table 3 R.5 Thank you for the reminder. We have added the results of these two papers under the same testing procedure in Table 3 of the global response PDF, and we will also include them in the final version. Q.6 A discrepancy between many of the reported numbers in their respective original papers and in this paper on X-TEST for 2k and 4k R.6 First, we would like to respectfully remind you that the results in different papers may have used different quantization methods (whether to round the network output) and test functions (the SSIM function in sklearn and the MATLAB-style SSIM can produce different results). To make a fair comparison, we have re-tested all the open-source models under the same testing procedure, which may lead to minor differences in performance compared to their original results. As for EMA-VFI-S, it used a model trained on the Vimeo90K septuplet dataset for testing X-TEST, while our method and all the other methods used models trained on the triplet dataset for testing. For the fair comparison, we tested the EMA-VFI-S model trained on the triplet dataset using our iterative frame interpolation method as in Q.1, resulting in a performance of 29.91 dB instead of 30.89 dB. Regarding the performance of BiFormer, BiFormer did not publicly release its test code and we used the test code of SGM [2]. The results were consistent with the SGM, so we initially thought we had successfully reproduced the correct results. Unfortunately, after careful examination and comparison, we identified an inconsistency between the input processing of SGM's testing procedure and BiFormer's training procedure, making the results of BiFormer lower than it should be. After re-conducting the proper testing, we also obtained a result of 31.32 dB, and we have also updated the results of BiFormer on other datasets in Table 3 of the global response PDF. We also rechecked all the results in Table 3 and ensured their accuracy. We are truly grateful for your reminder, and we will also remind the authors of SGM to correct the BiFormer results. Q.7 Performance of SoftSplat R.7 Thank you for the reminder. We sincerely apologize for not finding that SoftSplat has already been open-sourced. The performance of Softsplat under the same testing procedure has been updated in Table 2 of the global response PDF, and you are welcome to check it. Q.8 Minor errors R.8 We will further improve the color distinction in Figure 1 and carefully check the grammatical details of the paper. Thank you for the reminder. > [1] Reda, Fitsum, et al. "Film: Frame interpolation for large motion." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022. > [2] Liu, Chunxu, et al. "Sparse Global Matching for Video Frame Interpolation with Large Motion." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. --- Rebuttal Comment 1.1: Comment: I thank the authors for the answers, and highly appreciate the extensive additional results and verification of the numbers. Minor detail, it would be nice if in the final version the FLOPs and runtime will also be added for 2K/4K. The additional provided details on training and additional ablations provided in the rebuttal pdf confirm the contribution of the paper and I still recommend the paper for acceptance. --- Reply to Comment 1.1.1: Comment: We will be sure to further polish the final version based on your helpful suggestions. Thank you so much for your kind recognition of our work.	Summary: The paper presents a novel approach for video frame interpolation using Selective State Space Models (S6). The authors introduce VFIMamba, a method designed to efficiently and dynamically model inter-frame information. This method features the Mixed-SSM Block (MSB), which rearranges tokens from adjacent frames in an interleaved manner and applies multi-directional S6 modeling. Additionally, the paper proposes a curriculum learning strategy to progressively improve the model's ability to handle varying motion magnitudes. Experimental results show that VFIMamba achieves state-of-the-art performance on various benchmarks, especially in high-resolution scenarios. Strengths: Originality: The introduction of the S6 model into video frame interpolation tasks is a novel contribution. The use of Mixed-SSM Blocks and the interleaved token arrangement are creative solutions to enhance inter-frame modeling. Quality: The paper presents thorough experiments and comparisons with state-of-the-art methods. The quantitative results demonstrate significant improvements in performance, particularly in high-resolution and large-motion scenarios. Clarity: The methodology is clearly explained, with detailed descriptions of the proposed model components and training strategies. The visualizations and tables effectively support the claims made in the paper. Significance: The VFIMamba method addresses key challenges in video frame interpolation, such as the need for large receptive fields and efficient computation. Weaknesses: Real-time Application: Although VFIMamba achieves high performance, it still falls short of real-time requirements. The paper could benefit from a discussion on potential strategies to improve inference speed. Performance Gap Analysis: On low-resolution datasets, VFIMamba-S fails to yield second-best scores on most benchmarks and only outperform baselines with comparable FLOPs (e.g., EMA-VFI-S) by a relatively small margin. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What specific optimizations could be applied to VFIMamba to make it suitable for real-time applications? Are there trade-offs between speed and accuracy that need to be considered? 2. Could the authors elaborate on the specific factors contributing to the performance gap, and suggest potential directions to address these issues? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have acknowledged several limitations of their work, including the resource-intensive nature of training VFIMamba and the current inability to meet real-time requirements. The authors suggest future work on designing more efficient SSMs and exploring the application of SSMs in the frame generation module. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for the recognition and suggestions regarding our work, and our responses to your questions are as follows: Q.1 What specific optimizations could be applied to VFIMamba to make it suitable for real-time applications? Are there trade-offs between speed and accuracy that need to be considered? R.1 Thank you for this insightful question. The running speed of SSMs is mainly related to two aspects: the length of the input sequence and the number of scans performed on the sequence. Regarding the sequence length, we can reduce the number of tokens by fusing the neighboring tokens in the spatio-temporal position (e.g., through ToMe [1]). However, this may also have an impact on the performance of fine-grained spatio-temporal modeling. As for the number of scans, we currently follow VSSM [2] and use 4 scan directions for modeling. In the future, we can reduce the number of scans by dynamically selecting the necessary scan directions for video frame interpolation. Q.2 Could the authors elaborate on the specific factors contributing to the performance gap, and suggest potential directions to address these issues? R.2 Thank you for your question. First, we would like to kindly remind you that except for EMA-VFI-S, VFIMamba-S has a significant performance advantage over other models with similar FLOPS (AdaCof, XVFI, M2M, RIFE, etc.). As for why the performance gap with EMA-VFI-S cannot be widened, EMA-VFI-S belongs to the local attention-based method, and as shown in Table 1 of the paper, the biggest advantage of Mamba over local attention-based methods is the global receptive field, which may not have a significant impact in low-resolution cases. In the future, perhaps combining SSMs with some more fine-grained local modeling methods can further improve the performance at low resolutions. > [1] Bolya, Daniel, et al. "Token merging: Your vit but faster." arXiv preprint arXiv:2210.09461 (2022). > [2] Liu, Yue et al. “VMamba: Visual State Space Model.” ArXiv abs/2401.10166 (2024). --- Rebuttal Comment 1.1: Comment: Thanks to the authors' reply, I raise my score to Accept since I am satisfied with the authors' response regarding performance gap. I am looking forward to the author's future work to further address on accelerating SSM. --- Reply to Comment 1.1.1: Comment: We are truly grateful for your kind recognition of our work.	Summary: The paper introduces Mamba-based video frame interpolation. To fully incorporate the power of Mamba, the paper proposes an interleaving rearrangement method. Using this method, the SSM scans the same location tokens of 2 frames together instead of processing each frame separately. The paper also proposes curriculum learning with large motion data. The experiments show the improvement of using these ideas. Strengths: This paper includes three contributions: 1) Use of Mamba for image interpolation, 2) Interleaved rearrangement, and 3) curriculum learning. - The idea is simple but effective for video frame interpolation. All these contributions are justified throughout the paper (especially Sections 3 and 4). - The paper is easy to follow. Weaknesses: I have comments and questions about the final comparisons (Tables 2 and 3). >Most methods training models exclusively on the Vimeo90K (Xue et al., 2019). .... (2) Sequential Learning: To mitigate the limitations of training solely on Vimeo90K, some methods (Liu et al., 2024a; Park et al., 2023) further train the model on X-TRAIN (Sim et al., 2021) after initial training on Vimeo90K. - Based on the description above, the proposed model used both Vimeo90K and X-TRAIN for training, but not all methods use both datasets. It's unclear which models are trained only on Vimeo90K or Vimeo90K+X-TRAIN. The improvement can be due to the increase of training datasets especially since the improvement is marginal compared to some models such as EMA-VFI (for low resolution), AMT-L, AMT-G, and SGM-VFI. I understand that curriculum learning improves performance compared to the simple data mix, but the current tables are not comparable if the trained datasets are not the same across the models. - I ask authors to add which models are trained only on Vimeo90K or Vimeo90K+X-TRAIN in the tables. - If some comparable models are trained only on Vimeo90K, can authors add the comparisons of these models with Vimeo90K+X-TRAIN? I'm not sure if it's possible during the rebuttal period. - As some comparing models are better or similar to the proposed model, it's difficult to compare which ones are better. Adding the average of each model on the right column would be helpful. - How can the interleave rearrangement be efficiently implemented? It would be good to add details of how it is implemented in the paper. Also, is there any speed issue of using the interleave rearrangement instead of the sequential one? Technical Quality: 3 Clarity: 3 Questions for Authors: See the weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The limitation is discussed in the paper and is reasonable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your feedback on the experimental and implementation details of our work. We would like to provide the following responses: Q.1 Regarding which models are trained only on Vimeo90K or Vimeo90K+X-TRAIN in the tables. R.1 Thank you for your suggestion. We would like to kindly remind you that the recent published papers (e.g., BiFormer [1] and SGM [2]) also did not specify the exact training dataset used for the compared methods. We have followed their settings, but we agree that adding this detail will make the comparison more informative for the readers. We will add the relevant information to Tables 2 and 3 as shown in the PDF of the global response. Q.2 Add the comparisons of these Vimeo90K-only models trained on Vimeo90K+X-TRAIN. R.2 As discussed in Table 6 and Appendix A.4 of the paper, we have already retrained the RIFE [3] and EMA-VFI-S [4] models on Vimeo90K+X-TRAIN. The results show that through curriculum learning, these two methods have achieved performance improvements on the high-resolution dataset, while maintaining their performance on Vimeo90K, demonstrating the generalization of our proposed curriculum learning training strategy. Meanwhile, VFIMamba still exhibits the best performance under the same training setting, indicating the higher upper bound of the VFIMamba structure. Q.3 Add the average of each model on the right column. R.3 Thanks for your suggestion. We will add the relevant information to Tables 2 and 3 as shown in the PDF of the global response. Q.4 Regarding the implementation of the interleave rearrangement. R.4 The implementation of the interleave rearrangement is very simple and efficient, which can be achieved through the following few lines of PyTorch code: ```python def interleave_merge(self, x1, x2): # x1 is the first frame, x2 is the second B, C, H, W = x1.shape N = H * W # Flatten the 2D tokens and transpose to B x N x C format x1 = x1.view(B, C, -1).transpose(1, 2) x2 = x2.view(B, C, -1).transpose(1, 2) # Concatenate in the C dimension to get B x N x 2C, and then reshape to get B x 2N x C, and PyTorch will automatically interleave the tokens of the two frames x = torch.cat([x1, x2], dim=-1).reshape(B, 2*N, C) # Reshape back to B x C x 2N return x.transpose(1, 2).contiguous() ``` Both the interleave rearrangement and sequential rearrangement only involve reshaping and rearranging the token, which are linear-time operations. Therefore, the running efficiency of interleave rearrangement and sequential rearrangement can be considered approximately the same. > [1]Park, Junheum, Jintae Kim, and Chang-Su Kim. "Biformer: Learning bilateral motion estimation via bilateral transformer for 4k video frame interpolation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. > [2]Liu, Chunxu, et al. "Sparse Global Matching for Video Frame Interpolation with Large Motion." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. > [3]Huang, Zhewei, et al. "Real-time intermediate flow estimation for video frame interpolation." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022. > [4]Zhang, Guozhen, et al. "Extracting motion and appearance via inter-frame attention for efficient video frame interpolation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. --- Rebuttal Comment 1.1: Title: response to the rebuttal Comment: Thank you for the answers. My major concern was the unfair comparisons in Tables 2 and 3 as the paper uses an extra dataset. The authors' answers and additional information in Tables 2 and 3 (also Table 6) resolve my concern. Please add these details to the revised version. Also, please move Table 6 to the main paper; it's an important ablation. As the paper includes clear contributions and improvements, I increase my rating. --- Reply to Comment 1.1.1: Comment: We will be sure to incorporate your suggestions into the final version. Thank you so much for your kind acknowledgement of our work.	Rebuttal 1: Rebuttal: We sincerely appreciate all reviewers' efforts in reviewing our paper and giving insightful comments as well as valuable suggestions. We are glad to find that the reviewers generally acknowledge the following contributions of our work. * Framework. We are the first to adapt the S6 model to the VFI task [YfRu,A9TL,Bwpp,9Hn5] for addressing key challenges in video frame interpolation, such as the need for large receptive fields and efficient computation. The use of Mixed-SSM Blocks and the interleaved token arrangement are effective solutions for enhancing inter-frame modeling [A9TL,Bwpp,9Hn5]. * Training strategy. We propose a novel curriculum learning strategy that is beneficial for VFI methods when having to deal with large motions [A9TL, 9Hn5]. * Experiments. Our experimental results achieve state-of-the-art performance on most datasets, particularly in high-resolution and large-motion scenarios [Bwpp,9Hn5,YfRu]. Furthermore, we have conducted exhaustive ablation studies to demonstrate the effectiveness of our proposed method [A9TL, Bwpp, 9Hn5]. As suggested by the reviewers, we include the following contents in the revised manuscript to further polish our paper. The major revision is summarized as follows. Our detailed responses can be found in each response section to the reviewers. * Adding more detailed information to the main comparison Table 2 and Table 3. This includes: 1) The training dataset for each method [A9TL, 9Hn5]. 2) The average performance of each model [A9TL]. 3) Results for M2M, FLDR, BiFormer and SoftSplat [9Hn5]. 4) Runtime [YfRu]. We have provided the updated Table 2 and Table 3 in the PDF below. * More visualization comparisons. This includes visualizations on how the FLOPs/memory requirement scales with resolution [9Hn5], as well as error maps [YfRu]. We have provided these in the PDF below. * Transferring important information from the appendix to the main paper. This includes detailed testing procedures and the experiment on the generalization of curriculum learning [A9TL, 9Hn5]. Pdf: /pdf/fa2b56b1c597ac7cf46496663138b522d89c089e.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
FNP: Fourier Neural Processes for Arbitrary-Resolution Data Assimilation	Accept (poster)	Summary: This paper proposes an innovative Fourier Neural Process (FNP) model designed to address the limitations of existing data assimilation methods when handling observational data of varying resolutions. The FNP model combines the characteristics of neural processes and Fourier transforms to effectively assimilate observational data of arbitrary resolutions. Extensive experiments were conducted on the ERA5 dataset, demonstrating the superior performance of FNP in assimilating observational data of different resolutions. Strengths: 1. The FNP addresses the limitation of existing data assimilation methods that cannot assimilate observational data of varying resolutions. Additionally, the proposed FNP, based on neural processes, provides uncertainty estimates compared to deterministic data assimilation. 2. The smoothing of high-frequency information is addressed by using the Fourier neural operator, which preserves high-frequency information. Experimental results show that the proposed Neural Fourier layer improves performance both visually and in terms of metrics. 3. FNP achieves state-of-the-art performance in arbitrary-resolution data assimilation for key weather variables. Weaknesses: 1. The ERA5 dataset is global meteorological data with limited high-frequency information. Validation on datasets with richer high-frequency information (e.g., HRRR) could better demonstrate the effectiveness of the Neural Fourier layer and provide more confidence. 2. The role of FNP in weather forecast tasks is not thoroughly discussed or experimentally validated. Technical Quality: 3 Clarity: 3 Questions for Authors: none Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: see above Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, Thank you very much for your thorough review, highly constructive comments, and feedback! We sincerely appreciate your recognition of the effectiveness and significance of our method. Below, we will address each of your questions and concerns in sequence. > The ERA5 dataset is global meteorological data with limited high-frequency information. Validation on datasets with richer high-frequency information (e.g., HRRR) could better demonstrate the effectiveness of the Neural Fourier layer and provide more confidence. Excellent suggestions! Conducting experiments on the HRRR dataset can better demonstrate the effectiveness of NFL and enhance the experimental content of this paper. Unfortunately, we regret that we are unable to download the HRRR dataset and train models on it within the limited time frame, resulting in the unavailability of experimental results here. We will strive to include experiments on the HRRR dataset in our future work. Nevertheless, we have conducted additional experiments to showcase the effectiveness of NFL from a different perspective. Table A from the PDF in global response provides ablation study on the generalization of different modules at various resolutions. As the resolution increases, the impact of NFL on performance becomes more pronounced. Similarly, in terms of visual perception, FNP's ability to capture high-frequency information strengthens with higher resolutions. This indirectly reflects the correlation between NFL and the high-frequency information gain brought by our method. > The role of FNP in weather forecast tasks is not thoroughly discussed or experimentally validated. Your feedback is highly professional! Data assimilation aims to improve weather forecasting results by reducing initial errors. Therefore, it is crucial to explore the impact of different methods on forecast error reduction. Figure A from the PDF in global response provides results on the forecast RMSE improvement of z500 variable over the next ten days through data assimilation, where lead time 0 corresponds to the reduction of initial errors. Darker colors indicate stronger improvements, meaning a greater reduction in forecast errors compared to not using data assimilation. Similar to the results of data assimilation, FNP consistently achieves state-of-the-art results in most cases, with its advantage becoming more pronounced as the resolution increases. Moreover, FNP is the only model that strictly enhances forecast improvement with increasing resolution and observational information. Additionally, apart from the accuracy of initial values affecting forecast errors, the physical characteristics of the initial states (e.g., physical balance) also influence the rate of forecast error growth. FNP demonstrates greater improvements in forecast errors at all lead times compared to improvements in initial errors, indicating that FNP not only reduces forecast errors but also slows down the growth rate of forecast errors. Other models do not exhibit the same trend, further highlighting the superior characteristics of the initial states produced by FNP. We appreciate your insights once again and will include this part in the appendix of the final version. We sincerely hope that we have addressed your concerns and questions and look forward to your reading and response. If you have any further questions, please feel free to let us know and we will do our best to answer them.	Summary: The paper introduces a new approach (Fourier Neural Processes, FNP) for weather data assimilation using data from different resolutions. The authors show that the new approach improves the results over similar data assimilation networks from earlier papers. Strengths: The paper demonstrates well the advantages of the proposed model to the similar ConvCNP network. The authors perform a thorough analysis of the ability of the method to assimilate the most relevant atmospheric variables. The paper also addresses an important area in the atmospheric ML data chain that has received less attention than e.g. weather forecasting. Weaknesses: I found the paper rather hard to follow. Concepts like "spatial variable functional representation" and "dynamic alignment and merge" are not well explained. As a result, even having previous experience of Fourier neural operators it is difficult to get a sense of what the network actually does. The authors refer the reader to previous papers on the subject, but I feel the current paper would need to better explain these ideas to a reader from outside the specific niche of research to be accessible for the NeurIPS audience. Technical Quality: 3 Clarity: 2 Questions for Authors: Figure 4 and section 4.2 refer to fine-tuning the model and present better results for the fine-tuned version. What does the fine-tuning involve? I don't get a good picture of this from the paper. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The paper does an adequate job of explaining the limitations of the current model, and makes useful suggestions for further development. However, the difficulty of understanding the paper makes it harder to assess if the work has additional limitations that are not well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, Thank you very much for your detailed review, thoughtful comments, and feedback! We sincerely appreciate your recognition of the effectiveness of our method and its relevance to the research field. At the same time, we deeply regret and apologize for any confusion our writing may have caused you. Below, we will address each of your concerns in sequence. > I found the paper rather hard to follow. Concepts like "spatial variable functional representation" and "dynamic alignment and merge" are not well explained. As a result, even having previous experience of Fourier neural operators it is difficult to get a sense of what the network actually does. The authors refer the reader to previous papers on the subject, but I feel the current paper would need to better explain these ideas to a reader from outside the specific niche of research to be accessible for the NeurIPS audience. Spatial-variable functional representation: All neural process methods share a common core idea of encoding contextual information to model a representation $R$, which is then used for decoding the target values. In CNP, $R$ is a static global representation, while in ANP, $R$ is dynamic and related to the absolute position. In ConvCNP, the encoder's mapping is translation equivariant, resulting in $R$ that is dependent on relative position, which is termed as functional representation (FR). FNP adopts the ConvCNP concept, utilizing SetConv to encode context coordinates and values separately to capture density and signal, which are then concatenated and fed into deep feature extraction module composed of NFL blocks to model FR. Tailored to meteorological data characteristics, we model a distinct spatial FR for each variable and a variable FR for all variables instead of a mixed FR. This is called spatial-variable functional representation, and the modeling process for each FR follows the aforementioned procedure. This decoupled modeling offers several advantages. Firstly, the individual spatial FRs diversify and specialize the model's spatial relationship modeling, as different variables may exhibit distinct spatial patterns. Secondly, the variable FR can capture the interrelationships among different variables. Modeling the correlations in these two dimensions is also crucial in traditional data assimilation techniques (i.e., the role of error covariance). Thirdly, it clarifies the objectives of each FR and reduces complexity, thereby easing model training, accelerating convergence, and enhancing performance. Finally, this decoupling allows for lower data embedding dimensions for each FR (i.e., number of channels), which can significantly reduce computational resource consumption. In the ablation study in Table 3, the FLOPs with and without SVD are 67.872G and 167.932G, respectively, corresponding the data embedding dimensions of 128 and 256. The SVD makes the model achieve better performance with lower FLOPs, thus enhancing computational efficiency. Dynamic alignment and merge: The DAM aligns the FR of the background and observations in the spatial dimension and then merges them. Alignment is achieved through interpolation operations, ensuring that the spatial dimensions of observations are consistent with the background in all circumstances for subsequent information fusion. The merge is implemented by selecting based on similarity to shared features, with the calculation for similarity and selection rules provided in the formula in Section 3.3. > Figure 4 and section 4.2 refer to fine-tuning the model and present better results for the fine-tuned version. What does the fine-tuning involve? I don't get a good picture of this from the paper. In our experiments in Sections 4.2 and 4.3, we present the performance of models without fine-tuning and with fine-tuning. "Without fine-tuning" refers to training in one setting and directly testing in another, showcasing the model's generalization capability. "Fine-tuning" involves further training for a certain number of epochs in the testing setting, demonstrating the model's optimal performance in that specific context. Specifically, in Section 4.2, "without fine-tuning" denotes assimilating observations all at 1.40625° during training, while assimilating observations at 0.703125° and 0.25° in testing. The performance after fine-tuning refers to the testing results after the model is further trained at the corresponding resolutions. In Section 4.3, "without fine-tuning" means using the model weights trained from the data assimilation task directly for reconstructing the observational information. We infer the FR of the observations using only the encoding weights of the observation branch, and then infer the output based on the FR using the decoding weights. The performance after fine-tuning refers to the testing results after continuing to train the weights of these two parts in observational information reconstrution task. We sincerely hope that we have addressed your concerns and questions and look forward to your reading and response. If you have any further questions, please feel free to let us know and we will do our best to answer them. --- Rebuttal 2: Title: Hoping for your feedback Comment: Dear reviewer, Thank you for your question. We believe we have adequately addressed all of the issues you raised in your review. We would like to emphasize that data assimilation is a very important area that has taken AI-based weather forecasting one step further toward practical operational applications. Compared to existing work, our approach has greater potential and a wider range of applications to advance the meteorology field and benefit society. If your question has been satisfactorily addressed, we hope you will review and possibly update your score. We are willing to address any additional questions or concerns you may have. Thank you.	Summary: This paper proposes a new variant of neural processes called Fourier Neural Processes (FNPs) to solve the data assimilation problem with arbitrary solution, which is an important component in modern weather forecast system. The proposed method based on FNP has better computational efficiency, and achieves state-of-the-art performance in assimilating observations with varying resolutions on large-scale simulated data. Strengths: This paper focuses on applying neural processes to the data assimilation problem in weather forecast system. Different from many previous work in this area which mainly focus on toy examples, or hand designed problems, this work solves a real application problem that has high impact. I really enjoy this paper and would love to see more work like this in our area. Weaknesses: The primary issue with this paper is its technical clarity. The majority of the technical nuances are conveyed through textual descriptions, and at times, the terminology is not adequately explained. It would be beneficial if the authors could provide the precise mathematical formalisms for key modules such as the “Spatial-variable decoupled representation,” the Neural Fourier layer, and DAM. Although one can how these modules work, a definitive mathematical expression would greatly alleviate any confusion. I am also slightly confused by the following usage of words: 1. Conditional points/conditional domain. Do you mean "context" as supposed to "target"? 2. Variable dimension. Do you mean the dimension across different meteorological variables at the same spatial location? 3. What do you mean by "dynamics" at line 175. 4. Line 228, the term RMSE is a statistical metric used way beyond geospatial analysis and atmospheric science. So you might want to use something else to refer to the special "latitude-weighted" version. Technical Quality: 3 Clarity: 2 Questions for Authors: See weakness. I would like to also ask the following: 1. What do you think is the main reason that FNP outperforms ConvCNPs? What do you think is the most important innovation of FNP for the data assimilation problem, given that ConvNPs (https://proceedings.nips.cc/paper/2020/file/5df0385cba256a135be596dbe28fa7aa-Paper.pdf) already has applications for Environmental Data. 2. Why the neural Fourier layer (NFL) is more computational efficient than standard conv layers, if NFL contains three branches, one one branch is Conv operations? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have thoroughly discussed the limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, Thank you very much for your thorough review, highly constructive comments, and feedback! We greatly appreciate your positive reception of our work and recognition of its practical value. At the same time, we deeply regret any confusion or difficulties you may have encountered regarding our methodological expressions and terminologies. We will address each of your points in the following responses. > It would be beneficial if the authors could provide the precise mathematical formalisms for key modules such as the “Spatial-variable decoupled representation,” the Neural Fourier layer, and DAM. Although one can how these modules work, a definitive mathematical expression would greatly alleviate any confusion. Excellent suggestion! Following the description in Figure 1, we will endeavor to provide mathematical formulations for the key modules in FNP. - Functional representation (FR): $FR(x^c,y^c)=Concate(SetConv(x^c),SetConv(y^c))$ - Spatial-Variable decoupled representation (SVD): As described in Section 3.2, we model a separate spatial FR for each variable and a variable FR for all variables instead of a mixed FR, with each FR following the formula above. - Neural Fourier layer (NFL): $NFL(·)=\mathcal{F}^{-1}(Linear(\mathcal{F}(·)))+Conv(·)+Identity(·)$ - Dynamic Alighment and Merge (DAM): The DAM module aligns the functional representations of the background and observations in the spatial dimension and merges them. The alignment is achieved through interpolation operations, while the merge is implemented by selecting based on similarity to shared features, with the calculation for similarity and selection rules provided in the formula in Section 3.3. > Conditional points/conditional domain. Do you mean "context" as supposed to "target"? Conditional points and conditional domain refer to the context points and context domain, corresponding to the input of the model, whlie target points and target domain correspond to the output of the model. > Variable dimension. Do you mean the dimension across different meteorological variables at the same spatial location? Yes, it is also the channel dimension, because the data of different variables are concatenated on the channel dimension. > What do you mean by "dynamics" at line 175. It means the ability to support inputs of varying sizes, i.e., assimilating observations with different resolutions. > Line 228, the term RMSE is a statistical metric used way beyond geospatial analysis and atmospheric science. So you might want to use something else to refer to the special "latitude-weighted" version. You are right, in atmospheric science, we usually use RMSE to refer to the latitude-weighted version rather than the standard version. We appreciate and respect your rigor, and we will use the term "WRMSE" to replace "RMSE" in final version. > What do you think is the main reason that FNP outperforms ConvCNPs? The main difference between FNP and ConvCNP lies in our tailored design for data assimilation task. The SVD reduces the complexity of model training and computational resource consumption while achieving better performance. Unified coordinate transformation and dynamic alignment ensure that the model comprehends the spatial correspondences between data of different resolutions, while dynamic selection and merge mechanism enhances the effectiveness of information fusion. The visualization in Figure 2 demonstrates the enhancement in the model's ability to extract deep features and capture high-frequency information through the NFL. In our supplementary experiments (see Table A from the PDF in global response), as the observational resolution increases, the impact of the NFL in ablation study on generalization becomes increasingly evident, further validating its effectiveness. >What do you think is the most important innovation of FNP for the data assimilation problem, given that ConvNPs (https://proceedings.nips.cc/paper/2020/file/5df0385cba256a135be596dbe28fa7aa-Paper.pdf) already has applications for Environmental Data. FNP has the capability to assimilate observational data with arbitrary resolution without the need for prior interpolation of observations that typically have higher resolutions. It avoids information loss, thereby enhancing the performance of data assimilation. The types and resolutions of observational data in practical applications are highly diverse and complex. FNP can directly assimilate such data without pre-processing, offering greater practical value. Additionally, the challenge of dealing with high-dimensional data is a common issue faced by traditional methods and AI approaches. As our experiments have shown, the outstanding generalization ability of FNP enables training at low resolutions and direct application to high resolutions. This not only significantly reduces computational resource consumption but also provides a pathway for data assimilation at higher resolutions. > Why the neural Fourier layer (NFL) is more computational efficient than standard conv layers, if NFL contains three branches, one one branch is Conv operations? In the ablation study presented in Table 3, FNP utilized a structure encoded with 4 NFL blocks, amounting to 67.872G FLOPs. In contrast, the FNP without NFL opted for a replacement comprising 12 convolutional blocks (consistent with the official implementation in [1]), totaling 167.932G FLOPs. FNP's ability to achieve better performance with lower computational complexity demonstrates its higher computational efficiency. - [1] Yann Dubois, Jonathan Gordon, and Andrew YK Foong. Neural process family. http://yanndubs.github.io/Neural-Process-Family/, 2020. We sincerely hope that we have addressed your concerns and questions and look forward to your reading and response. If you have any further questions, please feel free to let us know and we will do our best to answer them. --- Rebuttal Comment 1.1: Title: Response to Author Rebuttal Comment: Thanks a lot for the explanation! They have addressed my confusion and questions. It would be good to incorporate them in the writing. My original rating will remain unchanged. --- Reply to Comment 1.1.1: Comment: Thank you again for your recognition of our work and your efforts to help us improve the paper！	Summary: The authors propose a method for arbitrary-resolution data assimilation, called Fourier Neural Processes (FNP). This approach improves generalization by addressing resolution limitations in existing methods. Key features include unified coordinate transformation, spatial-variable functional representation, and dynamic alignment and merge (DAM). The method effectively integrates diverse observational data without the need for fine-tuning. Experimental results demonstrate its relatively good performance in handling varied data sources. Strengths: (+) This method is the first to address the challenge of data assimilation with arbitrary resolutions. (+) The authors introduce the neural processes to arbitrary-resolution data assimilation. (+) Demonstrating excellent generalization, the proposed method can perform observational information reconstruction directly from training on data assimilation, without requiring fine-tuning. (+) Each proposed component is simple yet effective, as shown in the ablation study. Weaknesses: Since the authors aim to address arbitrary-resolution data assimilation with good generalization ability, it would be beneficial to clearly highlight the key observations or components for this problem upfront. The authors quickly delve into the pipeline description without first presenting the key and unique idea of their method for this problem. It takes readers some time to figure out, 'How does this method solve the issue of arbitrary resolution in data assimilation?' Technical Quality: 3 Clarity: 2 Questions for Authors: Could the authors clarify the unique contributions of their method in addressing the issue of arbitrary resolution from a design perspective? It seems that the core of your approach relies on interpolation in the feature space. If the generalization ability for arbitrary resolutions is primarily due to this interpolation, does this mean that integrating similar interpolation techniques into any deep learning method could achieve similar improvements in generalization? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors point out the key limitation of their method: it has not been tested on a real-world dataset due to the lack of relevant benchmarks and large-scale datasets in the data assimilation community. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, Thank you very much for your detailed review, thoughtful comments, and feedback! We appreciate your recognition of the effectiveness of our method, while deeply regretting and apologizing for any confusion or concerns we may have caused you regarding the motivation and design of our approach. Below, we address each of your questions in turn. > Since the authors aim to address arbitrary-resolution data assimilation with good generalization ability, it would be beneficial to clearly highlight the key observations or components for this problem upfront. Weather forecasting is of paramount importance to both science and society. In recent years, AI-based medium-range weather forecasts have advanced rapidly, yet they still rely on traditional data assimilation techniques from conventional numerical weather prediction (NWP) systems to provide initial states. Consequently, data-driven data assimilation methods have garnered increasing attention as a crucial component in constructing end-to-end weather forecasting systems based on AI. These AI-based approaches not only significantly reduce resource consumption but also offer new possibilities for overcoming bottlenecks in traditional numerical methods. Existing work is primarily designed and trained based on specific resolutions (often matching the background resolution), resulting in models that can only be utilized in the same settings. In this paper, we propose the Fourier Neural Processes (FNP) for arbitrary-resolution data assimilation, whose necessity and importance lie mainly in the following points. Firstly, FNP eliminates the need to interpolate observational data, typically with higher resolution, which can avoid the information loss and improve the performance of data assimilation. Secondly, observational data in practical applications vary greatly in type and resolution; FNP can directly assimilate the diverse data without preprocessing, enhancing its practical application value. Lastly, the challenge of dealing with high-dimensional data is a common issue faced by traditional and AI methods; As our experiments demonstrate, the outstanding generalization ability of FNP enables training at low resolutions and direct application at high resolutions, thereby significantly reducing computational resource consumption. > The authors quickly delve into the pipeline description without first presenting the key and unique idea of their method for this problem. It takes readers some time to figure out, 'How does this method solve the issue of arbitrary resolution in data assimilation?’ From an implementation perspective, the ability to support inputs of arbitrary resolutions lies in the fact that fundamental operations (such as convolution, MLP, etc.) are independent of input size. Neural processes operate on paired (context or target) coordinate-value units, making them highly suitable for data assimilation tasks where observational data may take irregular forms. Therefore, we introduced a convolutional version of neural processes (i.e., ConvCNP) and incorporated custom modules independent of input size to achieve data assimilation with arbitrary resolutions. However, as indicated by our experiments, the performance of ConvCNP is unsatisfactory. Therefore, from a performance standpoint, the key to effectively addressing the issue of arbitrary resolutions lies in the efficacy of our module design (e.g., unified coordinate transformation and dynamic alignment to ensure the model comprehends spatial correspondences between data of different resolutions, and the neural Fourier layer providing significant high-frequency information gain when assimilating high-resolution observations). > Could the authors clarify the unique contributions of their method in addressing the issue of arbitrary resolution from a design perspective? Based on the aforementioned ideas, we can categorize the core contributions of our approach in design into two aspects. Firstly, we astutely recognized the unique advantages of neural processes in addressing data assimilation challenges, introducing neural processes into data assimilation tasks to enable data assimilation with arbitrary resolutions. Secondly, through efficient module design, we achieved a significant improvement in performance and generalization, making FNP the only model whose performance demonstrates consistency with theoretical understanding (with the increase of resolution and the increase of observational information, the model's performance continues to improve significantly). > It seems that the core of your approach relies on interpolation in the feature space. If the generalization ability for arbitrary resolutions is primarily due to this interpolation, does this mean that integrating similar interpolation techniques into any deep learning method could achieve similar improvements in generalization? Based on the above summary, interpolation in feature space is only a small part of our method. In fact, in the results we presented, ConvCNP also employed interpolation and simple fusion instead of DAM, yet its generalization was poor. Here, we further provide ablation study on generalization of different modules, as shown in Table A from the PDF in global response. The models in the table are all trained at 1.40625° resolution and directly tested at other resolutions to validate their generalization. It can be observed that all modules contribute to enhancing generalization to varying degrees. In addition, not all deep learning methods can support inputs of arbitrary sizes, thereby combining interpolation to achieve data assimilation with arbitrary resolutions. Even if they can, simple interpolation does not necessarily lead to improved generalization. We sincerely hope that we have addressed your concerns and questions and look forward to your reading and response. If you have any further questions, please feel free to let us know and we will do our best to answer them. --- Rebuttal 2: Title: Response to Author Rebuttal Comment: Thank you for the detailed responses. I’ve reviewed the comments from other reviewers as well as your replies. I’m curious about the performance if both NFL and SVD are removed simultaneously. Additionally, if SVD alone is removed, does the generalization capability (arbitrary resolution) of the proposed method primarily stem from the unified coordinate transformation? --- Rebuttal Comment 2.1: Comment: I'm sorry that we can't give the quantitative performance when removing both NFL and SVD simultaneously due to the lack of time left. But based on the above experimental results and our experience, all modules are helpful to the performance, so its performance should be better than ConvCNP and worse than other FNP models. As we mentioned above, all modules, including DAM and NFL, can help improve the generalization capability. From an experimental point of view, the results in Table A prove that both DAM and NFL can improve generalization. From a theoretical point of view, the dynamic alignment in DAM can help the model capture the spatial relationship between data, and the frequency-domain analysis in NFL provides a global perspective, which can better reflect the overall characteristics of the signal at various resolutions. We would like to emphasize that data assimilation is a very important field that brings AI-based weather forecasting one step closer to practical applications. Compared to existing work, our approach addresses the challenges of arbitrary-resolution assimilation. It not only significantly improves the performance of data assimilation, but also has greater potential and a wider range of applications to advance the field of meteorology and benefit society. If your question has been answered satisfactorily, we hope you will review and possibly update your rating. We are willing to address any additional questions or concerns you may have. Once again, we would like to express our sincere gratitude to you!	Rebuttal 1: Rebuttal: Dear reviewers and meta-reviewers, We greatly appreciate the considerable time you have dedicated to providing us with constructive comments and feedback to further enhance our paper. It is gratifying to see that all reviewers acknowledge the effectiveness and contribution of our method. We have received many thoughtful comments, including some highly professional suggestions, which we find both excited and grateful. Simultaneously, we sincerely apologize for any confusion or doubts we may have caused you. We have addressed all reviewers' queries with additional explanations and experiments, aiming to offer further insights and alleviate any concerns. In response to the insightful suggestions provided, we will revise and expand the manuscript. We welcome any follow-up discussions! Last but not least, we thank the PCs, ACs, and all the reviewers again for devoting time and effort to this review! Pdf: /pdf/bb64b863f97576ba2f13be4f32cbd47afbee91b0.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Do Finetti: On Causal Effects for Exchangeable Data	Accept (oral)	Summary: The paper generalizes the traditional iid settings in casual inference to exchangeability settings by de Finetti theorem, and proposes a new model named the casual Polya urn model to illustrate the new scheme and to catch more relationship. The experiments show when the number of environment is less than 5000, the new schemes performs well. Strengths: First of all, I am sorry that I do not know much about the casual inference. But the paper uses exchangeability instead of iid settings, which seems an improvement. Weaknesses: 1. Aldous 1985 shows many (not all) conclusions in iid can be naturally transformed into those under exchangeability. So the theoretical improvement seems not much. 2. Only simulated experiments for the casual inference problems. 3. De Finetti Theorem is ‘iff’. So it is inappropriate to use methods based on iid settings on the exchangeable but not iid data to compare. 4. For the experiment, what about the larger number of environment? It seems the original one performs better. Technical Quality: 4 Clarity: 4 Questions for Authors: Besides above, 5. In casual inference problems, is it easy to identify the exchangeability, especially for the real data? Confidence: 2 Soundness: 4 Presentation: 4 Contribution: 2 Limitations: Besides above, 6. Some typos even in reference; for example, the first reference is not well-written. 7. Could see more complicated structure between theta psi and X. In the paper, psi is independent of X. 8. Typos, like ‘Nature’ should be ‘nature’. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time and address the questions below. > Aldous 1985 shows many (not all) conclusions in iid can be naturally transformed into those under exchangeability. So the theoretical improvement seems not much. Indeed, many i.i.d. results transfer to the exchangeable case, however, the situation is more complex. Aldous (1985) presents a lecture on exchangeability in probability theory, covering topics such as de Finetti’s theorem, its consequences, extensions, analogues, distributions invariant under group transformations, and exchangeable sets. However, Aldous (1985) does not focus on causality. Our work shows that causal effect estimation in certain types of non-IID data (specifically those meeting an ICM assumption) enables us to draw conclusions that cannot be inferred from IID data alone. > Only simulated experiments for the casual inference problems. The experiments are designed to show that the truncated factorization formula developed for i.i.d. data fails to apply for exchangeable non-i.i.d. data even given knowledge of the true graph. Fig. 4 shows even with near infinite data, the dotted blue line (i.i.d. with the true graph) cannot reach 0 mean squared error in contrast to do-finetti with the true graph. The simulated experiment thus merely creates a controlled setup to demonstrate that conclusions in i.i.d. (e.g. truncated formula) fail to apply for exchangeable non-i.i.d. data. > De Finetti Theorem is ‘iff’. So it is inappropriate to use methods based on iid settings on the exchangeable but not iid data to compare. We agree. The standard i.i.d. methods aren’t appropriate. Maybe we have not expressed this well, and a statement akin to the above would help put the experiment in perspective? > For the experiment, what about the larger number of environment? It seems the original one performs better. We are confused with the statement. In Figure 4 (both left and right), we observe that do-Finetti algorithm outperforms the i.i.d. baseline in the large number of environments. The left plots show that do Finetti achieves near zero mean squared error in causal effect estimation compared to the i.i.d. baseline which has high errors. The right plot also shows that do Finetti simultaneously identifies the correct graph in contrast to the i.i.d. baseline with low graph accuracy. > 5. In causal inference problems, is it easy to identify the exchangeability, especially for the real data? We thank the reviewer for the question. There might be empirical tests one could run based on permutations and there exists some work for testing exchangeability in real-world data, e.g., [1], [2]. Likely, this may not be a closed question, especially in terms of causal inference and exchangeability. Though we’d hope to leverage the above work as a promising first step to study it for future directions. > 6. Some typos even in reference; for example, the first reference is not well-written. We apologise and will thoroughly go over this for the revision. > 7. Could see more complicated structure between theta psi and X. In the paper, psi is independent of X. This could be interesting future directions, however we see this as a starting point, and we found ICM assumption was ideal for us in that It allows us to derive non-trivial results (Theorem 2) It is an assumption which is common in the causality community [3]. > 8. Typos, like ‘Nature’ should be ‘nature’. Here we deliberately choose the capital letter “Nature” to show respect and reverence towards the governing laws of the universe. This though common in the literature, we acknowledge it is more a personal preference. Overall, we thank the reviewer for taking the time and if we adequately addressed your concerns over * theoretical improvement (exchangeable non-i.i.d. data reveals important properties current causal literature does not cover), and * fair experimental comparisons and results (i.i.d. methods fail to apply for exchangeable non-i.i.d. data and hence demands a new causal effect estimation function for exchangeable non-i.i.d. data and do Finetti algorithm offers a solution), we invite the reviewer to consider raising the score. [1] Vovk, V., Gammerman, A., Shafer, G. (2022). Testing Exchangeability. In: Algorithmic Learning in a Random World. Springer, Cham. https://doi.org/10.1007/978-3-031-06649-8_8 [2] Aw, Alan J., Jeffrey P. Spence, and Yun S. Song. "A simple and flexible test of sample exchangeability with applications to statistical genomics." The annals of applied statistics 18.1 (2024): 858. [3] Schölkopf, Bernhard, et al. "Toward causal representation learning." Proceedings of the IEEE 109.5 (2021): 612-634. --- Rebuttal Comment 1.1: Comment: Thank you for your reply! Since I am not an expert of casual inference, my questions focus on the exchangeability. Since I don't find many literature about exchangeability on causal inference structure, combining your rebuttal, I agree that a non-iid structure is non-trivial and an interesting topic. Since in the exchangeability settings, more relationship with Bayesian methods could be explored more. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for responding and pushing us to be clear on the paper's contributions! We agree with the reviewer that there could be exciting areas to explore on the connection between causality and Bayesian methods.	Summary: The paper studies causal effect identification and estimation in exchangeable data. The main result here is theorem 1, which shows that causal effects are identifiable in ICM generative processes. Strengths: - The paper provides a great framework to think about interventions in exchangeable data. Starting from what interventions should be considered (Definition 3) to identifying a procedure for computing the post-interventional distributions. - The paper presentation, at least in the first part, was simple and intuitive. I always found myself asking a question and then find it being answered in the next paragraph. However, probably due to space constraints, this did change in the latter parts of the paper. Weaknesses: - The latter parts of the paper is rushed and left me confused. For example, it is unclear how causal de Finetti theorems apply to the Causal Pólya Urn Model, Theorem 2, and the entirety of section 4 is very rushed and I have struggled to understand what theorem 2 say exactly. - I have felt that the algorithm could have taken more of real-estate in the presentation of the paper. Also, it is unclear how the graph structure is learned in the algorithm - This is more of a nit pick, but the appendix contains a few typos and is in a worse state in general than the main text. For example, the use of index i in equation 51, 53, ... Technical Quality: 4 Clarity: 3 Questions for Authors: - In the experiments, it seems to me that the model generating the synthetic dataset is different from that described in section 3.2. In particular, in the experiments, X_i is sampled from a Ber(theta) and hence P(X_i = 1) = P(X_2 =1) = ... = theta, whereas if I understood the model in 3.2, then the probability P(X_n =1) will be much greater than P(X_1 =1) if for example all X_m =1 for all m < n. Are they actually different? Or did I misunderstood? And how can the model described in 3.2 be represented by equation 4? (I read F.2 but it seems to me that equation 51 follows the model in Section 5). - In the experiments, can the authors elaborate on the IID baseline? Do you run the algorithm on the "full" graph G which have nodes X_1 Y_1 X_2 Y_2? I assume this is what's been done as it is the fairest baseline, but I'm not sure. Appendix K seems to imply that and the main paper did not make it clear. - In the description of ICMs, the author mention the expression: > Causal mechanisms are independent of each other in the sense that a change in one mechanism P(Xi \| PAi) does not inform or influence any of the other mechanisms P(Xj \| PAj) What would be a concrete example where such condition is violated? Confidence: 2 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: It has been addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and their appreciation of our work. We hope to clarify their questions below: > The latter parts of the paper is rushed and left me confused. For example, it is unclear how causal de Finetti theorems apply to the Causal Pólya Urn Model, Theorem 2, and the entirety of section 4 is very rushed and I have struggled to understand what theorem 2 say exactly. We apologise and will try to clarify. Appendix F shows the causal Pólya urn model can be equivalently modelled as in the causal de Finetti theorem, i.e., $\int \int \prod_i p(y_i \| x_i, \psi) p(x_i \| \theta) p(\theta) p(\psi) d\theta d\psi$, where $p(\theta), p(\psi)$ are beta-distributions and $p(x_i \| \theta), p(y_i \| x_i, \psi)$ are Bernoulli distributions. This is a bivariate version of equation 4 when we only consider two variables X and Y. We will include a more detailed discussion on Appendix F in the main text for the next version. Theorem 2 says that for ICM generative processes, both causal graphs and causal effects can be identified simultaneously. This is in contrast to an i.i.d. process, where causal effect identification often requires access to aspects of the causal graph which itself is not identifiable from observational data, and thus it is assumed that the causal graph is provided in addition to the observational data. > Also, it is unclear how the graph structure is learned in the algorithm For learning the graph structure, we refer to Algorithm 1 in Guo et al. 2024 [1], as the present paper focuses on the study of causal effects. We will make this point clearer in the next version. > This is more of a nit pick, but the appendix contains a few typos and is in a worse state in general than the main text. For example, the use of index i in equation 51, 53, ... Thank you for pointing out the typos, we will correct them in next version. > In the experiments, it seems to me that the model generating the synthetic dataset is different from that described in section 3.2. In particular, in the experiments, X_i is sampled from a Ber(theta) and hence P(X_i = 1) = P(X_2 =1) = ... = theta, whereas if I understood the model in 3.2, then the probability P(X_n =1) will be much greater than P(X_1 =1) if for example all X_m =1 for all m < n. Are they actually different? Or did I misunderstood? And how can the model described in 3.2 be represented by equation 4? (I read F.2 but it seems to me that equation 51 follows the model in Section 5). The model described in 3.2 can be represented by equation 4, because Appendix F.2 shows that the joint distribution $P(x1, y1, x2, y2, …) $ in the causal Pólya urn model can be modelled as $\int \int \prod_i p(y_i \| x_i, \psi) p(x_i \| \theta) p(\theta) p(\psi) d\theta d\psi$, where $p(\theta), p(\psi)$ are beta-distributions and $p(x_i \| \theta), p(y_i \| x_i, \psi)$ are Bernoulli distributions. This corresponds to equation 4 as this is the equivalent bivariate version where X is the parent of Y, and $\theta, \psi$ are statistically independent $\theta_i$’s in equation 4. Therefore as equation 51 follows the model in section 5 (as the reviewer suggested) and it is the representation for the causal Pólya urn model (due to the arguments above and in Appendix F.2), we argue that it is the same as the one described in section 3.2. We hope it clarifies things and thank you for going into the Appendix. > In the experiments, can the authors elaborate on the IID baseline? Do you run the algorithm on the "full" graph G which have nodes X_1 Y_1 X_2 Y_2? The IID baseline is taking into account the full graph x1, x2, y1, y2 and treats the variables as $(x_i, y_i) \sim_{i.i.d.} (X, Y)$. This means there are no bi-directed edges connecting X1, X2 and Y1, Y2. The causal effect estimand for the i.i.d. case is analogous to Eq. 10. The experiment is designed to show that the truncated factorization developed for i.i.d. data indeed does not apply for exchangeable non-i.i.d. data, hence the need for the generalised truncated factorization introduced in Theorem 1. > In the description of ICMs, the author mention the expression: Causal mechanisms are independent of each other in the sense that a change in one mechanism P(Xi \| PAi) does not inform or influence any of the other mechanisms P(Xj \| PAj) What would be a concrete example where such condition is violated? This condition will be violated when we decompose a distribution into non-causal conditionals. For example, suppose that for weather stations, altitude (A) causes temperature (T) but not vice versa, i.e. building a greenhouse effect on top of a mountain will not increase the height of the mountain. In that case, P(T \| A) and P(A) will be independent causal mechanisms that can be changed independently in the generative process, but P( A \| T) and P(T) will not be. This is described in more detail for instance in Peters et al., Elements of Causal Inference. The causal de Finetti theorem formalises ICM mathematically: suppose $\theta$ represents a mountain and $\psi$ represents seasons. With random measurements given fixed mountain and season, we have $T_i, A_i$. The causal de Finetti theorem says $A \to T$ characterizes by $T_1 \perp A_2 \| A_1$. Equivalently, it means $P(T_1 \| A_1, A_2) = P(T_1 \| A_1)$, i.e. knowing the altitude of measurements in other locations will not help the prediction of the temperature measured at location 1. If $T \to A$, causal de Finetti theorem says $A_1 \perp T_2 \| T_1$, equivalently expressed as $P(A_1 \| T_1, T_2) = P(A_1 \| T_1)$. However we know it does not hold as if $T_1 = -10$, and $T_2 = 30$, then one can infer it is likely to be in hot season and given observed a low temperature $T_1$, one could infer $A_1$ will be high in altitude. [1] Guo, S., Tóth, V., Schölkopf, B. and Huszár, F., 2024. Causal de Finetti: On the identification of invariant causal structure in exchangeable data. Advances in Neural Information Processing Systems, 36. --- Rebuttal 2: Comment: I thank the reviewer for their comprehensive and very well-explained response. No further questions from me! --- Rebuttal Comment 2.1: Comment: Thank you for taking the time and helping us to improve the paper! We will include the clarification on causal Pólya urn model in the main text for the next revision.	Summary: The paper formalizes the observational and interventional distribution under the ICM generative process, of which iid is the special case. It provides an identifiability result for the causal effect given that the causal graph is known. Then, it shows that both the causal graph and the causal effect can be identified simultaneously. Strengths: 1. Problem: The problem is important as it will bring the causal effect estimation literature closer to real-world scenarios. 2. Theory: The theoretical results are strong, especially Theorem 2, which shows that both the causal graph and the effect can be estimated simultaneously. I have not checked the proofs, though. 3. Experiment: The experiment on the simulated data verify the theoretical claim. 4. Presentation: The paper is well-written and easy to follow. All the notation and definitions are clear. Weaknesses: 1. Experiments: I understand the main purpose of the work is to establish the theoretical foundation of causal effect estimation for exchangeable data, but it would be interesting to apply the method to some real-world datasets (not necessary for the rebuttal). Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Definition 3: We should also break the edge from the de-finnetti parameters to the intervened variable, right? Or do we not need any graphical operations? 2. I am slightly confused by the statements in Lines 153-154 and Line 197. They seem to contradict each other. Is it due to conditioning on x1 and x2 that Eq10 and eq11 are not equal since, in ICM, they are not iid? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Yes, the authors have addressed the limitation in the Conclusion section and Appendix L. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and their recognition of the importance of relaxing the i.i.d. assumption, a general problem in the ML community. Please see below answers to the questions: > Definition 3: We should also break the edge from the de-finnetti parameters to the intervened variable, right? Or do we not need any graphical operations? Yes, we need to break the edge from the de-Finetti parameters to the intervened variables when performing an intervention. This definition aims to clarify different implications when performing graph surgery on SCM and ICM processes. > I am slightly confused by the statements in Lines 153-154 and Line 197. They seem to contradict each other. Is it due to conditioning on x1 and x2 that Eq10 and eq11 are not equal since, in ICM, they are not iid? Lines 153-154 state that IID is a special case of exchangeability whenever p(ψ) = δ(ψ = ψ0), and line 197 states that the causal effect differs whenever p(ψ) does not equal δ(ψ = ψ0). These statements are consistent in that IID is a special case of ICM when $ p(\psi) = \delta(\psi = \psi_0) $. However, our focus is on causal effects for the ICM exchangeable non-IID case, where $p(\psi) \neq \delta(\psi = \psi_0)$. We show that the causal effects in the ICM exchangeable non-IID case (Eq. 11) differ from those in the ICM IID case (Eq. 10) due to the dependency among observations. --- Rebuttal Comment 1.1: Comment: I thank the authors for their response. I will keep the score.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Seek Commonality but Preserve Differences: Dissected Dynamics Modeling for Multi-modal Visual RL	Accept (poster)	Summary: The paper proposed a method, namely Dissected Dynamics Modeling (DDM), for multi-modal environment dynamics modeling in visual RL. The core idea is to adopt additional modules to extract separate modality-consistent and modality-inconsistent features in each modality stream with designated losses as the regularization. During training, the model tries to maximize mutual information between modality-consistent features from different modalities at the current and the next timestamp. For modality-inconsistent features from, the model tries to enforce orthogonality. Experiments on CARLA and DMControl indicated the superiority of DDM over other state-of-the-arts and further analyses suggested the effectiveness of DDM v.s. DeepMDP. Ablation studies supported some of the important design choices. Strengths: - The general idea of attempting to decouple modality-consistent and modality-inconsistent features to improve environment modeling makes sense, which could provide some insights to future work to this particular field - Experiment results show good improvements over some state-of-the-arts, suggesting DDM's effectiveness for environment modeling in visual RL - Good ablations, analyses and visualizations provided a comprehensive understanding of the method to some extent justifying DDM's soundness and supporting its superiority for environment modeling in visual RL Weaknesses: - The proposed method is only showed to be working on a subset of modalities that are visual only, which greatly limits its generalizability and contribution as a method for multi-modal learning. However, visual RL is not my expertise and thus I cannot evaluate the contribution on this particular field. - Following above, from the experiments it seems that the method also only works on visual inputs from the same camera perspective. i wonder if there are experiments demonstrating the model also works for visual inputs from multiple camera positions. - For experiments on DMControl, how is the masking done for frames at different timestamps? Also, besides 20% masking, do the authors have results with other masking ratio? - Some places are not written in a technically accurate way, though it's hard to tell whether it is because of technical misunderstanding or simply misuse of words. For example: - Line 46 - 48, "Firstly, modality-correlated features provide a foundational perspective by capturing shared and complemented information across different sensory inputs." The modality-consistent features that are shared among modalities does not COMPLEMENT each other. Instead, those modality-inconsistent features that are unique for each modality themselves complement other modalities. - Line 50-51, "these inconsistencies are typically deemed less critical and are filtered out through modality alignment". Modality alignment should not have the effect of filtering out modality-inconsistent features, in my understanding. If the authors have some evidence, I'd like to see. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the above Weaknesses part and address my concern accordingly Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes, limitations are discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are truly grateful for your thoughtful remarks and experimental suggestions. These remarks shed light on what we can improve and are crucial for refining our work. We address your main concerns as follows: > The method is only tested on visual modalities, limiting its generalizability and contribution. We appreciate the comment and recognize the importance of generalizing to different modality types. We think our method’s contributions are twofold: the fundamental idea and the method itself. While our method mainly focuses on visual modalities, the core idea—that modality-correlated and distinct features are both crucial—might also provide initial spark of inspiration in non-visual fields. For instance, consider a photo of a sea and a corresponding text description like “Vibrant sea under a clear blue sky, with fluffy white clouds.” The photo may show details that the text misses, such as distant ships. Similarly, the text may also capture high-level concepts that lack in the image feature, such as the sensory concept of “fluffy”. These distinct features between modalities can be valuable for the task at hand, which is worth exploring and might motivate further research on image-text learning. We hope the potential inspiration brought by our central idea can benefit the non-visual fields and help address some of the concerns. We are thankful for your reflective remark, which suggests a promising extension of our work to fit other modalities (e.g., audio, text) and benefit the entire multi-modal research community. We sincerely appreciate this insight and are committed to work on these extensions. > Experiments for visual inputs from multiple camera positions? Thanks for the experiment advice. We have further verified the ability of our model on multiple camera positions. First, we switch the camera view of RGB modality in DMControl. Second, we test on CARLA with RGB and LiDAR BEV as input modalities. LiDAR BEV is a bird view map, whose perspective is very different from RGB. Due to the limited time budget, we only compare our method with the most competitive methods on these two environments. The results and illustrations of different camera views are presented in Table 1,2 and Fig. 1(e),(f) in the rebuttal PDF. These results show that our method also works on multiple camera positions. > How is the masking done? Do the authors have results with other masking ratio? Our masking operation is performed independently at each timestamp. Therefore, both the masked modality type and the masking locations vary randomly across different timestamps, which simulates a challenging occlusion scenario. For other masking ratios, we further test the ratio of 0%, 40%, 60%, and compare our method with SPR, the most competitive baseline on DMControl. The new results are in Table 2 of the rebuttal PDF. The results show that as the masking ratio increases, both SPR and our DDM experience performance drops. However, DDM still outperforms SPR at different ratios. > Some places are not written accurately, such as Line 46 - 48 and Line 50-51. We apologize for the confusion caused. For the first issue in Line 46 – 48, we confirm it was a misuse of the word "complemented," which inaccurately described modality consistencies. We meant to convey that the modality-consistent features contain shared and common information and create a unified description for the environment. We appreciate the detailed attention to this error and will correct it. For the second issue in Line 50-51, our use of the phrase “filter out” may have been overly definitive. We intended to convey that modality alignment aims to encourage consistency and, consequently, mitigate inconsistencies. To clarify this, we provide evidence both in the literature and with experiments. Specifically, we discuss several alignment methods as follows: Imposing consistency constraints is a straightforward method for modality alignment, typically achieved by minimizing cross-modality feature distances [1,2]. Ideally, if two modality features have zero distance, it means perfect consistency and the absence of inconsistencies. We reference multiple works as evidence to support this design goal, as quoted below: “... the states expressed by different modalities can be consistent at the same time... to achieve this, we use a similarity measurement...” (Sec. IV.A in [1]) “...to ensure the consistency of the latent embeddings of different modalities in the shared latent space, we develop two parallel cross-alignment schemes...” (Sec.3.2 in [2]) Mutual information optimization [3] is also designed to enhance feature consistency (and thereby decrease inconsistencies), we quote [3] as follows: “The key innovation is a specially-designed density ratio estimator that encourages consistency between the latent codes of each modality.” (Abstract in [3]) For experimental evidence, we calculate the CKA index [4] between modality features. CKA is a metric to quantify similarity between network features for the same input. A higher CKA indicates greater feature correlation. Specifically, for RGB modality on CARLA, we extract features $z$ from a baseline SAC model without modality alignment, and the modality-consistent feature $\overline{z}$ and inconsistent feature $\hat{z}$ from our DDM. We Then compare the CKA between these features across 1K input samples. The results are as follows: ① CKA($z$,$\overline{z}$)=0.865 ② CKA($z$,$\hat{z}$) = 0.897 ③ CKA($\overline{z}$,$\hat{z}$) = 0.561 It can be seen the both ① and ② are relatively high, showing that without modality alignment, $z$ retains both common and unique modality information. Further, ③ is notably lower than ②, indicating that compared with the non-aligned feature $z$, the inconsistent information in the aligned $\overline{z}$ is indeed reduced. References: [1,2,3] correspond to [32,27,4] in our paper [4] Similarity of neural network representations revisited, ICML 2019 --- Rebuttal 2: Comment: I appreciate the authors' responses, which addressed my concerns well. After reading other reviews and responses, I decided to update my rating from Borderline Accept to Weak Accept. Please make necessary modification and add the relevant clarification, discussion, and experiment results to your final paper. --- Rebuttal Comment 2.1: Title: Thank you for the positive support Comment: We appreciate the comments that provide new perspectives on both experimental design and future research directions. We are glad to hear that your concerns have been addressed satisfactorily. We will ensure that the discussed changes and results are integrated into the revised manuscript to enhance its clarity and robustness. Thank you again for your positive support.	Summary: The paper presents a solution for better multimodal dynamic modeling in visual RL. The paper claims that existing works only emphasize consistent (aligned) information across modalities, leaving out the opportunity for the model to benefit from the inconsistent features. The work introduces a new consistency loss, where the (cross-modal) mutual prediction happens dynamically, meaning a modality has to infer the other modality's features for the next step rather than the current one. Furthermore, the work introduces a "soft objective" for cross-modal feature orthogonalization. The work shows consistent improvement over the existing state-of-the-art on CARLA and DMControl benchmarks. Strengths: 1) The paper is well-motivated, well-written, and easy to follow. The state-of-the-art results provide stats with several runs. The design of each loss component is straightforward. 2) The paper achieves significant improvement over the reported state-of-the-art results. 3) The cross-modal transition prediction loss is interesting and results in the most significant performance boost, as shown in Table 3. Overall, each loss component is shown to benefit the model's performance. 4) The method is generalizable beyond two modalities. Weaknesses: 1) The authors report the episode to return and driving distance but do not report DS/RC/IP, which is reported in other methods mentioned in the papers and could be more informative for evaluating driving performance. Is there any reason why the authors do not report those metrics? 2) L265-267 states that $ \mathcal{L_{fp}} $ and $ \mathcal{L}_{r} $ do not benefit much to the existing DeepMDPandSPR. Are there any supporting references/results for this claim? Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Do the authors think that incorporating a component of their modeling, e.g., $ \mathcal{L_{tp}} $, into existing models could result in a boost? 2) In Section 3.2 [L166], the authors suggest replacing a stronger objective (Eq. 6) with Eq. 7. Did you observe in your experiments that, indeed, a model trained with Eq.7 is better than Eq.6? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors discussed potential limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for the time and effort invested in reviewing our work. The positive remarks are truly appreciated, and we feel encouraged by the feedback. We address the points raised in the comments as follows: > The authors report the episode to return and driving distance but do not report DS/RC/IP, is there any reason why the authors do not report those metrics? Thanks for the thoughtful question on the metrics. Our evaluation protocol on CARLA follows the commonly adopted one in many existing RL works [1,2,3,4]. Different from the setting in TransFuser [5] (which reports DS/RC/IP), the RL agent's goal in this protocol is to drive as far as possible in limited timesteps without colliding into other moving vehicles or barriers. Each episode (i.e., evaluation trial) immediately ends when the agent vehicle collides or reaches the timestep limit. There is no predefined route during evaluation, so RC cannot be calculated. Because each episode immediately ends after a collision, the calculation of IP also becomes impossible. Since RC and IP are both impractical to be obtained, DS cannot be provided. Instead, the RL agent’s performance is primarily evaluated by Episode Reward (ER), which accounts for driving distance and driving stability, such as less abrupt steering. Note that although we compare with TransFuser in our paper, we evaluate it under the RL protocol rather than its original protocol. As we explained in the paper, this involves integrating the modality fusion and alignment modules of TransFuser with a baseline RL algorithm (SAC). The original setting in TransFuser and the RL setting in our work represent distinct approaches to training autonomous driving agents. The evaluation protocols and metrics for these two settings are also different, which is why we did not report DS/RC/IP scores. [1] SECANT: Self-Expert Cloning for Zero-Shot Generalization of Visual Policies, ICML 2021 [2] Learning Better with Less: Effective Augmentation for Sample-Efficient Visual Reinforcement Learning, NeurIPS 2023 [3] Model-Based Reinforcement Learning with Isolated Imaginations, TPAMI 2023 [4] Pre-training Contextualized World Models with In-the-wild Videos for Reinforcement Learning, NeurIPS 2023 [5] TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving, TPAMI 2023 > L265-267 states that $\mathcal{L}\_{fp}$ and $\mathcal{L}\_r$ do not benefit much to the existing DeepMDP and SPR. Are there any supporting references/results for this claim? Thanks for identifying this issue. There might be a misunderstanding here. Specifically, we are not trying to use $\mathcal{L}\_{fp}$ and $\mathcal{L}\_r$ to benefit DeepMDP and SPR. In L265-267, we state “However, this enhancement does not bring a significant advantage over conventional methods such as DeepMDP and SPR”. We mean that although $\mathcal{L}\_{fp}$ and $\mathcal{L}\_r$ can improve the performance of the baseline model, the improved results are still not significantly better than the results of DeepMDP and SPR (reported in Table 1 of our paper). So, we are describing a cross-table comparison, by comparing the $+\mathcal{L}\_{fp}$ and $+\mathcal{L}\_r$ rows in Table 3 of our paper and the DeepMDP and SPR rows in Table 1 of our paper. We did not mention Table 1 in L264-267, which may cause confusion. We appreciate your detailed examination and will correct this issue in the revised paper. > Do the authors think that incorporating a component of their modeling, e.g., $\mathcal{L}\_{tp}$, into existing models could result in a boost? We are confident that our modeling such as cross-modality transition prediction $\mathcal{L}\_{tp}$ can boost existing methods. This is because our method is flexible and is not limited to any particular RL model or network architecture. To verify the effectiveness of our method on other models, we further apply $\mathcal{L}\_{tp}$ to DrQ [1] and evaluate its performance. The results are given in Table 2 of the rebuttal PDF, which show that $\mathcal{L}\_{tp}$ can also boost the performance of existing RL models. [1] Image augmentation is all you need: Regularizing deep reinforcement learning from pixels, ICLR 2020 > In Section 3.2 [L166], the authors suggest replacing a stronger objective (Eq. 6) with Eq. 7. Did you observe in your experiments that, indeed, a model trained with Eq.7 is better than Eq.6? Yes, we have compared Eq.6 and Eq.7 in Sec.4.4 (L304-L313) of our paper. The results are presented in Fig.7 of our paper, which show that models trained with Eq.7 consistently perform better than models trained with Eq.6. --- Rebuttal Comment 1.1: Comment: Dear authors, I greatly appreciate your responses and the additional results presented in the PDF. I think the authors addressed all my comments and I think this work should be accepted. --- Reply to Comment 1.1.1: Title: Appreciation for the supportive feedback and recommendation Comment: We are truly grateful for your positive comments and for recognizing the efforts in our response and additional results. Thank you once again for your support and the time you invested in reviewing our work thoroughly.	Summary: The paper proposes Dissected Dynamics Modeling (DDM), a dynamics modeling framework for learning latent features in multi-modal visual RL. The methodology focuses on capturing both the shared and distinct information contained across input modalities. The paper presents a multi-modal architecture and training loss designed to accomplish this, and experimentally demonstrates the benefits of DDM on tasks from CARLA and the DeepMind Control Suite. Strengths: [S1] Novel approach to multi-modal RL: The paper proposes a novel methodology for multi-modal visual RL that accounts for both the similarities and differences across input modalities, compared to prior works that largely focus only on the similarities across input modalities. Experiments demonstrate the benefits of this novel approach. [S2] Detailed experimental analysis: The paper provides detailed experimental analysis of the proposed DDM method. This includes comparisons across several types of baselines (although some choices of baselines are not state-of-the-art), ablation studies that justify design choices, robustness analysis, and visualizations for qualitative understanding. [S3] Important problem: The paper focuses on how to effectively leverage information from multiple input modalities for decision making. This addresses an important consideration for the successful deployment of RL in real-world systems where multi-modal data is common. Weaknesses: [W1] Clarity of implementation / experimental details: The high-level idea of the DDM framework is clear, but some of the implementation details are not clearly described. Some components of the experiments were also not described in detail, which made it difficult to interpret some of the results. Please see the questions below. [W2] Connections to related work: DDM appears to build upon previous works in visual / multi-modal RL for its dynamics modeling and consistency extraction components, but these connections are not made clear. I think it would be useful to better emphasize what parts of DDM build upon existing works vs. what parts are novel contributions. (i) The dynamics modeling approach looks similar to reconstruction-free approaches in visual RL [a,b,c,d], but the similarities / differences are not discussed. (ii) It is mentioned that existing multi-modal approaches have focused on aligning modalities, but the paper does not discuss how the proposed consistency extraction method relates to these approaches. [W3] Heuristic design choices: Design choices of the proposed architecture are justified through experimental results, but no theoretical support is provided (or connections to existing approaches that provide theoretical support). --- References: [a] Gelada et al. (2019). DeepMDP: Learning continuous latent space models for representation learning. [b] Schwarzer et al. (2021). Data-efficient reinforcement learning with self-predictive representations. [c] Okada et al. (2021). Dreaming: Model-based reinforcement learning by latent imagination without reconstruction. [d] Zhang et al. (2021). Learning invariant representations for reinforcement learning without reconstruction. Technical Quality: 3 Clarity: 3 Questions for Authors: Methodology: [Q1] Are gradients with respect to the latent feature encodings $z$ stopped in any components of the loss function in (13), or are they propagated through the loss function everywhere that $z$ appears? [Q2] Are the next-step target values $z_{t+1}$ in (5) and (11) calculated using the same feature encoder used for $z_t$? It would be helpful to make this more clear. Are gradients taken through $z_{t+1}$ in (5) and (11)? --- Experiments: [Q3] What do the single modality baseline results represent? Do they only use RGB images, or do they consider a combined input that incorporates information from all modalities? If these baselines are not restricted to RGB images, it would be useful to include results (baselines or ablation of DDM) using only RGB inputs to demonstrate that the use of multi-modal inputs leads to improved performance over standard visual RL. [Q4] DreamerV3 [e] and DrQ-v2 [f] are strong model-based and model-free visual RL algorithms, respectively. Why were DeepMDP and SPR chosen as baselines instead of DreamerV3 (which has reported better head-to-head performance vs. SPR in Atari 100k)? Why was DrQ chosen as a baseline instead of DrQ-v2 (which has reported significant improvements over DrQ on DeepMind Control Suite)? [Q5] Has performance converged by the end of training for the experimental results, given the number of training steps used (100k and 500k in CARLA and DeepMind Control Suite, respectively)? In prior works on visual RL, DeepMind Control Suite tasks are typically trained for longer than this (1M steps in DreamerV3 [e], 3M steps in DrQ-v2 [f]), and these baselines had not converged to final performance after 500k steps. --- Minor: Spelling: Inon. --> Incon. (p. 2, Figure 1 caption) --- References: [e] Hafner et al. (2023). Mastering diverse domains through world models. [f] Yarats et al. (2021). Mastering visual continuous control: Improved data-augmented reinforcement learning. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have addressed the limitations of the current work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to express our sincere gratitude for the insightful comments. We also deeply appreciate the suggestions regarding presentation and experimentation. Below are our responses to your concerns: > [W1] Some implementation and experiment details are not clear. Please refer to our responses to [Q1-Q5] for detailed clarification. > [W2] Connections to related work need to be emphasized. Thank you for the suggestion. Here we discuss the connections in detail: (i) Connections with dynamics modeling methods [a-d] The primary similarity between [a-d] and our DDM is the state and reward prediction strategy commonly used in RL dynamics modeling. However, DDM differs as it is tailored for multi-modal RL, introducing novel approaches in both modality decomposition and dynamics modeling. Specifically, both DeepMDP [a] and DDM predicts latent states and rewards. However, DDM does not address modality relationships or learn state transitions in a modality-aware manner. SPR [b] uses multi-step state prediction but it overlooks modality interactions. Dreaming [c] mainly focuses on learning over different timesteps and samples. Differently, DDM aims at dissecting features across different modalities. DBC [d] employs a bisimulation metric to model dynamics, which differs from our dissected modeling strategy. (ii) Relation to existing consistency extraction methods Although the goal of drawing modality features closer for consistency extraction is similar between DDM and existing methods, the techniques vary significantly. MAIE [32] directly minimizes modality features at the same timestep. In contrast, DDM predicts cross-modality transition across adjacent timesteps. TransFuser [5] and EFNet [40] utilize attention, MUMMI [4] optimizes mutual information, and HAVE [20] employs attention and hypernetworks. Like MAIE, these methods do not create a synergy between dynamics modeling and modality alignment, which is a unique feature of the cross-modality transition prediction in our DDM. (Reference numbers are the same as in our paper) > [W3] No theoretical support is provided. Thank you for emphasizing the need for a theoretical analysis. We have conducted initial research to find theoretical insights that support our approach. Our investigation suggests that information theory [1] might provide a suitable theoretical framework for our study. For instance, research in [2] shows how multiple sources provide structured multivariate information about a target variable. Analogously, in multi-modal RL, each modality can be treated as a source and the agent's action as the target. Based on this, we aim to establish a lower bound on the information gap between using only modality-consistent features versus all available information. Deriving this lower bound will underscore the importance of both modality consistencies and inconsistencies in decision-making, thus supporting our method design and empirical findings. However, despite our earnest efforts on this idea, establishing a definitive result remains challenging for us within the limited rebuttal timeline. We sincerely value your feedback and will deepen this theoretical analysis in future work. [1] The Mathematical Theory of Communication, Bell System Technical Journal, 1948 [2] Nonnegative Decomposition of Multivariate Information, arXiv, 2010 > [Q1] Are gradients of $z$ stopped in (13)? Sorry for not clarifying this. We stop the gradient of all prediction target features in (4), (5), (8), and (11). That is, for each loss term of the form $\|\|X(z_a)-z_b\|\|^2$, where $X$ represents prediction head and $z_a$, $z_b$ are modality features, we stop the gradient of $z_b$. > [Q2] 1. Are $z_{t+1}$ in (5) and (11) used the same encoder as $z_t$? 2. Are gradients taken through $z_{t+1}$? 1. Yes, they are. We do not utilize a moving average encoder like in SPR [b] but rather use the same online encoder while stopping the gradient of the target values like in DBC [d]. 2. The gradients are stopped as explained in Q1. This avoids potential model collapse, which is observed in SPR paper when no stop gradient operation is applied to the target values. ([b,d] are the same as in [W2]) > [Q3] What are single modality baseline results? Any results using only RGB inputs? The results consider a combined input of all modalities. Following the advice, we have tested the case where only RGB inputs are used. The results are in Table 1 in the rebuttal PDF, which shows that the baseline with multi-modal input indeed outperforms RGB. > [Q4] Why use DeepMDP, SPR and DrQ as baselines instead of DreamerV3 and DrQ-v2? For single-modality baselines, we would like a direct comparison of the dynamics modeling technique itself and reduce the impact of other factors (e.g., the RL algorithm and the use of planning mechanism or not). Therefore, we choose DeepMDP, SPR and DrQ, all use SAC as RL algorithm without planning, just like our method. We appreciate the recommendation of DreamerV3 and DrQ-v2 and have conducted experiments with them. As shown in Table 2 of the rebuttal PDF, DrQ-v2 obtains stable performance, while DreamerV3 gets inferior results on the Cartpole swingup_sparse task. This is probably caused by the sparse reward and modality inconsistencies, which hampers the learning of its transition model. > [Q5] Has the performance converged? Our setting of training steps follows DBC [d] and SECANT [1]. To verify the convergence, we further train different methods with double steps (i.e., 200K/1M for CARLA/DMControl). The results are given in Fig.1 (a)-(d) in the rebuttal PDF. The figure shows that the improvement beyond 100K/500k steps becomes exceedingly marginal. Therefore, our current training can achieve a satisfactory level of performance. [d] is the same as in [W2] [1] SECANT: Self-Expert Cloning for Zero-Shot Generalization of Visual Policies, ICML 2021 > Minor: Inon.-> Incon. Thanks for spotting this, we will correct it. --- Rebuttal Comment 1.1: Title: Response Comment: Thank you for the detailed responses and additional experimental results. Incorporating some of these discussions / results into the paper will improve clarity and further strengthen the experimental analysis. My main questions and concerns have been addressed, and I have increased my overall review score to reflect this. --- Reply to Comment 1.1.1: Title: Thank you for the supportive feedback Comment: We appreciate the guidance provided in your detailed and informative review. We will incorporate these discussions and results into our revised manuscript. Thank you again for your supportive feedback.	null	null	Rebuttal 1: Rebuttal: We sincerely appreciate all reviewers for the valuable time and effort dedicated to reviewing our work. The comments have been highly constructive, and the positive evaluations from all reviewers are immensely encouraging. We have carefully considered each remark and have responded accordingly. For a quick and convenient overview, we briefly outline our responses as follows: 1. Additional experiments. Following the reviewers’ suggestions, we have conducted additional experiments, such as evaluating more baselines, employing our design to other RL models, and test on different camera perspectives. 2. Clarification of details. We have clarified several details regarding our technical design and paper presentation, such as the gradient flow during training and the explanations of certain words and phrases in our paper. 3. Further discussions. Based on the reviewers’ comments, we have conducted additional discussions on several key aspects, such as the relationship of our method with existing methods and the analysis of modality alignment methods. Please see our detailed responses below and the attached rebuttal PDF for addressing the individual comments from each reviewer. In addition to the rebuttal, we will incorporate these responses into our revised paper to further enhance its clarity. We thank you once again for your insightful feedback. Pdf: /pdf/7ed2f4b907bef70e57ef6c6bf12993d004e2315d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering	Accept (poster)	Summary: The paper proposes a cross-view reconstruction-based multi-view clustering algorithm to address the issue of anchor shift in missing data scenarios. Specifically, The method guides anchor learning by reconstructing the missing parts of the data. It uses an affine combination-based reconstruction strategy, rather than a convex combination, to avoid the negative impact of blind reconstruction areas. Theoretical analysis demonstrates the superiority of affine combination constraints over convex combination for this problem. Comparative experiments on multiple datasets with varying missing rates show the effective clustering capability of the proposed algorithm. The key innovations are the cross-view reconstruction approach, the use of affine combination constraints, and the theoretical justification for this choice. The experimental results validate the effectiveness of the proposed method for handling missing data in multi-view clustering tasks. Strengths: The anchor shift problem studied in this paper is a challenging issue that has been rarely addressed in the multi-view clustering field. The proposed method effectively addresses the observed problems through the collaboration of two modules, yielding good experimental results. The theoretical analysis provided in the paper intuitively demonstrates the superiority of the affine combination-based framework over previous convex combination approaches. Weaknesses: However, the paper contains some inaccurate descriptions, such as stating that matrix decomposition has n^2 complexity in the introduction, which is not entirely accurate. The paper includes many symbols, but the lack of a symbol table reduces readability. There is a contradiction regarding the final step of Algorithm 1, where it states k-means is directly applied to Z, while Section 3.2 describes applying k-means to the left singular vector matrix of Z. Technical Quality: 3 Clarity: 3 Questions for Authors: Regarding Algorithm 1, the paper does not discuss the possibility of using spectral clustering instead of k-means in the final step. It's unclear how this substitution would affect the performance of the proposed algorithm. Additionally, the paper does not clearly explain the relationship between the cross-view learning approach used in this work and the broader multi-view learning field. The connection between these concepts is not well-articulated. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: See Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Inaccurate statements Thanks for the comment. Due to the additional introduction of regularization constraints on the entire graph, such as manifold constraints, most existing matrix decomposition methods exhibit an $O(n^2)$ complexity [1][2]. We will revise this statement in the final version. 2. Notation table Thanks for the comment. We have added a symbol table to explain the key symbols used in our paper in the global rebuttal file. This symbol table will also be added in the final version to enhance readability. 3. Contradiction in Algorithm 1 Sorry for our mistake. In our method, the final clustering results are produced by applying $k$-means to the left singular vector matrix of $\textbf{Z}$. We will correct it in our final version. 4. Final process of clustering Thanks for the comment. The anchor graph $\textbf{Z}$ obtained from our method is $n \times m$, which cannot be directly used for spectral clustering. According to [3], applying $k$-means to the left singular vectors of $\textbf{Z}$ is equivalent to performing spectral clustering on the doubly stochastic matrix constructed from $\textbf{Z}$. Our method differs from [3] in the constraints applied to $\textbf{Z}$. Therefore, in Section A.4, we prove that the anchor graph $\textbf{Z}$ obtained from our method can also be recovered as a doubly stochastic matrix. 5. Relationship of cross-view learning and multi-view learning Thanks for the comment. Cross-view learning is a subset of multi-view learning. Unlike most previous multi-view learning methods, which build relationships between samples within the single view, cross-view learning aims to explore the relationships between samples across different views. In multi-view alignment methods, cross-view learning is a key technique for aligning samples between different views. In our approach, based on the assumption that samples will always appear in at least one view under missing data scenarios, introducing cross-view learning helps mitigate the impact of missing samples within a single view on anchor point learning. References [1] Rai N, Negi S, Chaudhury S, et al. Partial multi-view clustering using graph regularized NMF. 2016 23rd International Conference on Pattern Recognition (ICPR). IEEE, 2016: 2192-2197. [2] Zhao H, Ding Z, Fu Y. Multi-view clustering via deep matrix factorization. Proceedings of the AAAI conference on artificial intelligence. 2017, 31(1). [3] Wang S, Liu X, Liu L, et al. Highly-efficient incomplete large-scale multi-view clustering with consensus bipartite graph. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 9776-9785.	Summary: By employing cross-view anchor learning and affine combination-based reconstruction，the authors propose an incomplete multi-view clustering method to alleviate the anchor-shift problem. Besides, the authors theoretically analysis the advantages of affine combination-based reconstruction, which help to explore blind spots in sample reconstruction. Experimental results on several datasets validate the effectiveness of the proposed method. Strengths: 1.The affine combination-based reconstruction module is a simple yet effective novel approach. 2.The authors provide thorough theoretical and experimental validation of the proposed module's effectiveness. Weaknesses: 1.The authors fail to provide precise definitions for some symbols used in the paper, such as n_p. 2.Some inconsistencies are present in the experimental figures. For example, the x-axes of the first three subfigures in Fig. 3 are spaced at intervals of 20, whereas Fig. 3d uses intervals of 50. Additionally, Fig. 4c and 4d need to be rotated for better presentation. 3.The writing quality needs improvement. The conclusion section is too brief and does not adequately summarize the paper. Technical Quality: 3 Clarity: 3 Questions for Authors: 1.Can the authors provide a clearer explanation of Fig. 1, especially in its caption? 2.Why do the last three datasets in Table 4 have identical dimensions? Is this an oversight by the authors? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have acknowledged some limitations of their work but have not discussed future work directions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Definition of symbols Thanks for the comment. $n_p$ represents the number of samples in the $p$-th view. In the global rebuttal file, we have provided a notation table that explains the main symbols used in the paper. Please refer to the Table 2 in the global rebuttal file. 2. Inconsistencies of figures Thanks for the comment. We have made corrections to the inconsistencies in Fig. 3 and the obscured parts in Fig. 4. Please refer to Figure 1 and Figure 2 in the global rebuttal file. 3. Rewrite conclusion Thanks for the comment. We have rewritten the conclusion as follows to better summarize: In this paper, we introduce an AIMC-CVR method to alleviate the anchor-shift problem in anchor-based incomplete multi-view clustering. Specifically, we introduce a cross-view learning module and a reconstruction module to fix the influence of incomplete data. Additionally, we explore the blind spots in sample reconstruction with affine combination. Experiments and theoretical analysis validate the effectiveness of the proposed AIMC-CVR method. We will replace the conclusion in our final version. 4. Clearer explanation of Fig. 1 We further explain Fig.1 as follows: (a)Anchors learned in complete data: The true anchors are generated with k-means performed in the complete data. (b)Anchors initialized in incomplete data: The initial anchors are generated with k-means performed in the incomplete data, which is shift the origin point compared with the true anchors. (c)Data reconstructed with convex combination: The convex combination-based reconstructed data is restricted in the convex hull of anchors. (d)Data reconstructed with affine combination: The affine combination-based reconstructed data is breaking through the limitations of convex hull, which can represent any position in affine space. 5. Explanation of the dataset All three datasets consist of image data. We used four operators to extract features for each image, namely LBP (Local Binary Pattern), HOG (Histogram of Oriented Gradient), Gist, and Gabor. These operators provide four different views, each with the same dimensionality on different datasets.	Summary: This paper proposes an anchor-based incomplete multi-view clustering with cross-view reconstruction (AIMC-CVR). To tackle the anchor-shift induced by incomplete multi-view data, AIMC-CVR reconstructs missing samples with learned anchors. The traditional convex combination is replaced with affine combination for more reconstruction accuracy. Two theorems is proposed to demonstrate the advantages of affine combination. Strengths: The motivation of this paper is clear, focusing on the anchor shift problem in missing data scenarios, and the corresponding solution is theoretically and experimentally validated. Weaknesses: The paper emphasizes the missing rates in the experimental data but lacks a detailed description of how the missing data is constructed. The descriptions of the five variants in the ablation study are not clear enough. Understanding these variants is crucial for validating the effectiveness of the proposed algorithm. There are some symbol errors in the paper, such as "KTT conditions" which should be "KKT conditions". Technical Quality: 3 Clarity: 3 Questions for Authors: Why use v*v projection matrices? In fact, v matrices projecting data from each view to the same dimension should be sufficient to solve the dimensional inconsistency problem. Could the authors provide further explanation? In 2021, Yin et al. studied the reconstruction of missing datap[1]. How does the reconstruction method in this paper compare to theirs in terms of advantages? [1] Yin J, Sun S. Incomplete multi-view clustering with reconstructed views. IEEE TKDE, 2021. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have acknowledged some limitations of their work in the appendix and suggested potential solutions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Incomplete data construction Thanks for the comment. For the datasets mentioned in our method, we remove some instances on each view randomly to get their incomplete versions. Specifically, with the principle that each instance is present in at least one view, we generate missing datasets at missing rates in intervals of 0.1 from 0.1 to 0.9. 2. Descriptions of the five variants Thanks for the comment. The five variants of our proposed method are constructed as follows: (1) AIMC-CVR-v1 removes the cross-view anchor learning module by setting $\lambda = 0$. (2) AIMC-CVR-v2 removes the affine combination-based reconstruction module. (3) AIMC-CVR-v3 removes the sparsity regularization term by setting $\beta = 0$. (4) AIMC-CVR-v4 keeps the initialized $\mathbf{A}^{(p)}$ fixed and does not update it during subsequent optimization. (5) AIMC-CVR-v5 replaces affine combination with convex combination by adding non-negative constraints to $\mathbf{Z}^{(p)}$. 3. Symbol errors Sorry for our mistake. We will correct the symbol errors in our final version. 4. Different projection strategy In fact, we only used $v(v-1)/2$ projection matrices because we project lower-dimensional data to higher-dimensional spaces when one view's data dimension is smaller than another's. Unlike previous methods that use $v$ projection matrices to project data into the same lower-dimensional space, our approach aims to preserve high-dimensional information and reduce information loss. To validate the effectiveness of our method, we replace the projection matrix $\mathbf{W}^{(pq)}\in \mathbb{R}^{d_p \times d_q}$ between the $p$-th and $q$-th view with $\mathbf{W}^{(p)}\in \mathbb{R}^{k \times d_p}$ to get the variant AIMC-CVR-V6. The clustering performance of our proposed AIMC-CVR and its variant AIMC-CVR-V6 is shown in Table 1 on the global rebuttal file, which demonstrates the superiority of our projection strategy. 5. Novelty of imputation strategy The imputation method proposed by Yin et al. is constructed based on an $n \times n$ full graph, which incurs $O(n^2)$ space complexity and $O(n^2 log(n))$ time complexity. In contrast, our method achieves imputation based on the anchor graph, with both time and space complexity being linear with respect to $n$, offering a significant efficiency advantage, as demonstrated in Section A.6. Additionally, we expand the reconstruction space of the samples into the affine space of the anchors, which provides a larger search space compared to Yin's method.	Summary: This paper proposes a novel anchor-based IMVC method called AIMC-CVR to address the anchor-shift caused by missing data. AIMC-CVR consists of two modules: cross-view anchor learning and affine combination-based reconstruction. The former helps in learning a complete anchor graph, while the latter aims to recover the missing data. Carefully designed experiments validate the effectiveness of AIMC-CVR in clustering tasks. Strengths: -The concept of anchor shift is novel and introduced for the first time in this paper. -Experiments demonstrate the effectiveness of the proposed method in clustering performance. Weaknesses: -The paper introduces two modules: cross-view anchor learning and affine combination-based reconstruction. However, it lacks a description of the relationship between these two modules. Are both modules necessary in the proposed method? -The description of the experimental setup is unclear, particularly the settings in Figure 1. Clear descriptions of the data sources and the anchor generation method are crucial to support the motivation of this paper. Technical Quality: 3 Clarity: 2 Questions for Authors: -What do the gray areas in Figure 1(c) represent? There seems to be a deviation between the reconstructed points in Figure 1(d) and the real points in Figure 1(a). Does this deviation affect subsequent results? -Why do different datasets show different convergence states in Figure 3? I observed that some converge earlier while others converge later. Can you provide further explanation? Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. The necessity of the proposed modules: Thanks for the comment. In AIMC-CVR, both modules are essential for alleviating the anchor-shift problem caused by incomplete data. The cross-view anchor learning module mitigates such problem by leveraging available data across views to learn complete anchor graphs and more accurate anchors. Meanwhile, the affine combination-based reconstruction module focuses on reconstructing missing samples based on the anchors, thus preventing the anchors from being affected by the missing data. These modules cooperate towards the same goal: the former utilizes cross-view complementary information, and the latter employs imputation techniques to alleviate anchor-shift. However, in practice, the reconstruction module depends on the initial anchors learned by the cross-view anchor learning module. Ablation experiments show significant performance drops in versions V1 and V2, where each of these modules is removed, underscoring their necessity. 2. Experimental setup in Fig.1: Thanks for the comment. When generating data, we initially uniformly select four points on a circle with a radius of 1 centered at the origin in a two-dimensional coordinate system as cluster centers. Subsequently, we randomly generate 40 two-dimensional vectors following the standard normal distribution. Each group of 10 vectors is added to the corresponding cluster center, resulting in four classes with a total of 40 data points. Notably, the two-dimensional vectors generated on the two views are distinct but equally important. In Fig. 1a, the true anchors are obtained by applying k-means on the complete data. Fig. 1b shows the initial anchors derived from k-means on the incomplete data. Fig. 1c presents the anchors based on AIMC-CVR-V5, where the affine constraint is replaced with a convex constraint. Fig. 1d displays the anchors learned using the proposed AIMC-CVR method. We will add the above setting in the final version. 3. Further description of Fig.1: Thanks for the comment. In Fig. 1c, the white area represents the convex hull formed by four anchors, indicating that all points within this area can be expressed as a convex combination of the anchors. The gray area represents the blind spots where points cannot be expressed by the convex combination of these anchors. The points in Fig. 1a are the original sample points, whereas the black points in Fig. 1d are reconstructed with our proposed method. Although there are some differences in representation, the reconstructed points in Fig. 1d are evidently closer to the true points compared to those restricted within the convex hull in Fig. 1c. Our objective is not to precisely recover the missing samples, but to mitigate the impact of missing samples on the anchors. Compared to the anchors generated in Fig. 1b, the anchors in Fig. 1d are closer to the true anchors. Thus, while our method can not completely recover the missing samples, it effectively alleviates the anchor-shift problem in incomplete multi-view clustering, achieving state-of-the-art clustering performance. 4. Difference in convergence states on different datasets: Thanks for the comment. We speculate that the differences may arise from the following three factors: Firstly, different types of data naturally have inherent differences, leading to varying numbers of iterations required to meet the convergence criteria during optimization. Secondly, when fusing multi-view data, the similarity between views varies across datasets. Data with more similar views converge faster during integration, while data with greater view differences take longer. Thirdly, the alternating optimization algorithm is highly influenced by the initial value settings, with different initial settings resulting in different numbers of iterations. --- Rebuttal Comment 1.1: Comment: Thanks for your responses. Most of my concerns have been addressed. I will raise my score.	Rebuttal 1: Rebuttal: We thank the SAC, AC, and PCs for their efforts and constructive comments, which are helpful in further improving the quality of our manuscript. We respond to your questions carefully one by one carefully, and we hope our responses can address your concerns. Note that there are two tables and two figure in the attached PDF, corresponding to RQ4 for Reviewer fq57, RQ1 and RQ2 for Reviewer 4LaZ, and RQ2 for Reviewer mjvb. Pdf: /pdf/afbf2716d4af2c901fadaad5005796d37ea17bc8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Key-Grid: Unsupervised 3D Keypoints Detection using Grid Heatmap Features	Accept (poster)	Summary: This paper presents a novel unsupervised network for keypoint detection. The method can be applied to handle both rigid and deformable objects. It follows an autoencoder framework where the encoder predicts keypoints and the decoder utilizes the generated keypoints to reconstruct the objects. The main contribution lies in the grid based representation which is suitable to capture the shifted geometric structure of the keypoints brought by deformations. Strengths: 1 This paper presents an unsupervised method for consistent keypoint detection. "Unsupervised" means the method can be applied in larger-scale datasets without 3D annotations. 2 The paper is easy to follow. The writing is good and clear. 3 Extensive experiments show that this method achieves sota performance on many public datasets Weaknesses: 1 The visualized results presented in this paper all feature uniform and complete point clouds. Can this method handle partial point clouds, such as those obtained from back-projecting depth maps? Can it deal with point clouds that contain severe outliers, not just those with some Gaussian noise? This is very important for practical use. 2 In 289, However, if using a SPRIN module [41] which is a SOTA SE(3)-invariant backbone 289 to replace PointNet++ [22] directly, the training process of Key-Grid does not converge. Why? How about using other backbones? like Vector Neurons? Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., & Guibas, L. J. (2021). Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 12200-12209). 3 This proposed representation is based on Euclidean distance between the grid point towards each line segment, right? Why not first establish a graph and then aggregate features along the graph? I saw many papers using graph to model deformable objects. Technical Quality: 2 Clarity: 3 Questions for Authors: I have listed some questions in the weakness. The first question is the most important. In L 14, 'Meanwhile, we incorporate the information from each layer of the encoder into the decoder section', Is this your contribution? I recommend to move Fig 6 in the appendix to the main text for clarity. Confidence: 2 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The author has discussed potential limitations in the paper Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your great suggestion. In the rebuttal phase, we address the reviewers' concerns about the effectiveness of our method on partial point clouds, such as those obtained from back-projecting depth map sand point clouds containing outliers. We also provide a detailed explanation of the convergence issues with the SE(3)-invariant model and highlight the advantages of using grid heatmaps over graphs. $\color{Indigo}Q1$: The visualized results presented in this paper all feature uniform and complete point clouds. Can this method handle partial point clouds, such as those obtained from back-projecting depth maps? Can it deal with point clouds that contain severe outliers, not just those with some Gaussian noise? This is very important for practical use. $\color{red}A$: In Figure 1(a) and (b), we present the visualization results of keypoints detected by Key-Grid on both point clouds with outlines and partial point clouds sampling from the depth map. We observe that Key-Grid identifies keypoints with strong semantic consistency, even when dealing with point clouds containing ten outliers in Figure 1(a). For the point clouds obtained from depth maps, we first generate depth maps of objects from multiple-angle photographs and then sample point clouds based on the depth maps. We find that when facing the low-quality/partial point clouds, Key-Grid identifies keypoints whose locations are similar to those detected in the complete point clouds. $\color{Indigo}Q2$: In 289, However, if using a SPRIN module which is a SOTA SE(3)-invariant backbone to replace PointNet++ [22] directly, the training process of Key-Grid does not converge. Why? How about using other backbones? Like Vector Neurons? $\color{red}A$: Figure 3 in our new submission material provides the training loss of Key-Grid using SPRIN and Vector Neurons instead of PointNet++ to detect keypoints on the fold pants. By examining the curves, we observe that Key-Grid does not converge by using SPRIN and Vector Neurons in the training phase. We think the SE(3) backbone model is not converging, primarily because of the stringent training conditions. Our designed loss function fails to provide sufficient information to enable the SE(3) model to converge. Like USEEK, to achieve the convergence of the SE(3) model, the authors compute the difference between features outputted by SPRIN and features from the pre-trained PointNet++ in SM as a loss function to optimize the SE(3) model. If the SE(3) model is directly used as the backbone in SM and Key-Grid, we only optimize it by locating the keypoints instead of providing a standard feature for the SE(3) model to learn from. Therefore, we think this is the main reason why the SE(3) model fails to converge. $\color{Indigo}Q3$: This proposed representation is based on Euclidean distance between the grid point towards each line segment, right? Why not first establish a graph and then aggregate features along the graph? I saw many papers using graph to model deformable objects. $\color{red}A$: Thank you for your valuable perspective. Currently, applying graph networks to deformable objects typically involves constructing a graph structure from point clouds, where nodes represent points on the object and edges capture mesh information between these points. Subsequently, graph neural networks input graph structures and output the required points [1]. However, the role of the grid heatmap is to use keypoints to build the skeletal structure of the deformable object, aiding the network in reconstructing the original point cloud. This process is contrary to the current application of graph neural networks for deformable objects. Therefore, we think that replacing grid heatmaps with graph structures is not suitable for our approach. [1] Learning language-conditioned deformable object manipulation with graph dynamics. arXiv:2303.01310, 2023. --- Rebuttal Comment 1.1: Title: Final Rating Comment: Thank you for the response. The rebuttal is very good and has addressed my major concerns, and I raise my rating to "weak accept." --- Rebuttal 2: Comment: Thank you for your feedback on our paper, particularly regarding the applicability of the Key-Grid to partial and noisy point clouds. We will include this section in the next version of the paper. We greatly appreciate your consideration of the paper as ready for "weak accept". We notice that the rating is not updated. We kindly hope that you update the rating of our paper to reflect the improvements made to the paper.	Summary: This paper presents Key-Grid, an unsupervised keypoint detection network designed for 3D point clouds. Unlike previous methods that emphasize leveraging various priors on 3D structures, this paper converts keypoints into a grid heatmap. This heatmap forms a continuous feature landscape across the entire 3D space, providing richer and more stable geometric descriptions of objects, particularly for deformable objects. Meanwhile, this paper achieves state-of-the-art (SOTA) results on the ShapeNetCoreV2 and ClothesNet datasets. Strengths: 1. The proposed grid heatmap is effective in the task of 3D keypoint detection and represents a novel approach to introducing 3D priors for deformable objects. 2. The authors conducted extensive experiments, achieving SOTA performance on both rigid and deformable objects. 3. The overall design follows the mainstream architecture of 3D keypoint detection, providing convenience for subsequent research in this field. Weaknesses: Q1. My main concern is whether it can be empirically or theoretically justified why the grid heatmap can represent a dense extension of the skeleton, and why the current method of generating the grid heatmap is reasonable. Could using the density of the 'skeleton' or the density of the point cloud directly represent the grid heatmap instead? Q2. In the ablation study (Table 4), the accuracy drop after "No Grid Heatmap" is minimal, yet the Grid Heatmap is the most significant contribution of this paper. Could the authors explain this phenomenon and provide ablation studies across more categories? Q3. Can this paper adaptively select the number of detected keypoints for different categories of objects? Q4. The quality of figures 1 and 2 could be enhanced. For example, the information content of figure 1 is too minimal to serve as a teaser. Technical Quality: 2 Clarity: 3 Questions for Authors: If the authors address the issues mentioned in the weaknesses, I am willing to increase the rating. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Yes, the authors discussed the limitations and potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your insightful comments. We address your concerns as follows: We explain the rationale of the grid heatmap from a theoretical perspective and highlight its advantages over the skeleton structure proposed by SM. Subsequently, we demonstrate that the grid heatmap is more effective for keypoint detection than skeleton structure on deformable objects through experimental validation. Furthermore, we clarify some confusions, such as the importance of the grid heatmap to the Key-Grid and whether Key-Grid can adaptively generate keypoints. $\color{Indigo}Q1$：My main concern is whether it can be empirically or theoretically justified why the grid heatmap can represent a dense extension of the skeleton, and why the current method of generating the grid heatmap is reasonable. Could using the density of the 'skeleton' or the density of the point cloud directly represent the grid heatmap instead? $\color{red}A$: Skeleton structures proposed by SM does not accurately represent the skeletal structure of deformed objects. For example, in Figure 2 of our paper, if you connect the red keypoint with the black keypoint, this skeleton does not rationally represent the structure of the folded pants. Unlike SM, which directly uses the lines connecting keypoints as the object's skeleton, we characterize the overall skeletal structure of deformed objects more delicately by depicting the distances between grid points and keypoints pair line. Figure 4(C) of our paper shows the visualization results of the grid heatmaps and skeleton structures proposed by SM， denoting that the grid heatmap provides a more accurate depiction of the underlying skeletal structure of deformed objects. We also report the performance of Key-Grid which replaces the grid heatmap with the skeleton on the fold deformation in the following table. We can conclude that the grid heatmap helps Key-Grid identify keypoints with more semantic relevance on the fold deformation of clothes. \| \| Folded Shirt \| Folded Pant \| \|--------------\|----------------\|--------------\| \| Grid Heatmap \| 92.0 \| 100.0 \| \| Skeleton Structure \| 83.9 \| 92.7 \| $\color{Indigo}Q2$: In the ablation study (Table 4), the accuracy drop after "No Grid Heatmap" is minimal, yet the Grid Heatmap is the most significant contribution of this paper. Could the authors explain this phenomenon and provide ablation studies across more categories? $\color{red}A$: Key-Grid utilizes the farthest point keypoint loss to pre-select several initial positions for keypoints. Subsequently, we employ the Grid Heatmap and leverage encoder information for hierarchical decoding to assist the model in finalizing these pre-selected positions. In other words, the Grid Heatmap fundamentally aids the model in selecting keypoint positions rather than directly generate their positions. Therefore, when removing the grid heatmap, its overall impact on the model's effectiveness is less significant compared to removing the farthest point keypoint loss, which provides initial candidate keypoint positions. However, in Table 4 from our paper, “No Grid Heatmap” actually has a greater impact on reducing the semantic relevance of keypoints compared to “No Encoder Information”, which also illustrates the significant role of grid heatmap in the decoder. In the following table, we present an ablation study of more diverse categories of deformable and rigid objects under the DAS metric and also show the average impact of ablating different modules on the performance of Key-Grid. \| \| Table (Rigid) \| Motorbike (Rigid) \| Hat (Drop) \| Long pant (Drop) \| Hat (Drag) \| Long pant (Drag) \| Average \| \|-------------------------\|---------------\|-------------------\|------------\|------------------\|------------\|------------------\|--------------\| \| Key-Grid \| 78.6 \| 63.7 \| 100.0 \| 83.7 \| 51.5 \| 99.0 \| 79.4 \| \| No Grid Heatmap \| 63.8 \| 51.3 \| 87.9 \| 70.6 \| 36.7 \| 86.1 \| 66.1(-13,3) \| \| No Encoder Information \| 65.3 \| 51.8 \| 89.1 \| 74.9 \| 39.8 \| 88.9 \| 68.3(-11.1) \| $\color{Indigo}Q3$: Can this paper adaptively select the number of detected keypoints for different categories of objects? $\color{red}A$: Thank you for your suggestion. Currently, in Key-Grid, we need to manually set the total number of predicted keypoints beforehand. In future research, we will propose an adaptive keypoint detection method that can generate varying numbers of keypoints with accurate positions for different samples. $\color{Indigo}Q4$: The quality of figures 1 and 2 could be enhanced. For example, the information content of figure 1 is too minimal to serve as a teaser. $\color{red}A$: Thanks for your advice. In the come ready version, we will submit the new version of Figure 1 and Figure 2. The new version of Figure 1 not only provides the visualizations of keypoints detected on both deformable and rigid objects but also adds the grid heatmap and skeleton structure visualizations like Figure 4(c) in our article, which demonstrates that grid heatmap can more accurately represent the geometric information of the object. --- Rebuttal Comment 1.1: Title: Additional comments Comment: Thank you for the response and the additional experiments. My main concerns regarding the grid heatmap representation and ablation study have been resolved. The quality of the Figures 1 and 2 has also improved. I tend to accept this paper. However, I still have another question: can the distances between key points or the density of key points express the dense extension of the deformable object? --- Reply to Comment 1.1.1: Title: Thanks for your respone Comment: In my view, using the distance from a point to keypoints to represent the deformable object’s structure is less effective than using the distance from a point to the skeleton. For example, in the main paper of Figure 2, if we use the distance from a point to keypoints, in the tail section of the pants (i.e., the area between the red and blue points), the value near the red point differs from the value at the center of this section. However, the geometric information represented by these two points should be consistent. By using the distance from a point to the skeleton, the point values are derived based on the skeleton spreading outward from the center. This results in different values for points near the skeletons compared to those at the object’s edges, reflecting their distinct geometric information. Therefore, we believe that using the distance from a point to the skeleton more accurately characterizes the structure of deformable objects than using the distance to keypoints.	Summary: The authors propose a novel unsupervised method to detect 3D point clouds key points by producing an intermediary heatmap based on a grid and points distances from the skeleton and connected key points. They achieve state-of-the-art accuracy and semantic consistency and easily achieve Se-(3) invariance with minimal adaptation of the method. Strengths: The method's description, results, and a very in-depth ablation study, together with many supplementary materials, make this paper pleasant to read and easy to understand, with impressive results. Although simple at first glance, the method has the merit of being solid and achieving good results. Great results section and analysis allow the reader to understand in depth the method and results. Weaknesses: A small neat-picking addition would be to add one or two visual examples where the method does not work super well, to contrast with the excellent results presented in the paper Technical Quality: 3 Clarity: 4 Questions for Authors: Perhaps I missed this point, but how does the author ensure that the key points are semantically the same across different folding/deformation of the objects? ( and give them the same colour) Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Perhaps the only limitation I can pick up is the consistency across more than two deformations of an object. Would the method perform well over long-term deformations? It ties with showing one or two failure cases of the method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable feedback. We address your concerns as follows: we provide demonstrations of our method on some less successful objects and investigate whether keypoints maintain semantic consistency across various deformations of objects. $\color{Indigo} Q1$：A small neat-picking addition would be to add one or two visual examples where the method does not work super well, to contrast with the excellent results presented in the paper $\color{red}A$: We present the visual results of the drag deformation for 'Mask' and 'Skirt' in Figure 4 of our submission material, which exhibit lower DAS values than other categories on the ClothesNet dataset. Although the semantic consistency of keypoints detected on Masks and Skirts is relatively lower than other categories, keypoints identified by Key-Grid uniformly distribute across the object's surface, located at positions rich in geometric information compared to other methods. For instance, keypoints detected by Key-Grid are evenly distributed along the straps of the mask. $\color{Indigo} Q2$：Perhaps I missed this point, but how does the author ensure that the key points are semantically the same across different folding/deformation of the objects? $\color{red}A$: Figure 2 of our submission material shows the visualization results of keypoints detected by Key-Grid on long pants under the dropping, pulling, and dragging deformation. We can observe that keypoints detected by Key-Grid keep great semantic consistency during this long-term deformation process. At the same time, we observe that keypoints identified by Key-Grid are uniformly distributed across deformable objects, with their positions containing substantial geometric information, such as the hem and waist of long pants.	Summary: Novel PointNet-based autoencoder method called KeyGrid, that predicts semantic keypoints on objects, even when objects are subject to deformations. Similar to previous approaches, keypoints are produced as a linear combination of inputs points, according to a learned weight matrix. The key novelty over related works using PointNet++-based auto-encoders such as SkeletonMerger (SM) [26] is to pass a densely sampled grid of features (3d “point heatmap”) at each decoder layer (which are hierarchical). In this heatmap grid, each cell holds the maximum of the weighted distances to the 'keypoint skeleton segments'. Also, KeyGrid appears to pass richer encoder information to the decoder than the SM work, although the differences here are less clear. The method is shown to outperform previous baselines such as SK3D and SM on ShapeNetV2 dataset (more right objects) and on ClothesNet (deformable objects). Strengths: - Beats KeypointDeformer (KD), SkeletonMerger (SM), SC3K on almost all categories on the ClothesNet dataset and ShapeNetCoreV2. Deformable objects keypoints seem to improve more. - Ablations in Tab. 4 indicate the importance on 4 (among dozens) of semantic classes how encoder-info, a grid heatmap, farthest point sampling loss, and similarity (i.e. chamfer loss) each boost performance. Weaknesses: # Significance - Is the hierarchical decoder setup (excluding the heatmap grid itself) novel, esp compared to SkeletonMerger [26]? Right now there are no specific claims or ablations related to this. - The primary novelty relative to SM [26] seems to be the use of keypoint grid heatmaps in the decoder, storing max distances from grid cells to the weighted keypoint segments. However, this is just one possible shape descriptor feature, and intuitively it should be quite a weak and ad hoc feature, even if the authors gave some intuition why it may be better than the min function. Would other shape feature with yet more information be more useful could (e.g. why not have K channels with distances to all segments or to the keypoints as opposed to just taking the max, etc)? In general it would help if the paper provided more intuition or experimental results giving more insight on the information contained in this heatmap. - This class of approaches seems generally limited to cases where entire point clouds were visible. How would you handle cases when some parts of the objects were occluded, such as in real-world scenes? This has not beed discussed. # Related work Could discuss other dense representations for correspondence like the pointmaps used in “DUSt3R: Geometric 3D Vision Made Easy” https://arxiv.org/pdf/2312.14132 ) # Clarity - I found Sec 3.3 particularly difficult to follow and understand. A rewrite, or further details in the appendix clarifying better how these quantities are 'composed' and aligned would help. - Fig 1 illustrates points heatmap in the decoder by simply showing a point cloud. I was expecting pictures like in Fig 4c. - L13, L58 Unclear where keypoint pairs come from, when discussed for first time. - No discussion of how correspondences are established, e.g. as shown in Figure 3. - The loss terms in Equation 11 are not mathematically defined, would be helpful in the Appendix. # Experimental results - KeypointDeformer, SkeletonMerger, and SC3K are tested on KeypointNet dataset (https://arxiv.org/pdf/2002.12687 ). Why do the authors omit it here? Other works such as SC3K state “We use KeypointNet dataset [36 ] in our experiments, considering that this is the standard and most recent dataset used for keypoints estimation”. If the results are not as good on that dataset, that is okay, as long as we provide them and state that our method is specifically designed for datasets with deformable/soft objects. This would provide confidence that the authors ran their eval of official source code correctly on the two new datasets. - Why no comparisons with supervised baselines? - Figure 4D: Sun3D keypoints are so small in the figure it appears as if they are randomly distributed # Some Grammar & Style Nits: - L453 SC3K bibtex should be ICCV 2023, not Arxiv “ - L148 typo “giving a total of 4096 gird points.” -> “…4096 grid points” - L229 “mean Intersection over Unions” -> “mean Intersection over Union” - Grammar: Figure 3: caption: “Different methods on the Hat and Long Pant during” -> prefer “Different methods on the Hat and Long Pant categories during” - Table 3 caption is too close to Table 3, whitespace seems too shrunken - L75 “Deformable object dataset” -> “Deformable object datasets” - Table 1 caption “codes” -> should be singular, not plural (“the results we reproduced based on their official codes” -> “the results we reproduced based on their official code.”) - Color scheme is jarring in Tables 1-4, prefer tango colors that have a more attractive colormap (https://sobac.com/sobac/tangocolors.htm) - L3 extraneous article: “focus on the rigid body objects” -> “focus on rigid body objects” - L6 extraneous article: “for both the rigid-body and deformable objects” -> “for both rigid body and deformable objects” - L300 “the importance of encoder information and grid heatmap for reconstruction process,” -> “...for the reconstruction process,” - L302: “Table 4 show that the keypoints detected by Key-Grid which utilizes both two strategies to reconstruct the point cloud have better semantic consistency.” -> “Table 4 shows that when using both input streams to reconstruct the point cloud, the keypoints detected by Key-Grid exhibit better semantic consistency” - L9-10 Instead of “Unlike previous work, we leverage the identified keypoint information to form a 3D grid feature heatmap 1called grid heatmap, which is used in the decoder section”, prefer to say something like “..to form a 3D grid feature heatmap which we refer to as a grid heatmap. A grid heatmap is…” - L14 “Into the decoder section” -> “into the decoder model”? Ambiguous what “section” refers to - L51 “aiming at the semantic consistency” -> “aiming for semantic consistency”, e.g. “an unsupervised keypoint detector on 3D point clouds aiming for semantic consistency under shape variations of both rigid-body and deformable objects.” - L74 section should be named “Related Work” not “Related Works” Technical Quality: 3 Clarity: 2 Questions for Authors: - Compared to SM[26], is the keypoint heatmap the main difference, how about the use of hierarchical decoder. - Wouldn't a richer shape descriptor heatmap do even better than a max function? (e.g. K channels, distances to the keypoints, etc). - How do you get keypoint correspondences across shapes for your model, do they emerge naturally? - What are the method results on KeypointNet dataset? Can you provide some additional comparisons to supervised baselines on your datasets, if there are such? - Can this method be applied on datasets where shapes are significantly occluded (e.g. viewed from a specific direction)? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Limitations have been sufficiently addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing thoughtful and detailed feedback, which greatly enhances the quality of our article. Based on your comments regarding related work, clarity, and grammar & style nits, we will revise our paper in the came ready section. Regarding your inquiries about experimental results and significance sections, we address each one individually. Additionally, we provide the new visualization results of keypoint detected by Key-Grid on the SUN3D dataset in the new submission material. $\color{Indigo}Q1$: Compared to SM, is the keypoint heatmap the main difference, how about the use of hierarchical decoder. $\color{red}A$: The main distinction between SM and Key-Grid lies in the decoder design. We incorporate information from the encoder into the decoder to reconstruct point clouds through hierarchical decoding. Additionally, we integrate grid heatmap information during the reconstruction process. Therefore, we consider the hierarchical decoder and the grid heatmap as the two significant innovations of our method. In our paper, we conduct ablation experiments on the hierarchical decoder in Section 4.5. In Table 4 from our paper, the label “No Encoder Information” means the ablation of the hierarchical decoder. We observe that when we do not use a hierarchical decoder, our method's performance will decrease on both the deformable and rigid objects. $\color{Indigo}Q2$: Wouldn't a richer shape descriptor heatmap do even better than a max function? (e.g. K channels, distances to the keypoints, etc). $\color{red}A$: In the following table, we display the impact of grid heatmaps with different distance definitions on keypoint recognition in deformable objects. These definitions are the K-channel distances from each grid point to all segments or keypoints. For the folding deformation, the distance defined in our paper achieves great performance on the semantic consistency of keypoints. We think that adopting a K-channel distance will result in some overlap of distance information between the point clouds outside the deformable objects and those inside. This condition reduces the ability of the grid heatmap to depict the structure of the deformable objects. \| \| Folded Shirt \| Folded Pant \| \|--------------------------------------\|----------------\|--------------\| \| Distance to keypoints \| 86.4 \| 93.1 \| \| Distance to segment \| 88.7 \| 95.8 \| \| Our \| 92 \| 100 \| $\color{Indigo}Q3$: How do you get keypoint correspondences across shapes for your model, do they emerge naturally? $\color{red}A$: The positions of keypoints are outputted in an orderly manner by Key-Grid and other baselines(SC3K, SM, and KD). In our article's visualization of keypoints, we use different colors to represent keypoints outputted in different orders. During the entire deformation process, if keypoints of the same color (with the same output order) maintain their positions unchanged, we refer to these keypoints as having good semantic consistency. The output of keypoints with good semantic consistency can trace various positions that contain essential geometric information in objects when objects undergo deformation. This is crucial for robot manipulation of deformable objects and understanding the deformation process of objects. In our paper, Key-Grid naturally outputs keypoints with semantic consistency. $\color{Indigo}Q4$: What are the method results on KeypointNet dataset? Can you provide some additional comparisons to supervised baselines on your datasets, if there are such? $\color{red}A$: Thank you for pointing out this issue. In the following table, we present various supervised networks and self-supervised methods to compare Key-Grid on the KeypointNet dataset based on mIoU metric. Several standard networks, PointNet [1], SpiderCNN [2], and PointConv [3], are trained on KeypointNet to predict the probability of each point being a keypoint in a supervised manner. We can observe that compared with other self-supervised methods, Key-Grid recognizes the more accurate location of keypoint, and it performs even better than some supervised methods using PointNet and SpiderCNN as backbones. \| \| Airplane \| Chair \| Car \| Average \| \|-----------\|----------\|-------\|-------\|---------\| \| PointNet \| 45.4 \| 23.8 \| 15.3 \| 28.2 \| \| SpiderCNN \| 55.0 \| 49.0 \| 38.7 \| 47.6 \| \| PointConv \| 93.5 \| 86.0 \| 83.6 \| 87.0 \| \| SM \| 79.4 \| 68.4 \| 63.2 \| 70.3 \| \| SC3K \| 82.7 \| 38.5 \| 34.9 \| 52.0 \| \| Key-Grid \| 80.9 \| 75.2 \| 69.3 \| 75.1 \| $\color{Indigo}Q5$: Can this method be applied on datasets where shapes are significantly occluded (e.g. viewed from a specific direction)? $\color{red}A$: Thanks for your insightful suggestion. Keypoints detected by Key-Grid estimated on the deformation of objects from side views are depicted in Figure 1 from the new submission material. Meanwhile, we also show the keypoint detection results when objects' shapes are occluded in Figure 1 from the new submission material. We find that Key-Grid exhibits robustness on the occluded point clouds and point clouds obtained from different views. [1] Pointnet: Deep learning on point sets for 3d classification and segmentation. CVPR 2017 [2] Spidercnn: Deep learning on point sets with parameterize convolutional filters. ECCV 2018 [3] Pointconv: Deep convolutional networks on 3d point clouds. CVPR 2019 --- Rebuttal Comment 1.1: Title: rebuttal response Comment: The authors have provided helpful answers and results that largely address my questions. I also saw the newly attached figure, which shows performance under some occlusion - which is ok even if the occlusion amount shown is quite small. In light of this, I am willing to raise my rating to 'weakly accept'. --- Rebuttal 2: Comment: Thank you for your detailed review and comments on our paper. We will revise our paper according to your suggestions. We are also pleased to hear that you believe our paper has reached the ’weak accept‘ stage. However, we observe that the rating is not updated. We would be grateful if you could revise the rating to better reflect the improvements made to the paper.	Rebuttal 1: Rebuttal: Thanks to the esteemed reviewers for your insightful feedback, which has significantly enhanced the quality of our paper. Based on your suggestions, we provide the corresponding visual results in the new submission material. Figure 1(a): In response to Reviewer YtUf and Reviewer LZF7's inquiries regarding our method's capability to handle the occluded, partial, and outlier-laden point clouds, we show the visualization results of Key-Grid's performance on these types of point clouds. Figure 1(b) : Regarding the Reviewer LZF7's inquiry about whether Key-Grid can handle point clouds obtained from depth maps, we generate a depth map from multi-angle images and then illustrate keypoints identified by Key-Grid on the point cloud sampled from this long pant depth map. Figure 2: For Reviewer U8aQ's concern about the effectiveness of Key-Grid on objects undergoing long-term deformation, we demonstrate the KeyGrid's capability to capture the keypoints with high semantic consistency during the long-term deformation processes containing dropping, pulling, and dragging. In response to the YtUf reviewer's concern about the visualization result of keypoints in the Sun3D dataset is not clear, we improve the visibility of the keypoints and present additional visualization results with more samples from the Sun3D dataset. Figure 3: Regarding the LZF7 reviewer's concern about the non-convergence of Key-Grid with the SE(3)-invariant backbone, we present the training loss curves of the Key-Grid method with the PointNet+/SPRIN/Vector Neurons models. Figure 4: In response to the U8aQ reviewer's request to show some examples where our method does not work well, we select the visualizations of keypoint detected on the Skirt and Mask under the drag deformations. Pdf: /pdf/0548ca2ef06a19b9fb9bbdaf83f59521a56846b7.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction	Accept (poster)	Summary: This paper proposes a paged-based KV cache manager that identifies and recalls important tokens for LLM inference, termed ArkVale. Results show that ArkVale achieves 2.2x latency and 4.6x throughput improvement on various long context tasks. Strengths: 1. The paper is easy to follow, with clear writing and presentation. 2. Evaluation results are comprehensive. Weaknesses: 1. How does the page size affect the memory consumption for the KV cache? Would smaller page size lead to potential fragmentation issues? 2. In the related work section, it would be nice if the authors could the relationship between ArkVale and some concurrent works [1,2,3]. [1] Keyformer: KV Cache Reduction through Key Tokens Selection for Efficient Generative Inference, MLSys 2024. [2] Q-Hitter: A Better Token Oracle for Efficient LLM Inference via Sparse-Quantized KV Cache, MLSys 2024. [3] ALISA: Accelerating Large Language Model Inference via Sparsity-Aware KV Caching, ISCA 2024. Technical Quality: 3 Clarity: 4 Questions for Authors: Please see the weaknesses above. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Please see the weaknesses above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >How does the page size affect the memory consumption for the KV cache? Would smaller page size lead to potential fragmentation issues? We allocate pages from a pre-allocated memory pool, and all pages have the same page-size, thus avoiding the issue of memory fragmentation. However, a smaller page-size results in an increased number of pages, leading to a greater number of page digests and consequently increased memory usage. --- >In the related work section, it would be nice if the authors could the relationship between ArkVale and some concurrent works [1,2,3]. Keyformer [1] are similar to works like H2O in that they use historical information to perform eviction on the KV-Cache, and they have improved the eviction score function by leveraging the Gumbel distribution. Q-Hitter [2] combines KV-cache quantization with KV-cache eviction. But neither of these works can dynamically assesses the importance of tokens. ALISA [3] uses a post-training dynamic sparse attention approach, similar to SparQ, which dynamically evaluates the importance of each token in the KV-cache and selects a subset of tokens for attention calculation. However, this token-level approach is too granular and incurs significant additional overhead (requiring all KV-cache tokens to be involved in assessing importance), and it also poses challenges for memory management if it requires token-level eviction & recall. In contrast, our approach operates at the page-level for importance estimation, eviction, and recall, achieving a good balance between accuracy and performance overhead. [1] Keyformer: KV Cache Reduction through Key Tokens Selection for Efficient Generative Inference, MLSys 2024. [2] Q-Hitter: A Better Token Oracle for Efficient LLM Inference via Sparse-Quantized KV Cache, MLSys 2024. [3] ALISA: Accelerating Large Language Model Inference via Sparsity-Aware KV Caching, ISCA 2024. --- Rebuttal 2: Comment: Thank you for your detailed response. I don't have any further questions. My only suggestion would be to incorporate the discussion of concurrent works into the final version if the paper is accepted.	Summary: ARKVALE presents a page-based key-value (KV) cache manager designed to address the challenges associated with long-context processing in large language models. The main contribution is its ability to recognize and recall important tokens that were previously evicted, thereby optimizing memory usage and improving throughput and latency without significant accuracy loss. The method involves asynchronously backing up filled pages in external memory and using bounding-volume techniques to summarize and estimate the importance of these pages for efficient recall and eviction. Experiments show strong performance in terms of both accuracy and efficiency. Strengths: - ARKVALE empirically identifies the dynamism of token importance and proposes efficient token eviction and recall methods. This approach ensures good memory efficiency while maintaining accuracy. - The efficient system design and implementation improve decoding latency by up to 2.2x and batching throughput by up to 4.6x. - ARKVALE performs well across different tasks by comprehensive benchmarking, shows its ability in various long-context scenarios. Weaknesses: - The benchmark is only conducted on LongChat, lacking evaluations on different models such as Mistral and LLaMA-3. - There is a lack of comparison with other methods focusing on efficiency, and some other baselines like [1,2]. - The paper lacks discussion and ablation studies on some hyper-parameters, such as top-k, page-size and the relationship between top-k, page-size and the cache budget. [1] Infllm: Unveiling the intrinsic capacity of llms for understanding extremely long sequences with training-free memory [2] Snapkv: Llm knows what you are looking for before generation Technical Quality: 3 Clarity: 3 Questions for Authors: - What is the performance of ARKVALE on Mistral-7B-Instruct-v0.2 and LLaMA-3-8B-Instruct? - How does the performance (accuracy, latency/throughput) of ARKVALE compare to other methods [1, 2]? - What is the process for selecting the hyperparameters, and what their influence on performance? - Does the prefill phase use all tokens? For example, for the first token generated, is full attention used? Additionally, is this method compatible with FlashAttention/FlashDecoding? [1] Infllm: Unveiling the intrinsic capacity of llms for understanding extremely long sequences with training-free memory [2] Snapkv: Llm knows what you are looking for before generation Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weaknesses and questions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >What is the performance of ARKVALE on Mistral-7B-Instruct-v0.2 and LLaMA-3-8B-Instruct? We conduct experiments on our method adapted to Mistral-7B and Llama-3-8B, as detailed in Global Response. --- >How does the performance (accuracy, latency/throughput) of ARKVALE compare to other methods In the Global Response, we compare the accuracy and latency of ArkVale with other baselines. --- >The paper lacks discussion and ablation studies on some hyper-parameters, such as top-k, page-size and the relationship between top-k, page-size and the cache budget. >What is the process for selecting the hyperparameters, and what their influence on performance? For the cache-budget $c$ and page-size $p$, we empirically set the top-k to $\min(40*32, c/2) / p$. In the Global Response, we discuss the impact of different cache-budgets and page-sizes on accuracy and latency. --- >Does the prefill phase use all tokens? For example, for the first token generated, is full attention used? Yes, currently we use the full kv-cache during the prefill phase. However, it is also feasible to perform eviction early in the prefill phase, although the impact on accuracy needs to be determined through experiments. We plan to explore this in future work. --- >Additionally, is this method compatible with FlashAttention/FlashDecoding? Yes, but it requires using the Paged version of FlashAttention/FlashDecoding, such as the kernels implemented in FlashInfer. --- Rebuttal 2: Comment: Thanks for the reply, this answers my questions.	Summary: The paper proposed a method to minimize the risk of KV cache eviction by efficiently and soundly offloading some of them into external memory, which is realized by page organization, page digest, and digest ranking/scoring. The method gets much better performance in context retrieval tasks compared to other KV eviction methods. Strengths: The paper proposes a reliable way to do KV cache eviction with minimal risk. The system is sound and intuitive. It seems compatible with real world frameworks like vLLM. Experiments show good results. Weaknesses: It's not very clear if the dynamics of importance will change along the decoding in other tasks, so the observations may be limited. It would be interesting to see how the page size affect the methods, as vLLM's default page size is 16. Is there any reason to scale to 32? Would the method still work when TP>1? Technical Quality: 3 Clarity: 3 Questions for Authors: I'm wondering if there will be an open-sourced implementation/PR to vLLM? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: It's desirable to see normal length tasks, e.g., GSM8K. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > It would be interesting to see how the page size affect the methods, as vLLM's default page size is 16. Is there any reason to scale to 32? We briefly discuss the impact of different page-sizes on accuracy and latency in Global Response. A page-size of 32 is a compromise between these two aspects (though in fact, a page-size of 16 would also be perfectly acceptable). --- >Would the method still work when TP>1? Yes, ArkVale can be applied to distributed scenarios. But in the context of distributed inference for LLMs, it is common for different heads of a KV-cache to be distributed across various GPUs. Thus, head-wise page eviction and recall may be necessary. --- >I'm wondering if there will be an open-sourced implementation/PR to vLLM? Yes, we plan to introduce serving technologies such as continuous batching into ArkVale and adapt it to serving frameworks like vLLM and SGLang. --- >It's desirable to see normal length tasks, e.g., GSM8K. We evaluate our method on six normal length benchmarks using `lm-eval-harness`. The results are presented in the table below (with page-size=32): \| Cache Budget \| \\| \| Full \| 4096 \| 2048 \| 1024 \| 512 \| \| ------------------- \| --- \| ----- \| ----- \| ----- \| ----- \| ----- \| \| GSM8K (5-shot) \| \\| \| 0.092 \| 0.092 \| 0.092 \| 0.086 \| 0.078 \| \| HellaSwag (0-shot) \| \\| \| 0.544 \| 0.544 \| 0.544 \| 0.544 \| 0.544 \| \| WinoGrande (0-shot) \| \\| \| 0.684 \| 0.683 \| 0.683 \| 0.683 \| 0.683 \| \| PIQA (0-shot) \| \\| \| 0.760 \| 0.761 \| 0.761 \| 0.761 \| 0.761 \| \| OpenBookQA (0-shot) \| \\| \| 0.306 \| 0.306 \| 0.306 \| 0.306 \| 0.306 \| \| MathQA (0-shot) \| \\| \| 0.253 \| 0.253 \| 0.253 \| 0.253 \| 0.253 \| The data from the table indicates that, in most cases, ArkVale maintains comparable accuracy to the original model when the cache budget does not exceed 4096. This is likely because the context lengths for these tasks are not long enough, and they primarily test the model's fundamental capabilities rather than its contextual memory. However, in tasks like GSM8K, a notably small cache budget can lead to an obvious drop in accuracy.	Summary: The paper introduces ARKVALE, a novel page-based key-value (KV) cache management approach designed to optimize the performance of Large Language Models (LLMs) when dealing with long context lengths. As the demand for higher context lengths in tasks such as multi-turn chats and content generation increases, the management of extended key-value caches becomes crucial due to memory constraints and the impact on computation latency and throughput. ARKVALE addresses these challenges by dynamically managing the cache, selectively evicting less important tokens while recalling those that become relevant again at different decoding steps. This strategy leverages a mechanism that organizes tokens into pages and uses digests to estimate the importance of these pages. By summarizing pages into smaller, more manageable units, ARKVALE efficiently decides which pages to recall from external memory and which to evict, enabling focused attention computations on only the most relevant subsets of data. The paper's experiments demonstrate that ARKVALE effectively handles various long-context tasks with minimal loss in accuracy, even under constrained cache sizes between 2K and 4K tokens. The results indicate significant improvements in model decoding latency and batching throughput, specifically up to 2.2x faster latency and 4.6x higher throughput. These enhancements are achieved by applying attention to a reduced subset of pages, thus decreasing the per-sample memory usage of the KV cache. This system not only improves the operational efficiency of LLMs but also maintains high accuracy levels, suggesting a scalable solution for managing extensive data in real-time language processing applications. Strengths: Originality: ARKVALE introduces a novel page-based system for KV cache management that dynamically adjusts which data is retained or discarded based on the evolving importance of tokens throughout the decoding process. This approach creatively combines ideas from memory management and attention mechanisms in LLMs, setting it apart from previous methods that often permanently discard tokens without the ability to recall them. Quality: The methods proposed are rigorously tested across various benchmarks and scenarios, demonstrating minimal loss in accuracy while substantially improving efficiency in terms of decoding latency and throughput. The experimental setup is thorough, using several datasets to ensure robustness and reproducibility of results. Clarity: The paper is well-organized, presenting complex ideas in a structured and understandable manner. The use of diagrams to illustrate the KV cache management process helps in demystifying the approach and makes the operational details of ARKVALE accessible to readers. Significance: ARKVALE's impact is multifaceted. For practical applications, it allows for the deployment of LLMs in environments where memory and latency are constraints, thus broadening their usability in real-world applications. Theoretically, it advances our understanding of efficient memory management in models requiring extensive context, potentially influencing future developments in LLM architectures and optimization techniques. Weaknesses: - Insufficient Model Evaluation: The paper evaluates the proposed method using only the LongChat-7b-v1.5-32k model, which is relatively outdated in the current landscape. It does not demonstrate the method's generality and robustness across different model architectures (e.g., MoE, GQA) and scales (70B models). Technical Quality: 3 Clarity: 3 Questions for Authors: Clarification on Model Choice: The paper primarily utilizes the LongChat-7b-v1.5-32k model for evaluating ARKVALE. Can the authors provide specific reasons for choosing this model over others? Additionally, how do the authors anticipate ARKVALE would perform with other, potentially larger or more recent LLM architectures? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have not discussed the limitations of their work in the submission. It would be great to add sections discussing the limitations and societal impacts of ARKVALE. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: >Clarification on Model Choice: The paper primarily utilizes the LongChat-7b-v1.5-32k model for evaluating ARKVALE. Can the authors provide specific reasons for choosing this model over others? Additionally, how do the authors anticipate ARKVALE would perform with other, potentially larger or more recent LLM architectures? We primarily choose the LongChat-7b-v1.5-32k model because it is a relatively state-of-the-art long-text extension of Llama2 at the time. We conduct experiments on our method adapted to the relatively newer models Mistral-7B and Llama-3-8B that utilize GQA. The results are shown in Global Response. --- >The authors have not discussed the limitations of their work in the submission. It would be great to add sections discussing the limitations and societal impacts of ARKVALE. Chiefly, the limitation of our method lies on storing a backup for each KV cache page in external memory (typically the CPU memory). On the one hand, although the latency of data transfer for backup can be hided by asynchronous copy, the energy consumption cannot be eliminated. On the other hand, it may occupy much CPU memory, potentially impacting the performance of other applications under some extreme conditions. Furthermore, when the CPU's memory capacity is insufficient, these backups may need to be offloaded to disk storage. --- Rebuttal Comment 1.1: Title: Thank you for the response Comment: I appreciate the authors' responses; most concerns have been addressed. I will keep my evaluation for acceptance. I also noticed that this paper is quite similar to a paper [1] presented at ICML 2024, published after the NeurIPS submission deadline. While not obligatory, discussing and comparing this work with that paper would be beneficial. [1] Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference --- Reply to Comment 1.1.1: Title: Comparison with Quest Comment: Quest [1] shares similarities with ArkVale (and some other works like InfLLM [2]) in estimating and performing topk-filtering on kv-cache at the page/block granularity. The main differences are as follows: - Their approach aligns more closely with post-training dynamic sparse attention methods like SparQ [3] and IceFormer [4]. These methods do not involve eviction of kv-cache and cannot save GPU memory. - The estimation method we employ is based on bounding-volume, whereas their estimation method is similar to the "cuboid-max" (a subclass of our approach) introduced in our paper. [1] Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference [2] InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory [3] SparQ Attention: Bandwidth-Efficient LLM Inference [4] IceFormer: Accelerated Inference with Long-Sequence Transformers on CPUs	Rebuttal 1: Rebuttal: ## Adaption to other models We adapt ArkVale to both `MaziyarPanahi/Llama-3-8B-Instruct-64k` (with 64k context-length extended from Meta Llama-3-8B) and `mistralai/Mistral-7B-Instruct-v0.3` (with a 32k context-length) in hugginggface, and test them on the datasets used in our paper, with the results shown in the table below (with page-size=32): \| Model \| \\| \| Llama3 \| \| \| \| \| \\| \| Mistral \| \| \| \| \| \| ---------------- \| --- \| ---------- \| --------- \| ------ \| --------- \| --------- \| --- \| ---------- \| --------- \| --------- \| ------ \| ----- \| \| Cache Budget \| \\| \| *Full* \| 4096 \| 2048 \| 1024 \| 512 \| \\| \| *Full* \| 4096 \| 2048 \| 1024 \| 512 \| \| HotpotQA \| \\| \| 42.94 \| 40.9 \| 38.01 \| 37.39 \| 32.54 \| \\| \| 49.37 \| 48.72 \| 49.92 \| 49.14 \| 48.91 \| \| NarrativeQA \| \\| \| 16.77 \| 17.2 \| 18.03 \| 17.8 \| 18.14 \| \\| \| 28.74 \| 28.61 \| 25.87 \| 25.34 \| 23.59 \| \| Qasper \| \\| \| 11.0 \| 10.6 \| 10.03 \| 9.5 \| 8.71 \| \\| \| 41.58 \| 42.2 \| 41.44 \| 39.24 \| 36.9 \| \| GovReport \| \\| \| 28.91 \| 27.34 \| 25.62 \| 24.38 \| 21.96 \| \\| \| 34.91 \| 32.84 \| 31.59 \| 29.69 \| 25.85 \| \| TriviaQA \| \\| \| 89.91 \| 89.91 \| 89.91 \| 90.12 \| 89.35 \| \\| \| 88.59 \| 88.94 \| 88.94 \| 89.38 \| 88.7 \| \| PassageRetrieval \| \\| \| 52.75 \| 59.75 \| 61.25 \| 62.0 \| 64.91 \| \\| \| 98.0 \| 95.0 \| 94.5 \| 90.5 \| 82.0 \| From the data in the table, we can observe that in most cases, ArkVale can approach or even surpass the accuracy of the original model when the cache budget does not exceed 4096. --- ## Additional Baselines We add comparisons with two open-sourced concurrent works, SnapKV [1] and InfLLM [2]. SnapKV, similar to works like H2O and TOVA, evicts KV-cache based on historical token information, but it only performs eviction during the prefill phase, thereby reducing overhead in the decode phase. InfLLM is a context expansion work that employs block-level memory units, which share similarities with ArkVale's page-level eviction/recall method. However, it generalizes memory units using several representative tokens (defaulting to 4) and, like streaming-llm, it retains a fixed number of initial-tokens and recent-tokens, employing the same positional encoding for tokens outside the recent-tokens. We test these two baselines on LongBench based on the experimental setup and evaluation method detailed in Section 6.3 of our paper. The experimental results are presented in Figure 1 of the attached PDF. SnapKV nearly outperforms all other baselines, yet it is still not better than ArkVale overall. We attribute this anomaly to our dynamic selection of important tokens. InfLLM, similar to ArkVale, inspects and recalls some important tokens, and additionally preserves initial-tokens and recent-tokens, thus performing comparably or even better than ArkVale on datasets such as Qasper, GovReport, and TriviaQA. However, since it apply the same positional encoding to all the tokens other than recent-tokens, its performance is unstable, as evidenced by its poor performance on NarrativeQA and PassageRetrieval. Furthermore, using representative tokens as page digest may not be good enough for page importance estimation. Figure 2 of the attached PDF shows the average latency per decode step for ArkVale (page-size=32) and baselines under different sequence lengths in the passkey retrieval task. Due to the inconvenience of setting batch-size greater than 1 for some baselines, experiments were conducted with batch-size=1. In all cases, ArkVale achieves the shortest latency, which gradually increases with the increasing cache budget. H2O, TOVA, and StreamingLLM incur token eviction during every decode step, introducing additional overhead, whereas SnapKV only performs eviction during the prefill phase, resulting in better performance than the others. Although importance estimation and page recall may introduce some additional overhead to ArkVale, ArkVale is slightly faster than SnapKV, mainly due to its efficient paged memory management with page-attention. The poor performance of InfLLM is likely due to its suboptimal implementation of memory management and hand-crafted attention kernel. [1] Infllm: Unveiling the intrinsic capacity of llms for understanding extremely long sequences with training-free memory [2] Snapkv: Llm knows what you are looking for before generation --- ## Impact of Different Page-sizes Generally speaking, the smaller the page-size, the more accurate the estimation of page importance, the higher the model's accuracy. However, a smaller page-size will also increase the number of pages for the same number of tokens, thereby increasing the space occupied by the page digest (1~2 token per page), and simultaneously increasing the latency overhead for page importance estimation. We test the impact of different cache-budgets and page-sizes on model accuracy and decode latency on LongChat-7b-v1.5-32k. We select PassageRetrieval as the benchmark. The results are presented in Figure 3 of the attached PDF. It can be observed that as the page-size decreases/cache-budget increases, both accuracy and latency gradually increase. Notably, the accuracy for page-size=8 is slightly worse than that for page-size=16, possibly due to the smaller page-size causing fewer adjacent tokens to be selected simultaneously, which impacts attention locality. Considering accuracy and latency & memory overhead, we typically set the page-size to 16 or 32. Pdf: /pdf/1e565e6c673b586666a6c2db5b26d6ac2e054f4a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Rethinking Weight Decay for Robust Fine-Tuning of Foundation Models	Accept (poster)	Summary: The paper proposes a method for robust fine-tuning of deep networks. The proposed method is a variant of L2-SP. The original formulation of L2-SP applied a uniform L2 penalty on the difference between the weights of the pre-trained model and the fine-tuned model. The proposed method (L2-SPD) aims to make this penalty non-uniform for different layers. This is achieved by only applying weight decay to layers where the direction of descent and weight decay are sufficiently similar. The efficacy of the method is demonstrated over various image classification, segmentation and reasoning tasks, showing better ID and OOD performance than baselines. Strengths: * The projection interpretation is very interesting. * The exposition of the method and its insights is very nice. * The insight behind the method is simple, but the method is widely applicable for an important problem. * The method is well compatible with LoRA style approaches, increasing its practical applicability. * The method is stable across a wide range of hyper-parameters. * Wide range of benchmarking is performed, and the empirical results are strong. Weaknesses: There are some minor issues with the presentation - * It would be nice to have a short preliminaries section to define all notations used in the paper, along with their dimensions. * It is not clear from Alg 2 how SPD is selective to different layers. It appears that $\theta_t$ represents the entire model’s parameters, and $c_t$ is a single scalar for the entire model. Hence, it seems like the “weight decay” only happens for the entire model if the total gradient is in the opposite direction of the weight decay. I might be incorrect here, but I am unable to see where the group structure comes in (possibly $c_t$ is separate for each layer?) Apart from this, some of the experiments have the following weaknesses - * The Himmelbau example is not fully clear to me either. It would be interesting to compare the trajectories against L2-SP to actually understand the “selective penalization”, since regular L2-SP will also penalise deviations similarly. I tried running the code that the authors provided, and with some tuning of the hyper-parameters I was able to get similar trajectories with L2-SP as well. * Lack of ablation studies on the two components of the method. It would be good to understand the gains from selective projection and adaptive regularization separately as well. * I am unable to understand some of the trends in the results on ImageNet. It is not immediately clear to me how regularizing the model to minimize distance from the Zero-Shot model can outperform both the Zero-Shot model and the vanilla FT model on one domain (Im-Sketch), but is in between the two on another domain (Im-Rendition). I had the intuition that L2-SPD will pull the weights to be closer to the Zero-shot model, hence inheriting its robustness and biases, but I could be mistaken. * Minor - The caption for Table 3 states that "SPD... beats L2-SP by 8.8%", however, I cannot understand how this was computed. * Minor - While the paper compares against L2-SP, it misses Elastic Weight Consolidation, a similar method using Fisher information, which also adaptively sets the regularization parameters for different layers. Technical Quality: 3 Clarity: 2 Questions for Authors: * Is there a way to disentangle the gains from selective projection and adaptive regularization? Related to this, how often does the selective projection kick in in practice, and how does it change as training progresses (this is not crucial to the paper, but is a good-to-have analysis)? * I am unable to understand some of the trends in the results on ImageNet. It is not immediately clear to me how regularizing the model to minimize distance from the Zero-Shot model can outperform both the Zero-Shot model and the vanilla FT model on one domain (Im-Sketch), but is in between the two on another domain (Im-Rendition). I had the intuition that L2-SPD will pull the weights to be closer to the Zero-shot model, hence inheriting its robustness and biases, but I could be mistaken. * Can this technique of selective regularization be done at a more granular level beyond layers? * In section 4.1, why are all domainnet domains not used, omitting the infograph domain? The paper will be helped by a justification of this. * For table 2a, the range of deviation is quite different for L2-SP and L2-SPD. While L2-SPD does dominate performance for similar deviations, I wonder if it is possible to use even larger $\lambda$ and make the deviation very low? Will the same correlation with OOD accuracy still hold? * Hyper-parameter selection is not clear for ImageNet. Was k-fold cross validation used? Is the ID val used to select hyper-parameters? * On quickdraw, will more aggressive regularization yield better OOD accuracy? Since the pretrained model is supposed to be closer to other domains? * For ImageNet, what is the linear layer for final classification? * How does the method compare against a simpler baseline like Wise-FT? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have addressed limitaions of their method, pointing out worse performance if the base model is not aligned with the OOD distribution. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Comment: We appreciate the reviewer's taking the time to examine the paper and the provided code closely. We clarified your concerns and added Infograph experiments as requested. If anything is not clear, we are happy to discuss it during the rebuttal period. The Himmelbau example. We appreciate the reviewer's effort in running our code! In this simple example, you are correct that regular L2-SP would generate a similar trajectory. However, to see the difference between SPD's selective regularization and L2-SP, we suggest that the reviewer print out which direction is regularized during the optimization process. Add the following printing statement to line 40 in the code before the return. * print(self.condition_buffer) * return new_param_list The condition buffer contains the selection condition $c_t$ for the X and Y directions. Whenever an entry is negative, regularization is applied. This demonstrates the underlying mechanism for SPD. Is there a way to disentangle the gains..? Sorry for the confusion! SPD introduces only one component that affects performance, selective projection $c_t$ (Section 3.3), which selectively applies regularization during fine-tuning. The selectivity is the only complement that provides the performance gain. Section 3.4 introduces a re-parametrization of the regular L2 regularization as a projection for better hyper-parameter tuning (see a comment from the reviewer Ex5N). It's a different way of writing L2 regularization and does not introduce performance gain. Selectivity kicks in very often. However, it is hard to quantify, and it changes depending on the task, model, data, optimization hyper-parameters, etc. An analysis of the selectivity's behavior will be insightful. Trends in the results on ImageNet. This is a very good question. Indeed, L2-SPD will pull the weights closer to the zero-shot model, as shown in Table 1 and Table 2 in the main paper, thus inheriting its robustness and biases. However, the extent of performance gain has a more subtle answer. It depends on the interplay between the pre-trained model, the fine-tuning (in-distribution) dataset, and the OOD datasets. Specifically, the ID dataset can provide useful information to the OOD datasets. In this benchmark, the useful information is that all datasets share the same label space. If fine-tuning leads to more gain, it's possible that after fine-tuning (assuming with proper regularization), the fine-tuned model can outperform both zero-shot and vanilla FT. Can this technique of selective regularization be done at a more granular level beyond layers? Yes, theoretically, the method can be applied at a parameter level. However, the storage cost is significantly increased because we would need to store a different $c_t$ for each parameter in the model. In section 4.1, why are all domainnet domains not used, omitting the infograph domain?. Infograph was omitted because none of the methods achieved good ID performance on this domain. We added the domain back to answer the reviewer's question (Table 1 in the rebuttal PDF). The superiority of SPD remains unchanged with the addition of this domain. Larger regualrization and make the deviation very low? Will the same correlation with OOD accuracy still hold? We have tried larger hyper-parameters. At extremely large values, the model will eventually underfit and perform poorly on ID and OOD data like L2-SP. We intend to show that L2-SPD outperforms L2-SP for similar deviations as the reviewer suggested. Hyper-parameter selection is not clear for ImageNet. We used the ID validation dataset following prior works. ImageNet-V2 can be seen as the clean test set for ImageNet. On quickdraw, will more aggressive regularization yield better OOD accuracy? Yes, we conducted a similar experiment as in Table 2 in the main paper. More aggressive regularization yields better OOD on quickdraw. We swept hyper-parameters from 1.0 to 2.3 and calculated the correlation coefficient between the hyper-parameters and OOD performance. The correlation is 0.88, which means a robust positive correlation between large regularization and good OOD performance. Linear layer for For ImageNet Because we used a CLIP ViT-Base model, we used a zero-shot linear head to initialize and fine-tune the linear layer. This practice follows from prior works such as WISE-FT. Comparison with Wise-FT? Technically, WISE-FT is an orthogonal method. Therefore, SPD should further improve its performance. To validate this, we added experiments combining WISE-FT and SPD in our Imagenet experiments (Table 1 in the rebuttal PDF). Our results show that SPD further improves the performance of Wise-FT. It's also worth pointing out that SPD is more general than WISE-FT, which only applies to models capable of zero-shot. It does not apply to the segmentation experiment (Table 4 in the main paper), whereas SPD does not have this constraint. --- Rebuttal Comment 1.1: Title: Official Comment Comment: I thank the authors for their response. The response has addressed most of my concerns. I believe that incorporating these into the paper will greatly improve the presentation, and hence I keep my rating of acceptance.	Summary: This paper proposes a new weight decay technique to adapt foundation models to target tasks, focusing on fitting the target data while maintaining the pre-trained knowledge. Specifically, the method, Selective Projection Decay (SPD), selectively imposes a strong penalty on certain layers while allowing others to change freely. Experimentally, the method consistently provides better in-distribution generalization and out-of-distribution robustness across multiple popular vision and language benchmarks. Strengths: - This is a well-motivated approach for fine-tuning foundation models on target tasks. By selectively imposing regularization on certain layers, it better fits the target data while retaining the pre-trained knowledge. - The method has also proven to be effective for parameter-efficient fine-tuning, which is particularly beneficial. - Extensive experiments conducted on both vision and language benchmarks demonstrate the method's strong performance. Weaknesses: There are no major weaknesses, please respond to my questions and issues. Technical Quality: 4 Clarity: 3 Questions for Authors: - Why is L2-SP not compared in the experiments shown in Tables 1 and 5? The authors are requested to provide the results for this method as well. - Tables 2a and 2b should be converted into figures to enhance readability. - Minor Typo: On line 8, "retraining" should be corrected to "retaining". - Several of the results are wrongly highlighted in Table 3. - The appendix sections referred in the paper are missing. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The limitation of this work is clearly highlighted: the method's performance depends on the quality of the foundation model. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Comment: We thank the reviewer for the positive comments and pointing our our typos. We clarified your concerns and added L2-SP experiments as requested. If anything is unclear, we are happy to discuss it during the rebuttal period. Why is L2-SP not compared in the experiments shown in Tables 1 and 5? The authors are requested to provide the results for this method as well. * Table 1: Originally, we designed this table to present the improvement over a vanilla AdamW. For the rebuttal, we added L2-SP to Table 1 (Table 1 in the rebuttal PDF). SPD outperforms L2-SP in all domains. * Table 5: In the LLaMA PEFT experiments, we compared our method with Adam $+$ weight decay, as indicated in the table title. In this PEFT setting, L2-SP reduces to weight decay, as explained in section 3.5. In other words, the L2-SP baseline is already included. Tables 2a and 2b should be converted into figures to enhance readability. Thank you for the great suggestion. We will update the two tables to heat maps for better readability. --- Rebuttal Comment 1.1: Comment: Thank you to the authors for addressing my questions. I agree with the other reviewer that the theoretical perspective is lacking in this work. However, I believe the empirical results highlight the significance of the study, so I will keep my score.	Summary: The authors propose Selective Projection Decay, a modification on the AdamW optimizer where a regularization penalty is applied that controls for an overall drift from the prior weights such that as long as the overall shift is controlled for but individual layers can potentially adapt differently. The authors provide experiments on OOD generalization tasks and PEFT language modelling fine-tuning. Strengths: - The explanation of the method is clear and detailed. The method itself is straightforward and intuitive to understand. - There are experiments on a number of different tasks and domains that show some improvement upon prior methods in terms of raw performance metrics. Weaknesses: - The way the authors frame it, there appears to be limited novelty with the proposed method, with it appearing to simply involve an additional conditional check before applying a potential regularization penalty that has been explored quite a bit in existing literature. I do think that there is something of merit here, however it's hard to pinpoint it under the current framing of the work. - The statement "We recommend starting with $\lambda= 1$ and adjusting the strength according to the specific needs" is rather unsatisfying and doesn't provide the reader with a good idea of whether or not $\lambda$ is easy to tune. This can probably be resolved by adding additional analysis along this end compare do just Table 2. - Experimentally, the results feel somewhat lacking and somewhat misleading. For example, in the Domain-Net experiments, the only comparison is with AdamW, which inherently should lack the exploration capabilities the authors are attempting to account for. There are quite a few baselines that are missing as a result in my opinion, such as sharpness aware minimization (SAM) and/or other exploration based optimizers. - For language modelling, it would be more useful to show performance over a multitude of models and sizes, as the results appear to suggest for example that the gap between AdamW and Adam-SPD narrows quite significantly between sizes 7B and 13B, as well as when the PEFT method changes. Technical Quality: 3 Clarity: 3 Questions for Authors: - I think a more interesting way to use the proposed method would be to have $\lambda$ to be a parameter or layer-specific hyper-parameter rather than a global hyper-parameter. This way, having individual $c_t$ for different components might make the method more robust and in fact help the model adapt relevant weights while keeping specific weights close to the pre-trained model. In many cases, different layers can begin to exhibit different functionalities and therefore there might be more use in having some layers change more than others based on the features that might be captured there. - Although not a deal-breaker, I would appreciate a perhaps more in-depth analysis of the language modelling experiments, in particular without PEFT. While I understand the relationship with PEFT and the proposed optimizer, I believe that it would be quite valuable to show how AdamW-SPD works even under naive fine-tuning scenarios. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We really appreciate the thoughtful comments, especially on suggesting a layer-specific hyper-parameter configuration. This was what we thought would be interesting as well. In the response, we provided clarifications to your concerns and a new experiment. If anything is not clear, we are happy to discuss it during the rebuttal period. I think a more interesting way to use the proposed method would be to have to be a parameter or layer-specific hyper-parameter rather than a global hyper-parameter. This way, having individual for different components might make the method more robust and in fact help the model adapt relevant weights while keeping specific weights close to the pre-trained model. In many cases, different layers can begin to exhibit different functionalities and therefore there might be more use in having some layers change more than others based on the features that might be captured there. Sorry for the confusion! SPD is a layer-specific method, as the reviewer suggested. SPD has a different selecting condition $c_t$ for each layer. In other words, algorithm 2 is displayed for each layer. SPD behaves precisely like the reviewer suggests, and we will update the writing to clarify this. The layer-specific $c_t$ allows specific layers to have small regularization and other layers to have large regularization. This gives SPD the edge over other methods by fitting the new dataset (better ID performance) and retraining knowledge from the pre-trained model (better OOD performance). We will update the description to make this point more clear. The statement "We recommend starting with and adjusting the strength according to the specific needs" is rather unsatisfying and doesn't provide the reader with a good idea of whether or not is easy to tune. This can probably be resolved by adding additional analysis along this end compare do just Table 2. Thank you for mentioning the ease of tuning this method. This is a contribution of this work. Practically, the method is just as easy to tune as regular weight decay and L2 regularization because the hyper-parameter $\lambda$ in SPD has the same effect as in those methods, i.e., adjusting regularization strength. In some cases, it can be much easier to tune as shown in the sensitivity analysis in Table 2 of main paper. SPD is much less sensitive to the hyperparameter, consistently yielding good performance on both ID and OOD datasets, whereas L2-SP is more sensitive to the choice of hyperparameter. Moreover, SPD is even simpler to understand as described in Section 3.4. We re-parametrized L2 regularization (weight decay) as a projection. This gives an intuitive interpretation of the regularization hyper-parameter. For example, in a conventional weight decay setup, we can set the hyper-parameter as $\lambda=0.01$. We do not necessarily know what $0.01$ means as a hyper-parameter. However, with the re-parametrization in Section 3.4, we can set it to be $\lambda=1$; this intuitively means that the regularization projects the current model weights back to the deviation of the last update. Experimentally, the results feel somewhat lacking and somewhat misleading. For example, in the Domain-Net experiments, the only comparison is with AdamW, which inherently should lack the exploration capabilities the authors are attempting to account for. There are quite a few baselines that are missing as a result, in my opinion, such as sharpness aware minimization (SAM) and/or other exploration-based optimizers. Thank you for bringing up this work. We have added a comparison to SAM on ImageNet (Table 2 in the rebuttal PDF). SPD outperforms SAM in all experiments. SPD is designed explicitly for robust fine-tuning of pre-trained foundation models, whereas SAM does not consider pre-trained initialization. SAM focuses on finding a good local minimum with uniformly low loss. This property could lead to better generalization for regular training but not necessarily a more robust fine-tuned model. Fine-tuning requires careful balancing between fitting the new data and retaining knowledge in the pre-trained model. SPD explicitly considers this problem. There may be other exploration-based optimizers to achieve better generalization. However, they are not designed for fine-tuning and generally do not consider the pre-trained initialization in their formulation. Instead, our experiments focus on comparing baselines for robust fine-tuning (Tables 3 and 4 in the main paper). In our setup, AdamW is the best example of an exploration-enhanced optimizer. It is equipped with an adaptive learning rate and momentum. Both mechanisms encourage AdamW to escape local minima and explore further. However, our understanding of exploration may differ from that of the reviewer. We welcome further discussion. Although not a deal-breaker, I would appreciate a perhaps more in-depth analysis of the language modelling experiments, in particular without PEFT. While I understand the relationship with PEFT and the proposed optimizer, I believe that it would be quite valuable to show how AdamW-SPD works even under naive fine-tuning scenarios. We provide a new experiment for VQA tasks. Specifically, we fine-tune the PaliGemma-3B model on the VQAv2 dataset with LoRA and test it with nine additional VQA datasets consisting of different types of distribution shifts across vision, question, and answer (IV-VQA, CV-VQA, VQA-Rephrasings, VQA-CE, AdVQA, TextVQA, VizWiz and OK-VQA). In Tab.~\ref{tab:vqa} of the rebuttal PDF, SPD achieves the best ID and average OOD performance. We also show the performance evaluation for both near and far OOD datasets. SPD is consistently more robust under different types and degrees of distribution shifts. We acknowledge that full fine-tuning of large language models using SPD would be very interesting. However, our experiments are limited by computation resources and time. It will be an interesting next step in our exploration. --- Rebuttal Comment 1.1: Comment: I appreciate the response provided by the authors. I believe the authors have addressed most of my concerns and I have adjusted my score accordingly. While I am now more convinced about the merits of the work, I still believe that increased analysis of the proposed method (from a more theoretical perspective rather than completely empirical) would be more meaningful given that the nature of the work.	Summary: This paper proposes a novel weight decay strategy called Selective Projection Decay (SPD). SPD selectively imposes a stronger penalty on certain layers, and is designed to improve both in-distribution (ID) and out-of-distribution (OOD) performance. The paper demonstrates the effectiveness of SPD through experiments on several benchmarks. Strengths: - SPD is well-motivated and simple to implement. - The paper is well-written and easy to follow. It clearly lays out intuition and motivation. I especially liked sections 3.3 and 3.4, which relate the condition to online hyperparameter optimization and the deviation ratio to a re-interpretation of L2-SP as a projection. I thought the line of logic here was quite clear. - The paper demonstrates strong performance on several standard benchmarks for robust fine-tuning. Weaknesses: - Overall, I think this is a strong paper, and its strengths outweigh its weaknesses. I think the paper could benefit from an ablation experiment, for example, trying Adam-SPD without the condition or benchmarking SPD's performance with optimizers other than Adam. - [1] reports that ensembling with the initial weights (whether in model space or weight space) is a simple strategy that improves OOD robustness, and a few recent works [2, 3] report that ensembling continues to improve performance after their proposed fine-tuning strategies. Could the authors validate how the benefits from SPD extend to the ensembling setting? [1] Robust fine-tuning of zero-shot models [2] Finetune like you pretrain: Improved finetuning of zero-shot vision models [3] AutoFT: Learning an Objective for Robust Fine-Tuning Technical Quality: 3 Clarity: 4 Questions for Authors: Please see the weaknesses section. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Please see the weaknesses section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Comment: We really appreciate the positive comments and suggestions for ensemble methods. In the response, we clarified your concerns and included new experiments. If anything is not clear, we are happy to discuss it during the rebuttal period. Overall, I think this is a strong paper, and its strengths outweigh its weaknesses. I think the paper could benefit from an ablation experiment, for example, trying Adam-SPD without the condition or benchmarking SPD's performance with optimizers other than Adam. Thank you for the positive comments on this work. The ablation study without the condition is shown in Table 2. Essentially, without the selecting condition (line 96-102), SPD reduces to a regular L2 regularization, which is shown ineffective in our comprehensive analysis (line 255). Furthermore, SPD is a general method and applicable to other optimizers. We hypothesize that it should have positive effects on optimizers using regular weight decay and L2 regularization. [1] reports that ensembling with the initial weights (whether in model space or weight space) is a simple strategy that improves OOD robustness, and a few recent works [2, 3] report that ensembling continues to improve performance after their proposed fine-tuning strategies. Could the authors validate how the benefits from SPD extend to the ensembling setting? Thank you for mentioning these works, especially the ensemble technique such as Wise-FT. Technically, ensemble techniques are orthogonal methods. SPD should further improve the performance of them such as WISE-FT. To validate this, we added experiments combining WISE-FT and SPD in our Imagenet experiments (Table 1 in the rebuttal PDF). Our results show that SPD further improves the performance of Wise-FT. It's also worth pointing out that SPD is more general than WISE-FT, which only applies to models with linear connectivity and capable of zero-shot classification. It does not apply to the segmentation experiment (Table 4 in the main paper), whereas SPD does not have this constraint.	Rebuttal 1: Rebuttal: We thank all of the reviewers for their positive comments on this work's adaptability (dDi9), performance (Ex5N, CVio, Jrpf, 33UR), and insights (Ex5N, 33UR). We aim to provide concrete responses to your questions and clarify your confusion. The rebuttal PDF includes new experiments and studies requested by the reviewers. The new experiments are summarized below. * (Jrpf,33UR) We added L2-SP and Infograph to Table 1 in the main paper. Our method remains the best-performing method. * (CVio) We added a comparison to Sharpness-Aware-Minimization (SAM) on ImageNet . SPD outperforms SAM on ImageNet because SAM is not explicitly designed for robust fine-tuning. * (CVio) We added a new VQA fine-tuning experiment with 9 additional VAQ OOD datasets with distribution shifts across vision and language. SPD is consistently more robust under different types and degrees of distribution shifts. We hope our response answers your questions and are open to further discussion during the rebuttal period. Pdf: /pdf/b3b51152392697595db27ec998b9380125b9fdac.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This work proposes a new regularization scheme called Selective Projection Decay (SPD), which selectively imposes penalties on layers where the current progress direction does not align with the vanilla update direction. This method is compatible with PEFT methods, including LoRA-type algorithms, making it practical for end users. Experiments show that ADAM equipped with SPD consistently outperforms baselines, including L2 regularization, in terms of in-distribution generalization and out-of-distribution robustness on multiple popular vision and language benchmarks. Strengths: The simplicity of the proposed method is its major strength, making it easy for many users to adopt. Compatibility with parameter-efficient fine-tuning is a significant advantage. Additionally, its simplicity ensures good reproducibility. Moreover, the experiment details are well described, further supporting reproducibility. The presentation is very clear. Mostly, it’s easy to follow the main idea, interpretations, and experiment results. Specifically, the difference between L2-SP and SPD is clearly explained by comparing the pseudo-codes and providing easy-to-understand interpretations. Weaknesses: No obvious weaknesses are observed. One minor issue is that the performance improvement, while consistently observed, is not significantly high. Additionally, I have noted some unclear justifications in the manuscript, as described in the questions section. Technical Quality: 2 Clarity: 2 Questions for Authors: #1. How does the performance compare when using EWC regularization for fine-tuning, since EWC is generally better than L2 regularization? #2. Comparing the performance of L2-SP and Adam-SPD in hyper-parameter sweeping experiments seems challenging since these are reported in tables. Perhaps heatmaps could be a clearer alternative to visualize and highlight the results. #3. The authors argue that SPD selectively imposes penalties on large deviations in a layer-wise manner. However, in Algorithm 2, there is no layer-wise information. It appears that all parameters are updated based on the alignment score denoted by c_t. From my understanding of Algorithm 2, if the alignment is high (i.e., c_t is smaller than zero), the updated parameters are further adjusted by the interpolation-like equation. If c_t is larger than zero, do we use the vanilla ADAM update, since \theta_t has already been updated before calculating c_t? Am I correctly understanding Algorithm 2? In such cases, it seems that the parameters shouldn’t be updated, since SPD will impose a penalty to slow down updates for those layers with negative c_t. #4. In the LLaMA PEFT experiments, there is no baseline with L2 regularization or its advanced versions (such as EWC). While I expect that ADAM with SPD is better than AdamW, I would like to see if there are also improvements compared to ADAM with L2 regularization in PEFT settings. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors state that the proposed approach only works well if the pretrained model is good enough to provide good initial parameters. No clear negative societal impacts have been observed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Comment: We thank the reviewer for the positive comments and for suggesting insightful new related works. In the response, we clarified all your concerns. If anything is not clear, we are happy to discuss it during the rebuttal period. How does the performance compare when using EWC regularization for fine-tuning, since EWC is generally better than L2 regularization? Thank you for mentioning this work! However, EWC does not apply to general fine-tuning and was explicitly developed for continual learning (CL). Under the CL setting, EWC requires the gradient information from the previous task, which is not available for fine-tuning a pre-trained model. Therefore, we cannot fairly compare to EWC in the fine-tuning setting. Nevertheless, EWC points out that uniform regularization can be ineffective, which is also the main motivation of this paper for fine-tuning. Comparing the performance of L2-SP and Adam-SPD in hyper-parameter sweeping experiments seems challenging since these are reported in tables. Perhaps heatmaps could be a clearer alternative to visualize and highlight the results. Yes, this is a very good suggestion. We will update the tables into heat maps in the main paper. The authors argue that SPD selectively imposes penalties on large deviations in a layer-wise manner. However, in Algorithm 2, there is no layer-wise information. It appears that all parameters are updated based on the alignment score denoted by $c_t$. From my understanding of Algorithm 2, if the alignment is high (i.e., $c_t$ is smaller than zero), the updated parameters are further adjusted by the interpolation-like equation. If $c_t$ is larger than zero, do we use the vanilla ADAM update, since $\theta_t$ has already been updated before calculating $c_t$? Am I correctly understanding Algorithm 2? In such cases, it seems that the parameters shouldn’t be updated, since SPD will impose a penalty to slow down updates for those layers with negative $c_t$. SPD uses layer-wise information in the calculation of $c_t$. Specifically, each layer has a different $c_t$. Here, $\theta_t$ denotes the weight matrix of a layer, not an individual parameter or the entire network. We will make the notation clear in the update. Yes, when $c_t>0$, the layer $\theta_t$ is updated using vanilla Adam. If this does not clarify your confusion, we are happy to discuss this more in the following rebuttal period. In the LLaMA PEFT experiments, there is no baseline with L2 regularization or its advanced versions (such as EWC). While I expect that ADAM with SPD is better than AdamW, I would like to see if there are also improvements compared to ADAM with L2 regularization in PEFT settings. In the LLaMA PEFT experiments, we compared our method with Adam $+$ weight decay, as indicated in the table title. In this PEFT setting, L2 regularization reduces to weight decay as explained in section 3.5. In other words, the L2 regularization baseline is already included. We will make this clearer in the text; thanks for bringing this up.	null	null	null	null	null	null
Magnet: We Never Know How Text-to-Image Diffusion Models Work, Until We Learn How Vision-Language Models Function	Accept (poster)	Summary: The paper investigates the text embedding representation of text-to-image models in the context of stable diffusion. In particular, the authors find that object features often bind to their commonly associated attributes and propose the Magnet approach that interpolates object features with their designated attributes (positively) and attributes designated with other objects (negatively) specified in the prompt. To determine the strength of the interpolation, the authors leverage the similarity between the EOT and the last PAD token. In addition, for any object word, the authors retrieve their neighboring words based on feature and semantic similarity to further twist the binding vectors, resulting in enhanced concept disentanglement. Strengths: 1. The paper is novel in that it mitigates attribute binding problem from the text-encoder perspective instead of iteratively refining the cross-attention activations that most previous works focused on. This brings both enhanced results and improved efficiency comparatively. 2. The paper's idea is well-motivated with empirical evidence (e.g., using the cosine similarity of PAD and EOT to decide the strength of embedding interpolation by showing that there's more decay in attribute-specific information in later PAD tokens when such attributes are uncommon). 3. The paper is well-written with many design choices are well-justified (e.g., use of human evaluators, strength formula, necessity of both positive and negative binding vectors, etc). Weaknesses: 1. The scope of the work is limited -- it only covers fixing attribute binding problems in text-to-image generation, one downstream task that is commonly encountered in compositional generation [1]. 2. The method is model-specific. While it alleviates text-to-image models with a CLIP-based text encoder to a significant degree, it is unknown if this method also improves text-to-image models that use multiple text encoders [2] or different text encoders [3]. 3. I find the neighbor finding procedure a bit awkward. For example, the authors have to manually gather 614 object nouns as feature neighbors and have to prompt GPT-3 to gather semantic neighbors, limiting the applicability of the work toward a completely automatic pipeline (e.g., without manual curation or LLM-assisted selection). [1] T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation [2] SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis [3] PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Could authors add reference to Stanza’s dependency parsing module in line 99? 2. Given attribute binding is a core task in compositional text-to-image generation, could authors also discuss more recent and/or relevant in related works, such as [4, 5, 6, 7]? 3. Can authors also evaluate on image quality metrics such as FID [9] and compare with other methods? 4. I wonder if authors could compare Magnet to this concurrent work [8]? As its preprint came out this March, there's no need to experimentally compare and contrast. Some discussions would be appreciated. [4] Training-Free Layout Control with Cross-Attention Guidance [5] TokenCompose: Text-to-Image Diffusion with Token-level Supervision [6] RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models [7] Compositional Visual Generation with Composable Diffusion Models [8] Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions [9] GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The paper discussed limitations in missing objects, possibilities of over- and under-manipulation as well as failures in correcting position relationships of the objects with specific examples, which are justifiable given the scope of their method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Firstly, we would like to express our sincere gratitude for reviewing our manuscript and providing valuable feedback. Below are our responses to the weaknesses (W) and Questions (Q). W1: We acknowledge that attribute binding, as our main focus, is part of compositional generation. However, since Magnet performs outside the U-Net, it is plug-and-play for layout-based methods (see Fig. 2(c), layout-guidance+Magnet) or add spatial control (see Fig. 2(d), ControlNet+Magnet) to address the object relationships. Additionally, we have compared Magnet with GORS [1] (see Fig. 2(e) in PDF). Whether trained on color or complex datasets, GORS cannot demonstrate the anti-prior ability and disentangle concepts like Magnet. Meanwhile, it fine-tunes both the CLIP text encoder and the U-Net with LoRA is time-consuming to adapt to different T2I models, while Magnet is plug-and-play. Although the focus of this work is on attribute understanding, we sincerely hope that our work can provide new insights to the community and motivate further work investigating compositional understanding of T2I models. W2: We have applied Magnet to SD 2.1 (another version of the CLIP text encoder), SDXL [2] (multiple text encoders), and PixArt [3] (T5 model) in Fig. 2(a)(b) in the attached PDF. Magnet also improves text alignment and image quality, showing the anti-prior ability. We will add these results to our final version. W3: Magnet is completely automatic during evaluation. Though we describe that we gathered 641 nouns "manually", this procedure is one-for-all. We compute 641 embeddings once for each new text encoder and save them to the local path. In practice, given any prompt, only need to load this local file to search neighbors and encode a few new prompts (e.g., a prompt with 2 concepts, $\sim5\times2\times2$ new prompts are automatically constructed and need to be encoded if $K=5$). This process is fast as shown in Tab. 2 where Magnet has a negligible increase in runtime and memory usage compared to StructureDiffusion and Attend-and-Excite. Q1: Your careful review is much appreciated. We will cite Stanza correctly in the main paper, the same as line 466. Q2: Here is our justification for Magnet different from other works: [4], [6] rely on the layout constraints, which is an additional condition and not included in the vanilla diffusion models, while Magnet only requires the input prompt. We cited two similar works in lines 252-253. [5] controls objects in prompts, cannot improve the text alignment for attributes like Magnet. [7] composes different noise latents and needs multiple diffusion processes to obtain the target latents. Magnet operates the text embedding outside the U-Net and performs only one diffusion process, which adds negligible cost to the original model and is more efficient. We will discuss these mentioned works in related works in the final version. Q3: We have evaluated on FID for two SD versions (v1.4 and v2.1). We follow the standard evaluation process and generate 10k images from randomly sampled MSCOCO captions. The result is SD v1.4 (19.04), +Magnet (18.92); SD v2.1 (19.76), +Magnet (19.20). This shows that Magnet will not deteriorate the image quality while improving the text alignment. Q4: Thanks for recommending this impressive work. We found some ideas in this paper aligned with ours. But here are some major differences to point out: 1. It lacks interpretability. In our paper, we analyze the CLIP text encoder and the diffusion model by the 4-case experiment in Fig. 2(a), pointing out the entanglement of the padding embeddings. This has motivated our method and explains why Magnet works (i.e., using the binding vector to enhance the distinction between concepts and improve disentanglement); 2. It defines "positive and negative prompts" that are semantically contrastive, e.g., "young" vs. "old" for the "age" attribute (may need to manually specify per attribute). In Magnet, positive and negative attributes are obtained from the given prompt automatically, and no need to be semantically opposite. We also introduce unconditional prompts as pivots, which is suitable for attributes that are hard to define their opposite attributes (e.g., what is the contrastive adjective of the color "yellow"?); 3. Its direction vector is learned with a loss function to ensure robustness. According to the experiment setting, it uses fixed strength (edit delta). Differently, we propose the neighbor strategy and adaptive strength to improve the vector estimation. Encountered with a new attribute, [7] needs a training process to seek this direction, while Magnet can be directly applied. In the final version, we will discuss this concurrent work [8] in more detail. --- Rebuttal Comment 1.1: Title: Response to author's rebuttal Comment: I would like to thank the authors for the rebuttal -- it has addressed all of my concerns to a considerable extent. I'm also impressed that Magnet works just as well on different types of models (e.g., SD 2.1, SDXL, PixArt) and/or for different constrained generation (e.g., Layout-Guidance, ControlNet) based on the provided rebuttal file. Since SDXL and PixArt use different text encoding strategies (e.g., multimodal text encoder or different text encoder), I wonder if the authors could discuss any necessary modifications (e.g., hyperparameters, Magnet formula, etc) to adapt your original setting for the CLIP encoder to work for two encoders (i.e., SDXL) or a T5 encoder (PixArt) to achieve considerable improvement in attribute binding? This would help me better assess the flexibility of your method as well as the scope of impact. Thanks! --- Reply to Comment 1.1.1: Title: Response to Comment about Flexibility of Magnet Comment: Thank you for your thoughtful follow-up and your interest in the flexibility of Magnet. When applying Magnet to SDXL, we did not modify any hyperparameters or formulas, keeping all settings consistent with the main paper (e.g., $K=5$, $\lambda=0.6$). The adaptive strength formula (Eq. 3) is designed based on our analysis of the CLIP encoder and naturally adapts to SDXL. However, the T5 encoder in PixArt operates differently by modeling bidirectional context. Consequently, it is necessary to conduct further analysis of T5 and redesign the strength formula. Due to the time constraints of the rebuttal process, we were unable to conduct a detailed analysis and therefore used fixed strengths ($\alpha=2$, $\beta=0.5$) for all objects. We also found that $K=10$ is more robust for PixArt. We will incorporate these considerations in future work to further demonstrate the flexibility and impact of our method.	Summary: - This work studies how CLIP text embeddings commonly used in text-to-image diffusion models affect attributes in generated images, and how attributes can be bound to the correct objects during generation. - There is an analysis of the (a) CLIP text encoder and how it interacts with the padding used during T2I diffusion (b) how text embeddings of different nouns relate to different attributes w.r.t to distance (c) how the previous two observations interact during diffusion-based generation. - An algorithmic innovation — Magnet — is proposed. Magnet introduces a binding vector that can be applied to the embedding of a noun-object to bind an attribute to it so that it is faithfully applied during generation. - There is a human evaluation (amongst other evaluations) which shows that Magnet improves attribute-object bindings. Strengths: - The paper is organized very well. I appreciated the bolding. - The in-depth analysis of what causes the attribute binding problem was interesting and would be widely useful to the community. - Human evaluation shows that attribute alignment is substantially improved by Magnet. I appreciated the use of a human evaluation rather than merely using automated metrics. - Magnet is much cheaper w.r.t runtime and memory usage than competing methods. Weaknesses: I had great difficulty understanding most of the figures. Even after understanding the method, I cannot understand what Fig 3. is showing and how it relates to the method. Similarly, I found the equations and the notations cumbersome, illegible, and frustrating. I am honestly unsure if I understand the method, because the writing is very unclear (though well organized) and the notation is completely overwrought. One minor weakness is that the analysis is highly specific to the CLIP text encoder. I don't know how much of a weakness this is practically since the CLIP text encoder is a de-facto standard, but I am writing it here anyway for completeness. Technical Quality: 3 Clarity: 3 Questions for Authors: - Have at least one simple figure that explains the intuition behind the approach. - Rewrite 3.1 to be more clear or at least provide examples of all of the $\mathcal{P}$-terms, I am not even sure the notation is correct here. - More generally, provide a comprehensible explanation of the method so I can confirm my understanding is correct. I'm uncomfortable providing a higher score otherwise, though I would like to, given the amount of work that seems to have gone into the paper. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Not applicable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Your positive comment on the contribution of this work is much appreciated. We apologize for any difficulty you may have experienced in following this paper. Hope the following example and the illustration in Fig. 3 in the attached PDF will help you to understand these $\mathcal{P}$-terms and our method. Consider an input prompt $\mathcal{P}=$"a blue book and a red cup", we use an off-the-shelf parser to extract concepts "blue book"($A_1\\&E_1$) and "red cup"($A_2\\&E_2$). The term $\mathcal{\tilde{P}}$ refers to a new prompt that involves only one object (in line 110, described as "out of the current context of $\mathcal{P}$"). For instance, when operating the object "cup" (i.e., $i=2$), we define positive concept $\mathcal{\tilde{P}}^{pos}_2$="red cup"($A_2\\&E_2$), negative concept $\mathcal{\tilde{P}}^{neg}_2$="blue cup"($A_1\\&E_2$), and unconditional concept $\mathcal{\tilde{P}}^{uc}_2$="cup"($\varnothing\\&E_2$). The function $\mathcal{F}$ extracts the embedding w.r.t. a word in one specific prompt (illustrated by the red box). Notice that the same word in different prompts with varied contexts will produce different word embeddings, i.e., $\mathcal{F}(E_2,\mathcal{\tilde{P}}^{pos}_2)\neq \mathcal{F}(E_2,\mathcal{P})$. Based on this fact, we introduce Eq. (1)(2) to estimate the binding vectors $v^{pos}_2,v^{neg}_2$ for "cup". The vector $v^{pos}_2$ identifies the direction towards "red", while $v^{neg}_2$ is the direction towards "blue", specifically for "cup". Similarly, we obtain the binding vectors $v^{pos}_1,v^{neg}_1$ specifically for the object "book"($E_1$). When involving neighbors, we compute $K$ objects close to the "cup" embedding in the feature space (lines 129-132), denoted $\\{B^{(2)}\_k\\}^{K}\_{k=1}$, where superscript $(2)$ refers to the current object "cup" $i=2$. Next, we replace "cup" in three constructed prompts with each neighbor object. For instance, the second neighbor "mug" ($B^{(2)}_2$, where subscript refers to $r=2$) will produce the positive concept $\mathcal{\tilde{P}}^{pos}_2$="red mug"($A_2\\&B^{(2)}_2$), negative concept $\mathcal{\tilde{P}}^{neg}_2$="blue mug"($A_1\\&B^{(2)}_2$) and unconditional concept $\mathcal{\tilde{P}}^{uc}_2$="mug"($\varnothing\\&B^{(2)}_2$). The same process is for the remaining neighbors to obtain $K$ positive and $K$ negative vectors. The final binding vectors $v^{pos}_2,v^{neg}_2$ used to modify "cup" are averaged as Eq. (5)(6). The adaptive strength in Eq. (3) improves robustness for practical use. For the object "cup" ($i=2$), we extract the first [EOT], i.e., $\mathcal{G}(\mathcal{\tilde{P}}^{pos}_2)$, and the last padding embedding, i.e., $\mathcal{H}(\mathcal{\tilde{P}}^{pos}_2)$, to compute $\omega_2$ and two strengths $\alpha_2,\beta_2$. Note the used prompt to obtain $\omega_2$ is not the input prompt $\mathcal{P}$ but the constructed positive concept $\mathcal{\tilde{P}}^{pos}_2$. Our motivation for this strategy is described in lines 116-119. Similarly, $\omega_1, \alpha_1,\beta_1$ are calculated for the object "book" ($i=1$). Finally, we modulate the original object embedding $c\_{E\_i}$ using the adaptive strengths $\alpha_i,\beta_i$ and the estimated positive and negative vectors $v^{pos}_i,v^{neg}_i$, denoted $\hat{c}\_{E\_i}$ as Eq. (4). Note Magnet does not manipulate the cross-attentional activations as existing works do. As described in Section 3.3, Magnet only modifies the embedding of each object word in the input prompt and treats the denoising U-Net as a black box to generate the image. Magnet provides a simple but effective way to identify the attribute direction for each object, as we said in line 98, to attract the target attribute (i.e., positive vector) and repulse other attributes (i.e., negative vector). Experiments show Magnet disentangling different concepts and improving attribute alignment (see Fig. 4,5 in the main paper). In Fig. 2(b) in the attached PDF, we show Magnet is not limited to the CLIP text encoder, improving the text alignment and synthesis quality of PixArt [1], which uses T5 as the text encoder. We sincerely hope that the above explanation will clear up the confusion. We will improve the method section to make this paper more readable. [1] PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis --- Rebuttal Comment 1.1: Comment: Thank you for your effort. Unfortunately, your explanation in the comments did not help. I strongly encourage you to choose different notation. I will keep my rating; I think there is a clear technical contribution in this paper, but the comment did not clear anything up for me. I have put my confidence at a 2; the AC can disregard my review at their discretion.	Summary: The authors propose _Magnet_ to solve the attribute binding problem. (1) Initially, they specifically analyze how the improper binding problem occurs in text embeddings. By comparing embeddings for each token, they demonstrate the attribute bias phenomenon where attributes do not bind well to the object token $c_{object}$ in rare concepts like "blue apple." Additionally, through cosine similarity analysis of padding embeddings, they hypothesize that the latter padding tokens, e.g., $pad_{73}$, tend to represent prior bindings (like red apple) rather than binding the target concept (blue) with the object (apple). (2) Magnet suggests a method to slightly edit the embedding of the object to align with the desired concept, based on these observations. They demonstrate their method's efficacy through ARC-6k and CC-500 datasets. Strengths: - Writing quality The writing quality is good for understanding the core motivation and method. Analyzing embeddings using cosine similarity is intuitive and effectively demonstrates how attribute bias negatively impacts attribute binding. The flow from the initial analysis to the method is also easy to follow. - Novelty Resolving attribute binding in the text space, as done by Magnet, seems quite plausible and novel. Identifying the limitations of the T2I diffusion model from the VLM model has been necessary but has not been extensively conducted until now. Weaknesses: Overall, I think it's a good paper. It would be better if the evaluation were a bit more solid. - Evaluation I agree that a user study w.r.t performance comparison as shown in Table 1 is necessary, but for this task, it is necessary to evaluate whether each method achieved binding rather than comparing different methods. Since we don't know the precise performance of each technique, it's difficult to understand the actual performance of the Magnet and how much it has improved. The results in Table 2 are weak. It's disappointing that the automatic results are quantitatively worse compared to attend-and-excite. While I agree that manual inspection is the most accurate, trying out a recent VLM like GPT-4o, which can discern characteristics that image encoder-based evaluations fail to capture, might be worthwhile. Technical Quality: 4 Clarity: 3 Questions for Authors: Q1: How is the analysis in lines 64-78 and Figure 2-(a) connected to the method? Q2: Is there no need to modify the padding? If the padding tokens positioned at the end contain prior knowledge, it seems they might also interfere with binding. Q3: How did you perform detection using GrondingDINO in Table 2? I am curious whether you detected "blue apple" or just "apple." It might be somewhat out-of-scope, but I am also interested in how different the results are between detecting "blue apple" and "apple." Q4: As mentioned in the limitations (Figure 20), Magnet seems relatively weak at handling positional information. I suspect this issue arises not so much from a problem with Magnet itself but rather because CLIP's embedding does not effectively bind positional information to the desired token. Is there an analysis similar to Figure 1 for positional information? It seems only color is present in the appendix. Showing this could further demonstrate, as per the title of your paper, an understanding of VLM’s function, providing clarity on what can and cannot be bound. --- - Relevant references Before this paper was submitted, methods for editing text tokens were proposed in papers like [1, 2]. Although these papers also deal with a different task, and attribute binding, and were not considered in the evaluation due to being on Arxiv, I recommend citing them in the camera-ready version. [1]: Uncovering the Text Embedding in Text-to-Image Diffusion Models, https://arxiv.org/abs/2404.01154 [2]: TexSliders: Diffusion-Based Texture Editing in CLIP Space, https://arxiv.org/abs/2405.00672 Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: The authors describe the failure cases of Magnet in the appendix. They have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your thorough review of our manuscript and the insightful comments provided. Here are our detailed responses to the identified Weaknesses (W) and Questions (Q). W1: Our comparison experiment has been designed to evaluate whether each method achieved binding. In Tab. 1, attribute disentanglement is assessed by asking "Which image shows different attributes more clearly?". Only the method achieving the best binding will be voted, otherwise evaluators tend to choose no winner. In Tab. 2 attribute alignment (Attr.), human annotators annotate a precise number of successful bindings (whether one object presents the desired attribute). This is done separately for each method, without comparing the others. Lines 161-178 have explained each criterion in detail. W2: Automatic results in Tab. 2 only assess object existence (see Q3). Magnet is inferior to Attend-and-Excite (Attend) since it optimizes the noisy latent to encourage the model to attend to all objects. As discussed in limitations, we acknowledge Magnet suffering from missing objects. Though designed to enhance attribute alignment (has outperformed all baselines in Tab. 2), Magnet shows improvement in object existence (Det. $+5$ on SD), which is a bonus of disentangling different concepts (see Fig. 6(b)). This still holds when integrating Magnet with Attend. We tested Attend+Magnet on CC-500 via GroundingDINO. The Det.($\uparrow$) result is 87.7, compared to 84.3 w.r.t the original Attend. Attend+Magnet can encourage all objects and improve attributes simultaneously (see Fig. 8). Since there is no available API at this moment, we cannot evaluate Magnet with GPT-4o. But we will definitely try out GPT-4o whenever possible. Q1: The 4-case experiment is deeply connected to our method. Lines 64-78 and Fig. 2(a) focus on comparing how the context information in the word embedding affects generation (i.e., from case 1 to 2, or case 3 to 4), and how paddings' context information affects (i.e., from case 1 to 3, or case 2 to 4). Appendix A.2 provides a detailed analysis of these fine-grained cases. We discern that the attribute information is rich in word embeddings but entangled in padding embeddings, and modifying these word embeddings will not change the image layout as significantly as modifying padding embeddings. These have motivated us to introduce the binding vector, adaptive strength, and neighbor strategy (see lines 439-443). Q2: We did consider modifying the padding tokens, but for the following reasons it didn't work: 1. These padding embeddings strongly entangle different concepts, especially given complex prompts. It is hard to manipulate one specific object from these paddings; 2. As shown in Fig. 2 and Fig. 12 in the main paper, modifying the padding (from cases 1 to 3) will change the image significantly compared to modifying the word embedding (from cases 1 to 2). We consider not changing the non-attribute part can better maintain the pre-trained model's capability; 3. As shown in Fig. 1 in the attached PDF, the padding embeddings are less activated than word embeddings, especially in U-Net's last two upsampling blocks and the later diffusion steps, which are more crucial for generating semantic features. Q3: In our comparison experiment, we input all object words in the prompt to GroundingDINO. For example, given the prompt "a blue apple and a green vase" with two objects, GroundingDINO's input is "apple . vase ." without any adjective. We have considered inputting adjectives, i.e., "blue apple . green vase .". However, no matter what color the apple is (e.g., blue, red, or green), GroundingDINO in both settings will detect it. That's why we use GroundingDINO to assess object existence only, but not attribute alignment (see lines 177-178). Appendix Fig. 15(c) presents GroundingDINO's failure cases. We further conducted the following experiment to make our judgment sound. Given 50 images (generated by the prompt "a blue apple and a green vase"), GroundingDINO detected $38$ (w/o adjective), $38$ (w/ adjective), human annotated $\sim 20$ "blue apple", maximum is 1 per image. This comparison shows that GroundingDINO is not reliable for measuring attribute alignment. Q4: We agree with your idea that the CLIP's embedding does not effectively bind positional information to the desired token. However, this makes an analysis like Fig. 1 difficult - we don't know which token to identify. We conjecture that the positional information is used in the early diffusion steps, where all tokens have relatively small attention values (see Fig. 1 in the attached PDF). In this case, the entangled padding embeddings may affect generation. This also explains why Magnet modifying the word embeddings cannot address the positional problem. However, we can integrate Magnet with layout-based methods to handle position (see Fig. 2(c) in the PDF). Except for colors, Magnet is capable of binding other attributes (see Fig. 2(a) row 3 and 2(b) row 2 in the PDF). The final version will include a detailed discussion of positional relationships and more examples. Finally, the two textual-based image editing works are impressive. We will cite the given papers and discuss the differences in the final version. [1] T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation --- Rebuttal 2: Comment: Thank you to the authors for their detailed response. I have carefully considered the points raised. > [W1] "Our comparison experiment has been designed to evaluate whether each method achieved binding. In Table 1, attribute disentanglement is assessed by asking, 'Which image shows different attributes more clearly?' Only the method achieving the best binding will be voted for; otherwise, evaluators tend to choose no winner." While I understand the authors' approach, I still believe that this method inherently involves a comparison. Evaluating A vs. B seems likely to yield different insights than evaluating A and B separately. The authors suggest that both approaches could lead to similar results, but I respectfully disagree, as I believe pairwise comparisons and individual method evaluations may not produce the same outcomes. > [W2] "Since there is no available API at this moment, we cannot evaluate Magnet with GPT-4o. But we will definitely try out GPT-4o whenever possible." I appreciate the authors' intent to evaluate with GPT-4o when possible. However, isn't the GPT-4o API already available? > [Q3] "This comparison shows that GroundingDINO is not reliable for measuring attribute alignment." I understand this concern, and I recognize the challenges it presents. It is unfortunate that there currently isn't a more robust evaluation method beyond manual inspection. I had hoped that GPT-4o might offer some supplementary capabilities in this area. Given these unresolved concerns regarding the evaluation, I will maintain my score as a weak accept.	Summary: The paper analyzes and improves upon the “(attribute) binding problem” in VLMs such as CLIP, and the focus is primarily on the text encoder side. First the authors analyze how the individual text embeddings behave in a diagnostic setting when encoding a two-word text “COLOR OBJECT”. With these insights they propose to modify the original embedding of an object by calculating a “binding vector” that uses a positive attribute (the one actually in the prompt, i.e. “yellow”) and negative attributes (other colors). They show improvements to previous methods on top of which they build such as Structure Diffusion. Strengths: The paper is very practical and intuitive, motivating the final method with analysis of the embedding space. It has a clear contribution and the whole story ties around this contribution, providing additional insights. On top, the evaluation is thorough and uses human judgment. (there is no automatic metric here, I agree). It is also good to see that the neighbor strategy was quantitatively ablated with human judgment, and section 4.6 adds additional nice insights, albeit not beyond a few examples as far as I can tell. Overall a very interesting and creative exploration, and the empirical results do show that the method works! That is exciting, given that this is a persistent major problem many papers have addressed. Weaknesses: While the method clearly improves upon previous work, it requires pre-computing or manually defining a lot of components such as the negative attribute set or the neighbor set, with additional computation of the adaptive strength. With pre-computation this is not too expensive but it feels somewhat hacked together. The abstract mentions too many terms where it is unclear whether they are established terms and if not, the abstract is not the right place to mention them all: blended text embeddings, attribute bias, object embedding, binding vector. Since this papers main focus is a thorough analysis and interpretability insights, it has to be clearer with its definition, i.e.: “The above phenomenon, which we call attribute bias, occurs when one object prefers or dislikes a specific attribute” → refer/dislike are not well-defined. The 4 experiments with swapping various embeddings seem very interesting but it is not motivated why we need all 4, maybe 2 cases would have been more insightful and easier for the reader to understand what’s going on? Finally, having more quantitative statistically significant experiments and less showing of examples would have strengthened the paper's contributions further! Technical Quality: 2 Clarity: 2 Questions for Authors: It says in the abstract “blended text embeddings as the cause of improper binding… “. Is “blended text embedding” embeddings a term people are expected to know? I work in vision-and-language and have not heard of it. Again in abstract “we observe the attribute bias phenomenon”: was this meant to say attribute binding? How are negative attributes exactly determined? Most things are not as easy as color. Is there an explanation why the Structure Diffusion baseline is so bad? Does your method differ to Stucture Diffusion only in the proposed Magnet method and nothing else? Same exact SD version, CLIP version etc.? This is important to clarify in the paper so that it is clear that your method alone leads to this improvement. Note: Fig. 1b is very small and also low-resolution, please fix! Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 4 Limitations: Yes, addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your careful review and the valuable suggestions provided. Our responses to the Weaknesses (W) and Questions (Q) are outlined below. W1: Magnet is fully automatic during evaluation and does not require manual definition of these positive/negative attributes. We have introduced these components to ensure robustness. Various ablation experiments have proved the effectiveness of each component, including the parameter $\lambda$ (see Appendix Fig. 17), the neighbor strategy (see Fig. 18), and the positive and negative vectors (see Fig. 19). Meanwhile, our method can be simplified by using fixed strengths and estimating the binding vector without neighbors (i.e., by the object itself). Furthermore, as you pointed out, it is not expensive at all. W2&Q1&Q2: Your comments on our abstract are very valuable to us. The following is our explanation of the terms used in the abstract: 1. the term "blended text embeddings" in other words means "text embeddings are blended", with reference to [1] "tokens in the later part of a sequence are blended with the token semantics before them"; 2. the term "attribute bias phenomenon" appears in line 60, i.e., our discovery based on the analysis of the CLIP text encoder, rather than the "attribute binding problem". We will check all the terms in the abstract, and improve it to be clear and professional. W3: This sentence can be explained with an example in Fig. 1(b) (apology for its blur, we have provided a high-resolution one in Fig. 1 in the attached PDF). For the object "banana", both word and EOT embeddings show high similarity on the color "yellow" but low similarity on "blue", as if "banana" prefers the attribute "yellow" but dislikes "blue". We will improve the analysis section and clarify all definitions in the final version. W4: The 4-case experiment is deeply connected to our method. As discussed in lines 439-443, our motivation is based on several observations between 4 cases. Specifically, case 1 is a reference case w.r.t. standard generation, case 2 shows how the context information of word embeddings affects generation, case 3 shows how padding embeddings' context information affects generation, and case 4 tests whether the model can capture attribute information in adjectives (described in lines 428-429). Using all 4 cases would be more convincing to introduce our proposed binding vector and adaptive strength. In addition, we add 3 new cases in the Appendix to justify the information-forgotten problem in the latter padding embeddings (see Fig. 12, cases A-C). We will add more descriptions of this experiment to the main paper and improve readability. Q3: We define "negative attributes" as adjectives in the given prompt that do not belong to the current object. For instance, given the prompt "a green apple on a wooden table and a red chair" with three concepts, when operating the object "apple", its positive attribute is "green", then negative attributes are "wooden" and "red" (belong to "table" and "chair", respectively). It should be noted that negative and positive attributes do not need to be semantically opposed to each other. Magnet is not limited to colors and can work on prompts with extensive attributes (see Fig. 2(a) row 3 and (b) row 2 in the attached PDF). We will add more non-color examples in the final version. Q4: We assure that our experiments are completely fair through consistent control of all settings (e.g., the same SD&CLIP version and seeds). Its poor performance can be explained by our discovery in the text encoder: the improper binding is caused not only by the word embeddings, but also by the entangled padding embeddings. StructureDiffusion simply separating the word embeddings of different concepts is insufficient to offset the entanglement in the padding embeddings. Conversely, we introduce binding vectors to reinforce the difference between each concept. This improves disentanglement. Additionally, StructureDiffusion also performed poorly in [2]'s comparison experiments - Structure Diffusion may even get worse scores than Stable Diffusion. Finally, we sincerely apologize for Fig. 1 and will fix it in the final version. Your feedback is very helpful to us. We will improve our paper to strengthen our contribution, reducing visual examples and adding quantitative results (e.g., FID-10k: SD 19.04, Magnet 18.92). [1] Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis [2] Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models --- Rebuttal Comment 1.1: Comment: Thank you for the explanations and the new PDF! It is good to see you provided new results and engaged with all the questions. I still think my score of 6 is appropriate and think this paper is a weak accept.	Rebuttal 1: Rebuttal: In the attached PDF file, we provide a new perspective on how the word and padding embeddings affect generation (see Fig. 1), additional examples applying Magnet to different T2I models and other techniques (see Fig. 2), and a new Magnet pipeline figure for ease of understanding (see Fig. 3). We have read all the papers mentioned by the reviewers and would like to highlight some differences as follows: 1. Existing works analyzing both word and EOT/pad embeddings emphasize the semantic effect of the word embedding, while we point out the entanglement of the pad embeddings to explain existing problems of SD (e.g., incorrect attributes, indistinguishable objects, see Appendix Fig. 13); 2. Related works require additional inputs or fine-tune the denoising U-Net, whereas Magnet takes only the given text as input and can be applied directly to any prompt; 3. Operating outside the U-Net, Magnet provides a plug-and-play capability and can be readily integrated to existing T2I models (e.g., SD, SDXL [1], PixArt [2]) and controlling techniques (e.g, Attend-and-Excite, layout-guidance [3], ControlNet [4]); 4. Magnet shows the anti-prior ability, i.e., to generate high-quality images with unnatural concepts, while related works are incompetent (see Appendix Fig. 24, and Fig. 2(e) compared to GORS [5] in the attached PDF). Last but not least, we sincerely hope that our work will motivate the community to explore generative diffusion models and discover other interesting phenomena. [1] SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis [2] PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis [3] Training-Free Layout Control with Cross-Attention Guidance [4] Adding Conditional Control to Text-to-Image Diffusion Models [5] T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation Pdf: /pdf/333f9af7397a136d1a2db32c7c975f3c3d07a960.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Motion Forecasting in Continuous Driving	Accept (spotlight)	Summary: This paper proposes a model for motion forecasting that models the continuous stream of world state over sequential timesteps. This is in contrast to most/all other works which do an indepedent prediction of the future for each timestep based on a fixed window of history, without context of previous model belief states or outputs from previous time steps. One stream of data is scene context, represented as embeddings, where context is fused with current accounting for transformation of the scene due to ego motion. The second stream is agent trajectories, where a history of n previous predicted trajectories for each agent is maintained explicitly in a memory bank, and fused into the current time steps prediction as additional signal. They do thorough ablations on Argoverse benchmark dataset, showing compelling evidence that the streaming architecture is beneficial. Furthermore, they exceed SOTA on most metrics in the benchmark. When I checked today (mid July), they are ranked 7th overall in the leaderboard - https://eval.ai/web/challenges/challenge-page/1719/leaderboard/4098 Strengths: This paper addresses an important problem often ignored in these sanitized datasets: that these models in real world application in an AV system run in a sequential streaming fashion, and (to my knowledge) all other works operate on each time frame independently (given a fixed window of historical state). The proposed solution itself is a good one, transforming old belief state into current reference frames in both embedding space and with spatial transformations where appropriate. At the same time, the complexity of the architecture is kept reasonably low, with each contribution justified. Results are very strong and well justified with thorough ablations. Weaknesses: Minor, but I had some difficulty understanding some of the explanations: - Data reorganization / Figure 2: "we select several split points Ti along historical frame steps" --> how are they selected? what are "historical frame steps"? Can you more simply say this as, for example, "We split sequences evenly into sub-sequences"? - Around L158: you started using the term "modes" for the first time, where I think you should just say trajectories, or be more careful defining the tensors F_{mo} and Y_{mo} making clear their dimensionality. - You never clearly state the output representation of your model (which I assumed to be a weighted trajectory set). Without being more clear here, Section 3.3. Model training is sort of meaningless - for example, what is the classification loss classifying? (I'm 99% sure it's a softmax distribution over trajectory modes, but this is not said anywhere.) Technical Quality: 4 Clarity: 2 Questions for Authors: See weaknesses Confidence: 4 Soundness: 4 Presentation: 2 Contribution: 4 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed review as well as the suggestions for improvement. Our response to the reviewer’s comments is below: Apologies for these unclear explanations. + Data reorganization: As shown in Fig. 2(b), the dashed part refers to historical time steps (50 historical steps in AV2), and we select several split points on this part. We consider an original trajectory (the gray arrow) as a sequence, and our processing approach generates several shorter trajectories (sub-sequences). Besides, the split points are manually specified in our method and discussed in ablation Tab.4(b). + Mode: All correct! Multiple trajectories are predicted for each agent to enable comprehensive motion possibilities, in which case $mode$ is commonly used to represent each prediction. We will make it clear. Besides, the dimension of $Y_{mo}$ is $[B, N_a, N_{mode}, 2]$. + Model output: The output is several predicted trajectories with corresponding probabilities, as the reviewer understood. While providing the types of loss functions in Line 177, we indeed overlooked the concrete input items. We will clarify this.	Summary: This submission tackles the task of trajectory forecasting in autonomous driving. In particular, it proposes improvements in two aspects: 1/ Data reorganization: Current datasets are artificially split into non-overlapping segments. The authors propose to reorganize the data to have overlapping windows. 2/ Relaying context and predictions: since the sub-scenes now overlap, they can use context and predictions from the previous timesteps (corresponding to a different sub-scene) to improve predictions at the current timestep. Relaying context serves to provide more information about the surroundings (an agent could have been missed in the current frame for instance), and relaying the predictions would smooth out the predictions across time. They demonstrate the validity of their approach on Argoverse benchmarks, achieving state-of-the-art performances. Strengths: - The proposed improvements are interesting, sound, and motivated. - The data reorganization, I believe, is a simple idea that would that should be pushed in the motion forecasting community to build a new standard that better leverage the data. It can also immediately be applied to other models and hopefully improve performances. - The relaying through time helps achieve strong results either with their architecture or on top of QCNet. Weaknesses: - The results are not very conclusive: - Compared to QCNet, except for $ADE_1$/$FDE_1$ (which are not the main metrics usually considered for motion forecasting) the improvements are small. It would be useful to give more insight into why only these metrics improve. We also don't have the $ADE_1$/$FDE_1$ values of QCNet + RealMotion so we don't know if these findings are consistent across models. - The comparisons are not fair in terms of available information. Because of the propagation, RealMotion has access to more past information both in the past trajectories and agents, and the past map. A naïve way to integrate more past trajectories would be to just use longer past history. This should be a minimal baseline or ablation in the paper. If we also wanted to add agents seen in the past but not at the current timestep, it could be done with some masking or extrapolation scheme for missing frames. Remains the past map information, for which it's indeed more complicated to have a naïve baseline for a fair comparison, as simple aggregation could drastically raise the computation footprint so I can understand why the authors wouldn't want to test this baseline. - The loss functions, and in particular $\mathcal{L}_{refine}$ are not detailed enough. I can guess where they apply but it would be good if that was stated explicitly. - Figure 2 is unclear to me. Perhaps it could need more captions or a rework. - I think it would be easier to interpret the results if the architecture of RealMotion would be more clearly delivered, perhaps in a Figure (in Appendix?) that represent the full architecture detailing in particular how the different elements interact in the encoder and the decoder that are not given in Figure 3. Technical Quality: 2 Clarity: 2 Questions for Authors: - Why do we see drastic improvement on ADE/FDE_1 but the other improvements are very tame? Does that imply most of the improvements are to the scoring function so that RealMotion gets a better top 1? If that's the case, what could be the mechanisms that explain this behavior? Otherwise, I think this paper rightfully questions some practices of motion forecasting in autonomous driving. Generalizing something along the lines of the proposed data reorganization would be very helpful for the community. However, I think the experiments on relaying might not be fair enough and need to be compared to simple baselines with more/better past history. We would also need a few more numbers to see if most of the findings and recommendations have a good chance at being consistent accross models. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: - Limitations are discussed in the Appendix. - Broader societal impacts do not seem explicitly addressed despite what has been reported in the checklist. I'm not sure there are obvious uncontroversially negative societal impacts to be discussed though. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed review as well as the suggestions for improvement. Our response to the reviewer’s comments is below: Q1: Drastic improvement on $ADE_1/FDE_1$. By comparing QCNet and other methods (such as ProphNet), it can be seen this phenomenon is common. This is due to different calculation of metrics: the metrics@1 is the error for the one selected trajectory according to the score, while the metrics@6 is the minimum error among 6 predicted trajectories. Hence, there is more room for improvement for the metrics@1. Regarding the comparison between QCNet and ours, the improvement in $FDE_{6}$ is still salient. Besides, we provide the complete comparisons below. Since we can only integrate our restructured data, please refer to Tab.3 row 1 and 2 for comparison. \|Method\|$minFDE_1$\|$minADE_1$\|$minFDE_6$\|$minADE_6$\| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|QCNet\|4.34 \| 1.69 \| 1.27 \| 0.73\| \|QCNet w/ data\|4.26 \| 1.64 \| 1.24\|0.71\| Q2: The comparisons are not fair. We do NOT use more past information compared to existing methods, but just process past information in a different way with the same datasets. Taking existing methods on Argoverse 2 as an example, they directly process each independent scenes with 50 historical frames, while we split each scene into several continuous small scenes with less past information (30 historical frames), then processing them in a streaming fashion. Hence, existing methods and RealMotion all use the same past information, making a fair comparison. Q3: Loss functions. Apologies for unclear expression. We will add the formulation for these losses. Especially, $L_{\rm refine}$ is utilized to supervise the refined trajectories $Y_{\rm mo}$ mentioned in Eq(4), which can be formulated as $L_{\rm refine} = {\rm smoothL1}(Y_{\rm mo}, Y_{\rm gt})$, where $Y_{\rm gt}$ refers to the ground truth. Q4: Fig.2. Fig.2 depicts how we process existing datasets with independent scenes into a continuous format to simulate real-world situations. We will remove some redundant elements to make it clearer. Q5: The architecture of RealMotion To ensure the generalizability of our proposed modules, as most current methods we adopt a standard forecasting framework, which consists of an agent encoder and a map encoder to respectively encode agent and map features, a Transformer encoder to implement interaction, and a decoder to generate trajectories and corresponding probabilities. The figure is shown in PDF file Fig.1. --- Rebuttal 2: Comment: Thank you for your answers and your clarifications. After going through the paper in light of the provided answers, it seems I indeed misunderstood an important part of the submission. To make sure I'm getting it right this time, would you agree your working hypothesis could be sumarized as follows? - Given a large history (of 50 frames), it is better to have a model that operate on a smaller sequence (30 frames) + an embedding of the past that you can obtain using the same model on earlier frames, than directly on the full (50 frames) history like QCNet does. --- Rebuttal 3: Comment: Thanks for the quick response. Exactly. This is more than just simply considering longer historical frames like previous methods, otherwise, our method would not exceed previous alternatives. Our formulation allows to capitalize the sequential data more effectively in real world scenarios, not only emphasizing a longer historical scene context but also leveraging historical predictions, which cannot be achieved by simply increasing historical frames. We will further clarify in the revised version. --- Rebuttal 4: Comment: Dear Reviewer GAo1 Thanks again for the valuable comments. We believe our responses addressed all the questions/concerns. It would be great if the reviewer can kindly check our responses and provide feedback with further questions/concerns (if any). We would be more than happy to address them. Thank you! Best wishes, Authors --- Rebuttal 5: Comment: Indeed, after going again through the submission with the new information from the rebuttal and the reviews, I find my concerns has been adressed properly in the rebuttal. Moreover, I believe the reasons for my initial wrong impression have also been mostly adressed thanks to the discussions between the authors and Reviewer fgWb. Since the authors have commited to make the necessary change, I'm happy to increase my rating. --- Rebuttal Comment 5.1: Comment: We appreciate the reviewer's time for reviewing and thanks for the recognition.	Summary: This paper proposes a framework named "RealMotion" highlighting the importance of continuous motion forecasting in autonomous driving. From the formulation perspective, RealMotion investigates predicting trajectories in a continuous sequence of timestamps instead of previous independent predictions. From the methodology perspective, RealMotion proposes the "relaying" modules that propagate historical scene and agent contexts into the current frame to enhance prediction. Finally, RealMotion achieves top performance on Argoverse benchmarks and proves the importance of continuous contexts. Strengths: 1. This paper investigates a foundational challenge in autonomous driving that is overlooked by previous people -- how to model the continuous contexts in motion forecasting. This problem is reasonable and is also something I would love to follow and work on. 2. The methods proposed by RealMotion are intuitive and reasonable. Relaying the historical context to current frames to enhance motion forecasting is indeed a novel contribution to autonomous driving. 3. The experiments prove the effectiveness of continuous contexts. In addition to the state-of-the-art performance, the comparison between RealMotion-I and RealMotion in Table 1 makes it a strong argument. Weaknesses: 1. The authors have missed some closely related works, such as streaming motion forecasting (e.g., Pang et al.) and other end-to-end autonomous driving (e.g., Gu et al.), where motion forecasting is continuous. For instance, Pang et al. also discover the setbacks of predicting trajectories on independent timestamps, formulate the "streaming" forecasting task, and propose a differentiable filter to refine the predicted trajectories. I acknowledge the newly designed modules by RealMotion, but I would expect the authors to have a better discussion of these relevant efforts. Pang et al. Streaming Motion Forecasting for Autonomous Driving. IROS 2023. Gu et al. ViP3D: End-to-end Visual Trajectory Prediction via 3D Agent Queries. CVPR 2023. 2. RealMotion is also closely related to QCNet and based on its code. QCNet already has the interface of continuous motion forecasting and leverages a GRU to deal with the temporal contexts in queries. First of all, I would like the authors to clarify or discuss in the paper about the difference/improvement to QCNet more explicitly. In addition, I am also curious if it is possible to support that the "relaying" modules in RealMotion is better than QCNet's end-to-end mechanism. For instance, is it possible to plug the relaying module to QCNet and make a comparison? 3. I would like the authors to develop some metrics to analyze the improvement in the trajectories. More specifically, we wish to analyze "where does RealMotion improve the trajectory quality?" Since the main argument of this paper is "continuity," I am curious if the author can use a "fluctuation" measure like Pang et al. to prove that the trajectory smoothness is indeed improved. It is also encouraged to show some visual comparisons if space is allowed. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed limitations. No additional limitations from my side. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed review as well as the suggestions for improvement. Our response to the reviewer’s comments is below: Q1: Missing related work. Thanks. We will add them. Q2: Comparison with QCNet. To ensure efficiency, our method adopts an agent-centric design different from QCNet. As mentioned with QCNet only the query-centric design is good at continuous motion forecasting and GRU is employed for fusing continuous predictions in the current scene. Hence, it should still be considered an independent method. We have tried to integrate our stream modules with QCNet, but unfortunately, it cannot be trained in our environment due to huge memory overhead. We hence just provided the results using our reorganized data in Appendix B.3. Q3: Continuity metrics. Following Pang et al., we evaluate the $fluctuation$ for RealMotion and baseline using our reorganized data as shown below. It can be observed that our method certainly improves the $fluctuation$ metric, with the Trajectory Stream playing a particularly important role. \|\|baseline\|baseline+scene stream\|baseline+traj stream\|RealMotion\| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|$fluctuation$\|0.385\|0.371\|0.354\|0.347\| We also provide some qualitative results in PDF file Fig. 2, which demonstrates that RealMotion can better maintain temporal consistency of trajectories. --- Rebuttal Comment 1.1: Comment: I have checked the rebuttal and the reviews from other reviewers. The added analytical experiment makes sense to me, and the added results of improving QCNet with RealMotion are also reasonable. I think the added visualization is good. For the revisions, I suggest the authors add some highlights pointing to the frames/branches where RealMotion has better continuity than the baseline. I maintain my original rating of weak accept for now. --- Rebuttal 2: Comment: Thanks. We will further highlight the crucial parts.	Summary: The paper introduces an approach to iteratively process temporal scene context for the purposes of agent motion prediction. This differs from traditional approaches that ingest the whole context directly. The scene is processed in temporal chunks, where in each chunk contains agent-centric context consisting of other agents and map within 150m. Information is aggregated across chunks (i.e. over time) using two novel components: - "Scene context stream", where agent/map features from the past frame are transformed using Motion-aware Layer Normalization [40] and cross-attended by current frame features. - "Agent context stream", where agent predictions from the current frame are modified by attending predictions from past frames (stored in a memory bank and properly aligned). This 'streaming-type' method obtains very strong results on the Argoverse 1 and 2 datasets, and is shown to improve two different baseline model architectures based on [5] and [51] that consume the whole historical context in a single step. Some helpful ablations and latency studies of the proposed components is provided. Strengths: - Strong idea to stream scene context into the model, shown to be effective even with short 2-3 'chunks' of context over 2-3 seconds. - Intuitive and well designed scene and agent context stream components that aggregate information across frames (or 'chunks'). The components are designed to model local information in agent-centric coordinates and aggregate it in a geometry-aware manner. - Strong overall results on Argoverse 1 and 2, and when comparing the baseline models the new technique is extending. - Mostly comprehensive related work section. Weaknesses: - In the abstract/intro, I find some of the framing to be a bit misleading. Examples: - "7:This significantly simplifies the forecasting task, making the solutions unrealistic to use in practice". I do not find this claim to be well supported. The current methods still ingest a significant amount of relevant context. To me the contribution of the paper is to follow a streaming paradigm and design more domain-appropriate fusion mechanisms across time than the vanilla MLP or transformer ops that are used traditionally. - "31 - 36: existing methods all tackle an artificially narrowed and unrealistic motion forecasting task".... The methods still ingest 2-3 seconds of context, but perhaps not as effectively as the proposed work. That does not confirm this overly strong claim however. Also the paper only demonstrates good performance with up to 3s of context, that does not prove yet that much longer context is helpful or practical. - To a similar point, Fig 1 is also a bit misleading: current methods would directly ingest the available scene history in the current datasets, and fuse the information, even if the information fusion is not done with the same geometric and streaming intuitions. - In the method description, some details are unclear or could be explained better: - 83: "generate trajectory segments of identical length" Length can be misconstrued to be the length of the trajectories in meters, not that they are the same across chunks. This whole paragraph could be rewritten for clarity and a more helpful diagram could be shown -- illustrating what exactly is done for Argoverse 2 (one now needs to read all the way to Sec 4 to understand how this works). - "114: Additionally, we forecast a singular trajectory for auxiliary training purposes, focusing on the movement patterns of other agents. " Unclear what this means exactly. Do you produce only a single trajectory prediction in past steps? - "“Equal length into past and future” → confusing, assume fixed time interval in both directions. - The L_refine loss in Eq 5 is not explained, unclear what is being refined. My guess is that this is predictions in earlier chunks - What is exactly RealMotion without the context stream components? This is present both in the Ablation and in Appendix B.3. How do you fuse information across time then? - In the experimental results, some SOTA methods from the leaderboard seem to omit SEPT (ICLR 2023), which has higher numbers than RealMotion, even though it is cited. Why did you you omit it? - A couple more ablations or experiments would be helpful. - It seems that the method is similar to a filter, where the current chunk only attends to previous chunk (but not to the ones before). Is my reading of the paper correct? If so, an experiment where the method can attend longer context memory with multiple frames would be interesting. - It would also be interesting to try pretraining this method on more data examples (those can be produced by predicting shorter futures, then you can use more of the segments). And then fine-tune on 6s futures. Technical Quality: 3 Clarity: 2 Questions for Authors: Main question is, do you agree with my reading that "the contribution of the paper is to follow a streaming paradigm and design more domain-appropriate fusion mechanisms across time than the vanilla MLP or transformer ops that are used traditionally. " As opposed to the strong claims in the current abstract / intro. Why did you omit SEPT from the leaderboard results (Table 1)? Please see my clarification or ablation/more experiment comments in the Weaknesses. I am particularly interested to understand why you did not try to attend to longer temporal context in the last chunk (if I read the paper correctly). Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 4 Limitations: The same RealMotion methods tried here could be tried on the Waymo Open Dataset which also has 9 second segments, but one could take the first 3 seconds as history and predict 6 seconds, as opposed to the standard 1s / 8s split that is mentioned in the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed review as well as the suggestions for improvement. Our response to the reviewer’s comments is below: Q1: Some of the framing to be a bit misleading. Apologies for clarity issue in abstract and introduction. We intent to stress the significance of streaming design in real-world settings. We will improve the clarity. As for Fig. 1, we discuss in real-world setting where scenes should be continuous, going beyond the datasets. In this case, current methods can only process independent scenes with limited history as set up in current datasets, while we advocate to utilize more abundant historical information. Q2: Method description. Apologies for unclear expression. + 83: Indeed, the $length$ refers to time steps instead of meters. We will clarify this part. + 114: It is not for past steps but for other agents. In addition to predicting the trajectories of agents of interest for evaluation, we also predict a single trajectory of other surrounding agents (not involved in evaluation) for auxiliary supervision. That is typical and beneficial for learning better motion pattern. + Equal length: We assume this refers to Line 83 -- that means the newly generated sub-trajectories have equal total steps, but the number of steps in the history and future can differ. + refine loss: $L_{\rm refine}$ is utilized to supervise the refined trajectories $Y_{\rm mo}$ mentioned in Eq(4), which can be formulated as $L_{\rm refine} = {\rm smoothL1}(Y_{\rm mo}, Y_{\rm gt})$, where $Y_{\rm gt}$ refers to the ground truth trajectories. + RealMotion: In ablation, RealMotion without stream modules can be considered as a one-shot forecasting models similar to current methods. This variant cannot fuse temporal information, but utilizing our reorganized data as a data augmentation method is still possible. In Appendix B.3, it should refer to only using our reorganized data, which we will clarify. Q3: Comparison with SEPT. We consider that SEPT is a pretrained self-supervised method orthogonal to ours. For example, it utilizes all sets (including test set) to address the disparity between train set and test set, leading to more data used than current methods and making the direct comparison unfair. A property way of using SEPT as model pretraining is likely to improve general methods like ours, and we could include them for further improvements. Q4: More ablations + Actually, the current chunk has attended to several preceding chunks. For the trajectory stream, we explicitly maintain a memory bank for more history predictions from a longer history. For the scene stream, since previous chunk involves earlier context, the current can also utilize a longer context to some extent through implicit propagation. Besides, We have tried to propagate two frames of scene context as shown below: \|\|Latency\|Memory\|$minFDE_6$\|$minADE_6$\| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|2 frames\|23ms\|1.8G\|1.28\|0.64\| \|1 frames\|20ms\|1.4G\|1.31\|0.66\| This strategy can indeed bring slight improvements, but we finally did not adopt more frames due to considerations of simplicity, efficiency and memory overhead, especially in complex scenarios such as intersections. + It is an excellent insight! To make the data more aligned with the forecasting task, we did not process future trajectories (such as Waymo mentioned below) in this paper, but it is indeed worth integrating and exploring distinct motion tasks (simulation, prediction and planning) and datasets. We intend to further extend our model in future work. --- Rebuttal 2: Comment: Authors have answered my technical and clarification questions in a satisfactory manner (Q2 - Q4). Please try to address them in a future manuscript version, some of these details are key to the method (e.g. trajectory stream attending to multiple past frames and L_refine definition etc) but are not sufficiently well described. Wrt my Q1 I am not particularly reassured yet. Please explain what your new pitch is in 2 sentences. Also, while in theory your method can benefit from longer historical information, in practice you have not tried it, so it is not proven. What seems proven to me is that presenting the information in a streaming fashion may be superior. Fig 1 is confusing because it shows that for standard methods, each frame attending only its own timestep independently. This is not correct since it attends X frames of previous data but it is not shown (yet the past frames are shown in yellow for RealMotion). --- Rebuttal 3: Comment: Thanks for the recognition. We cannot provide the evaluation of longer history due to there is no temporal relationships between different scenes in existing benchmarks. Once such longer history data becomes available, it is definitely insightful to test our approach then. That being said, we have now tested the longest possible history with the existing benchmark sampled from the real-world driving data. Regarding Fig.1, please note each blue circle represents a specific scene rather than a single frame, which consists of historical trajectories and HP map within a certain range. In this case, current methods independently process every individual scene, while RealMotion is designed to model the temporal relationships across successive scenes. We will further clarify to minimize similar misunderstanding. --- Rebuttal Comment 3.1: Comment: > We cannot provide the evaluation of longer history due to there is no temporal relationships between different scenes in existing benchmarks. Once such longer history data becomes available, it is definitely insightful to test our approach then. This is understandable. But this is an important point/disclaimer, and it is not actually mentioned in the intro etc. I also had asked: "please explain what your updated intro contribution statement relative to existing methods is, in 2 sentences" to see how it would address my overclaiming concerns. I note that your response does not address this ask. > Regarding Fig.1, please note each blue circle represents a specific scene rather than a single frame. I suggest at the very least you augment the figure caption to clarify that each blue scene circle contains X seconds of past history, or some such. --- Rebuttal 4: Comment: > please explain what your updated intro contribution statement relative to existing methods is, in 2 sentences" to see how it would address my overclaiming concerns. Thanks. Existing methods tackle motion forecasting task in a scene-independent manner, which could limit the model performance in real-world settings where motion is typically continuous while ego-car drives on. Upon this observation, we propose temporal relationship modeling across scenes and introduce an effective approach, RealMotion, to realize this idea in a two stream framework. > I suggest at the very least you augment the figure caption to clarify that each blue scene circle contains X seconds of past history, or some such. Thanks for the suggestion and we will revise as suggested. --- Rebuttal Comment 4.1: Comment: Thank you for all the clarifications. Overall I am leaning to accept, as the ideas and results generally warrant it, and will maintain my score. In doing this, I assume you will temper claims that existing solutions are "unrealistic to use in practice" even if you seem to be able to improve on them, and will explicitly call out the fact that you have not actually validated benefits from being able to process longer histories, per se, which seems left for future work. --- Reply to Comment 4.1.1: Comment: We appreciate the reviewer's time for reviewing and thanks again for the valuable comments. We will revise and refine the paper as suggested in the revision.	Rebuttal 1: Rebuttal: We provide additional figures in the PDF files for more intuitive demonstration. Pdf: /pdf/26653ba2e4ff6a0bb5db293d1817ed78e33b874c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Is Cross-validation the Gold Standard to Estimate Out-of-sample Model Performance?	Accept (poster)	Summary: This is a primarily theoretical paper describing the bias and confidence intervals formed from different methods of assessing the true out-of-sample error of various models. In particular, this paper compares cross-validation (CV) to the "plug-in" estimator (i.e., the training loss). The authors assess the methods based on two criteria: how quickly the converge towards the true out-of-sample error as one gathers more data $n$, and whether or not the confidence intervals (formed as the usual $1-\alpha$ intervals of a CLT approximation to the folds of cross-validation / loss on each training point) achieve the proper coverage of $1-\alpha$. The authors make two main very practical conclusions, regardless of model type (parametric or nonparametric): 1. That there is essentially no point to using $K$-fold CV for some fixed $K$, as it has equal bias to the plug-in estimate and achieves coverage for strictly fewer models than does the plug-in. 2. The utility of leave-one-out CV (LOOCV) is questionable, since although it has somewhat better astymptotic performance than plug-in, the gains may not justify its usually enormous computational cost. Strengths: Contribution: CV is an extremely widely used methodology that had, until somewhat recently, very slim theoretical underpinnings. I think this paper is a great help in ameliorating that problem, and it even makes some surprising and well-backed-up arguments about the use of CV in practice. In particular, I don't think many attendees of NeurIPS would believe the statement "using the training loss is pretty much just as good as using $K$ fold CV"! But, with this paper (hopefully) at the conference, people may start to ask themselves interesting questions about the use of $K$ fold. In terms of a more specific contribution, the authors note that previous theoretical work has been light in many cases where the convergence rate of the model is asymptotically less than $1/n^{1/4}$, whereas their paper details this case (as well as the $>1/n^{1/4}$ case) clearly. I'm not deeply familiar enough with the theoretical CV literature to validate this claim about previous work, but I think this is an important contribution in and of itself if it's correct. Clarity: In general, the paper is pretty well written. It's a fairly technical paper, but the authors do a good job of making it straightforward to read. I even found the first chunk of the proofs pretty easy to read through (although I got tripped up eventually, see weaknesses). Overall, I found all concepts completely clearly and precisely defined, which I think is not true of many theoretical papers submitted to NeurIPS. Weaknesses: I have a few miscellaneous comments. I don't think any of them are particularly important with the exception of the first two. Claimed strength of results The introduction says "[we] identify necessary conditions, in contrast to merely sufficient conditions in the literature, under which the plug-in and CV variants exhibit low biases and valid coverage." I think a paper truly doing this would be incredibly exciting. However, I do not think this paper does this. In particular, there are eight enumerated assumptions that are used throughout the paper (Assumptions 1-4 in particular), and I did not see any analysis that these assumptions are necessary. The paper does completely break down how the bias and coverage depend on the convergence rate of the model at hand. And I think that's great. But I think it means these statements in the introduction and abstract aren't quite correct and should be clarified. Utility of CV versus plug-in (i.e., training loss) The paper kind of implies in its intro / abstract that it is going to show $K$-fold and LOOCV are actually not as good as the plug in estimator. I think Theorems 1 and 2 (the main theoretical results of the paper) show a more nuanced picture than this, though: 1. $K$-fold. Theorem 1 shows that $K$-fold (for $K = O(1)$) has bias that is equal in magnitude to plug-in, and Theorem 2 shows it has equivalent asymptotic coverage. The conclusion from this is very clearly spelled out: "plug-in is always no worse than $K$-fold CV." This might make it sound like $K$-fold is a waste of time to a practitioner, but I would argue it's. In particular, plug-in is biased to be below the true out-of-sample error, whereas $K$-fold is biased to be above the true error. Think of the perspective of a risk-adverse user or a regulator deciding whether to allow a model in a high-stakes situation (self-driving cars, medical procedures, etc.). I really can't see such users ever seeing plug-in as equivalent to $K$-fold; in fact, I would argue that they would hugely prefer to use $K$-fold precisely because its evaluation of the model will be conservative. 2. I thought some of the discussion in the paper was critical of LOOCV's statistical performance, however, LOOCV seems to outperform plug-in in both Theorem 1 and 2. I appreciated the point that one might see the statistical gains as minor in exchange for the computational expense. But I think it could be clearer that the takeaway is "maybe don't bother with LOOCV, especially if you have a parametric model that's nearing it's 'asymptotic regime'". I'm calling this out here because I think this is a really interesting paper that might get highlighted at the conference. Many non-theoretically oriented people attend NeurIPS, and I really wouldn't want them to get the wrong idea based on reading the introduction / abstract. A doctor making the medical device that will diagnose any of our future tumors might be in attendance -- do you really want them to use plug-in over $K$-fold to assess the quality of their algorithms? Small issues in theorems 1. I think it should be made clear when $K$ is a constant, and when it can vary with $n$. I think Theorem2 wants it to be a constant whereas Theorem 4 allows it to scale with $n$. But this should be clear. 2. In Theorem 2, why do we care about the coverage of $c(z^)$? Is this just a fun fact? Some interpretation would be good. 3. Theorem 4 seems to give a much more specific description of LOOCV's bias than does Theorem 1 -- why not state Theorem 4's result in Theorem 1? 4. I was a little confused by Theorem 3/4 being labeled as "Theorems"; they seem like Lemmas to me given that they are building blocks used for Theorems 1/2. 5. In Theorem 2, when $\gamma_v \leq 1/4$, the asymptotic coverage of plug-in is given as $\leq 1-\alpha$. The text then states that equality holds (i.e. coverage is $= 1-\alpha$) when $\gamma_v > 1/4$. But the point is that we're covering $\gamma_v \leq 1/4$, so I think the $\leq$ should be a $<$. 6. I often thought that restating the theorem would be helpful before the proofs in the Appendices; I had to open the paper in multiple windows to flip through everything. 7. There are a lot of random quantities floating around the paper, and it wasn't always clear what expectations were being taken over. It would be good to subscript expectations with what they're with respect to wherever possible. 8. Line 130: I think the upper bound on the Hessian here is from Assumption 2, not Assumption 3. 9. I got a little lost in the proof of Theorem 3 when it suddenly jumped into discussing Assumption 8. First, Assumption 8 isn't an assumption of the Theorem, and Assumption 8 seems to be about nonparametric models, whereas Theorem 3 is generic. I think it could be more clear what's going on here. High dimensions* I think a comparison in high dimensions is a major missing piece to the story here -- at least for parametric models. In particular, the use of cross-validation in the fixed dimension $n \to \infty$ case is a lot less relevant, as in this case it's well known (e.g. from empirical risk minimization theory) that the plug-in estimate will perform well. In high dimensions, say where $n/d \to$ some constant, the plug-in often does not perform well, whereas cross-validation can. I don't think this is a significant direction for change in the paper, but I do think a little discussion (maybe in the future work section) would be good. Technical Quality: 4 Clarity: 3 Questions for Authors: I'm curious how the results hold up in high dimensions, but I don't think this is a critical direction for the paper. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: I think the authors could use a little more discussion of the limitations of their results (see the first section under "Weaknesses"). I don't think there are any societal impacts of this paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for recognizing our contributions both on the theoretical and practical fronts, and also for the detailed and very helpful suggestions. We address your comments point-by-point as follows. Claimed strength of results: We would follow your suggestion and change that claim in the introduction to "This analysis helps us provide a complete breakdown of how bias and coverage depend on the convergence rate of the model at hand. This in turn fills in the gap in understanding which methods outperform which others, in regimes that have appeared challenging for previous works." We will also make similar changes in the abstract. Actually, our original thought was similar to your comment in that we regarded the complete breakdown as a provision of both sufficient and necessary conditions, with the list of assumptions being used to prescribe some regularity in the setting. Nonetheless, we see your very valid point that the use of the term ``necessary conditions" would cause confusion, and so we would change that as you suggested. Utility of CV versus plug-in: We agree with the reviewer on these additional helpful insights, and would add them into our discussion in Section 3. Specifically, - In the "Comparing plug-in and K-fold CV" part, we would add "In terms of the magnitude of bias and interval coverage, plug-in is always no worse than K-fold CV. However, in terms of the direction of bias, plug-in gives an optimistic evaluation of the true model performance, while K-fold CV gives a pessimistic evaluation. In high-stake scenarios where a conservative evaluation is preferred, e.g., evaluating the treatment effect of a new drug, K-fold CV can be more desirable than plug-in. " - In the "Comparing plug-in and LOOCV" part, we would add "In these cases where LOOCV and plug-in give similar evaluation intervals statistically, plug-in should be preferred as it is computationally much more efficient." Small issues in theorems: First, we would incorporate your suggestions in Points 4 (change to lemmas), 6, 7, 8. - For point 1, indeed K can scale with n in the result of Theorem 4, and we would clarify the dependence of K there. - For point 2, the coverage of $c(z^)$ is of interest to some problems in stochastic optimization, namely in estimating the optimality gap to decide whether one should stop an optimization algorithm. We have mentioned some of this literature [10, 42, 45, 53, 54] in Appendix A. We would move this related discussion to the exposition after Theorem 2. - For point 3, indeed Theorem 4 implies Theorem 1. Nonetheless, our Theorem 1 highlights the difference among the three types of estimators with respect to the sample size $n$ and model convergence rate $\gamma_v, \gamma$, which we think is helpful for the reader. We are happy to change and merge Theorems 1 and 4 if the reviewer thinks this would be better. - For point 5, the general condition in Line 155 is under $\gamma \leq 1/4$. However, the condition in line 156 - 157 is given by $\gamma_v$. Note that $\gamma = \min(\gamma_b, \gamma_v)$, so that equality (i.e., the valid coverage guarantee of the plug-in estimator) holds when $\gamma \leq 1/4$ and $\gamma_v > 1/4$. - For point 9, we apologize for the confusion in the proof of Theorem 3. First, in terms of Assumption 8, the models we study in this paper are covered in Definitions 1 and 2, while Assumption 8 governs the regularity of the (nonparametric) models under Definition 2. Next, for parametric models, existing theoretical results ([36, 46]) have shown that the bias is of order $\Theta(1/n)$ (since $\gamma = 1/2$ as we mention in Proposition 1) under our assumptions. We would incorporate the missing discussion on the relation with Assumption 8 and in terms of the parametric model in the proof in our revised version. High-dimensional Investigation: The reviewer certainly raises a good question about high-dimensional investigation. We provide some discussions in the Global Response. Essentially, both plug-in and $K$-fold CV may not perform well in this setting, and while there are works on addressing very specific problems (e.g., linear regression), to our best knowledge it remains an open problem to understand the comparisons among different evaluation methods for general high-dimensional problems. Lastly, regarding whether empirical risk minimization theory can give insights on the performance of plug-in (even in low dimension), our take is that, as this theory is mostly based on non-asymptotic bounds instead of asymptotically exact results (as we mentioned in Section 6), it could be difficult to know whether CV outperforms plug-in or vice versa. Besides, the performance of nonparametric estimators is not directly revealed from the standard empirical risk minimization theory which mostly focuses on parametric models (though with exceptions). Because of these, CV is still more commonly applied than plug-in, even though as our paper suggests plug-in could be superior when taking into account both statistical and computational benefits. --- Rebuttal Comment 1.1: Comment: Summary: after reading the reviews and other responses, I think the paper will be even stronger. I would vote for it being highlighted at the conference (e.g. an oral presentation); I think it has a lot of thought-provoking theoretical and* empirical implications. I think there was a lot of interesting discussion in the above reviews and responses! I especially thought JBLc’s comments on model selection —and the authors’ responses — were interesting, as this is a really straightforward failure case of plug-in that I hadn’t thought of while reading the paper. I think calling this out in the paper adds to the mystery a bit: plug-in is just as biased as K-fold for any fixed model, and yet, somehow, is horrible for model selection. This seems like an interesting direction for future work. I also appreciated the discussion of the high-dimensional work as an interesting direction for more future work. On merging Theorem 1&4 vs keeping them separate: I would just take the above comment as an $n=1$ sized observation that I found it confusing when reading the paper. I don't think it's a huge deal. --- Reply to Comment 1.1.1: Comment: We greatly thank the reviewer again for your constructive feedback and thoughtful reading for reviews. We also sincerely thank you for giving us a high evaluation! Specifically, we would emphasize the model selection and high-dimensional issues as our future work in our revised version.	Summary: The paper establishes asymptotic results for bias and coverage of three different kinds of validation schemes, namely k-fold cross-validation, plug-in validation, and leave-one-out cross-validation. The setup is general and includes both parametric and non-parametric models. There are experimental results that cover parts of the finite sample context. Strengths: - The research area is important and interesting. - The theoretical results are presented clearly and I have not found any errors in the derivations (but have not looked closely at the proofs). - The assumptions are clearly stated and not overly restrictive. Weaknesses: - The paper is very technical in nature and contains an appendix of 20 pages, most of which are also technical. (I count to seven theorems, 14 lemmas, and seven propositions.) It is not reasonable to expect reviewers to be able to cover this much material in the limited time frame of the review process in NeurIPS. I think the paper would be more appropriate for a journal submission. - The code in the supplement is not documented at all and does not contain any configuration or setup for reproducing the results. The project should come with a `pyproject.toml` file, `requirements.txt`, or similar configuration. The zip file also contains numerous files that shouldn't be there, such as cache files. As a result, it is hard to check the validity of the numerical experiments. The experiments also seem to depend on commercial software (MOSEK), which is not ideal. - The empirical results are questionable to me. You say that the typical number of folds in cross-validation is 5, yet the majority of your experiments use 2 folds, which I have never seen used in practice. I agree that the most common values are 5 and 10, and your experiments should reflect this, plus values above and below to show how the effect of $K$. Perhaps 5, 10, and 20. You even say you haven't "cherry-picked" your results, but this is exactly what this looks like to me, especially when you only show 5 fold results in a few of the cases. - The empirical results all rely on data with 10 features, making $n/p \geq 60$ everywhere. This is not really representative of the contexts in which these models are used. You should consider $p > n$ or at least lower ratios than this. [8] show that the coverage of the intervals depend on this ratio. - The results are based on one particular type of interval estimate for $k$-fold cross validation, but it is not clear that this is a good choice. (See Questions.) Technical Quality: 3 Clarity: 3 Questions for Authors: ### Questions - It seems to me that your results hinges on the interval estimates defined via (1) and (2) but, as you yourself say, these are not the only possible estimates. And in fact in [8] it is shown that these estimates in fact have as poor coverage as the naive estimates, which I suspect must influence your results as well. Why did you not use the more accurate estimates presented in [8] and why is there not more discussion of the results in that paper in general? ### Suggestions - Perhaps the title could be more informative. If I understand the paper correctly, you are stating that it is indeed _not_ the gold standard to evaluate model performance. So just say that in the title instead. - L29: It would help to be more descriptive about what the "plug-in" approach is. Many techniques "re-use training data for model evaluation". Cross-validation also fits this description, for instance. - L29: "Reuses" should be "reuse". - L31: "Model rates" could mean several things. Be more descriptive. - L33: Left quotation mark around "center" is not formatted correctly. - L258: Title escapes into margin. - L352: "as" should be "as they are". - Figure 1: Double y-axes are generally hard to read. I suggest you just divide the plot into two instead. - L953: You already have one example 3. Maybe you forgot to use the example environment here? - L973: Same as the previous comment. - Appendix: Equations escape into the right margin in several places. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - I am missing a discussion on optimistic vs. pesimistic bias - The listed limitations should include the limitation of looking at one specific way of constructing the interval estimates. - The results have no societal impacts as far as I can tell. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing the significance of our results and the helpful suggestions. We address the reviewers' main concerns as follows. Our choice of interval estimate, and not use the interval suggested in [8]: The reviewer raises the valid point that our interval is one particular choice and questions why we don't use the interval in [8]. First, we point out that the interval in [8], namely eq. (9) in [8], which uses nested CV (NCV) in both the point and variance estimates in the interval, is designed for (high-dimensional) parametric models. In contrast, our paper focuses on the distinct setting of nonparametric models, especially in the most challenging slow rate regime. That is, NCV is not designed for, and would not be competitive to our setting. In fact, NCV does not help theoretically and empirically: - NCV does not offer valid coverages for nonparametric models theoretically: The interval length of NCV is controlled by $\sqrt{MSE}$. Due to the restriction $\sqrt{MSE} \in [SE, \sqrt{K} SE]$ in Sec 4.3.2 in [8] (SE = $\Theta(n^{-1/2})$ is the standard error in their eq. (2)), this interval length is $\Theta(1/n^{1/2})$ for a fixed $K$. In nonparametric models with $\gamma < 1/4$, this length is overly small compared with the bias of NCV being $C n^{-2\gamma}$ (for a constant $C > 0$ smaller than that of eq. (2) in our paper). This is despite that the CV point estimator in [8] used bias correction in their eq. (9), that correction is for parametric models but not nonparametric models. Then the bias in [8] dominates the interval length and leads to an invalid coverage asymptotically. - NCV does not perform well empirically in our setting: This is expected from the above explanation. We implement NCV with 5 and 10 folds for randomforest in the regression problem (same setting as Figure 1 in our paper), shown in Table 2 of the attached pdf. While the bias of NCV is controllable and the interval gives nearly valid coverage when n is small, the larger bias order starts to exert effect when $n$ is large, leading to a significant drop in the coverage of NCV which becomes invalid. - NCV is computationally very demanding, requiring over 1000 random splits to stabilize the variance estimate (as mentioned in [8]) and needs to refit in total 1000K times. Note that [8] does not focus on computation expense as we do in our paper in comparing LOOCV with plug-in. - [8] does not provide rigorous theoretical guarantees on the valid coverage of their interval (9), even in their own setting. They only correct the variance estimate for parametric models. We will include the above comparison discussion with NCV in [8] in our future work part, and include the additional numerical results in Appendix in our revised version. In our comparison discussion, we would properly indicate that [8] focuses on a different setting and hence is disadvantageous for ours. Finally, back to our interval estimate (2), we stress that our choice is natural -- It is widely used with valid statistical properties and outperformance over earlier works as shown in [9]. [8] also considers this natural choice and shows that it performs well under large samples in their experiment (n>400 in Figure 10 in [8]). Our paper and [8] essentially point out two different problems by using (2): [8] corrects the variance estimate in high dimension but parametric models while we show that this interval does not yield valid coverage for nonparametric models with slow rates. Code Reproducibility: We apologize for not documenting the code well previously. We now provide clean documentation on the experimental configuration for reproducibility and also change the default solver to CVXOPT (open source package). Here is the anonymous link to our revised code: [Link](https://anonymous.4open.science/r/CV_GoldStandard-8E35). Empirical results using different fold numbers: We have shown 5-CV results for the regression problem in Table 5, and newsvendor problem in Table 7 in the paper. To alleviate further the reviewer's concern, we now also provide additional results of 5, 10, 20-CV for regression and portfolio tasks for more sample sizes in Table 1 of the attached pdf. We will include them in our revised version and use 5-CV throughout the main body. As shown in these results, larger K still suffers from coverage problems when n is large, even though it can help when $n$ is small. These continue to support the validity of our theoretical results. Results on different dimensional problems: As other reviewers mention, our main contribution and novelty are for the asymptotic regime with $n\to\infty$ and fixed $p$. This corresponds to most classical CV papers with statistical guarantees [5,6,9]. The models we choose, i.e., linear models, kNN, randomforest are common under this setup. Correspondingly, our empirical results are used to validate theoretical results in this regime. Nonetheless, we provide discussions on the high-dimensional case (including p > n) in the Global Response, but we politely point out that this is not our main focus and is worth a separate substantial work on its own (as Reviewer tGmJ also hints). Other suggestions: We use the question form in the title because the answer is not clear-cut - LOOCV is indeed the best statistically but computationally expensive. We will incorporate other great suggestions from the reviewer in our revised paper. Optimistic versus Pessimistic Bias*: Optimistic bias refers to underestimating the expected cost. Plug-in suffers such bias since it is evaluated on the same training data. In contrast, pessimistic bias refers to overestimating the expected cost. CV suffers such bias since CV is unbiased for the evaluation using fewer training samples than the whole dataset, and appears more erroneous than it should be compared to the true evaluation using the whole dataset. We will incorporate this discussion after stating the current Theorems 1 and 2. --- Rebuttal Comment 1.1: Title: Rebuttal Comment: Thanks! The additional experiments and comments regarding the choice of interval estimate definitely alleviate many of my concerns, and I will raise my score as a result. Apologies for my ignorance regarding your choice of regime. I still, however, contend that the paper is not a good fit for this format and would fit better for a journal submission. --- Reply to Comment 1.1.1: Comment: We greatly thank you for your constructive feedback and thoughtful reply to our response. We also sincerely thank you for increasing the score for our paper! We understand your consideration of the amount and technical nature of the materials, and we are doing our best to make our comprehensive results digestible and the key messages clear in the main body. Most importantly, we believe our results and messages would gain interests and be very useful to the general ML community -- and we are glad to see that you and other reviewers recognize this as well. Lastly, we also politely note that, compared to other published NeurIPS papers on related subjects in the learning theory field, our 20-page appendix perhaps is not that long according to our best understanding.	Summary: This paper considers model evaluation using plug-in method, CV and leave-one-out CV. The main contribution of this paper lies in the asymptotic bias and coverage performance analysis of these methods. The result is that in most cases, it turns out that the plug-in method is no worse than CV or leave-one-out CV. Strengths: The paper is clearly written and the message is clean. This paper provides new insights regarding CV and leave-one-out CV on model evaluation. I think this work contributes to a recent line of work on understanding the role of CV in model evaluation and model selection. Weaknesses: It has been recently revealed in [8] that the CV might be estimating the risk of the algorithm, i.e., E(c(\hat{z})), which probably explains why the performance of CV for estimating c(\hat{z}) is not so good. Even in terms of model evaluation, the recent results in [6,9] show that it might be better to treat CV as estimating the averaged performance of a couple of models, so CV might have a large bias for estimating c(\hat{z}). The same notation of z in l88 and l101 refers to the prediction function and values in the prediction domain, respectively, which is a little bit confusing. Technical Quality: 3 Clarity: 3 Questions for Authors: 1.In the overparameterized-regime, we would expect the predictor to interpolate, and the plug-in estimator would be zero, would we expect CV to perform better? In another way of speaking, does the result in this paper assume the dimension is fixed and the sample size goes to infinity? 2.Could the authors give an example where \alpha_n = o(n^{-1/2}) and \gamma \leq 1/4? 3.What happens for algorithm evaluation, i.e., estimating E(c(\hat{z})) and constructing a confidence interval for that? The algorithm evaluation task is an equally important problem. 4.Could we simply use sample splitting to perform model evaluation? For example, if n is large, then we can use half data for training and half data for evaluation, and I think it would give a pretty accurate evaluation for the model trained with n/2 samples. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This paper reveals that CV may not be a good choice for model evaluation. So a natural question is to ask is the plug-in method the most reasonable one we can use? Maybe the authors could provide some guidance for practitioners. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for recognizing our contributions and clarity, and also for raising the list of very helpful questions. Regarding "Weaknesses": Difference with literature investigating roles of CV: First, the reviewer mentions that [8] says CV estimates the risk of the algorithm which explains the poor performance of CV for estimating $c(\hat z)$. However, [8] is mostly interested in parametric models, where the difference between $c(\hat{z})$ and $E[c(\hat{z})]$ (or $c(z^)$) is generally of order $O_p(1/n)$ and negligible compared with the variability, as we mentioned in Section 1 and 3. Therefore, CV still yields valid coverage guarantees for these parametric models. On the other hand, it is unknown when such bias is significant and leads to invalid coverages in general nonparametric models* which is our main focus and novelty. The reviewer also indicates correctly that recent literature [6,9] say that CV accurately estimates the averaged performance and provides valid coverage guarantees for this quantity. However, like above, it is unclear when the difference between such an averaged performance and true model performance affects the coverage guarantee for general nonparametric models. Notations: We will change the notation $z$ to $\tilde z$ when referring to the prediction domain. Regarding "Questions": Overparametrized-regime: We assume the dimension is fixed and the sample size goes to infinity in this paper. We agree with you that the plug-in estimator would be zero under overparametrized-regime. However, whether CV would perform better in that regime is still an open question (e.g., K-fold CV does not perform well in our additional experiments; please refer to more details on the high-dimensional case in our Global Response). Examples under given conditions: $\alpha_n = o(n^{-1/2})$ is satisfied for general random forest and kNN learners from the algorithmic stability literature [14,15,40] discussed in our paper. Moreover, we provide the exact formula of $\gamma_v, \gamma_b$ ($\gamma = \min(\gamma_v, \gamma_b)$) for each learner in Examples 3 and 4 in Section 3. Specifically, in regression problems, $\gamma \leq 1/4$ is satisfied for general nonparametric models (including kNN learners in Example 3, Forest learner in Example 4, kernel estimators in Example 7) as long as $d_x \geq 2$ from Lemma 3 of Appendix B. Algorithmic evaluation task: We agree that algorithmic evaluation is also an important and related task. Our paper focuses on model evaluation but our analysis can shed light on constructing intervals for algorithmic evaluation, especially when the difference of the two targeted intervals is relatively small. In Corollaries 1 and 2 in Appendices E.1.4 and E.2 of our paper, when $\gamma > 1/4$, both plug-in and CV intervals (1) and (2) (regardless of $K$-fold or LOOCV) provide valid coverages for $E_{D_n}[c(\hat z)]$; Otherwise when $\gamma \leq 1/4$, for LOOCV, since $\sqrt{n}(\hat A_{loocv} - c(\hat z))\overset{p}{\to} N(0, \sigma^2)$ and $c(\hat z) = E_{D_n}[c(\hat z)] + O_p(\sqrt{Var_{D_n}[c(\hat z)]})$, the LOOCV interval (2) is valid for covering $E[c(\hat z)]$ when $Var_{D_n}[c(\hat z)] = o(1/n)$; For plug-in, the condition for the coverage validity for $E[c(\hat z)]$ is equivalent to $Var_{D_n}[c(\hat z)] = o(1/n)$ and $\gamma_v > 1/4$. However, it appears open whether there is a more explicit model rate representation for $Var_{D_n}[c(\hat z)] = o(1/n)$. Sample Splitting: We agree with you that such sample splitting gives an accurate evaluation for the model trained with $n/2$ samples. However, we are interested in the model trained with a full set of samples $\hat z(x) = A(D_n;x)$. The expected costs of these two models differ by $\Theta(n^{-2\gamma})$ due to the performance gap when using $n$ versus $n/2$ samples, which is similar to the evaluation bias using 2-fold CV. When $\gamma < 1/4$, the difference is $\omega(n^{-1/2})$, which is large and exceeds the interval length (of order $\Theta(n^{-1/2})$). That is, the interval used to evaluate the model with $n/2$ samples cannot provide valid coverage guarantees for $c(\hat z)$. Guidance to practitioners: First, in terms of the magnitude of bias, LOOCV is always smaller than plug-in, while plug-in is no larger than K-fold CV. Despite this bias ordering, the adoption of a method over another should also take into account the variability and computational demand, specifically: - For parametric models, and nonparametric models with a fast rate ($\gamma > 1/4$, which includes sieves estimators in our reference [18] when the true function $f(x)$ is 2$d_x$-th continuously differentiable in our Examples in Line 198-202), biases in all three considered procedures, plug-in, LOOCV and K-fold CV, are negligible compared to the variability captured in interval coverage. Correspondingly, all three intervals provide valid statistical coverages. Among them, plug-in is the most computationally efficient and should be preferred. - For nonparametric models with a slow rate but small variability ($\gamma_v > 1/4, \gamma \leq 1/4$), which include kNN with $k_n = \omega(\sqrt n)$ in Example 3 and the forest learner in Example 4 in our paper, the biases in plug-in and LOOCV are negligible but K-fold is not. Correspondingly, both plug-in and LOOCV provide valid coverages but $K$-fold CV does not. Since plug-in is computationally much lighter than LOOCV, it is again preferred. - For nonparametric models with slow rate ($\gamma_v \leq 1/4$), which include kNN with $k_n = \Theta(\sqrt n)$ in Example 3, only LOOCV has a negligible bias and provides valid coverages, and hence should be adopted. The above being said, we caution that, in terms of the direction of bias, plug-in is optimistic while K-fold CV is pessimistic, and so the latter could be preferred if a conservative evaluation is needed to address high-stake scenarios (as Reviewer tGmJ suggests). --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response and I will keep my positive score. --- Reply to Comment 1.1.1: Comment: We greatly thank the reviewer again for your constructive feedback and for giving us a high score.	Summary: The paper argues that Cross-Validation (CV), commonly used to evaluate machine learning models, may not be as statistically beneficial, especially in challenging nonparametric regimes. The paper shows that plug-in is always no worse than K-fold CV for models with any convergence rate. While leave-one-out CV can have a smaller bias than plug-in, this benefit is minimal compared to the variability of the evaluation. The theories are validated by numerical experiments. Strengths: 1. The paper addresses a well-motivated and significant problem to machine learning community: the statistical benefits of cross-validation in model evaluation. 2. The paper is well-written and well-structured, offering comprehensive background information and connecting its conclusions to existing analyses while highlighting its novelty. The framework presented is versatile, applicable to both parametric and nonparametric regimes. 3. The encouraging conclusion that the plug-in method performs no worse than CV is supported by robust theoretical analysis and convincing numerical experiments. Weaknesses: 1. While the paper acknowledges that it is unclear whether the analysis and conclusions can be extended to model selection, I am eager to see more discussion on this topic, possibly including some simple simulation explorations. 2. Although the theories are validated on several synthetic datasets, it would be beneficial to provide experimental results on real-world datasets as well. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. In eq (2), every datapoint $(x_{i}, y_{i})$ is used more than once ($k-1$ times) in k-fold CV, the empirical variance should be scaled of $\frac{1}{n(k-1)}$? 2. While the numerical experiments are well-designed, providing results from experiments on real-world data would be even more convincing. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors clearly state the limitations of the framework, leaving me eager to see the follow-up work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your recognition of the importance and clarity of our paper, and also for your very helpful comments. To respond to your main suggestions, we have run additional experiments on model selection and real-world data set. These experimental results will be incorporated into the Appendix. There we will also provide additional discussion on model selection that we leave for future work, which we reveal below as well. # Model Selection A difference between model selection and the model evaluation task we focus on in this paper is that in the former, we focus on the performance rank between different models instead of the absolute performance value. Intuitively, the accuracy of the performance rank depends on not only the evaluations of the compared models, but also the inter-dependence between these evaluation estimates. This adds complexity to understanding the errors made by each selection approach. There are some discussion for specific problems, such as linear regression, classification and density estimation (see, e.g., Section 6 of our reference [3]). However, the theoretical understanding of model selection remains wide open in general problems. That being said, there are some model selection problems where plug-in easily leads to a naive selection: - Selecting hyperparameters: Consider the best regularized parameter $\alpha$ in the ridge regression. Plug-in always chooses $\alpha = 0$ since it has the smallest training loss. - Selecting the best model class: Consider the regression problem $\hat f \in argmin_{f \in F_i} \sum_i^n (f(X_i) - Y_i)^2$ for the nested classes $F_1 \subset F_2 \subset F_3$, and we want to select the best $F_i$ among the three classes. Plug-in always selects the largest class, $F_3$, since it has the smallest training loss. In these problems, CV can select the regularization parameter or model class different from the naive choice. It hints that CV could be better than plug-in for such problem, but this is theoretically not well-understood. Per the reviewer's suggestions, we include some simulation experiments on model selection using plug-in versus CV as follows, where in each case we report the result averaged over 100 experimental repetitions: ## Case 1: Select the hyperparameter in ridge regression Suppose we have a "misspecified" linear data generating process: $Y = \sum_i^{50}(X^{(i)} + sin(X^{(i)})) + \epsilon$ with $X^{(i)}$ being the $i$-th component of $X$ and $\epsilon \sim U(0, 1)$. We want to find the best hyperparameter $\alpha$ from the regularized linear models $\beta_{\alpha} \in argmin_{\beta} (Y_i - \beta^{\top}X_i)^2 + \alpha \|\|\beta \|\|^2$ that minimizes the expected cost $E[(Y - \beta_{\alpha}^{\top}X)^2]$. We show the model selection results in the following table, where the first column of each procedure represents the best $\alpha$ that the procedure finds and the second column represents the true cost under such $\alpha$: \|n \| Plug-in \| \| 5-CV\| \| LOOCV \| \| \|--\|--\|--\|--\|--\|--\|--\| \|80\|0.0\|105.4\|10.1\|74.1\|12.1\|70.6\| \|100\|0.0\|57.2\|34.3 \|49.9\|24.2 \| 50.8\| \|200\|0.0\|36.8\|44.4\|36.9\| 18.2 \| 36.5\| When $n$ is small, both 5-CV and LOOCV find a better decision than the plug-in. but the accurate evaluation of LOOCV does not necessarily yield a better selection than 5-CV (e.g. $n = 100$). Plug-in always chooses $\alpha = 0$ and its relative performance depends on the power of regularizations, i.e., when $n = 200$, adding $\alpha$ does not improve much and thus the performance of plug-in to both 5-CV and LOOCV. However, when $n = 80, 100$, plug-in suffers from worse performance than 5-CV and LOOCV. ## Case 2: Select the best model class Given each $D_n$, we want to select the best among ridge, kNN, and random forest models. Continue the same data process as Case 1, but with 10 features. In each experimental repetition, we select $\lambda \sim U(0, 5)$ randomly and generate data from $Y\|X: \sum_i^{10} X^{(i)} + \lambda sin(X^{(i)}) + \epsilon$. We show the results of model selection procedures, where the first column of each procedure represents the probability of finding the correct best model, and the second column represents the true cost under the model selected by each procedure: \|n \| Plug-in \| \| 5-CV\| \| LOOCV \| \| \|--\|--\|--\|--\|--\|--\|--\| \|100\|0.52\|46.93\|0.88\|42.07\|0.90 \|41.64\| \|200\|0.57\|42.74\|0.88\|38.19\|0.89\|37.86\| \|400\|0.71\|39.94\|0.96\|37.26\|0.95\|37.26\| Here, both 5-CV and LOOCV help find a better decision than plug-in and there is not much difference between 5-CV and LOOCV. # Real-world Experiments We include one real-world dataset [puma32H](https://www.openml.org/d/1210) with 33 features and 1000000 samples as a regression task. The codes and instructions for reproducing are in the updated codebase link in the Global Response. We report the coverage probability of plug-in, 2-CV, and 5-CV for a kNN model with $k_n = n^{2/3}$, where each entry denotes the coverage probability estimated over 100 experimental repetitions for that procedure. Each experimental repetition is conducted differently by shuffling the entire dataset. We then select the first n rows as the training sample for each model and approximate the true model performance by averaging over the remaining samples in the dataset. \|n \| 10000\| 20000 \| 40000\| \|--\|--\|--\|--\| \|Plug-in \| 0.94\| 0.93 \|0.90 \| \|2-CV\|0.88\|0.81 \|0.76 \| \|5-CV\| 0.91\| 0.86\|0.81\| Plug-in provides valid coverages in this case while 2-CV, 5-CV do not, especially when $n$ is larger (20000 and 40000). These continue to validate our asymptotic theory. Scaling of the Variance: In equation (2), each datapoint $(x_i, y_i)$ is only used once for evaluation since the index of each datapoint is only counted once in one of $\{N_k, k \in [K]\}$, the collection of $K$ partitions of $[n]$. Note that this formula is the same as the standard all-pairs variance estimator in Theorem 5 in the reference [9]. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed discussions and all the experiments. The experiment results look good to me, so I have raised the score. --- Reply to Comment 1.1.1: Comment: We greatly thank the reviewer for your constructive feedback earlier, and for reading our reply closely and increasing your score!	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for their helpful suggestions. In this Global Response, we provide additional discussion on several aspects that address the reviewers' comments: practical guidance from our results, differences with existing literature regarding roles of CV, high-dimensional problems, and additional experimental results. # Guidance to Practitioners First, in terms of the magnitude of bias, LOOCV is always smaller than plug-in, while plug-in is no larger than K-fold CV. Despite this bias ordering, the adoption of a method over another should also take into account the variability and computational demand, specifically: - For parametric models, and nonparametric models with a fast rate ($\gamma > 1/4$, including sieves estimators in our reference [18] when the true function $f(x)$ is 2$d_x$-th continuously differentiable in our Examples in Line 198-202), biases in all three considered procedures, plug-in, LOOCV and K-fold CV, are negligible compared to the variability captured in interval coverage. Correspondingly, all three intervals provide valid statistical coverages. Among them, plug-in is the most computationally efficient and should be preferred. - For nonparametric models with a slow rate but small variability ($\gamma_v > 1/4, \gamma \leq 1/4$), which include kNN with $k_n = \omega(\sqrt n)$ in Example 3 and the forest learner in Example 4 in our paper, the biases in plug-in and LOOCV are negligible but K-fold is not. Correspondingly, both plug-in and LOOCV provide valid coverages but $K$-fold CV does not. Since plug-in is computationally much lighter than LOOCV, it is again preferred. - For nonparametric models with slow rate ($\gamma_v \leq 1/4$), which include kNN with $k_n = \Theta(\sqrt n)$ in Example 3, only LOOCV has a negligible bias and provides valid coverages, and hence should be adopted. The above being said, we caution that, in terms of the direction of bias, plug-in is optimistic while K-fold CV is pessimistic, and so the latter could be preferred if a conservative evaluation is needed to address high-stake scenarios. # Comparison with Existing Literature Investigating Roles of CV Our paper considers a distinct setting from [8] in that we consider general nonparametric models instead of (high-dimensional) parametric models, and provide corresponding theoretical guarantees on bias and coverage on competing procedures. To this end, it has been unknown when the biases of these procedures are significant and lead to invalid coverages in general nonparametric models which is our main focus and novelty. Additional discussions are provided in response to Reviewer PDH8. To address the variance estimate problem for high-dimensional parametric models, [8] proposes nested CV (NCV) in both the point and variance estimates in their interval. However, their remedy is not designed for, and also is not empirically competitive, on the nonparametric settings that we consider. Additional discussions are provided in response to Reviewer 8MSs. # Additional Experiments In our attached pdf, we include additional numerical results using different $K$ in CV in Table 1, which continues to provide the insights that K-fold CV suffers from poor coverage when n is large; We also include numerical comparisons between naive CV (in equation (2) of our paper) and nested CV (in our reference [8]) in Table 2; and some high-dimensional simulations in Table 3. Here is the anonymous link to our revised code: [Link](https://anonymous.4open.science/r/CV_GoldStandard-8E35) # High Dimensional Discussion For the high-dimensional setup, i.e., feature dimension $p$ and sample size $n$ both go to infinity such that $p/n$ converges to a nonzero constant, besides specific problems like (generalized) linear regression, the problem is wide open to our best knowledge. For linear models, it is known regarding the bias of our considered three procedures that: - For plug-in, when $p > n$, the predictor interpolates the training data so that the training loss becomes zero (unless we add regularization); when $p/n \to c \in (0, 1)$, plug-in still suffers a large bias from overfitting and cannot be used directly to construct the point estimator and confidence interval to evaluate model performance. Some bias correction procedures ([26]) are proposed to construct consistent point estimators in these high-dimensional scenarios. - For $K$-fold CV, the point estimate also suffers from a non-vanishing bias [W2018] and may not be a good choice for evaluating model performance. - For LOOCV, recent literature [W2018,P2021] shows that its bias goes to zero. Despite the known results above for the bias, no theoretically valid coverage guarantees exist for high-dimensional linear models in the literature, since almost all existing CV literature is based on some stability conditions and is only valid under low-dimensional asymptotic regimes ([6,8,37] and ours). For our three procedures, we validate claims for the point estimates above and also investigate the coverage performance of interval estimates (open question), using the same simulation data setup for ridge regression with $\alpha = 1$ as in our regression problems. We show the simulation results in Table 3 of the attached pdf, where both plug-in and 5-CV suffer from large bias and have zero coverage. The bias of LOOCV is small but the LOOCV interval suffers from poor coverage. These results indicate that constructing theoretically valid intervals remains open and challenging for high-dimensional problems, and is important to devise improved CV or bias correction approaches. [W2018] Wang et al. Approximate LOO for fast parameter tuning in high dimensions. ICML 2018. [P2021] Patil et al. Uniform consistency of cross-validation estimators for high-dimensional ridge regression. AISTATS 2021. Pdf: /pdf/a0108d06fbe3d6568b1555ab6c2661fd2f2f90a7.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Goal Conditioned Reinforcement Learning for Photo Finishing Tuning	Accept (poster)	Summary: This paper proposes a goal-conditioned reinforcement learning framework for photo finishing tuning. They introduce a novel state representation and treat the image processing pipeline as a black box, avoiding the need for differentiable proxies. The method can efficiently tune parameters to match various goals, including pixel-aligned target images and style images. Strengths: 1. The paper is well-written and easy to follow, clearly explaining the goal-conditioned reinforcement learning approach for photo finishing tuning. 2. The visualization in Figure 1 effectively demonstrates the superiority of the proposed method compared to existing approaches, showing rapid convergence and high-quality results. Weaknesses: 1. Compared to search-based methods, this proposed RL-based framework may not generalize well to unseen datasets. 2. The efficiency comparison in Table 2 may not be entirely fair. While the CMAES method runs on CPUs as required, the paper doesn't explore potential speed-ups through multi-processing. Technical Quality: 3 Clarity: 3 Questions for Authors: Is the search-based method (CMAES) run on a single CPU core to obtain the inference time? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The paper's proposed method may require training a new model for each new incoming dataset. This limitation could potentially impact the framework's adaptability and efficiency when applied to diverse or frequently changing data sources. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # To Reviewer WirW We sincerely thank you for reviewing our work. After reading your comments carefully, we summarize the following questions. ## Q1: Cross-Dataset Generalization * Please refer to Q1 in the To All section for a detailed answer to this question. To summarize, we conducted additional evaluations using the HDR+ dataset to demonstrate our RL-based framework's ability to generalize effectively to unseen datasets. Our method achieved a PSNR of 31.54 on the HDR+ photo-finishing tuning task, which is comparable to the 32.47 achieved on the FiveK dataset, and it outperforms all baselines. This highlights our method's strong generalization capabilities across different datasets, outperforming other methods such as CMAES, Greedy Search, Cascaded Proxy, and Monolithic Proxy. Our approach leverages a robust state representation that captures invariant features for photo finishing, enabling it to adapt to diverse inputs and target outputs beyond the training distributions. These results affirm that our RL-based framework is not only effective in the scenarios it was trained on but also exhibits superior performance on datasets it has not previously encountered. ## Q2: Efficiency Comparison of CMAEAS * Our implementation of CMAES takes full advantage of parallel computing capabilities on a high-performance 48-core CPU, ensuring a fair efficiency comparison for efficiency. Parallelization Enabled by `pymoo` Library: For the CMAES method, we implemented the baseline using the widely-used Python library `pymoo`, which supports parallelization. In our experiment, we enable the parallelization options provided by `pymoo`. This is done by setting ```classpymoo.algorithms.soo.nonconvex.cmaes.CMAES(... , parallelize=True, ...)``` and enabling `elementwise_renner` when calling the `pymoo` library. Hardware Configuration: Our speed testing experiments were conducted on a high-performance server-level system equipped with AMD EPYC 7402 (48 Cores) @ 2.8 GHz CPU, 8 NVIDIA RTX 4090 GPUs, 512 GB of RAM, and running CentOS 7.9. With this configuration, we utilized the full capabilities of the 48-core CPU to explore potential speed-ups for CMAES through multi-processing. This setup ensures that the CMAES method was not constrained by computational resources and accurately reflects the method's performance potential. Multi-Core Execution: To clarify, the reported inference times for the CMAES method were obtained by running the algorithm on all available 48 CPU cores, rather than a single core. By doing so, we provided a realistic comparison with other methods. --- Rebuttal 2: Comment: Dear Reviewer WirW, In our rebuttal, we test our model on an unseen dataset, HDR+. Results demonstrate that our method generalizes well to unseen datasets, outperforming all baselines by a large margin. We also explained that the CMAES baseline is implemented with a multi-core CPU with parallel acceleration. We want to follow up to see if our responses address your concerns. We would be very grateful to hear additional feedback from you and will provide further clarification if needed. Thank you again for your time and effort. --- Rebuttal Comment 2.1: Comment: Thank you for your response. It has successfully addressed all of my minor concerns. Based on this, I am willing to upgrade my score.	Summary: This paper presents a method by which RL is used to drive the optimization of ISP hyperparameters for two tasks: (1) recovering photo finishing parameters and (2) mimicking reference style photo characteristics. Strengths: Clearly written. Effective application of RL to photo finishing and stylization. Quantitative evaluation showing substantially improved results on photo finishing in terms of quality, number of queries, and run time. User study on photo stylization task. Ablation study on state representations. Weaknesses: I'm a little skeptical of whether the baselines for [26] and [27] were implemented correctly. What I don't really understand is that [26] does not provide code, [27] does provide code, but there's no mention as to whether they used the code from [27] and why the errors in the two methods look similar (almost all examples have a hue shift in a seemingly weird direction). I also don't see the terms "monolithic proxy" or "cascaded proxy" in either of these two references. I think these details at least need clarification for reproducibility. I don't really understand how CMAES was implemented. The method is cited as being from [9][18], but then the appendix states that it was implemented from [26] using pymoo. Some examples like the green image in Figure 4 make me skeptical that this baseline was implemented correctly. Technical Quality: 3 Clarity: 3 Questions for Authors: Can the authors provide full details on how the baselines were implemented? For me, acceptance is hinges on whether I can be convinced that the baselines were implemented correctly and are indeed that bad. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: seems fine Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # To Reviewer kK87 Q1 addresses the implementation details of three baselines. Q2 Q3 explain certain behavior (hue shift, green image) of baselines. Q4 extends extra details on baseline implementation and proves our baselines were implemented correctly. ## Q1: Full Details on Baseline Implementations, Terms "Cascaded & Monolithic" * Please refer to Q3 in the To All section for how we implement each baseline, including code used and full details. Please also refer to Q2 in the To All section for clarification of the terms "cascaded & monolithic proxy". Reviewer pointed out an incorrect citation of CMAES in the appendix. On line 487, we make a typo on the citation of CMAES; the correct citation should be [18], not [26]. We apologize for this misunderstanding and will correct it in our paper. Since [18] does not provide code, we reproduce it using `pymoo`. ## Q2: Similar Hue Shift for Proxy-Based Methods [26,27] * Hue shift is a common artifact. As shown in Figure 2 of the PDF file, snapshots from the result figures in the [27] paper also exhibit severe hue shift. In our paper, [26] and [27] have similar hue shifts in some cases, and we believe it was a result of inherent challenges in proxy-based methods, and is due to similar errors in proxy networks. The similar inaccuracies of proxy networks are due to the similar data bias introduced during proxy training data generation. [27] utilize a uniform strategy to generate proxy training data: >"We uniformly sample 100 points for each slider when generating the data for each intermediate." --Section 5,[27] This uniform sampling strategy introduces bias because sub-ranges of ISP parameters may not distribute evenly and often have non-linear effects. And ISP operations are black-box so the parameter range cannot be considered during data sampling. For example, the white balance parameter (RGB color channel coefficients) ranges from [0.6, 1.65], where 1 represents no gain. Uniform sampling results in more data points between [1, 1.65] than [0.6, 1], potentially leading to biased training data. Since the proxy network is purely data-driven, such data bias can lead to similar inaccuracies. As a result, both proxies tend to behave similarly, especially in complex ISP operations like color adjustments. Then, both methods use the same regression optimization approach with the Adam optimizer and identical hyperparameters, leading to similar optimization dynamics. It results in similar parameter shifts, particularly in white balance adjustments, where even slight inaccuracies can cause visible hue shifts. Importantly, these shifts do not occur in all cases, as shown in Figure 3 of our paper and Figure 1 of the rebuttal PDF. In conclusion, similar hue shifts are expected due to shared conditions and challenges in proxy-based ISP tuning. ## Q3: Green Image in CMAES Figure 4 * The color shift in Figure 4 is an occasional failure case specific to the stylization task, not the photo finishing tuning task. It occurs because style loss is too complex for the CMAES to explore the search space, causing the optimization stuck in local minimal. In contrast, our RL-based method has superior exploration capabilities under such cases. ## Q4: Extra Details on Each Baseline Implementation * We extend extra details on the performance and implementation of each baseline, to prove our baselines' correctness. ### Cascaded Proxy [27] Baseline [27] underperforms due to (1) our pipeline having more image processing operations, (2) the neural proxy cannot generalize well to unseen images, and (3) the unavailability of the dataset used in [27]. Accuracy Drop with More ISP Hyperparams: As shown in the table below, proxy approximation accuracy and finishing quality of [27] decreases with more ISP operations. This is due to error accumulation in the proxy pipeline with more ISP operations, which leads to inaccurate gradient and suboptimal parameter recovery, causing the finishing quality to decrease. \|Number of ISP Params\|1\|3\|5\|7\|9\| \|-\|-\|-\|-\|-\|-\| \|Proxy Approximation Accuracy of [27] (PSNR)\|51.96\|43.56\|39.07\|35.66\|28.10\| \|Photo Finishing Tuning Quality of [27] (PSNR)\|50.80\|36.54\|29.0\|26.33\|22.31\| Since [27] recovers only 4 parameters while our experiment recovers 9, the finishing quality of is naturally lower in our paper. Additionally, ISP operations like sharpening are difficult for neural networks to approximate, as mentioned in [27], which further impacts performance. Neural Proxy's Generalization: The neural proxy learned in [27] struggles to generalize to unseen images, as shown in proxy accuracy drops on validation data: \| \|Train\|Val\| \|-\|-\|-\| \|Proxy Approximation Accuracy (PSNR)\|34.78\|28.10\| This explains the poorer performance in our additional HDR+ experiments. Dataset used in [27] not released. Therefore, the numbers reported by [27] are not reproducible. We trained this baseline on the FiveK dataset, following the details in [27]. ### Monolithic Proxy [26] Our implementation reported a PSNR result above 20 on the FiveK dataset, comparable to the result reported by [27] when they reproduced this baseline. Like [27], [26] struggles with complex ISP operations and generalization. The vanishing gradient issue, analyzed in [27], further affects its performance. ### CMAES [18] Reported PSNR Comparable to [27] Paper: Our paper shows a PSNR above 28 for CMAES [18] on ISP tuning, comparable to the result reported by [27]. CMAES Performance with More Iterations: In the table below, the performance of CMAES improves with more iterations, eventually reaching a PSNR of 32.12 on the FiveK dataset, comparable to RL's quality. While this indicates that our reproduction of the CMAES baseline is correct, the increased iterations lead to slower performance due to more queries to the ISP. \|CMAES Iterations\|10\|100\|200\|500\| \|-\|-\|-\|-\|-\| \|PSNR\|18.1\|22.6\|28.3\|32.1\| --- Rebuttal 2: Comment: Dear Reviewer kK87, In our rebuttal, we address the implementation details of three baselines, explain the code used, and clarify terms in the paper, along with citations of CMAES (`Q1`). We also point out that certain behaviors of the baselines (such as the similar hue shift and green image in Figure 4) are due to the inherent flaws of the baselines (`Q2` `Q3`). Additionally, we provide more detailed experiments of the baselines to prove the correctness of our reproduction (`Q4`). We want to follow up to see if our responses address your concerns. We would be very grateful to hear additional feedback from you and will provide further clarification if needed. Thank you again for your time and effort. --- Rebuttal Comment 2.1: Comment: The rebuttal and implementation details are sufficient. I will raise my score to borderline accept.	Summary: Proposed Goal-Conditioned Reinforcement Learning for Photo Finishing Tuning. Specifically, the authors introduce a novel goal-conditioned reinforcement learning framework for parameter tuning in photo processing pipelines. Unlike existing methods, the proposed approach operates without relying on proxies and treats the pipeline as a black box. By leveraging a trained RL policy, it efficiently identifies optimal parameters in just 10 queries, contrasting with 500 queries typically required by zeroth-order methods. The proposed framework utilizes a goal image to guide iterative parameter tuning, allowing adaptation to diverse target images and styles. Experiments on photo finishing and stylization tasks validate the effectiveness and versatility of the proposed approach. Strengths: The paper is well written. It is easy to follow. The proposed method is well-motivated. The proposed method seems intuitive and effective for photo-finishing tuning tasks. Empirical evaluations on image-based datasets show the efficacy of the proposed method. Weaknesses: The motivations behind some algorithmic choices are not clear. Additional ablation studies would enhance the persuasiveness of the findings. It would be interesting to see the performance of the proposed method on other datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) Why TD3? why not any other off-policy RL algorithm? Are there any specific reasons behind this choice? (2) Enhancing the paper with more ablation studies would elevate its quality. For instance, exploring the impact of RL versus a greedy algorithm on performance would provide valuable insights. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # To reviewer syDC We sincerely thank you for reviewing our work. After reading your comments carefully, we summarize the following questions. ## Q1: Choice of RL Algorithm * In our task, the choice of RL algorithm is not the primary factor driving performance; instead, our proposed state representation plays the most crucial role. We selected the TD3 algorithm due to its common usage and robustness to hyperparameters, but alternative algorithms like SAC or PPO would yield similar results. To support this, we performed an additional ablation study comparing different RL algorithms. We trained our photo-finishing policy using SAC and PPO algorithms, keeping the state representation, policy network, reward design, and termination conditions identical to our original TD3 implementation. All implementations were based on the stable-baselines3 library in PyTorch. We evaluated each algorithm's performance on the photo-finishing tuning task using the FiveK-Random dataset. Table 1: RL Algorithm Comparison \| \| TD3 \| SAC \| PPO \| TD3 w/o state representation \| \| ---- \| ----- \| ----- \| ----- \| ---------------------------- \| \| PSNR \| 38.26 \| 37.10 \| 36.73 \| 32.17 \| As shown above, different RL algorithms achieve comparable results on this task. However, removing the proposed state representation from the TD3 algorithm results in a significant performance drop, highlighting its importance. The results demonstrate that the state representation, not the specific choice of the RL algorithm, is the key factor behind our method's superior performance. ## Q2: Additional Ablation over Greedy Algorithm * In our paper, we have shown that our RL-based method achieves higher photo-finishing quality and efficiency than optimization-based methods. We compared our method with baselines using zeroth-order search and first-order optimization with differential proxies but did not previously compare it with a simple greedy algorithm. Here, we provide an additional ablation study using a greedy algorithm. For the greedy algorithm, we followed Algorithm 1 from the [CRISP] paper, implementing a greedy search algorithm to find the optimal ISP parameters. This greedy algorithm mimics human behavior by incrementally adjusting each element towards the desired photo-finishing result until a stopping condition is met. > [CRISP] Learning Controllable ISP for Image Enhancement. TIP. 2023. As shown in Algorithm 1 in the Pseudo Code for Greedy Algorithm section below, the greedy algorithm mimics user behavior by incrementally improving image quality to achieve the desired results. The algorithm starts with the ISP parameters at $s_{init}$ and iteratively adjusts them with a step size $t$ until the stopping condition $K$ is reached. We set each element in $s_{init}$ to be 0, step size $t=0.1$, and iterations $K=200$ to be consistent with other optimization-based methods. We evaluated the greedy algorithm on the FiveK-Target and HDR+ datasets, with the results shown in Table 1 of the rebuttal PDF. We also present the results in the table below. \| Greedy Tuning Algorithm Evaluation \| PSNR \| SSIM \| LPIPS \| Queries \| \| ---------------------------------- \| ----- \| ------ \| ------ \| ------- \| \| FiveK Target Photo Finishing Task \| 26.14 \| 0.9250 \| 0.1380 \| 200 \| \| HDR+ Target Photo Finishing Task \| 25.79 \| 0.9212 \| 0.1542 \| 200 \| From the results, it can be observed that while the intuitive greedy algorithm achieves reasonable performance, it performs worse than the CMAES baseline under the same iterations. This indicates that the greedy strategy is not as effective as the evolutionary strategy in the CMAES algorithm. However, the greedy algorithm performs better than proxy-based methods [26, 27] because it requires no proxy training and searches directly in the parameter space. Note that like CMAES, the greedy tuning algorithm requires at least hundreds of queries to the ISP pipeline to converge, which are very time-consuming. Both greedy and CMAES are outperformed by our RL-based approach. This is because their searching processes are blind and brute-force, not conditioned on the input image, target image, or any prior on image processing operations. Our RL-based approach considers all these factors, requires no proxy training, and thus achieves better photo-finishing quality and efficiency. ## Q3: More Evaluation Dataset * Please refer to Q1 in the To All section for the answer to this question. To summarize, we conducted further evaluations using the HDR+ dataset, demonstrating our RL-based framework's ability to generalize effectively to unseen datasets. Our method achieved a PSNR of 31.54 on the HDR+ photo-finishing tuning task, comparable to the 32.47 achieved on the FiveK dataset, and outperformed all baselines. This highlights our method's generalization capability across different datasets and its superior performance compared to other baselines, including CMAES, Greedy Search, Cascaded Proxy, and Monolithic Proxy. ## Pseudo Code for Greedy Algorithm * $\textbf{Algorithm 1: Greedy Tuning Algorithm}$ * $\textbf{Inputs: } s_{\text{init}} \in \mathbb{R}^{D}, \text{ Step size } t, \text{ Stop condition } K, \text{ Input image } x, \text{ Tuning target image } \hat{x}$ $\textbf{Initialization: } e \gets \infty, s \gets s_{\text{init}}, d \gets 1, i \gets 0, k \gets 0$ $\textbf{While } k \leq K \textbf{ do}$ $\quad \textbf{if } e < \text{MSE}(f_{\text{PIPE}}(x, s), \hat{x}) \textbf{ then}$ $\quad \quad s_d \gets s_d - t, d \gets (d \bmod D) + 1, i \gets i + 1, k \gets k + 1$ $\quad \quad \textbf{if } i = D \textbf{ then}$ $\quad \quad \quad s \gets s + t, d \gets 1, i \gets 0$ $\quad \quad \textbf{end if}$ $\quad \textbf{else}$ $\quad \quad e \gets \text{MSE}(f_{\text{PIPE}}(x, s), \hat{x}), i \gets 0, k \gets 0, s_d \gets s_d + t$ $\quad \textbf{end if}$ $\textbf{end while}$ * --- Rebuttal Comment 1.1: Comment: I appreciate the authors' response and the additional results provided in the rebuttal. However, I remain unconvinced by the results presented in Table 1, which actually raise more concerns. 1. While I agree with the authors that state representation is crucial, the results showing variations with different RL algorithms suggest that the choice of RL algorithm is also important. It is unclear why TD3 performs better than the others. A stronger justification for the superiority of TD3 would be valuable.  2. In RL experiments, it's essential to evaluate each method across multiple seeds since results can vary significantly. The table does not provide enough information to determine the best method conclusively. Reporting average results across multiple random seeds would be beneficial.  3. The second table suggests that a naive greedy tuning algorithm performs well, though not as well as the RL-based method. It would be interesting to explore the possibility of developing a greedy tuning algorithm conditioned on the input image, target image, or prior image processing operations as a baseline. 4. Although the authors have added more ablation studies in the rebuttal, they remain insufficient. 5. While the authors have empirically demonstrated the method's generalization across different datasets, a detailed discussion on why the proposed method has superior generalization capability is necessary. In summary, the problem is intriguing, and the proposed method appears effective, but it requires more in-depth analysis, additional ablation studies, and comparisons against stronger baselines to validate its effectiveness. Consequently, I have updated my score to borderline reject. --- Rebuttal 2: Comment: Dear Reviewer syDC, In our rebuttal, we explain the reason for choosing TD3 and include further ablation studies to show that the choice of RL algorithm does not significantly impact performance. We also conduct more ablation over the greedy algorithm as suggested by the reviewer. Additionally, we test our method on the HDR+ dataset. Results demonstrate that our method generalizes well to unseen datasets. We want to follow up to see if our responses address your concerns. We would be very grateful to hear additional feedback from you and will provide further clarification if needed. Thank you again for your time and effort. --- Rebuttal 3: Comment: Dear Reviewer syDC, Thank you for your reply. We would like to further discuss the questions with you below: ### Q1. why TD3 performs better In our task, the RL algorithm is only a tool for optimizing the policy network. We analyze why TD3 performs better as follows: TD3 is an off-policy RL algorithm with a deterministic policy. The deterministic policy allows it to explore the entire action space, including actions near the boundary, which is crucial for recovering boundary ISP parameters in our photo finishing task. TD3 also employs tricks like target policy smoothing and delayed policy updates to enhance robustness. In our task, actions like significantly brightening an image can cause large changes in the reward function, so these robustness tricks are crucial to the stability of the RL algorithm. In contrast, SAC uses a stochastic policy, sampling actions from a Gaussian distribution, which reduces the likelihood of exploring boundary actions, impacting performance. PPO also faces this issue with its stochastic policy. Both SAC and PPO also lack robustness tricks like target policy smoothing. Moreover, PPO is an on-policy algorithm, making it less sample-efficient in our complex photo finishing environment, potentially leading to suboptimal results. \| \| deterministic policy \| robustness tricks \| off-policy \| \| ---- \| -------------------- \| ----------------- \| ---------- \| \| TD3 \| ✔ \| ✔ \| ✔ \| \| SAC \| ✗ \| ✗ \| ✔ \| \| PPO \| ✗ \| ✗ \| ✗ \| These differences explain the variation in performance, but the effect of the RL algorithm is still minor compared to state representation. As shown in the first table of the rebuttal, TD3, SAC, and PPO all yield a PSNR of around 37 with state representation, whereas TD3 without state representation achieves only 32.17. ### Q2. multiple seeds We set the random seed to 1 in all our experiments in the paper and the rebuttal, ensuring fair comparison. In our main comparisons with all baselines, our method outperforms them by a large margin (over 4 dB in PSNR). Even in the ablation study over different RL algorithms, TD3 still outperforms SAC by 1.16 dB in PSNR. As discussed earlier, the RL algorithm is just a tool within our framework, and we are confident that we have selected the RL algorithm that best fits our task. Additionally, we plan to include the average scores over multiple seeds in the camera-ready version of this paper. ### Q3. greedy tuning performs well & stronger baseline with conditioning There are no stronger baselines available. We followed the setup from [27] for baselines in the photo finishing task. We compared our method to existing work in the field of photo finishing tuning, including CMAES [18] and proxy-based methods [26, 27]. Currently, there are no stronger baselines for the photo finishing tuning task. Search-based methods (Greedy & CMAES) perform well because they optimize over more iterations. Both the greedy tuning algorithm and CMAES are search-based methods. The search-based methods, including the greedy algorithm added in the rebuttal, only performs well when using 200 searching iterations, while our method can achieve better results with only 10 iterations. The main advantage of our RL method is its fast convergence speed, as shown in Table 2 in our paper. Conditioning is not possible for greedy algorithm and is part of our contribution. Traditional search-based methods, including greedy algorithms, are driven purely by the loss function's output relative to the parameters being adjusted. These methods lack the ability to condition on complex features like the characteristics of the input image, the target image, or prior processing operations. Such conditioning requires an understanding capability and the ability to generalize from past experiences, which is beyond the capabilities of traditional search-based algorithms. Developing a novel search algorithm conditioned on these complex characteristics is one of our contributions. Our proposed state representation encodes information about the input image, target image, and prior image processing operations. The policy network can then leverage this information to make decisions that are conditioned on these inputs, which leads to more efficient and effective optimization. ### Q4. why the proposed method generalizes well We have analyzed why our method generalizes better in Q1 (cross-dataset generalization) of the To All section in our rebuttal.	Summary: This paper applies goal-conditioned reinforcement learning (RL) to photo finishing tuning. With only 10 queries, it demonstrates that goal-conditioned RL can achieve better performance than zeroth-order approaches that require 500 queries. Additionally, this method is non-differentiable, which might make it more suitable for commercial scenarios. Experiments show that this work can achieve better performance than previous efforts. Strengths: + The first great application to adapt goal-conditioned RL to the photo finishing tuning task. + A novel design for the photo finishing state representation. + Achieved good performance on the evaluation benchmark. Weaknesses: - Some details are missing: The methods used in this work were trained on the training set, but it is unknown whether the compared methods also underwent the training stage. - More evaluation on additional datasets is preferred. Technical Quality: 2 Clarity: 3 Questions for Authors: Please refer to the weakness section. Confidence: 2 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Please refer to the weakness section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # To Reviewer ASBs We sincerely thank you for reviewing our work. After reading your comments carefully, we summarize the following questions. ## Q1: Details on Baselines * In our main paper, we compare three baselines: (1) CMAES: a zeroth order optimization method, that does not need training, (2) Monolithic Proxy: first-order optimization through a single monolithic neural network proxy, that requires training a proxy network, and (3) Cascade proxy: first-order optimization through multiple cascaded neural network proxies, also requires training proxy networks. Each of the baseline's implementation details is available in Q3 of To All section. Here, We briefly summarize the principal of each of these baselines: 1. CMAES [18]: CMAES is a gradient-free search (zeroth order optimization) method using an evolution strategy. For our photo finishing tuning task, it directly optimizes the parameters of the image processing pipeline to match the target rendering style. Therefore, it does not undergo a training stage and optimizes directly on the validation set for multiple iterations instead. Such method faces no generalization challenges but converges very slowly as the searching processing is blind and brute-force. It also lacks exploration capability and is prone to stuck in local minimal with complex objectives. 2. Monolithic Proxy [26]: is a gradient-based optimization method (e.g. gradient descent). Since our task involves tuning an image processing pipeline (ISP) that is a non-differentiable black box, the gradient cannot flow through the ISP pipeline directly to drive the optimization of ISP parameters. Therefore, it requires training a neural network proxy to approximate the behavior of the ISP pipeline. The differential nature of the neural network proxy allows gradient flow to directly optimize ISP parameters during inference. However, for a complex image processing pipeline, this proxy is hard to train and may not fully reproduce original pipelines. 3. Cascaded Proxy [27]: is also a gradient-based optimization like [26], but the proxy neural network has a different architecture. The difference is that: the Monolithic Proxy uses a single UNet for the proxy, but the Cascaded Proxy improves the UNet with multiple small neural networks that are cascaded to address the vanishing gradient problem. Yet it still faces challenges with error accumulation, limited generalization capability of neural proxy, and cannot be applied to black-box pipeline. Please refer to Q3 in the To All section for further implementation details for these baselines. ## Q2: More Evaluation * Please refer to Q1 in the To All section for the answer to this question. To summarize, we conducted further evaluations using the HDR+ dataset, demonstrating our RL-based framework's ability to generalize effectively to unseen datasets. Our method achieved a PSNR of 31.54 on the HDR+ photo-finishing tuning task, comparable to the 32.47 achieved on the FiveK dataset, and outperformed all baselines. This highlights our method's generalization capability across different datasets and its superior performance compared to other baselines, including CMAES, Greedy Search, Cascaded Proxy, and Monolithic Proxy. --- Rebuttal 2: Comment: Dear Reviewer ASBs, In our rebuttal, we explain whether each baseline requires training. We also provide more evaluation on the HDR+ dataset. Results demonstrate that our method generalizes well to unseen datasets, outperforming all baselines by a large margin. We want to follow up to see if our responses address your concerns. We would be very grateful to hear additional feedback from you and will provide further clarification if needed. Thank you again for your time and effort.	Rebuttal 1: Rebuttal: # To ALL We sincerely thank all reviewers for your comments. We summarize the following questions and add more results to the rebuttal PDF file. ## Q1. Evaluation on more datasets (Reviewer ASBs syDC WirW) * We test our RL-based framework directly on an extra dataset (HDR+ dataset). The results show that our method generalizes well to unseen datasets and outperforms all baselines. On the photo finishing tuning task, our RL-based framework achieves a PSNR of 31.54, comparable to 32.47 on the FiveK dataset, and outperforms first-order and zeroth-order optimization methods by a large margin. Dataset: To further evaluate our method and demonstrate its generalizability, we used the HDR+ dataset: > [HDR+] Burst photography for high dynamic range and low-light imaging on mobile cameras. TOG. 2016. We used the official subset of the HDR+ dataset, which consists of 153 scenes, each containing up to 10 raw photos. The aligned and merged frames are used as the input, expertly tuned images serve as the photo-finishing targets. Evaluation result: To test the generalization capability of our RL-based framework, we directly tested our pre-trained model on the new HDR+ dataset. We compare to baselines including CMAES, Cascaded Proxy, Monolithic Proxy, and Greedy Search (requested by reviewer syDC). In Table 1 of the rebuttal PDF file, we report PSNR, SSIM, LPIPS, and queries to the ISP pipeline. We have also included test results from the FiveK-Target dataset in the table's left column for easy comparison. The results demonstrate that our RL policy generalizes effectively to unseen data, achieving higher photo-finishing quality than methods directly tuned on the test dataset. Qualitative comparisons in Figure 1 of the PDF file show our results are closer to targets, even with input and target images outside the training distribution. Cross dataset generalization: In Table 1 of the PDF file, our RL-based method achieves a PSNR of 31.54 on the HDR+ photo-finishing task, comparable to 32.47 on MIT-Adobe FiveK. This shows that our RL policy, trained on the FiveK dataset, effectively generalizes to the HDR+ dataset. Such out-of-distribution capability is facilitated by our proposed photo-finishing state representation, which extracts invariant features for photo finishing, allowing adaptation to diverse inputs and goals beyond the training distributions. The CMAES baseline shows consistent results on HDR+ compared to FiveK, as it is directly optimized on the test dataset without prior training. However, proxy-based methods [26, 27] perform worse because the proxy network trained on FiveK does not generalize well to HDR+, leading to incorrect gradients and poorer photo-finishing quality. ## Q2. Clarification for terms cascaded and monolithic proxy (Reviewer kK87) * Both [26] and [27] use neural network proxies to approximate ISP pipelines. The core difference is in the architecture of the neural network proxy: [26] uses a single UNet for the proxy, while [27] uses multiple small neural networks cascaded to address the vanishing gradient problem of UNet. The term "monolithic proxy" is borrowed from [27], which refers to [26] as a monolithic proxy. We refer to [27] as a "cascaded proxy" to differentiate their approaches. ## Q3. More details on the implementation of baselines (Reviewer ASBs kK87) * We provide details on the code used and implementation for each baseline. To summarize, we used as much of the provided code as possible and followed every publicly available detail. ### 3.1 Implementation details for cascaded proxy [27] Code used: Although [27] provides code, it is incomplete and cannot be run directly. For instance, only the UNet architecture from [26] is specified in the code, not the cascaded network proposed in [27]. There are also known bugs in the GitHub repository (in GitHub issue #2) that prevent the pipeline from running. We used as much of the provided code as possible and reproduced other parts by following detailed descriptions in [27], as shown below. Proxy network training: With no complete code available, we strictly followed the architecture in Tables 1 and 2 of [27] appendix, using 3 1x1 convolutions for pointwise ISP operations and 5 3x3 convolutions for areawise operations. We trained with the Adam optimizer, a learning rate of 1e-4, a batch size of 512, and 100 epochs, as recommended. Camera metadata was extracted from DNG files, as described in [27]. Training dataset: Since [27] does not release its dataset, we used the MIT-Adobe FiveK dataset, as in our method. Following Section 5 of [27], we used 1,000 raw images from FiveK and sampled 100 points for each ISP hyperparameter. ISP hyperparameter optimization: After training the proxy neural network, we used it as a differential proxy to approximate the ISP pipeline. As in [27], we use the gradient flow through the proxy to directly optimize ISP hyperparameters, following [27]'s optimization details with the same Adam optimizer, same loss, same learning rate, and same iterations. ### 3.2 Implementation details for monolithic proxy [26] Code used: [26] does not provide code, but [27] includes code for the UNet proxy used in [26]. We used this code for reproduction. Proxy network training:** The architecture uses a single UNet to approximate the ISP pipeline, with hyperparameters conditioned by concatenating extra planes to the features. We trained the proxy with the Adam optimizer, a learning rate of 1e-4, a batch size of 512, and 100 epochs. The training set generation and first-order optimization method are the same as for [27]. ### 3.3 Implementation details for CMAES [18] CMAES [9,18] is a zeroth order optimization method without a neural proxy. As [18] does not provide code, we implemented CMAES using the `pymoo` library, enabling parallel execution on multi-core CPUs and achieving reasonable performance. This baseline does not require training. Pdf: /pdf/e46455e9870594b2b2fc26bceee8b8b8157324be.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Poisson-Gamma Dynamical Systems with Non-Stationary Transition Dynamics	Reject	Summary: This work extends Poisson-Gamma Dynamical Systems (PGDSs) by considering non-stationary transition dynamics to effectively capture the evolving dynamics of observed count sequences. The authors propose a model where the underlying transition matrices evolve over time, based on three (gradually more complex and flexible) Dirichlet Markov chains. For inference of the model, the authors make use of the Dirichlet-Multinomial-Beta data augmentation to derive a fully-conjugate Gibbs sampler. Experiments showcase improved data-smoothing and forecasting performance of the proposed method across several real-world datasets. Strengths: - Extending PGDS models to accommodate time-varying transition dynamics is of interest and significant - The proposed variations of Dirichlet-Markov chains provide flexibility in capturing different modeling assumptions - Devising a closed-form Gibbs sampler for posterior inference of this model is significant. - The attained expressions seem correct to the best of my knowledge, although I did not carefully double-check the mathematical details of the derivation. Weaknesses: - The main limitation of this work is the assumption that the transition kernel is static within each sub-interval: i.e., the authors consider that the kernel can only change at discrete instants, while is constant within each sub-interval. Technical Quality: 3 Clarity: 2 Questions for Authors: - Can the authors justify and explain their choice for only allowing discrete-time transition kernel changes? - How can one determine the length of sub-intervals in practice? How did the authors determine these sub-intervals in their experiments? - Would it be possible to accommodate sub-intervals of varying length, $M$, and what would be the implications? - Would it be possible to consider a continuously changing transition kernel? What would be the implications for the model and/or the estimation procedure? - The different transition kernel evolution models proposed do not only differ in their flexibility to capture different phenomena, but also on their complexity: - Can the authors elaborate on the number of learnable parameters of each model? - What is the computational and statistical complexity associated with each? - Results do not seem to provide data-smoothing and forecasting performance improvements: is the added flexibility worth the complexity? - Sections 5 seems to have quite an overlap with some of the preliminaries introduced in Section 2: - Would it be possible to merge both, or is there a reason why these two should be self-contained in different sections? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors address the main limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the reviewer's constructive comments, our answers for the questions are as follows: 1. In practice, users can leverage the prior knowledge about the specific task to set the length of sub-intervals, or treat the length of sub-interval as a hyper-parameter, tuning it with part of the time series data. For the varying length of sub-intervals, as we mentioned in the conclusion section, we plan to adopt temporal point process for change point detection, and thus we can partition the whole time interval via the change points. We consider to simultaneously learn a set of change points to capture the irregulaly-spaced sub-intervals behind non-stationary sequential counts, and to learn the short dynamics underlying each sub-interval. To the end, we consider to work on an EM-type algorithm to maximize the model’s log-likelihood, in which the change points are formulated as latent variables. In principle, we can adopt Gaussian Process with Polya-Gamma augmentation technique to construct continuously changing transition kernel, and this will be our future work. 2. For the complexity of the three proposed transition kernels, the complexity of Dir-Dir construction is $\mathcal{O}\left(TK/M \right)$ and the complexity of Dir-Gam-Dir and PR-Gam-Dir constructions is $\mathcal{O}\left( TK^2/M\right)$. We believe it is worthy to proposed Dir-Gam-Dir and PR-Gam-Dir constructions though they are more complicated. First, in contrast to Dir-Dir construction, Dir-Gam-Dir and PR-Gam-Dir chains explicitly model the interactions among components and thus can improve the flexibility of the proposed model. And as shown in Fig.6 in our paper, the Dir-Gam-Dir and PR-Gam-Dir chains indeed capture more complicated non-stationary dynamics. Furthermore, the PR-Gam-Dir construction can induce sparse patterns. Besides, it is hard to find dataset that perfectly meet the models' assumption, hence the results of different models may be indistinguishable. As we claimed in the conclusion part, we consider to generalize Dirichlet belief networks by incorporating the proposed Dirichlet Markov chain constructions, and we also consider to capture non-stationary interaction dynamics among individuals over online social networks in the future research. For more complicated transition dynamics, the difference among the Dirichlet Markov chains will be amplified. 3. We will carefully revise our paper to merge or reduce some overlap contents in sec.2 and sec.5 to improve the readability of our paper for the final version. --- Rebuttal Comment 1.1: Title: Thank you! Comment: I thank the reviewers for their response to my questions, specifically, on the complexity details of each algorithm.	Summary: Existing PGDS models struggle with capturing the time-varying transition dynamics seen in real-world data. To address this, the submission proposed a non-stationary PGDS, allowing the transition matrices to evolve over time, modeled by Dirichlet Markov chains. Using Dirichlet-Multinomial-Beta data augmentation techniques, a fully-conjugate and efficient Gibbs sampler is developed for posterior simulation. Experiments demonstrate that the proposed non-stationary PGDS achieves improved predictive performance compared to related models. Strengths: The proposed non-stationary Poisson-Gamma Dynamical System offers several notable advantages. Firstly, its ability to allow transition matrices to evolve over time addresses the limitation of state-of-the-art PGDS models in capturing time-varying transition dynamics, making it more suitable for real-world count time series. Secondly, the use of specifically-designed Dirichlet Markov chains to model the evolving transition matrices enhances the model’s capacity to learn non-stationary dependency structures. Thirdly, the application of Dirichlet-Multinomial-Beta data augmentation techniques facilitates the development of a fully-conjugate and efficient Gibbs sampler for posterior simulation. Weaknesses: I did not find any obvious weaknesses. Technical Quality: 4 Clarity: 3 Questions for Authors: None. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors discussed some future work directions in the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's valuable time, positive comments for our manuscript.	Summary: The work extends Poisson-Gamma Dynamical systems (PGDS) to model non-stationary dynamics by replacing the constant transition matrix $\Pi$ with a time dependent one $\Pi^{(t)}$ and the original Dirichlet prior on the columns with three different Dirichlet Markov chain constructions. The manuscript describes am efficient Gibbs-sampler for inference. Strengths: The work addresses a relevant problem of modeling non-stationary dynamics in count time series. The provided extension relative to the original PGDS is sufficiently novel. I lack deep enough understanding of some parts related to the sampler, therefore I cannot assess if the construction of the sampler required new ideas or was a mechanical extension of the sampler for PGDS (this being the main reason for my lower confidence score.). I tend to assume new ideas were necessary. Weaknesses: My major problem is the experiment evaluation. In Table 1 in the NIPS dataset we can see results like $14.014 \pm 4.387$ bolded, over values like $14.706 \pm 4.414$, or $17.105 \pm 6.449$. In ICEWS values like $0.214 \pm 0.008$ over $0.215 \pm 0.007$ , in USEI $4.596 \pm 0.562$ over $4.703 \pm 0.538$, in COVID $6.969 \pm 1.107$ over $7.566 \pm 1.095$. These are mainly smoothing results. In the light of this I am not confident in the statement “As the experiment results shown in Table 1, the NS-PGDS exhibits improved performance in both data smoothing and forecasting tasks.”. We do not know how the confidence interval was computed, or how many repeats were made. The lack of statistical rigor in the evaluation stands in striking contrast with the sophisticated Bayesian model presented. Besides this, other possible problem with the evaluation is that the manuscript states that default paramerters were used for the benchmark methods “GP-DPFA, PGDS, GMC-RATE, GMC-HIER, BGAR” while the present method used specific K based on the dataset. It is very hard to tell if this is a fair comparison or not. Technical Quality: 2 Clarity: 3 Questions for Authors: 1) The time series are divided to equally-spaced sub-intervals, and a $\Pi$ transition matrix is infered for all interval. Why not use a diferent matrix for all time points with a very slowly varying Markow chain. Is this decision was made due to computational reasons or due to modeling assumptions? 2) Please provide the parameters for the benchmark methods in the appendix to facilitate comparison, it is very time consuming for the reader to search all default parameters from the references. 3) Please bold all indistinguishable results in Table 2 using a statistical test, and reassess your conclusions. 4) What is the reason of duplicated factors in the exploratory analysis like “neural-network-networks” and “network-neural-networks”. What the order of the words means? Strength of association? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: No specific limitation section was provided. The part on future work in the Conclusion can be interpreted as pointing out some limitations of the current model, but a specific limitation statement would be preferable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer's constructive comments. We clarify that in contrast to the reviewer's concern, our proposed methods outperform PGDS not only in data smoothing task. As shown in Table 1 in our paper, the proposed three models outperform the baselines in most tasks, even if we only consider NS-PGDS(Dir-Dir), it still outperforms PGDS in almost all tasks. The experimental results are computed via 10 random initialization points, and the means and standard deviations are listed in Table 1 of our paper. As for the hyper-parameter $K$, we set $K=100$ for the baselines, in fact, because of the sparsity nature of Dirichlet process, the rank of the model will be determined automatically and is not a big concern. We adopt relative small $K$ for our proposed methods because the time-varying transition matrices will increase the number of parameters of the models. Therefore, we adopt relatively small $K$ to reduce the computational burden. Besides, experimental results show that NS-PGDS can outperform baselines even with less parameters (much small $K$). Therefore, we believe it is a fair comparison between NS-PGDS and baselines. Answers to Questions: 1. In principle, allowing transition matrix changes at every time step is more natural than what we have done, however, as the reviewer pointed out, this approach will significantly increase the computational burden of the model. More importantly, if we assume the transition matrices change at every time step, then each matrix will be estimated via data point at only one time step, the estimation error will be huge, and the model can not converge. 2. For all baselines, we set $K=100$, and for PGDS, we set $\tau_0=1$, $\gamma_0=50$, $\eta_0=\epsilon_0=0.1$. For GP-DPFA, we wet $\gamma=1$, $c=1$, and $\theta_0=0.01$. For GMC-RATE, we set $\alpha=1$, $\beta=1$. For GMC-HIER, we set $\alpha_z=\beta_z=1$ and $\alpha_{\theta}=\beta_{\theta}=1$. For BGAR, we set $\rho=0.9$, $\alpha=1$ and $\beta=1$. And we will provide these information in the appendix of the final version. 3. We conducted Student's t-test to test the statistical significance. We evaluated the statistical significance of experimental results between PGDS and NS-PGDS(Dir-Dir) and the p-values are listed in below table. The results show that the proposed NS-PGDS significantly outperforms PGDS in forecasting task on four datasets. However, for data smoothing task, NS-PGDS only outperform PGDS marginally. That may because that data smoothing task is a much simpler task compared with data forecasting, PGDS is competent to this task very well, and thus it is hard for NS-PGDS to exceed PGDS by a large margin for smoothing task. \| \| ICEWS \| NIPS \| USEI \| COVID-19 \| \| :---: \| :-----: \| :------: \| :-----: \| :------: \| \| MAE-F \| 5.64e-6 \| 3.79e-10 \| 1.20e-7 \| 1.40e-3 \| \| MRE-F \| 3.76e-8 \| 1.11e-16 \| 1.46e-8 \| 1.16e-6 \| \| MAE-S \| 0.50 \| 0.36 \| 0.33 \| 0.12 \| \| MRE-S \| 0.50 \| 0.41 \| -- \| 0.27 \| 4. It is common for topic models to infer similar latent factors because topic models define a topic (latent factor) as frequency of words, there no reason for a word to appear in only one topic. The order means the frequency of words for a latent factor, for example, “image-sparse-matrix" means the top three frequent words of this topic are 'image', 'sparse' and 'matrix'. --- Rebuttal Comment 1.1: Title: Thank you for the clarification Comment: Thank you for the detailed response! My point was not that the method outperforms in only smoothing result, quite the opposite: that the listed questionable claims (that you yourself shown by the p-value analysis to be insignificant) i mentioned are smoothing results. --- Reply to Comment 1.1.1: Comment: Dear Reviewer: Sorry for misunderstanding your comments, and thanks for your response. Compared with the forecasting task, data smoothing is much simpler and PGDS is competent to this task very well. More importantly, for data smoothing task, we randomly masked 10 percents of the observed data over non-adjacent time steps, and predicted the masked values. The random selection of masked data also poses significant variance of the numerical results. Therefore, it is very hard for NS-PGDS to outperform PGDS in data smoothing task by a large margin. Besides, the main motivation of this work is to allow the transition matrix of PGDS to be time-varying. The results of forecasting task and exploratory analysis have partially validated the effectiveness of the proposed method. Though for data smoothing, the numerical results are not that significant in statistical, however, at least, the results of NS-PGDS are not worse than that of PGDS, and this does not conflict with our motivation. As the reviewer has pointed out, we will reassess our conclusions and give more detailed explanations about results of data smoothing task in the final version.	Summary: This paper introduces non-stationary Poisson-Gamma dynamical systems, an extension of Poisson Gamma dynamical systems with a dynamic transition matrix. Decomposing the time steps into equally spaced subintervals, the transition matrices evolve between sub-intervals, remaining static within sub-intervals. The authors introduce three options for transitions to occur. The authors derive a Gibbs sampling scheme for exact posterior inference using data augmentation techniques and showcase the effectiveness of their method through a series of predictive and qualitative results. Strengths: This is a well-written, organized paper that is easy to read. The proposed method allows for exact posterior inference through Gibbs sampling. The authors exhibit extensive predictive results across 4 datasets, although their method only exhibits marginal improvement as compared to Poisson Gamma dynamical systems. Weaknesses: I'm not convinced that the magnitude of the author's contribution, nor the significance of the paper is strong enough to warrant acceptance, and the methods produce only marginally better results than that of Poisson Gamma dynamical systems. The qualitative results are not groundbreaking. Technical Quality: 4 Clarity: 3 Questions for Authors: How does inference scale with the number of sub-intervals in each of the three methods? How does the user choose the number of sub-intervals? Can they be fit adaptively? How much extra training time do the proposed methods add, as compared to Poisson Gamma dynamical systems? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: The authors address the limitations of their work, stating intention to address these limitations (e.g. constant sub-interval lengths) in future work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: The authors thank the reviewer's valuable feedback. The main contributions of this paper are: (i) We extend state-of-the-art PGDS model such that the transition matrix of PGDS can evolve over time and thus better fit non-stationary environment. (ii) In order to model the time-varying transition matrices, we propose three Dirichlet Markov chains for capturing the complex transition dynamics behind sequential count data. (iii) We leverage Dirichlet-Multinomial-Beta augmentation technique to design the Gibbs sampler for the proposed Dirichlet Markov chains, which is not trivial. For the experimental results, we conducted Student's t-test to test the statistical significance. We evaluated the statistical significance of experimental results between PGDS and NS-PGDS(Dir-Dir) and the p-values are listed in below table. The results show that the proposed NS-PGDS significantly outperforms PGDS in forecasting task on four datasets. However, for data smoothing task, NS-PGDS only outperforms PGDS marginally. That may because that data smoothing task is a much simpler task compared with data forecasting, PGDS is competent to this task very well, and thus it is hard for NS-PGDS to exceed PGDS by a large margin for smoothing task. Besides, note that as shown in Fig 5 in our paper, the time-varying transition kernels indeed discover some interesting information about the time-varying interactions of research topics in NeurIPS conference, which could not be discovered via a constant transition matrix. \| \| ICEWS \| NIPS \| USEI \| COVID-19 \| \| :---: \| :-----: \| :------: \| :-----: \| :------: \| \| MAE-F \| 5.64e-6 \| 3.79e-10 \| 1.20e-7 \| 1.40e-3 \| \| MRE-F \| 3.76e-8 \| 1.11e-16 \| 1.46e-8 \| 1.16e-6 \| \| MAE-S \| 0.50 \| 0.36 \| 0.33 \| 0.12 \| \| MRE-S \| 0.50 \| 0.41 \| -- \| 0.27 \| Scalability with the number of sub-intervals: The complexity of the Gibbs inference algorithms for the three proposed methods scale linearly with the number of sub-intervals. How to choose the number of sub-intervals: For many real-world applications, the user will have prior knowledge about the specific application, and the prior knowledge can be leveraged to choose the length of each sub-interval. For ICEWS dataset, it contains international relations event of a year, and we assume the transition matrix is stationary within a month. Therefore we set $M=30$ for ICEWS. Similarly, we assume the interactions of research topics is stationary within 5 years for NeurIPS conference and the transition dynamics of COVID-19 in U.S. is stationary within 20 days. For USEI dataset, because $T \approx 340$, we heuristically split it into 10 sub-intervals and set $M=34$. In general, user can treat the length of sub-interval as a hyper-parameter, and set it via the user's prior knowledge or tuning it with part of the time series data. Adaptability: Indeed, to allow the proposed model determines the length of sub-interval adaptively is of great significance. As we mentioned in the conclusion section, we plan to adopt temporal point process for change point detection, We consider to simultaneously learn a set of change points to capture the irregulaly-spaced subintervals behind non-stationary sequential counts, and to learn the short dynamics underlying each subinterval. To the end, we consider to work on an EM-type algorithm to maximize the model’s log-likelihood, in which the change points are formulated as latent variables. Extra training time: The training time for PGDS and NS-PGDS are listed in below table, we set $K=100$ for all models. Below table shows that the proposed models achieve better results with little extra training time. \| \| ICEWS \| NIPS \| USEI \| COVID-19 \| \| :------------------: \| :-----: \| :-----: \| :-----: \| :------: \| \| PGDS \| 58.7min \| 18.7min \| 11.4min \| 15.8min \| \| NS-PGDS(Dir-Dir) \| 63.3min \| 19.8min \| 11.8min \| 16.3min \| \| NS-PGDS(Dir-Gam-Dir) \| 68.5min \| 21.4min \| 12.0min \| 16.5min \| \| NS-PGDS(PR-Gam-Dir) \| 69.2min \| 21.7min \| 12.1min \| 16.6min \| --- Rebuttal Comment 1.1: Comment: Thank you for your response. I believe these details add to the quality of the contribution, although it should be noted that the smoothing results are not significant! The superior forecasting results, however, are convincing. The little extra training time and superior forecasting results increase my confidence in this work, and I will change my recommendation accordingly.	Rebuttal 1: Rebuttal: We would like to extend our sincere gratitude to the reviewers for dedicating their time and expertise to evaluate our work. The main concerns of the reviewers are (i) the equally-spaced sub-intervals and the possibility for constructing time-varying transition kernels of other types, (ii) the experiment evaluation. We have carefully clarified these issues and responded to each reviewer's comments.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Improving Robustness of 3D Point Cloud Recognition from a Fourier Perspective	Accept (poster)	Summary: This work introduces a method called Frequency Adversarial Training (FAT) to improve the robustness of 3D point cloud recognition models and examines the robustness of models under 3D point cloud corruptions, including analysis on the power of different corruption effects in the frequency domain. FAT generates adversarial samples by adding perturbations to the frequency representations of point cloud data. The authors also provide a theoretical analysis demonstrating the effectiveness of FAT in improving OOD generalization performance of models. Strengths: 1. The paper is well-structured and organized, making it easy to understand and follow. 2. Introduces a novel concept of frequency augmentation for 3D point cloud data, while previous works are mainly mix-based and deformation-based on original point cloud data rather than their frequency representation. 3. The proposed method considers distribution shifts caused by adversarial perturbations and thus uses separate batch normalization layers, improving model robustness against both low-frequency and high-frequency corruptions. The authors also provide an ablation study demonstrating the effectiveness of each component in improving robustness. 4. The authors provide theoretical analysis showing that adversarial robustness in the frequency domain enhances real-world corruption robustness, and a frequency sensitivity measurement which offers valuable insights for analyzing model robustness from a frequency perspective. Weaknesses: 1. There is confusion in the caption of Figure 3, which suggests the Jacobian matrix for an input point cloud. According to the context, it should be the Jacobian matrix of the model regarding the input point cloud. 2. As FAT is based on adversarial training, it is expected to have comparisons with other adversarial training techniques rather than only mixing- and deformation-based approaches. 3. Minors: The figures are hard to read as the font size is too small. Missing space before the fourth sentence in the caption of Figure 3. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Adversarial augmentation approaches may require substantial computational resources, leading to inefficiency. How much computational resources does FAT require regarding computational time, memory usage, etc., and how does it compare with other augmentation approaches. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors address the limitation of their approach in term of the degraded standard accuracy. There might be a trade-off between corruption robustness and standard performance, left for future investigation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for appreciating our new contributions as well as providing the valuable feedback. Below we address the detailed comments, and hope that you can find our response satisfactory. *Question 1: There is confusion in the caption of Figure 3, which suggests the Jacobian matrix for an input point cloud. According to the context, it should be the Jacobian matrix of the model regarding the input point cloud.* Thanks for pointing out this issue. We will clarify this in the revision. *Question 2: As FAT is based on adversarial training, it is expected to have comparisons with other adversarial training techniques rather than only mixing- and deformation-based approaches.* We have already compared our proposed FAT with adversarial training technique on the ModelNet-C test set in Table 1 in Sec. 4.2. Our FAT outperforms all other methods in terms of mCE (mean corruption error). *Question 3: Minors: The figures are hard to read as the font size is too small. Missing space before the fourth sentence in the caption of Figure 3.* Thanks for pointing out these issues. We will correct them in the revision. *Question 4: How much computational resources does FAT require regarding computational time, memory usage, etc., and how does it compare with other augmentation approaches.* The computational cost of our FAT implementation primarily involves generating high-frequency and low-frequency adversarial examples. Compared to standard adversarial training and other data augmentations, FAT incurs approximately $1.7 \sim 3.2$ times higher computational costs and about 3 times more memory usage. Despite this limitation, such an increase in computational overhead is deemed acceptable for offline training scenarios. Moreover, our approach would not affect the efficiency of model inference, ensuring unhindered deployment of well-trained models in practical applications. We will add the discussion in the revision. --- Rebuttal Comment 1.1: Comment: Thank you for the responses. I've updated the rating. --- Reply to Comment 1.1.1: Title: Thank you for increasing the score Comment: Dear Reviewer ELyT, Thank you very much for increasing the score! We are glad to know that our response has addressed your concerns. We really appreciate your valuable comments and appreciation of our contributions. We will further improve the paper in the final. Best regards, Authors	Summary: This paper introduces Frequency Adversarial Training (FAT), leveraging Graph Fourier Transform (GFT) to enhance robustness against point cloud corruptions by training models with frequency-domain adversarial examples, demonstrating significant improvements in robustness across various architectures through extensive experiments. Strengths: 1. Innovative Approach: The use of the frequency domain for analyzing and improving the robustness of point cloud recognition models is novel and well-motivated. The application of GFT to understand the impact of corruptions on different frequency bands is a significant contribution. 2. Comprehensive Evaluation: The paper provides a thorough evaluation of the proposed method, comparing it with existing approaches and demonstrating its effectiveness across multiple models and datasets. 3. Theoretical and Empirical Validation: The authors provide both theoretical analysis and empirical evidence to support the effectiveness of FAT, strengthening the credibility of their claims. 4. Practical Relevance: Improving the robustness of 3D point cloud recognition models is highly relevant for safety-critical applications such as autonomous driving and robotics. 5. Writing quality: This paper is written and organized well Weaknesses: 1. Complexity of Implementation: The proposed FAT method involves several complex steps, including the generation of high-frequency and low-frequency adversarial examples and the use of multiple batch normalizations. This complexity might hinder the adoption of the method in practical scenarios. 2. Limited Impact on Clean Accuracy: The paper mentions a slight reduction in clean accuracy when using FAT, which might be a concern for applications where both robustness and accuracy are critical. 3. Generality of Results: While the experiments demonstrate the effectiveness of FAT on several models and datasets, additional experiments on more diverse datasets and real-world scenarios could further validate the generalizability of the approach, e.g.ShapeNet-C[1] and ScanObjectNN-C[2] 4. Lack of Updated Baselines in Experiments: In Table 1, the paper lacks experiments comparing some updated point cloud baselines, like RPC[3] and PointNeXt[4], which could provide a more comprehensive evaluation of the proposed method. 5. Missing Comparison with Updated Augmentation Methods: In Table 2, the paper lacks comparison with updated augmentation methods such as AdaptPoint[2], which could highlight the relative performance of FAT against newer augmentation techniques. [1]: PointCloud-C: Benchmarking and Analyzing Point Cloud Perception Robustness under Corruptions [2]: Sample-adaptive Augmentation for Point Cloud Recognition Against Real-world Corruptions [3]: Benchmarking and Analyzing Point Cloud Classification under Corruptions [4]: PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies Technical Quality: 3 Clarity: 3 Questions for Authors: Refer to the Weakness. I will improve my rates when the weaknesses are solved. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Refer to the Weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for acknowledging the novelty of our paper as well as providing the valuable feedback. Below we address the detailed comments, and hope that you can find our response satisfactory. *Question 1: Complexity of Implementation might hinder the adoption of the method in practical scenarios.* The computational cost of our FAT implementation primarily involves generating high-frequency and low-frequency adversarial examples. Compared to standard adversarial training and other data augmentations, FAT incurs approximately $1.7 \sim 3.2$ times higher computational costs. Despite this limitation, such an increase in computational overhead is deemed acceptable for offline training scenarios. Moreover, our approach would not affect the efficiency of model inference, ensuring unhindered deployment of well-trained models in practical applications. We will add the discussion in the revision. *Question 2: The paper mentions a slight reduction in clean accuracy when using FAT, which might be a concern for applications.* In our limitations section, we have already noted that while FAT slightly reduces clean accuracy, the magnitude of this reduction is minimal and deemed acceptable. This could be attributed to the inherent trade-off between accuracy and robustness [1]. Moreover, as shown in Table 2, existing data augmentation methods decrease clean accuracy by an average of 1.03\%, whereas our FAT exhibits a smaller average reduction of 0.55\% in clean accuracy. Additionally, Table 2 demonstrates that incorporating FAT results in an average improvement of 0.33\% in clean accuracy, indicating that FAT's impact on clean accuracy is not universally negative. *Question 3: The paper would be more convincing by additional experiments on more diverse datasets and real-world scenarios.* Thanks for the valuable suggestion. We further conduct experiments on the KITTI and ScanObjectNN datasets, both collected by LiDAR sensors. *Due to space constraints, detailed experimental results and analysis are presented in the Global Rebuttal.* (See more results on ScanObjectNN-C in response to Question 5). Experiments on these two real-world datasets [2][3] further validate the superiority and practicality of our FAT. *Question 4: The paper lacks experiments comparing some updated point cloud baselines, which could provide a more comprehensive evaluation of the proposed method.* Thanks for the valuable suggestion. We further conduct experiments for PointNeXt [4] on ModelNet-C and ScanObjectNN-C below. (See more results on PointNeXt in response to Question 5.) \|ModelNet-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointNext\|Vanilla Training\|0.932\|0.856\|1.460\|1.297\|0.904\|0.847\|0.957\|0.251\|0.276\| \|\|Adv Training\|0.924\|0.834\|1.593\|0.716\|1.025\|0.876\|1.144\|0.230\|0.251\| \|\|DUP Defense\|0.919\|0.840\|1.461\|0.838\|1.192\|0.715\|1.188\|0.224\|0.262\| \|\|FAT (Ours)\|0.930\|0.781\|1.412\|0.692\|0.986\|0.827\|1.082\|0.230\|0.241\| \|ScanObjectNN-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointNext\|Vanilla Training\|0.873\|0.921\|0.995\|1.079\|0.803\|0.807\|0.942\|0.944\|0.875\| \|\|Adv Training\|0.870\|0.901\|0.991\|1.027\|0.803\|0.833\|0.929\|0.912\|0.809\| \|\|DUP Defense\|0.859\|0.901\|0.980\|1.046\|0.826\|0.748\|0.973\|0.923\|0.809\| \|\|FAT (Ours)\|0.875\|0.877\|0.998\|0.916\|0.791\|0.786\|0.867\|0.938\|0.840\| The results demonstrate that FAT achieves consistent performance on the advanced network architecture PointNeXt, similar to observations on PointNet and more. Our FAT outperforms all other methods in terms of mCE. This indicates that FAT's performance is largely independent of the underlying model architecture, making it applicable to both traditional and modern networks. We will incorporate these additional experiments into the revision. *Question 5: The paper lacks comparison with updated augmentation methods such as AdaptPoint, which could highlight the relative performance of FAT against newer augmentation techniques.* Thanks for the valuable suggestion. We further conduct experiments for comparison with AdaptPoint [2] on ScanObjectNN-C. AdaptPoint follows the official experimental settings [2]. The results are shown below. \|ScanObjectNN-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointNet\|AdaptPoint\|0.743\|1.256\|1.359\|0.875\|1.519\|0.676\|1.112\|1.448\|1.804\| \|\|+ FAT\|0.744\|1.196\|1.370\|0.823\|1.446\|0.690\|1.125\|1.220\|1.701\| \|PointNext\|AdaptPoint\|0.885\|0.783\|0.767\|1.030\|0.810\|0.508\|0.628\|0.911\|0.824\| \|\|+ FAT\|0.885\|0.761\|0.748\|0.948\|0.833\|0.521\|0.648\|0.829\|0.802\| It is evident that incorporating FAT achieves a lower mCE, indicating its superiority. We will integrate these results into Table 2 and continue to include more comprehensive experiments as per your recommendations in the revised version. [1] Theoretically Principled Trade-off between Robustness and Accuracy, ICML 2019 [2] Sample-adaptive Augmentation for Point Cloud Recognition Against Real-world Corruptions, ICCV 2023 [3] Benchmarking Robustness of 3D Object Detection to Common Corruptions in Autonomous Driving, CVPR 2023 [4] PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies, NeurIPS 2022 --- Rebuttal Comment 1.1: Comment: Thanks, updated my ratings! Good luck ! --- Reply to Comment 1.1.1: Title: Thank you for increasing the score Comment: Dear Reviewer wa2k, Thank you very much for increasing the score! We are glad to know that our response has addressed your concerns. We really appreciate your valuable feedback. We will further improve the paper in the final. Best regards, Authors	Summary: This paper studies how to enhance the robustness of 3D point cloud recognition. The authors propose Frequency Adversarial Training (FAT) to improve the corruption robustness of 3D point cloud recognition models. FAT trains a model with adversarial examples that add perturbations to the frequency-domain representations of point clouds. Strengths: - The problem studied in this paper is important. - The authors propose generating the adversarial sample in the frequency domain, which is interesting. - This paper considers several baseline methods. Weaknesses: - This paper considers limited real-world applications. 3D point cloud recognition often fails due to LiDAR sensor inaccuracies and changes in the physical environment. However, this paper only utilizes one dataset, ModelNet40, where points are generated via 3D modeling techniques instead of being collected from real-world sensors, which is not convincing. If the paper aims to address real-world problems, it would be more convincing to include more real-world datasets, such as the KITTI dataset, where point clouds are collected from LiDAR sensors. - The defense performance does not significantly outperform baseline methods. As shown in Table 1, FAT only outperforms baseline methods by a few points. Moreover, the proposed method appears to be effective only when combined with other data augmentation techniques. Could the authors provide more details and an explanation of the phenomenon of the ‘surprising results’ mentioned in the paper? For example, why does mixing methods result in much better defense? - This paper does not analyze the relationship between the frequency and spatial domains in the context of the 3D point cloud. The two domains are mutually transformable, but the paper does not provide insights into their relationships or general observations. - The paper only studies the defense effects against general 3D data corruptions without considering specific attack methods. It remains unclear if this approach can also defend against specific attack methods targeting 3D perception, such as attacks using adversarial points with specific shapes, locations, and rotation angles. The following are a few works about attacks against LiDAR object detection in autonomous driving scenarios [1-4] for reference. [1] Yulong Cao, Chaowei Xiao, Dawei Yang, Jing Fang, Ruigang Yang, Mingyan Liu, and Bo Li. 2019. Adversarial objects against lidar-based autonomous driving systems. arXiv preprint arXiv:1907.05418 (2019). [2] James Tu, Mengye Ren, Sivabalan Manivasagam, Ming Liang, Bin Yang, Richard Du, Frank Cheng, and Raquel Urtasun. 2020. Physically realizable adversarial examples for lidar object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. [3] Yi Zhu, Chenglin Miao, Tianhang Zheng, Foad Hajiaghajani, Lu Su, and Chunming Qiao. 2021. Can we use arbitrary objects to attack lidar perception in autonomous driving?. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. [4] Shenchen Zhu, Yue Zhao, Kai Chen, Bo Wang, Hualong Ma, and Cheng’an Wei. 2024. AE-Morpher: Improve Physical Robustness of Adversarial Objects against LiDAR-based Detectors via Object Reconstruction. In Proceedings of the 33rd USENIX Security Symposium. Technical Quality: 3 Clarity: 3 Questions for Authors: Q1: As mentioned earlier in your paper, the corruption of 3D point clouds often occurs in real-world applications due to environmental noise and reflections. For instance, in autonomous driving scenarios, the LiDAR sensor may only capture partial points of the target vehicle because of occlusion and distance effects. However, this paper does not study those real-world data. I am curious whether your method would be effective on real-world datasets, such as the KITTI and Waymo driving datasets. Solely using ModelNet40 is not convincing. Q2: Can the proposed approach maintain good performance when considering specific attack methods, such as those using adversarial points with specific shapes, locations, and rotation angles? Q3: Modifications in the frequency domain can sometimes result in physically unrealizable transformations in the spatial domain. Will this limit the application of your work in real-world scenarios? Q4: Can the authors provide more details on the reasons for using GFT rather than DCT or other frequency transformation methods? Q5: What would be the outcome if medium frequency bands were modified? The proposed methods seem to only consider high and low frequencies. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for appreciating our new contributions as well as providing the valuable feedback. Below we address the detailed comments, and hope that you can find our response satisfactory. *Question 1: The paper would be more convincing by using real-world data from LiDAR sensors, such as KITTI dataset.* Thanks for the valuable suggestion. We further conduct experiments on the KITTI and ScanObjectNN datasets, both collected by LiDAR sensors. *Due to space constraints, detailed experimental results and analysis are presented in the Global Rebuttal.* Experiments on these two real-world datasets [1][2] further validate the superiority and practicality of our FAT. *Question 2: Why does the proposed method appear to be significantly more effective when combined with other data augmentation techniques?* FAT operates in the frequency domain and considers the underlying structure of compactly represented 3D point clouds, while traditional data augmentation methods focus on spatial transformations within the 3D data itself. Therefore, attributed to the complementary and compatible information from both the spatial and frequency domains, mixing methods result in much better defense. *Question 3: The paper does not analyze the relationship between the frequency and spatial domains in the context of the 3D point cloud.* Thanks for the suggestion. Similar to 2D images, the rough shape in point clouds’ spatial domain is represented by transformed low-frequency components while the fine details of objects are encoded in transformed high-frequency components. Furthermore, the frequency characteristics of point clouds represent higher level and global information than point-to-point relations in spatial domain. That is, the frequency representation encodes more abstract and essential contexts for recognizing the point cloud. This implies that in the frequency domain, point clouds are compactly represented, facilitating a better understanding of low-level distortions that are free of high-level semantics. We will incorporate this analysis in the revision. *Question 4: The paper only studies the defense effects against general 3D data corruptions without considering specific attack methods.* Thanks for the suggestion. Actually, FAT can enhance the models’ robustness against adversarial attacks to some extent. The enhanced PointNet model on ModelNet40 achieves the adversarial accuracy of 31.2\% under PGD-20 at $\epsilon$ = 0.05, compared to 0\% for the standard trained PointNet. The enhanced PointPillars model on KITTI achieves a defense accuracy of 65.5\% with IoU greater than 0.7 against Tu's attack method [3], and 56.8\% against Zhu's attack method [4], compared to 42.5\% and 26.0\% for the standard trained PointPillars model. We will include more comprehensive experimental results and discussions on defending against adversarial attacks in the revision, including references and analyses of specific attack methods [3][4][5][6]. *Question 5: Modifications in the frequency domain can sometimes result in physically unrealizable transformations in the spatial domain. Will this limit the application of your work in real-world scenarios?* No, it does not limit applicability. As a training method, FAT only needs to consider transformations in the digital world, which is always feasible regardless of the scenario, unlike attack methods that must consider physical implementation. *Question 6: Can the authors provide more details on the reasons for using GFT rather than DCT or other frequency transformation methods?* Images are typically transformed in the frequency domain with the 2D discrete Fourier transform (DFT) or discrete cosine transform (DCT). Different from images supported on regular grids, although 3D point clouds are highly structured, they reside on irregular domains without an ordering of points, hindering the deployment of traditional Fourier transforms. Specifically, an image represented as $\mathbb{R}^{n \times n}$ is regularly sampled on a grid, ensuring that pixels $(i, j)$ and $(i+1, j)$ are adjacent. In contrast, a point cloud represented as $\mathbb{R}^{n \times 3}$ lacks such ordering of points, where the Euclidean distance between the $i$-th and $(i+1)$-th points can be substantial. Traditional Fourier transforms cannot be directly applied to point clouds due to their unordered nature and loss of relative positional information. However, graphs provide a natural and accurate representation of irregular point clouds. Each point in a point cloud is treated as a vertex, connected to its $K$ nearest neighbors, with each point's coordinates serving as graph signals. Once a graph is constructed to represent the point cloud, the graph Fourier transform (GFT) can compactly transform it into the frequency domain by leveraging the edges of the graph to encode relative positional information. *Question 7: What would be the outcome if medium frequency bands were modified?* Thanks for the suggestion. We further conduct more ablation studies among FAT, as well as FAT variants: FAT with only medium-frequency modified, FAT with only low-frequency modified, and FAT with only high-frequency modified. *Due to space constraints, detailed experimental results and analysis are presented in the subsequent comments.* We will incorporate these additional experiments into the revision. --- Rebuttal Comment 1.1: Comment: *Question 7: What would be the outcome if medium frequency bands were modified?* Thanks for the suggestion. We further conduct more ablation studies among FAT, as well as FAT variants: FAT with only medium-frequency modified, FAT with only low-frequency modified, and FAT with only high-frequency modified. The results are shown below. \|ModelNet-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointNet\|Vanilla Training\|0.907\|1.422\|1.902\|0.642\|1.266\|0.500\|1.072\|2.980\|1.593\| \|\|FAT with only low-frequency modified\|0.906\|1.317\|1.702\|0.519\|1.234\|0.452\|1.043\|2.851\|1.415\| \|\|FAT with only medium-frequency modified\|0.905\|1.342\|1.729\|0.439\|1.468\|0.500\|1.144\|2.757\|1.354\| \|\|FAT with only high-frequency modified\|0.890\|1.306\|1.614\|0.373\|1.734\|0.504\|1.193\|2.627\|1.098\| \|\|FAT\|0.902\|1.237\|1.553\|0.370\|1.606\|0.448\|1.097\|2.583\|1.004\| Compared with other methods, FAT with only high-frequency modified has a lower mCE for high-frequency corruptions such as “Jitter”, while FAT with only low-frequency modified has a lower mCE for low-frequency corruptions such as “scale”. As discussed in Sec. 3.3, this is because adversarial training on high/low frequencies reduces the high/low frequency sensitivity, thus improving robustness to high/low-frequency corruptions. The performance of FAT with only medium-frequency modified falls between FAT with only high-frequency modified and FAT with only low-frequency modified. Compared with these methods, FAT achieves the lowest mCE, showing the effectiveness of our algorithm. We will incorporate these additional experiments into the revision. [1] Sample-adaptive Augmentation for Point Cloud Recognition Against Real-world Corruptions, ICCV 2023 [2] Benchmarking Robustness of 3D Object Detection to Common Corruptions in Autonomous Driving, CVPR 2023 [3] Physically Realizable Adversarial Examples for Lidar Object Detection, CVPR 2020 [4] Can We Use Arbitrary Objects to Attack Lidar Perception in Autonomous Driving? CCS 2021 [5] Adversarial objects against lidar-based autonomous driving systems, arXiv [6] AE-Morpher: Improve Physical Robustness of Adversarial Objects against LiDAR-based Detectors via Object Reconstruction, USENIX 2024	Summary: This paper introduces a novel approach to improving the robustness of 3D point cloud recognition models by introducing Frequency Adversarial Training (FAT). By analyzing the frequency space of point clouds through the graph Fourier transform, the authors found that models are sensitive against different frequency bands of corruptions. To reduce the effect of both low and high frequency corruptions, the authors propose FAT, which introduces adversarial examples during training that were generated by perturbing the point cloud’s frequency space and taking the inverse graph Fourier transform. From the authors’ experiments, FAT significantly improved the robustness of point cloud recognition models and was able to achieve a new state-of-the-art performance. Strengths: The authors introduce an original analysis of point cloud recognition models by quantifying the sensitivity of these models against low-frequency disruptions, such as rotations, and high-frequency disruptions, such as jittering. Frequency adversarial training is also an innovative approach to improving the robustness of these models against these disruptions, utilizing the graph Fourier transform to adversarially train against low and high-frequency distortions within the frequency space. The writing, extensive experiments, and theoretical analysis provided by the authors effectively and clearly show the positive impact FAT has on the robustness of point cloud recognition models. The use of a point cloud’s frequency space and application of FAT have the capacity to have significant impacts on future research in point cloud models beyond 3D recognition models. Weaknesses: While the paper mentions the importance of robust models for real-world applications, it exclusively evaluates models trained with FAT on synthetic datasets like ModelNet-C and ModelNet40-C. Evaluating FAT using data from sensors such as LiDAR would further show FAT’s applicability under real-world conditions (such as KITTI dataset, if possible). Additionally, the paper does not discuss the complexity and efficiency of FAT, which are important factors for real-world applications. Technical Quality: 4 Clarity: 4 Questions for Authors: It will be great to see some results on real-world datasets such as KITTI. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for appreciating our new contributions as well as providing the valuable feedback. Below we address the detailed comments, and hope that you can find our response satisfactory. *Question 1: The paper would be more convincing by using real-world data from LiDAR sensors, such as KITTI dataset.* Thanks for the valuable suggestion. We further conduct experiments on the KITTI and ScanObjectNN datasets, both collected by LiDAR sensors. Experiments on these two real-world datasets further validate the superiority and practicality of our FAT. The experimental settings and evaluation metrics on ScanObjectNN-C [1] are consistent with those on ModelNet-C. The KITTI-C dataset [2] includes four major types of corruptions: Weather-level (e.g., Strong Sunlight), Sensor-level (e.g., Density Decrease), Motion-level (e.g., Moving Object), and Object-level (e.g., Local Gaussian Noise). Following [2], we use $AP_{cor}$ (corruption average precision) at moderate difficulty as the evaluation metric on KITTI-C, where higher values indicate better performance. We employ the representative 3D object detection model PointPillars. The results are shown below. \|KITTI-C\|Method\|$AP_{clean}$\|mean $AP_{cor}$\|Weather-level $AP_{cor}$\|Sensor-level $AP_{cor}$\|Motion-level $AP_{cor}$\|Object-level $AP_{cor}$\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointPillars\|Vanilla Training\|78.34\|65.35\|63.39\|75.33\|49.61\|73.05\| \|\|Adv Training\|74.86\|62.96\|51.97\|78.68\|51.43\|69.77\| \|\|DUP Defense\|72.02\|63.98\|58.19\|78.05\|48.21\|71.46\| \|\|FAT (Ours)\|78.06\|67.02\|60.81\|79.19\|54.71\|73.39\| \|ScanObjectNN-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|DGCNN\|Vanilla Training\|0.858\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\| \|\|Adv Training\|0.843\|1.062\|1.146\|0.778\|1.097\|0.873\|0.960\|1.185\|1.396\| \|\|DUP Defense\|0.832\|1.029\|1.195\|0.833\|1.157\|1.197\|1.214\|0.773\|0.834\| \|\|FAT (Ours)\|0.856\|0.933\|0.968\|0.815\|0.959\|0.852\|0.950\|0.959\|1.026\| \|PointNet\|Vanilla Training\|0.739\|1.354\|1.610\|0.884\|1.427\|0.786\|1.264\|1.487\|2.022\| \|\|Adv Training\|0.725\|1.334\|1.532\|0.844\|1.403\|0.873\|1.333\|1.474\|1.881\| \|\|DUP Defense\|0.712\|1.348\|1.717\|0.825\|1.555\|1.157\|1.480\|0.829\|1.875\| \|\|FAT (Ours)\|0.734\|1.254\|1.393\|0.796\|1.465\|0.844\|1.264\|1.259\|1.759\| \|PointNext\|Vanilla Training\|0.873\|0.921\|0.995\|1.079\|0.803\|0.807\|0.942\|0.944\|0.875\| \|\|Adv Training\|0.870\|0.901\|0.991\|1.027\|0.803\|0.833\|0.929\|0.912\|0.809\| \|\|DUP Defense\|0.859\|0.901\|0.980\|1.046\|0.826\|0.748\|0.973\|0.923\|0.809\| \|\|FAT (Ours)\|0.875\|0.877\|0.998\|0.916\|0.791\|0.786\|0.867\|0.938\|0.840\| It can be seen that our FAT generally leads to lower mCE (mean corruption error) on ScanObjectNN-C and higher $AP_{cor}$ (average precision) on KITTI-C. The experimental results on both KITTI and ScanObjectNN datasets further validate the generalizability and applicability of our FAT under real-world conditions. We will add the results in the revision. *Question 2: The paper does not discuss the complexity and efficiency of FAT.* Thanks for the suggestion. The complexity of FAT implementation primarily involves generating high-frequency and low-frequency adversarial examples. Compared to standard adversarial training and other data augmentations, FAT incurs approximately $1.7 \sim 3.2$ times higher computational costs. Despite this limitation, such an increase in computational overhead is deemed acceptable for offline training scenarios. Moreover, our approach would not affect the efficiency of model inference, ensuring unhindered deployment of well-trained models in practical applications. We will add the discussion in the revision. [1] Sample-adaptive Augmentation for Point Cloud Recognition Against Real-world Corruptions, ICCV 2023 [2] Benchmarking Robustness of 3D Object Detection to Common Corruptions in Autonomous Driving, CVPR 2023 --- Rebuttal Comment 1.1: Comment: I am satisfied with the rebuttal and thus keep my rate unchanged. --- Reply to Comment 1.1.1: Title: Thank you for the appreciation of our contributions Comment: Dear Reviewer kjym, We are pleased to know that you find our response satisfactory. We really appreciate your valuable comments. We will incorporate the additional experiments and improve the paper in the final version. Best regards, Authors	Rebuttal 1: Rebuttal: We deeply appreciate all the reviewers for their insightful and constructive reviews of our manuscript. Delightfully, we are glad that the reviewers found that: - *The presentation of our paper is polished and easy to understand.* (Reviewers kjym, wa2k, ELyT) - *The problem studied in our paper is important* for safety-critical applications such as autonomous driving and robotics. (Reviewers kjym, WzcS, wa2k) - *Our idea is novel, interesting, and well-motivated.* (Reviewers kjym, WzcS, wa2k, ELyT) - *The performance of our apporach is well-justified, promising, and effective.* (Reviewers kjym, wa2k, ELyT) - *The theoretical analysis of our paper offers valuable insights* for analyzing model robustness. (Reviewers kjym, wa2k, ELyT) Below we address a common concern raised by reviewers kjym, WzcS, and wa2k regarding the need for additional experiments on real-world LiDAR sensor datasets to enhance our paper's persuasiveness. *Common Concern 1: The paper would be more convincing by additional experiments conducted on real-world LiDAR sensor datasets.* Thanks for the valuable suggestion. We further conduct experiments on the KITTI and ScanObjectNN datasets, both collected by LiDAR sensors. Experiments on these two real-world datasets further validate the superiority and practicality of our FAT. The experimental settings and evaluation metrics on ScanObjectNN-C [1] are consistent with those on ModelNet-C. The KITTI-C dataset [2] includes four major types of corruptions: Weather-level (e.g., Strong Sunlight), Sensor-level (e.g., Density Decrease), Motion-level (e.g., Moving Object), and Object-level (e.g., Local Gaussian Noise). Following [2], we use $AP_{cor}$ (corruption average precision) at moderate difficulty as the evaluation metric on KITTI-C, where higher values indicate better performance. We employ the representative 3D object detection model PointPillars. The results are shown below. \|KITTI-C\|Method\|$AP_{clean}$\|mean $AP_{cor}$\|Weather-level $AP_{cor}$\|Sensor-level $AP_{cor}$\|Motion-level $AP_{cor}$\|Object-level $AP_{cor}$\| \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|PointPillars\|Vanilla Training\|78.34\|65.35\|63.39\|75.33\|49.61\|73.05\| \|\|Adv Training\|74.86\|62.96\|51.97\|78.68\|51.43\|69.77\| \|\|DUP Defense\|72.02\|63.98\|58.19\|78.05\|48.21\|71.46\| \|\|FAT (Ours)\|78.06\|67.02\|60.81\|79.19\|54.71\|73.39\| \|ScanObjectNN-C\|Method\|OA\|mCE\|Rotate\|Jitter\|Scale\|Drop-G\|Drop-L\|Add-G\|Add-L \|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\| \|DGCNN\|Vanilla Training\|0.858\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\|1.000\| \|\|Adv Training\|0.843\|1.062\|1.146\|0.778\|1.097\|0.873\|0.960\|1.185\|1.396\| \|\|DUP Defense\|0.832\|1.029\|1.195\|0.833\|1.157\|1.197\|1.214\|0.773\|0.834\| \|\|FAT (Ours)\|0.856\|0.933\|0.968\|0.815\|0.959\|0.852\|0.950\|0.959\|1.026\| \|PointNet\|Vanilla Training\|0.739\|1.354\|1.610\|0.884\|1.427\|0.786\|1.264\|1.487\|2.022\| \|\|Adv Training\|0.725\|1.334\|1.532\|0.844\|1.403\|0.873\|1.333\|1.474\|1.881\| \|\|DUP Defense\|0.712\|1.348\|1.717\|0.825\|1.555\|1.157\|1.480\|0.829\|1.875\| \|\|FAT (Ours)\|0.734\|1.254\|1.393\|0.796\|1.465\|0.844\|1.264\|1.259\|1.759\| \|PointNext\|Vanilla Training\|0.873\|0.921\|0.995\|1.079\|0.803\|0.807\|0.942\|0.944\|0.875\| \|\|Adv Training\|0.870\|0.901\|0.991\|1.027\|0.803\|0.833\|0.929\|0.912\|0.809\| \|\|DUP Defense\|0.859\|0.901\|0.980\|1.046\|0.826\|0.748\|0.973\|0.923\|0.809\| \|\|FAT (Ours)\|0.875\|0.877\|0.998\|0.916\|0.791\|0.786\|0.867\|0.938\|0.840\| It can be seen that our FAT generally leads to lower mCE (mean corruption error) on ScanObjectNN-C and higher $AP_{cor}$ (average precision) on KITTI-C. The experimental results on both KITTI and ScanObjectNN datasets further validate the generalizability and applicability of our FAT under real-world conditions. We will add the results in the revision.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning	Accept (poster)	Summary: This work presents a novel approach for universal anomaly segmentation (AS) by framing AS as Change segmentation. This change of perspective motivates them to create a synthetic training set based on change detection (e.g. by stitching additional objects from a different dataset). With this synthetic dataset they train a change segmentation model which they later transfer directly for anomaly segmentation without fine-tuning. Strengths: The idea of framing Anomaly Segmentation as Change Segmentation is interesting and worth exploring. The results are encouraging, especially in the Gooods dataset. Weaknesses: The paper is confusing at times, with some important concepts not clearly defined: - For instance, the different method variants i.e. MetaUAS/MetaUAS/MetaUAS+ should be explained more thoroughly in the methodology section. - Also there should be consistency in the naming (MetaUAS vs MetaAS). - The highlighted numbers (red and blue) in the tables are excluding UniAD in Table 1 (which is marked in gray) and I could not find out why. Missing comparison with change segmentation methods. The paper proposes that AS can be tackled as change segmentation between one normal (reference) image and the query image with the anomaly, which sounds reasonable. However, I would expect then that existing Change Segmentation methods could be applied to AS (just like authors do with their method). Thus, it would be sensible to include a comparison with one or two baselines of recent change segmentation methods applied to AS the same way that MetaUAS is applied. Technical Quality: 3 Clarity: 3 Questions for Authors: Can the authors explain what are the differences between MetaUAS//+ ? Why is UniAD excluded from the comparison in Table1? Could all change segmentation methods be applied to anomaly segmentation "off-the-shelf"? If yes, why not include some existing methods as baseline? If no, please explain what are the particular differences that make MetaUAS applicable to AS while other change segmentation methods can not. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations and social impact are well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: + bbUz-Q1: What are the differences between MetaUAS, MetaUAS, and MetaUAS+? We apologize for the confusion about MetaUAS, MetaUAS, and MetaUAS+. In fact, we have given some details of our MetaUAS and its two variants (MetaUAS* and MetaUAS+) in the Sec. B of Supplementary Materials. Here, we provide some explanations, which might address your concerns. The MetaUAS is capable of segmenting any anomalies only given just one normal image prompt. Note that the normal image prompt is randomly selected for each testing class in MetaUAS. Different from MetaUAS, the MetaUAS searches the best-matched normal image as a prompt from the normal training set for each query image. The MetaUAS+ builds on MetaUAS, but introduces the visual prior from a CLIP model. The visual prior knowledge is obtained by computing the cosine similarity between the query feature and the corresponding prompt feature extracted from the vision encoder of the CLIP model, which is the same as WinCLIP+ and PromptAD. In addition, the MetaAS is a typo and the correct one should be MetaUAS. We will add these clarifications to the methodology section. + bbUz-Q2: Why UniAD is excluded from the comparison in Table 1? UniAD is a powerful unsupervised anomaly detection method, which trains a unified reconstruction model to detect multi-class anomalies using full-shot normal images from the target domain. In contrast to our training-free one-shot method, UniAD requires re-training the model using full-shot normal images. As reported in Table 1, we observe that our one-shot training-free method still exhibits strong performance compared to the full-shot UniAD. We will add a description for UniAD in the caption of Tabel 1. + bbUz-Q3: Could all change segmentation methods be applied to anomaly segmentation ``off-the-shelf''? It is challenging to apply existing change segmentation models to anomaly segmentation. Current change segmentation models are primarily focused on two scenarios: remote sensing and street scenes. The main characteristics of these scenarios and models can be summarized in four aspects. First, the scale of training images in these scenarios is usually small, typically hundreds or thousands of images. Second, the diversity of images is limited because they mainly consist of remote sensing images or street scene images. Third, these models primarily focus on semantic changes (such as buildings) while the background may undergo other changes due to factors like seasons and weather. Fourth, the prompt and query images are generally coarsely aligned in the change segmentation, which is different from anomaly detection scenarios where there are often large geometric variations. These characteristics limit the development of a general change detection model. Additionally, we evaluated the [off-the-shelf ChangerEx with s50 backbone](https://github.com/likyoo/open-cd/tree/main/configs/changer) on MVTec, and the corresponding results are reported in the following table. We can see that its generalization performance is poor compared to our method. \|Methods \|Backbone \|Training Data \|I-ROC \|I-PR \|P-ROC \|P-PR \|P-PRO\| \| :--- \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| \|ChangerEx \|S50 \|LEVIR-CD \|59.8 \|74.1 \|65.5 \|3.0 \|0.8 \| \|ChangerEx \|S50 \|S2Looking \|63.4 \|79.6 \|68.3 \|10.5 \|13.1 \| \|Ours \|E-b4 \|Synthesize \|91.3 \|96.2 \|94.6 \|59.6 \|82.6 \| --- Rebuttal Comment 1.1: Title: Thank you for the authors response Comment: The author response clarified my questions and after reading the other reviewers comments I will increase my score to weak accept.	Summary: The authors introduce a method for anomaly segmentation that relies solely on visual information and does not require any anomaly training data or language guidance. Their method is based on change segmentation, which allows for the synthesis of large-scale image pairs for training the anomaly segmentation model. The proposed MetaUAS framework demonstrates its generalizability across three industrial anomaly segmentation datasets: MVTec, VisA, and Goods, using one normal image per class as a prompt. Within their framework, the authors extensively experimented with different modules, including an image encoder, feature alignment, and a segmentation decoder. Notably, their feature alignment approach, which is based on a weighted combination of image prompt features, proves to be highly effective for anomaly segmentation by bringing together features of semantic segmentation and change segmentation. Strengths: - A novel approach is presented by formulating anomaly segmentation as change segmentation. - Extensive experiments on multiple datasets and in different model configurations demonstrate the effectiveness of the presented method, MetaUAS. In particular, the anomaly segmentation performance benefits from the specifically designed soft alignment module. - It is shown how efficiently a change segmentation dataset can be created with synthetic data, which is then used for training anomaly segmentation. - MetaUAS, trained on synthetic data, generalizes well to real-world problems. Notably, the method does not require any anomalous data for training but still outperforms existing methods that use such auxiliary data. - Unlike other existing methods, MetaUAS does not require guidance from language and solely relies on visual information. Weaknesses: - At least one normal image per class is always needed as a prompt to perform anomaly segmentation. All prompts must be processed by the encoder to extract their features, which could become expensive as the number of classes increases. - The prompt features can be stored offline as a prompt pool to reduce forward passes at test time when using MetaUAS. However, it is unclear whether a single example can represent all normal examples for a class effectively (Figure 4 highlights the issue of different prompts). Searching for the best matching normal image for each query image is obviously resource-intensive. - The anomalies shown in the industrial datasets relate to small changes in the appearance of the objects. The objects are also always displayed in the same scene. It is noted that the orientation of the object affects the anomaly segmentation performance. To further investigate the limitations, it would be interesting to see what would happen if the scenes changed more dramatically, which is realistic in open-world settings. For example, would MetaUAS be robust if the same normal object were shown with a different background? - Minor: The paper could benefit from better writing and structure. There are a few typos throughout the text, and the figures and tables could be placed differently. It is not always easy to follow. Technical Quality: 3 Clarity: 2 Questions for Authors: - Some more details on the training process would be appreciated. For instance, there is no information on how much training data was required to achieve the presented performance or how robust the model training was. The training data seems to be created in an automated fashion; to what extent does the amount of training data affect the anomaly segmentation performance? - One could use more than one image as a prompt to potentially obtain more robust anomaly scores. How efficient is the inference? In particular, compared to other existing approaches, could one afford a larger prompt pool or multiple forward passes of MetaUAS? - Relying on the changes in the visual appearance of objects seems to be a valid approach for anomaly segmentation in industrial settings. The same objects often appear in the same scene, and the anomalies are clear, small visual changes. Therefore, guidance with language seems unnecessary, as the semantic information is not crucial for this kind of anomaly segmentation. However, for other datasets on anomaly segmentation, semantic information seems crucial, such as the [SMIYC](https://segmentmeifyoucan.com/) benchmark, where any obstacle on the road is considered an anomaly. Could MetaUAS also be applied to such more challenging tasks where the normal "object" itself (e.g., the road) could have varying visual appearances as well? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have discussed limitations. Potential negative social impact have not been mentioned. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: + sm3a-Q1: The effects of the number of training data for the anomaly segmentation performance. Thanks for your interest in the effects of training scale. We are also very concerned about this issue, which has been analyzed in our ablation studies (Sec. 4.3). The corresponding experimental results are reported in Table 3(d), with the related analysis provided in Lines 361-366. The number of synthesized images and the split of training and validation are described in Sec. A of Supplementary Materials. Here, we provide further explanations. To investigate the influence of training scale on model performance, we conduct experiments with different training subsets where each one is generated by randomly sampling the original training set at various rates, such as {10%, 30%, 50%, 70%, 95%}. In Table 3(d), it can be seen that MetaUAS still works when the number of training images is small scale (e.g., 50%), and the performance can further improve when increasing the number of training samples. Unless otherwise specified, the default number of training images is 95% of the synthesized dataset. + sm3a-Q2: Compared to other existing approaches, could one afford a larger prompt pool or multiple forward passes of MetaUAS? We agree with you that using more normal image prompts can potentially enhance the robustness of the model. Additionally, we have observed that the performance gains tend to saturate when the number of normal image prompts increases to a certain level (e.g., 5) (see the response to gb5F-Q3). Assuming that each query image is provided with 5 normal image prompts, this means that the inference cost could increase by up to 5 times. However, considering that we use a lightweight model, even when 32 prompt-query pairs are processed in parallel during forward inference, the required time of each pair is only 3.2ms, as reported in Table 2. Therefore, our method can efficiently handle larger prompt pools compared to existing CLIP-based models. + sm3a-Q3: Could MetaUAS also be applied to such tasks where the normal ``object'' itself (e.g., the road, SMIYC) could have varying visual appearances as well? Our method is primarily towards open-world industrial anomaly detection, which has become an attractive yet challenging topic. It may not be suitable to directly apply our method in these scenarios where the normal ``object'' itself or its context exhibits large variations in appearance. This is because the core of our method relies on detecting changes to identify anomalies. Furthermore, we have carefully reviewed the suggested RoadAnomaly21, which contains only 100 testing images and does not provide normal images. In this dataset, the context changes of the road are more significant than that of the road itself. We believe that one feasible manner could be to first preprocess these images to remove backgrounds (i.e., the context of the road) and then apply our method to detect anomalies on the road surface. + sm3a-Q4: All prompts must be processed by the encoder to extract their features, which could become expensive as the number of classes increases. We apologize for any confusion this issue may have caused. Here, we provide some clarification that might address your concerns. For a class-specific query image, we can process prompt-query pairs online to perform anomaly segmentation without extracting offline features for normal image prompts. For a class-agnostic query image, we need to first extract offline features from a prompt pool, and then match the corresponding image prompt for the given query image to perform anomaly segmentation. In fact, in real-world industrial production lines, products are generally produced on a large scale according to specific models, so we could register normal image prompts according to actual needs. In addition, existing language-prompt-based anomaly detection methods (e.g., WinCLIP) also face similar issues, as they require offline computation of normal or anomaly textual features for all objects or textures. + sm3a-Q5: It is unclear whether a single example can represent all normal examples for a class effectively. Searching for the best matching normal image for each query image is resource-intensive. As shown in Table 1, our method requires only one normal image prompt to achieve competitive performance when faced with an open-world anomaly detection scenario. This indirectly supports that one normal image can essentially represent the normal pattern for a specific category. Undoubtedly, increasing the number of normal image prompts can further enhance performance, but this involves additional computational costs. In our MetaUAS, we have demonstrated that the optimal normal prompt yields better performance compared to a random image prompt, although the matching process does indeed bring computational cost. Exploring other efficient algorithms to obtain a better prompt is also worthwhile, such as using image-matching to pre-align the prompt image with the query image. + sm3a-Q6*: There are a few typos throughout the text, and the figures and tables could be placed differently. Thank you very much for pointing out the few typos. We will carefully review our manuscript and correct these typos in the revised version. --- Rebuttal Comment 1.1: Comment: Thanks for the efforts in the rebuttal. I will maintain my score.	Summary: The proposed method reformulates the one-shot anomaly detection task as a change detection task. The proposed method is trained with a synthesitzed change dataset, where objects are added, deleted or exchanged. Additionally, local changes are generated by pasting out-of-distribution textures on images in random locations. The proposed method is trained in a meta-learning manner where each meta-task is a prompt+query pair where a specific change must be detected and accurately segmented. The trained model is applied to industrial anomaly detection, where a single anomaly-free image is used as the prompt while a potentially anomalous image is used as the query. The model must then segment potential changes between the prompt and query image. Strengths: - The proposed method is interesting and it deviates significantly from the top performing few-shot anomaly detection methods since most recent methods are based on CLIP. - The ablation study is mostly well done and individual components seem to be evaluated properly. - The proposed method achieves good results - The change detection dataset generation is an interesting approach to creating a dataset and it is interesting that the method generalizes well to industrial inspection datasets even though the training dataset is from a different domain. Weaknesses: - Clarity issues in the paper (Exactly how prompt pooling is performed in Section 3.4.) - The way CLIP is added to the architecture in MetaAS+ is not clear. - Figure 4 - First column and the rest of the figure should be separated better. Currently it seems like the entire row depicts GT masks or queries. - The evaluation of backbones is limited to convolutional models. The only transformer is used with the MetaAS+ model. - MetaAS+ uses CLIP and achieves the best performance. Why this helps with the one-shot anomaly detection task is not intuitive and a further discussion would be beneficial for the reader. Technical Quality: 3 Clarity: 3 Questions for Authors: - Why are only convolutional networks used? Transformer networks should also work well here? A Transformer backbone in the ablation in Table 3b would be nice. - How is the visual prior from CLIP introduced to the framework. In Table 2 the complexity of MetaAS+ suggests that Eb4 and ViT are both run but this is not well explained. - Couldn’t this be extended to the few-shot setup at inference by just adding more prompts and aligning features? If no, why not? If yes, why limit the evaluation to one-shot? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: + gb5F-Q1: A Transformer backbone in the ablation in Table 3b would be nice. Thanks for your suggestion. As pointed out by DDkp and sm3a, our method is compatible with various pre-trained models, including Convolutional and Transformer architecture networks. Considering the efficiency, we employ EfficientNet-b4 (E-b4) as our encoder, which is the same as in the previous UniAD. Following your suggestion, we replace the convolution-based EfficientNet-b4 with the recent EfficientViT [1]. Specifically, three EfficientViTs with different capacities (b1, b2 and b3) are used as our encoders, and the corresponding anomaly detection results on MVTec are reported in the following table. We can see that their performance is still lower than that of EfficientNet-b4 in most metrics. Other more powerful and efficient backbones deserve further exploration. In the revised version, we will add these results. \|Backbone \|Total (M) \| Learnable (M) \|I-ROC \|I-PR \|P-ROC \|P-PR \|P-PRO\| \| :--- \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| :----: \| \|E-b4 \|22.1 \|4.6 \|91.3 \|96.2 \|94.6 \|59.6 \|82.6 \| \|EVit-b1 \|8.0 \|3.4 \|89.1 \|94.7 \|94.1 \|55.8 \|80.6 \| \|EVit-b2 \|19.5 \|4.5 \|88.5 \|94.9 \|93.0 \|56.3 \|75.1 \| \|EVit-b3 \|44.7 \|5.7 \|89.5 \|95.7 \|95.3 \|58.5 \|80.9 \| + gb5F-Q2: How the CLIP's visual prior is introduced into the framework. We apologize for the confusion about MetaUAS+. In fact, we have given more details of diverse state-of-the-art methods, our MetaUAS and its variants (MetaUAS and MetaUAS+) in the Sec. B of Supplementary Materials. Here, we give some explanations about MetaUAS, MetaUAS\ and MetaUAS+, and hope to address your concerns. The MetaUAS is capable of segmenting any anomalies only just by giving one normal image prompt. Note that the normal image prompt is randomly selected from a specific normal training set in MetaUAS. Different from MetaUAS, the MetaUAS searches the best-matched normal image as a vision prompt from the normal training set for each query image. The MetaUAS+ builds on MetaUAS, but introduces the visual prior from a CLIP model. The visual prior knowledge is obtained by computing the cosine similarity between the query feature and the corresponding prompt feature extracted from the CLIP vision encoder, which is the same as WinCLIP+ [2] and PromptAD [3]. + gb5F-Q3: Whether the evaluation can be extended from one-shot to few-shot? Our method can be flexibly extended to a few-shot setting, where a few normal image prompts are provided. However, we seek to push the limits of simple but effective methods for more general and challenging settings of anomaly segmentation. To this end, we explore universal anomaly segmentation from the perspective of a framework paradigm, where synthesized images are considered for one-prompt meta-learning, geometrical variations between prompt and query images can be effectively handled, and the inference is more efficient and simple without any guidance from language or fine-tuning on target datasets. Here, we present a straightforward manner to extend our approach from one-shot to few-shot. We use an average of all shots’ predictions as the final anomaly segmentation map. In the following table, we can observe a significant performance improvement from 1-shot to 3-shot. However, the performance tends to stabilize when reaching 5-shot. We leave it to future work on how to more elegantly extend MetaUAS from one-shot to few-shot. \|Shot \|I-ROC \|I-PR \|P-ROC \|P-PR \|P-PRO\| \| :--- \| :----: \| :----: \| :----: \| :----: \| :----: \| \|1 \|91.3 \|96.2 \|94.6 \|59.6 \|82.6\| \|3 \|92.7 \|96.9 \|95.6 \|63.4 \|85.5\| \|5 \|93.0 \|97.1 \|95.9 \|63.9 \|86.1\| + gb5F-Q4: How prompt pooling is performed? We obtain a feature representation by global average pooling on the highest-level feature from the encoder. For a class-agnostic query image, we match the corresponding normal image prompt using the cosine similarity between the query feature representation and all offline prompt feature representations. The above details have been given in Lines 265-270. + gb5F-Q6: A further discussion of MetaUAS+. Thanks for your suggestion. The MetaUAS+ can be seen as an ensemble of MetaUAS* and the visual prior knowledge of CLIP. Therefore, the performance improvement is intuitive from an ensemble learning perspective. We will add the discussion to the revised version. [1] EfficientViT: Lightweight Multi-Scale Attention for High-Resolution Dense Prediction, ICCV 2023. [2] WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation, CVPR 2023. [3] PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection, CVPR 2024. --- Rebuttal Comment 1.1: Comment: Thanks for the insightful rebuttal. The adaptation to the few-shot scenario is interesting. It would also be interesting to see the performance of the MetaUAS* on the few-shot setup due to the difference in performance to MetaUAS in the one-shot scenario. But otherwise my concerns have been addressed. --- Reply to Comment 1.1.1: Title: Response to Reviewer Concerns on Few-Shot MetaUAS* Comment: We are pleased to hear that our rebuttal has addressed most of your concerns, and we appreciate your continued engagement in this discussion. In response to your interest in the few-shot MetaUAS, we have included the results in the table below. \|Shot \|I-ROC \|I-PR \|P-ROC \|P-PR \|P-PRO\| \| :---: \| :---: \| :---: \| :---: \| :---: \| :---: \| \|1 \|94.2 \|97.6 \|95.3 \|63.7 \|83.1\| \|3 \|95.0 \|98.0 \|96.2 \|66.0 \|85.9\| \|5 \|95.2* \|98.1 \|96.4 \|66.4 \|86.4\| It is important to note that the one-shot MetaUAS* is based on the best-matched normal image prompt. Consequently, for the 3-shot and 5-shot MetaUAS, we utilized the top 3 and top 5 normal image prompts, respectively. As shown, the performance trend of MetaUAS is similar to that of MetaUAS (gb5F-Q3) as the shot number of normal image prompts increases. We acknowledge that our current extension from one-shot to few-shot is quite straightforward. We consider it a preliminary approach, leaving room for future work. Thank you for your feedback.	Summary: This paper considers the anomaly segmentation task as a change segmentation task. Then the large-scale image pairs with object-level and local region changes are synthesized to train a universal anomaly segmentation framework MetaUAS. This only needs one normal image as the prompt. The soft feature alignment module is proposed to handle geometrical variations between prompt and query images. This method achieves state-of-the-art performance on MVTec AD, VisA, and Goods datasets. Strengths: 1. This paper synthesizes large-scale image pairs with object-level and local region changes, and desiges a universal anomaly segmentation framework that only need one normal image as prompt. This idea is novel and has practical implications. 2. This paper propose the soft feature alignment module to handle geometrical variations between prompt and query images. This is a novel, simple and effective alignment method. 3. This method is compatible with various feature extractors, and is not limited to CLIP like some previous methods. 4. The authors conducted experiments on three anomaly detection datasets. Significant improvement has been achieved, which proves the powerful generalization ability of this method. Weaknesses: 1. The three fused features and two low-level original features are input into decoder. How these features are used in the decoder needs to be explained in more detail. 2. CLIPSeg [1] also uses one image as the prompt to perform segmentation tasks. It is beneficial to discuss and compare with CLIPSeg in this paper. [1] Lüddecke T, Ecker A. Image segmentation using text and image prompts[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 7086-7096. Technical Quality: 3 Clarity: 4 Questions for Authors: See the weaknesses. Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: None. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: + DDkp-Q1: How these features are used in the decoder needs to be explained in more detail. Our method implements the decoder using standard UNet. Furthermore, we also compare the UNet with the FPN decoder in Table 3(c). In our paper, we omit the details of UNet and FPN because they are two popular modules that have been standardized and integrated into the [PyTorch library](https://github.com/qubvel-org/segmentation_models.pytorch/tree/main/segmentation_models_pytorch). As we know, the decoder of UNet and FPN is widely used in dense visual prediction tasks, such as semantic segmentation and object detection. They use a top-down structure with lateral connections to fuse multi-scale input features and output a single high-resolution feature map. Considering reproducibility, we will make the codes and models of MetaUAS available. + DDkp-Q2: It is beneficial to discuss and compare with CLIPSeg. Thank you for pointing out a related work, CLIPSeg [1]. The CLIPSeg has one-shot semantic segmentation capabilities by providing one support image to the pre-trained CLIP model. However, its better performance still struggles with the powerful CLIP model and textual prompts. Different from the CLIPSeg, our method is compatible with various pre-trained models and does not require any guidance from language, as pointed out by you and Reviewer sm3a. [1] Image Segmentation Using Text and Image Prompts, CVPR 2022. --- Rebuttal Comment 1.1: Title: Response to Author Rebuttal Comment: Thank the authors for their response. I'm inclined to keep my rating. This is a very interesting and novel paper, but this method does not seem to be very robust to different normal image prompts, which makes it impossible to continue to improve the rating to Strong Accept. --- Reply to Comment 1.1.1: Title: Response to Reviewer Concerns on Robustness Comment: Thank you for your thoughtful feedback and for maintaining your rating (Accept). We appreciate your recognition of the novelty and interest in our work. We understand your concerns regarding the robustness of our method to different normal image prompts and would like to address these as follows to clarify our approach further. Firstly, as demonstrated in Figure 4, our MetaUAS shows robustness to different normal image prompts within the same category, particularly in those categories with significant geometric variations, such as grid and screw. This indicates that our method can effectively handle different prompts within the same category. Secondly, to quantitatively evaluate the robustness of our MetaUAS, we report performance metrics including the mean and variance based on the results from five different normal image prompts (generated with random seeds). It can be observed that our MetaUAS achieves higher means and lower variances compared to WinCLIP+ across most metrics, demonstrating its superior stability. Lastly, we have found that matching the optimal prompt (as shown in MetaUAS* in Table 1) or using few-shot normal prompts (e.g., 5-shot) can further result in significant performance improvements or enhance robustness (as detailed in the gb5F-Q3). We hope these explanations adequately address your concerns regarding the robustness of our method. We are grateful for the opportunity to discuss these aspects further and thank you once again for your valuable feedback.	Rebuttal 1: Rebuttal: We thank all reviewers (DDkp, gb5F, sm3a and bbUz) for your insightful comments. The reviewers believe that the proposed universal anomaly segmentation framework is novel and interesting (DDkp, gb5F and sm3a), simple, effective and efficient (DDkp and sm3a) and compatible (DDkp), the ablation study is well (gb5F and sm3a), our method achieves good and encouraging results, or powerful generalization ability (DDkp, gb5F, sm3a and bbUz). Next, we respond to the concerns of each reviewer one by one.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ZOPP: A Framework of Zero-shot Offboard Panoptic Perception for Autonomous Driving	Accept (poster)	Summary: The paper introduces a novel framework, termed Multi-modal Zero-shot Offboard Panoptic Perception (ZOPP), specifically designed for autonomous driving applications. This innovative approach integrates zero-shot recognition capabilities with 3D representations generated from point cloud data, enhancing the model's ability to interpret complex driving environments without the need for extensive labeled datasets. Strengths: 1. The proposed work by the authors is significant and unique. 2. The proposed ZOPP approach is more capable of handling diverse types of data compared to any existing model, making it more robust. Weaknesses: 1. The manuscript is severely deficient in addressing the issues arising from labelling constraints and offers no substantive solutions within the proposed model to mitigate these problems. 2. The authors have egregiously failed to provide any details regarding the computational resources utilized, including the computational cost and the necessary hardware specifications, leaving the readers in the dark about the feasibility and scalability of the proposed approach. 3. The manuscript is conspicuously lacking in structure, offering an incoherent presentation of the proposed work, which undermines the clarity and comprehensibility of the research. 4. Despite the purported significance of the proposed work, the evaluation is lamentably limited, providing insufficient empirical evidence to substantiate the claimed advancements and benefits. 5. The manuscript is grossly inadequate in offering detailed information about the training and testing procedures, thereby failing to provide the essential methodological transparency required for replication and validation of the results. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The authors mention that data labeling is a crucial factor and a primary hurdle in training within existing work. Could the authors elaborate on how they propose to counter this issue or if they have developed any specific labeling mechanism to address this challenge? 2. How have the authors handled data from multi-view cameras? Specifically, how do they manage instances of the same object appearing in different views? 3. (With respect to Fig. 3), how does the proposed convolution filtering operation for removing background 3D points from foreground pixels account for varying disparity values at the upper edges of foreground objects? What specific kernel design is employed to optimize this process? 4. In the context of disparity occlusion handling around the upper edges of foreground objects (as seen in Fig. 3), what is the mathematical formulation for projecting 3D background points into the pixel regions of foreground objects? Additionally, how does the proposed convolution filtering algorithm differentiate between true foreground and occluded background points at a sub-pixel level? 5. What is the likelihood of the model's success in real-world applications? Have the authors conducted any real-world testing, and if so, what were the results? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. The abstract is grossly inadequate for the proposed topic. 2. The manuscript fails to deliver a coherent and sequential exposition of the proposed work. 3. The authors have entirely neglected to address the computational cost associated with the work or necessary for deployment. 4. The manuscript is deficient in presenting a comparative analysis with the visual results of existing works. Flag For Ethics Review: ['Ethics review needed: Data privacy, copyright, and consent'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response 4.1 Data labeling issue. (Weakness 1 and Question 1) Previous literatures focus on generating high-quality 3D detection results as auto labels with offboard perception fashion. However, they still need high-quality human labels of the AD dataset as a prerequisite for training the whole pipeline. This is precisely the issue of data auto-labeling we aim to address. Therefore, we propose a unified framework with a compact structure to support various perception tasks of AD scenes in an offboard manner, without any requirements of human labels from AD dataset. To be specific, each proposed module and stage of our method do not rely on any human labels to generate the corresponding perception results. Naturally, we can leverage ZOPP as a cold-start paradigm for existing auto-labeling methods. ## Response 4.2 Computational resources and hardware specifications (Weakness 2) We have introduced the corresponding computational resources in Sec. C (Implementation Details) of Appendix, please kindly refer to it. The entire pipeline of our proposed method is a lightweight and compact framework, which does not rely heavily on many computational resources. Only the point completion module and the resconstruction module need to train the networks. We utilize 1 NVIDIA A100 to train the point completion network, and 4 NVIDIA A100 to accelerate the reconstruction with multi-processing settings. ## Response 4.3 Writing structure and incoherent presentation (Weakness 3) We will improve the presentation and polish the structure in the revised version. ## Response 4.4 Insufficient evidence (Weakness 4) Could you please kindly specify which aspect of the content "insufficient evidence" refers to? ## Response 4.5 Training and testing procedures (Weakness 5) We have presented the detailed training and resting procedures, hyper parameters, and settings for each module in Sec. C (Implementation Details) of Appendix, please kindly refer to it. ## Response 4.6 Instance management across multi-view cameras (Question 2) In section 3.1 of the main contents, we have proposed and introduced a multi-view mask tracking module to manage the object instances across multiple views. Specifically, we designed a simple yet effective similarity cost involving the computation of appearance and location similarities to facilitate object association. The appearance similarity is compared across different objects by computing the cosine distance of the visual features. The location similarity is derived by concatenating the images of all viewpoints in a panoramic order, followed by normalizing the pixel distances along the horizontal axis for each object. Therefore, objects with large similarity scores would be associated together with the same instance ID. We will then manage all the object instances with their corresponding unique IDs in the following stages of our approach. ## Response 4.7 Removing background 3D points at the upper edges of foreground objects (Question 3) Firstly, since LiDARs are always equipped higher than multi-view cameras on autonomous vehicles [1,2,3], the parallax occlusion issue will thereby arise at regions around the upper edges of foreground objects, rather than the center or bottom parts of foreground objects. Secondly, we have designed an experiential threshold of distance $\theta$ to determine whether the projected background pints should be filtered out. Regardless of the disparity differences among these projected background points, as long as they exceed the threshold, we will filter them out. ## Response 4.8 Specific kernel design of parallax occlusion filtering (Question 3) The design of the kernel is influenced by the configuration of the sensors, specifically determined by the beam numbers and rotating frequency of the LiDAR. They respectively determine the vertical and horizontal densities/resolutions of the point cloud projected onto 2D image plane. Therefore, if the LiDAR has a large number of beams and a high rotation frequency, we need to use a smaller kernel size to handle the dense projection points and improve the accuracy of the filtering operation. Conversely, if the density of projection points is lower, we can increase the kernel size and step size along the vertical and horizontal directions, thereby enhancing the algorithm's operational efficiency without compromising filtering accuracy. ## Response 4.9 Mathematical formulation of projection (Question 4) The process of projecting all 3D point clouds onto a 2D image plane follows the same projection formula, as discussed in Sec. 3.2.1. There is no separate projection formula for background points; rather, they appear in the region of foreground objects in the 2D image due to the parallax occlusion problem mentioned above. This issue is precisely what our proposed method aims to address. ## Response 4.10 Sub-pixel level results (Question 4) Currently, we do not support sub-pixel level calculations. All 3D point cloud projections onto the 2D image are rounded to the nearest integer pixel value coordinates, to obtain pixel-level semantic and instance ID information. Additionally, the semantic and instance mask results generated in the previous stage of our method are also at the pixel level. ## Response 4.11 Likelihood of applying in real-world applications (Question 5) In our experiments, we leverage WOD as the primary benchmark to assess the effectiveness of our method. WOD is one of the large-scale autonomous driving dataset collected from real-world commercial vehicles. In addition to evaluating the main results across various perception tasks, we also perform auto labeling applications. As shown in Sec. D.4 of Appendix, we generate 3D boxes on 5% of training set as auto labels to train the onboard detection model. This experiment demonstrates that our method could generate comparable auto labels and serve as an efficient cold-start approach for existing perception models. --- Rebuttal Comment 1.1: Comment: In response to the feedback provided, the authors have agreed to revise the manuscript and address several of its current limitations. However, certain critical aspects of the manuscript still require closer examination and further refinement. Notably, significant limitations remain in areas such as the evaluation phase, and more results proof will be required, which strengthens the proposed model's useability. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your efforts and valuable suggestions during the review and response periods. We are wondering whether you have any more specific questions or suggestions after our response. Your willingness to accept our paper is truly appreciated.	Summary: Offboard perception creates 3D labels for autonomous driving scenes. Current methods are limited and don't match human recognition levels. The authors developed a new framework called Zero-shot Offboard Panoptic Perception (ZOPP), which combines advanced recognition technologies with 3D point cloud data. ZOPP is a pioneering approach in automatic labeling for driving scenes. They validated its effectiveness on the Waymo dataset and downstream applications with good performance. Strengths: 1. Using foundation model to generate labels for autonomous driving is meaningful task. it can help to push a great advance in the realm of autonomous driving. 2. Experiments are extensive with good results. Weaknesses: 1. the proposed work just combines some off-the-shelf modules together, and use some basic mathematical thing to support the alignment, which looks not very innovative as NeurIPS requested, although the validation results is competitive. 2. The presentation need to be improved. sometimes, it is hard to guess what you refer to. Technical Quality: 3 Clarity: 3 Questions for Authors: refer to weaknesses Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: refer to weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful to you for recognizing our efforts in addressing your concerns during the reviewing process. ## Response 3.1 Lack of novelty (Weakness 1) Perception and understanding play a vital role in current data-driven autonomous driving. Previous literatures focus on alleviating the burdens if human labor and the cost of labeling. And we found several challenges in this field: - Only 3D object detection task is supported to generate auto labels in an offboard manner. - Still require huge amounts of data with high-quality human-labels. - Lack the capabilities of open-set and zero-short settings. Therefore, we tackle these challenges by proposing a unified framework for various perception tasks in an offboard manner without the requirements of human labels in AD scenes. Although our method incorporates several existing foundation models, previous research has not explored these aspects to address the practical needs of auto labeling in AD. To the best of our knowledge, we are the first to propose such kind of work. ## Response 3.2 Confusing presentation (Weakness 2) We will revise the writing carefully in the next version. --- Rebuttal Comment 1.1: Comment: I have read the rebuttal --- Rebuttal 2: Comment: We sincerely appreciate your efforts during the review and responses. We are wondering whether our response has addressed your concerns and do you have any more suggestions or questions. Moreover, if you find our response satisfactory, we kindly invite you to consider the possibility of improving your rating.	Summary: ZOPP proposes an offboard auto-annotation method to achieve lidar 3D detection as well as the occupancy label without any annotation data. The whole pipeline ensembles several models including the SAM-track and point cloud completion model. By using some post-processing to complete the Strengths: 1. The idea is straightforward. 2. The writing is clear. Weaknesses: Lack of novelty. It ensembles several SOTA methods but lacks in its contribution. Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: See weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are thankful to you for raising such important concerns and questions about our work, we highly appreciate your efforts during review process. ## Response 2.1 Performance of [0, 30]m on Waymo dataset (Weakness 1) We follow the experiment setting in prior zero-shot 3D detection work [1] to ensure consistency with other methods, to report the performance of [0, 30]m on Waymo open dataset in Table 2 of the main paper. Addtionally, results for full distances are presented in Table 1 of the main paper, Table 6 and Table 8 of Appendix. Please kindly refer to them. Furthermore, we evaluate the performance of several methods, PV-RCNN, Voxel-RCNN, DetZero and SAM3D [1], across different distance ranges and levels of occlusion. The distance ranges include the following segments: [0, 30]m, [30, 50]m, [50, +inf)m. The occlusion levels are categorized into three grades based on the extent to which the object is obscured in the image perspective. \| \| \| Total \| \| 0-30m \| \| 30-50m \| \| 50+m \| \| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \| Method \| Training Data \| L1 \| L2 \| L1 \| L2 \| L1 \| L2 \| L1 \| L2 \| \| VoxelRCNN \| train set \| 74.24 \| 65.91 \| 88.85 \| 87.52 \| 73.09 \| 66.71 \| 52.75 \| 40.43 \| \| PVRCNN \| train set \| 74.31 \| 65.89 \| 89.52 \| 88.19 \| 72.74 \| 66.24 \| 52.38 \| 40.12 \| \| DetZero \| train set \| 89.49 \| 83.34 \| 96.64 \| 95.90 \| 88.84 \| 84.37 \| 78.32 \| 66.77 \| \| SAM3D \| - \| 6.90 \| 5.88 \| 19.51 \| 19.05 \| 0.029 \| 0.026 \| 0.0 \| 0.0 \| \| ZOPP (ours) \| - \| 37.56 \| 35.61 \| 42.31 \| 41.16 \| 35.14 \| 33.86 \| 29.89 \| 28.67 \| From the table, as the distance increases, the performance of all methods are decreased. Specially, VoxelRCNN decreases (L1 AP) with a ratio of 17.74% and 40.63% on [30, 50]m and [50, +inf)m compared [0, 30]m, PVRCNN decreases with 18.74% and 41.48%, our method decreases with 16.94% and 29.35%. The reason is that our method could utilize the entire temporal information in the point cloud sequence by our mask tracking module to generate the unique object ID. Therefore, we can overcome the influence of distance compared to other onboard methods, especially at farther ranges. Reference: [1] Dingyuan Zhang et al. Sam3d: Zero-shot 3d object detection via segment anything model. arxiv preprint, 2023. ## Response 2.2 Performance of occlusions. (Weakness 1) We report the performance of the overall and the occlusion part on Waymo open dataset to compare with other methods. The occlusion levels are defined based on whether the objects are obscured in the image perspective, which are provided by WOD. \| Method \| Training Data \| All \| Occlusion \| \| :---: \| :---: \| :---: \| :---: \| \| VoxelRCNN \| train set \| 74.24 \| 58.39 \| \| PVRCNN\| train set \| 74.31 \| 58.47 \| \| SAM3D\| - \| 6.90 \| 4.74 \| \| ZOPP (ours) \| - \| 37.56 \| 33.42 \| As we can see, VoxelRCNN and PVRCNN decrease with a ratio of 21.35% and 21.32%, SAM3D decreases with a ratio of 31.30%, our method decreases with a ratio of 11.02%. Our method could overcome the influence of occlusion by leveraging the temporal contexts with our mask tracking module. ## Response 2.3 Failure pattern analysis (Weakness 1) We have briefly summarized some representative challenging scenarios in Sec. 6 (Limitations and Broader Impacts) of the main contents of our submitted paper. Firstly, our method would fail to effectively recognize similar object categories (e.g., construction vehicle, truck, trailer) and some uncommon object categories (e.g., tree trunk, lane marker) with the foundation models (Grounding-DINO). Since this is the first stage of our entire method, it will result in subsequent stages lacking the output of corresponding perception results, such as 3D segmentation and occupancy prediction. Secondly, neural rendering methods may encounter numerous challenges in street-view scenes, constrained by practice factors (adverse weather conditions, sensor imaging issues), such as camera overexposure. In these scenarios, where it is impossible to generate geometrically plausible 3D reconstructions, our occupancy decoding will fail. Please kindly refer to Fig. 2 of our global response PDF file to see the visualization. ## Response 2.4 Lack of novelty (Weakness 2) Please kindly refer to Response 3.2 for Reviewer Swqu. --- Rebuttal 2: Comment: We sincerely appreciate your valuable and helpful suggestions. We are wondering whether you have any more suggestions or questions after our response. Specifically, do you have any questions about the performance comparison based on distances and occlusions? Moreover, if you find our response satisfactory, we kindly invite you to consider the possibility of improving your rating. --- Rebuttal Comment 2.1: Comment: Thank you for the detailed explanation! It resolves my concerns on its performance. --- Reply to Comment 2.1.1: Comment: We are grateful to you for recognizing our efforts to address your concerns during the response process. Your feedback has been instrumental in enhancing the quality of our work, especially for the comprehensive comparisons based on the distances and occlusions. We look forward to continuing to meet your expectations in the final version of our paper.	Summary: This paper introduces ZOPP, a framework for zero-shot panoptic perception of autonomous driving scenes. Leveraging image foundation models, ZOPP is able to perform zero-shot 3D object detection, 3D semantic segmentation, 3D panoptic segmentation, and 3D occupancy prediction, the first zero-shot model of its kind. Experiments on the Waymo dataset achieve strong performance. Strengths: 1. The paper is well-written and easy to follow. 2. ZOPP, to the best of my knowledge, is the first zero-shot panoptic perception model for autonomous driving constituting a significant novelty. 3. Experiments and ablation studies are extensive. 4. A successful zero-shot panoptic perception framework is highly useful for autolabeling driving scenes, giving this paper a high likelihood for significant impact. Weaknesses: The most significant weakness is that zero-shot performance still significantly lags behind models trained using human-annotated labels. However, as this is the first work of it's kind, this is acceptable. Technical Quality: 4 Clarity: 3 Questions for Authors: Do the authors have any quantitative results for their point completion module? This would be interesting to see. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The authors adequately address limitations and broader impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate your positive acknowledgment of our work. We are pleased to provide the supplemental responses. ## Response 1.1 Performance gap between our method and fully supervised methods (Weakness 1) Yes. Indeed, our zero-shot method still exhibits a notable gap compared to fully supervised training methods when applied to datasets with abundant annotations. However, in scenarios where annotated data is scarce, our approach leverages autonomously generated perceptual outputs and can recognize objects whose classes were never labeled before. For instance, it can output 3D detection boxes for traffic signs and traffic lights, which are visualized in Fig. 1 of our global response PDF file. Moreover, our research demonstrates the great potential of integrating foundation models into the field of autonomous driving. This integration is poised to advance traditional tasks substantially. Looking forward, as foundation models continue to enhance their performance, we believe they can be seamlessly integrated into our framework. Through continuous performance iteration and optimization, we anticipate further enhancing the effectiveness of our approach. ## Response 1.2 Quantitative results of point completion module (Question 1) We have supplemented the quantification results for the point cloud completion module. On our experimental set, there are around 4106 object samples with complete point clouds (obtained by merging all the point clouds of each object across the entire scene sequence). We then sampled the partial clouds and used them as input to generate 4096 points as a completion. The L1 Chamfer distance performance is summarized below. It shows the great effectiveness of our point completion module. As an additional reference of its effectiveness, please kindly refer to Table 8 of Appendix in our submission to see the performance differences in 3D box interpretation before and after point cloud completion. \| \| Average \| Vehicle \| Pedestrian \| Cyclist \| \|---\|---\|---\|---\|---\| \| Samples \| 4106 \| 2891 \| 1215 \| 102 \| \| CD \| 7.19 \| 5.35 \| 6.39 \| 9.84 \| --- Rebuttal Comment 1.1: Comment: I have read the author's rebuttal. Their response adequately addresses the questions I raised, and given the promise I believe this paper holds, I keep my rating as is. --- Reply to Comment 1.1.1: Comment: We deeply value your meticulous review and are pleased that our responses have effectively addressed your questions and concerns. Your willingness to accept our paper is truly appreciated. We extend our profound appreciation for your insightful questions and invaluable suggestions, which undoubtedly contribute to elevating our manuscript's scholarly caliber.	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for their valuable comments and suggestions. We are sincerely grateful to the reviewers for dedicating their time and effort to review our work. We are delighted to see reviewers commenting on our paper with "significant novelty", "significant impact", "extensive and competitive experimental results", and "straightforward idea". In this rebuttal, we try our best to solve the concerns of reviewers. We summarize the important and common concerns in the following: ## Lack of novelty. (Reviewer aSzs, Reviewer Swqu) Perception and understanding play a vital role in current data-driven autonomous driving (AD). Previous literatures [1,2] focus on alleviating the burdens of human labor and the cost of labeling. And we found several challenges in this field: - Only 3D object detection task is supported to generate auto labels in an offboard manner. - Still require huge amounts of data with high-quality human labels. - Lack the capabilities of open-set and zero-short settings to generalize to new scenarios and dataset. Therefore, we tackle these challenges by proposing a unified framework for various perception tasks in an offboard manner without the requirements of human labels in AD scenes. Although our method incorporates several off-the-shelf foundation models, previous research has not explored these aspects to address the practical needs of auto labeling in AD. To the best of our knowledge, we are the first to establish such kind of work. ## Performance of [0, 30]m on Waymo and more comparisons of different distances and occlusions. (Reviewer aSzs) We follow the experiment setting in prior zero-shot 3D detection work [1] to ensure consistency with other methods, to report the performance of [0, 30]m on Waymo open dataset in Table 2 of the main paper. Additionally, results for full distances are presented in Table 1 of the main paper, Table 6 and Table 8 of Appendix (in our submission). Please kindly refer to them. Furthermore, we evaluate the performance of several methods, PV-RCNN, Voxel-RCNN, DetZerom and SAM3D, across different distance ranges and occlusions. The distance ranges include the following segments: [0, 30]m, [30, 50]m, [50, +inf)m. The occlusion levels are defined based on whether the objects are obscured in the image perspective, which are provided by WOD. Specifically, for different distances, VoxelRCNN decreases (L1 AP) with a ratio of 17.74% and 40.63% on [30, 50]m and [50, +inf)m compared [0, 30]m, PVRCNN decreases with 18.74% and 41.48%, our method decreases with 16.94% and 29.35%. For the occlusion part, VoxelRCNN and PVRCNN decrease with a ratio of 21.35% and 21.32%, our method decreases with a ratio of 11.02%. The reason is that our method could utilize the entire temporal information in the point cloud sequence by our mask tracking module to generate the unique object ID. Therefore, we can overcome the influence of distance compared to other onboard methods, especially at the farther ranges, and occlusion regions. ## Data labeling issues and real-world applications. (Reviewer 96Yj) Previous literatures focus on generating very high-quality 3D detection results as auto labels with offboard perception fashion. However, they still need high-quality human labels of the AD dataset as a prerequisite for training the whole pipeline. This creates a "chicken-or-egg" problem for auto labeling: when faced with a new dataset, how can it be automatically annotated with these methods? This is precisely the issue of data auto labeling we aim to address. Naturally, we can leverage ZOPP as a cold-start paradigm for existing auto-labeling methods. For example, the detected 3D bounding boxes and segmentation results can serve as the labels to train other onboard perception models. We also provide the corresponding experiment in Sec. D.4 of Appendix (please kindly refer to it), demonstrates that our method could generate comparable auto labels and serve as an efficient cold-start approach for existing perception models. ## Confusing writing structure and presentation. (Reviewer Swqu, Reviewer 96Yj) We will improve the presentation and polish the structure in the revised version. ## Failure pattern and visualization. (Reviewer aSzs) We have briefly summarized some representative challenging scenarios in Sec. 6 (Limitations and Broader Impacts) of the main contents of our submitted paper. Let's discuss the details of failure patterns below. Firstly, our method would fail to effectively recognize similar object categories (e.g., construction vehicle, truck, trailer) and some uncommon object categories (e.g., tree trunk, lane marker) with the foundation models (Grounding-DINO). Since this is the first stage of our entire method, it will result in subsequent stages lacking the output of corresponding perception results, such as 3D segmentation and occupancy prediction. Looking forward, as foundation models continue to improve their performance, we believe they can flexibly integrate into our framework. Secondly, neural rendering methods may encounter numerous challenges in street-view scenes, constrained by practice factors (adverse weather conditions, sensor imaging issues), such as camera overexposure. In these scenarios, where it is impossible to generate geometrically plausible 3D reconstructions, our occupancy decoding will fail. We also show the visualization of such failure cases in our global response PDF file. Please kindly refer to them. Please refer to the following rebuttals for other specific concerns and more details. We are looking forward to your further reply and discussion. Pdf: /pdf/101da41d6fa36f6781e49cdb6dc5d3fef9afdf8d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.