GAIA: Delving into Gradient-based Attribution Abnormality for Out-of-distribution Detection

Detecting out-of-distribution (OOD) examples is crucial to guarantee the reliability and safety of deep neural networks in real-world settings. In this paper, we offer an innovative perspective on quantifying the disparities between in-distribution (ID) and OOD data -- analyzing the uncertainty that arises when models attempt to explain their predictive decisions. This perspective is motivated by our observation that gradient-based attribution methods encounter challenges in assigning feature importance to OOD data, thereby yielding divergent explanation patterns. Consequently, we investigate how attribution gradients lead to uncertain explanation outcomes and introduce two forms of abnormalities for OOD detection: the zero-deflation abnormality and the channel-wise average abnormality. We then propose GAIA, a simple and effective approach that incorporates Gradient Abnormality Inspection and Aggregation. The effectiveness of GAIA is validated on both commonly utilized (CIFAR) and large-scale (ImageNet-1k) benchmarks. Specifically, GAIA reduces the average FPR95 by 23.10% on CIFAR10 and by 45.41% on CIFAR100 compared to advanced post-hoc methods.

Paper

References (42)

Scroll for more · 30 remaining

Similar papers

Peer review

Reviewer CiZ85/10 · confidence 4/52023-06-30

Summary

The work looks at leveraging gradient-level attribution information in order to detect semantically shifted OOD samples. In particular, the paper proposes two post-hoc OOD detection methods that leverage the extracted gradient attribution called GAIA-A and GAIA-Z. Both the proposed GAIA-A and GAIA-Z methodologies show strong empirical performance across a wide range of OOD detection tasks.

Strengths

1. The paper provides analysis on an underexplored domain relating to attribution gradient and how to leverage this information for OOD detection. 2. The resulting post-hoc OOD detectors are simple to implement and show strong empirical performance across a wide range of OOD detection tasks.

Weaknesses

There are several other post-hoc methods that have not been included in the empirical evaluation. For example, KNN[1] and Lee and AlRegib [2] would provide fair points of comparison. In addition, if possible, the reviewer would also encourage the authors to include deviations with each empirical result. [1] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, 2022. [2] Jinsol Lee and Ghassan AlRegib. Gradients as a measure of uncertainty in neural networks. In 2020 IEEE International Conference on Image Processing (ICIP), pages 2416–2420. IEEE, 2020.

Questions

1. On line 227 the authors hypothesize that "channel-wise average abnormality is better suited for application in scenarios with a large label space." However, the results from CIFAR-100 setting seem to indicate that the choice between GAIA-A and GAIA-Z is not simply based on how large the label space is. 2. On line 151, the authors discuss observations on the number of zero partial derivations for OOD samples. Are there empirical analyses that corroborate these observations?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The reviewer would recommend the authors consider adding additional points of comparison as stated in the weakness section above as well as additional runs of the method under models trained with differing seeds.

Reviewer CiZ82023-08-13

Response to Author Rebuttal

The reviewer would like to thank the author for providing the additional experimental evaluations and answering all the existing questions. **Fair comparison with KNN and Lee and AlRegib** These experimental results of Lee and AlRegib match with prior expectations and it is encouraging that GAIA is able to outperform all the prior gradient-based and KNN distance-based OOD detection methodologies. **Other comments and questions** I thank the reviewer for the clarification and the additional edits. Unfortunately, my general thoughts on the paper remain consistent so I won't raise the review score any higher but I want to encourage the authors to further organize the paper in an effort to improve clarity.

Authorsrebuttal2023-08-14

Thanks for the comments

We appreciate the response from the reviewer. In our rebuttal, it seems that we have addressed the suggestions and questions raised by the reviewer. Please kindly inform us if the reviewer has any other ongoing concerns or questions? Regarding the issue of the paper's organization, we attach great importance to it. We have already made adjustments and are continuing to refine it further. Our improvements: - We have relocated the "Related Work" section to the end of the manuscript to ensure a smoother flow of our idea. - In Section 4, we have repositioned Equations 3 and 4 to an additional theoretical analysis section at the end of the paper. Detailed explanations of these equations are provided in the appendix for greater clarity. Moreover, in order to enhance the reader's comprehension of our proposed idea, we have incorporated the visualization that elucidates the connection between the attribution phenomenon and the two abnormalities. - For the ablation experiments, we have restructured the sequence and introduced guiding statements at the beginning to clarify the logical flow of the ablation study. Our ablation study begins by validating the effectiveness of each step of the method, moving from outermost to innermost. Moreover, all Figures and Tables have been arranged coherently, following the sequence from GAIA-A to GAIA-Z. - We have carefully corrected the typos (grammar, table formatting, descriptive details, mathematical expressions, and so forth).

Reviewer qAeX5/10 · confidence 3/52023-07-02

Summary

The proposed gradient-based attribution method in this paper is a promising approach that can help distinguish between ID and OOD patterns. By analyzing the uncertainty that arises when models attempt to explain their predictive decisions, the method can provide a more robust and reliable approach to detecting OOD data, which is superior over previous works in gradient-based OOD detection. The authors test their approach on well-known OOD detection benchmarks such as ImageNet and CIFAR, which are widely used in computer vision research. The results demonstrate that the proposed approach is effective in detecting OOD data, outperforming state-of-the-art methods by a significant margin.

Strengths

1. Innovative perspective on quantifying disparities between in-distribution (ID) and out-of-distribution (OOD) data based on analyzing the attribution of embedding features. This approach offers a new perspective on detecting OOD data. 2. Introduces two forms of abnormalities for OOD detection, i.e., the zero-deflation abnormality and the channel-wise average abnormality, which may help to identify OOD data more accurately and effectively. 3. Proposes GAIA, a simple and effective approach that incorporates Gradient Abnormality Inspection and Aggregation, which can be readily applied to pre-trained models without further fine-tuning or additional training. 4. Demonstrates superior performance on both commonly utilized (CIFAR) and large-scale (ImageNet) benchmarks compared to competing approaches, reducing the average FPR95 by 26.75% on CIFAR10 and by 45.41% on CIFAR100.

Weaknesses

1. The authors claim that one can analyze the uncertainty raised when model making decisions via the gradient attribution, and it is the main contribution of this paper. However, I didn't find any theoretical explanation or heuristic justification about how the contribution value is related to uncertainty. With so many systematic analysis in Section 4.1, I think the only key point in supporting why the proposed method works is "Zero-deflation Abnormality" (forgive me if I misunderstand), but it does not explain why the suggest method is a good indicator of prediction uncertainty. 2. The authors claim using the softmax prediction across each class is a novely of their method. However, from my understanding, it is equivalent to the case in using the KL divergence between uniform distribution and model softmax prediction as the objective, which has been well discussed in GradNorm. Therefore, more or less, the author overclaim their contribution. 3. More advanced works, such as ASH, are not compared in the paper. Experimental results with larger models on ImageNet (such as ViT) should also be considered, following previous works such as [1]. 4. Since there are some previous works in studying the gradient-based OOD detection, a natural question is why the proposed method is superior over GradNorm. For example, it seems that the authors conduct gradient wrt. model outputs (instead of parameters as in GradNorm), is there any reason for your choice (either heuristically or theoretically); it seems that the authors use the number of non-zero gradients in OOD scoring, what is the superiority to GradNorm in using the L2 norm of gradients. [1] Yiyou Sun, Yifei Ming, Xiaojin Zhu and Yixuan Li. Out-of-distribution Detection with Deep Nearest Neighbors. ICML, 2022.

Questions

Please refer to the part of Weaknesses

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Please refer to the part of Weaknesses

Reviewer kfJE5/10 · confidence 4/52023-07-05

Summary

In this paper, the authors propose a novel perspective for quantifying the disparities between in-distribution (ID) and out-of-distribution (OOD) data by analyzing the uncertainty that arises when models attempt to explain their predictive decisions. They investigate the abnormality in gradient-based attribution methods when dealing with OOD data and introduce two forms of abnormalities. They further propose GAIA, a simple and effective approach based on gradient of attribution models for OOD detection. Experimental results demonstrate that GAIA outperforms state-of-the-art methods on CIFAR and ImageNet benchmarks.

Strengths

(1) The paper offers an innovative perspective on quantifying the disparities between ID and OOD data by analyzing the uncertainties in gradient-based attribution methods, based on zero-deflation abnormality and channel-wise average abnormality. (2) The proposed GAIA approach is simple and effective and does not require further fine-tuning or additional training, achieving superior performance to previous SOTAs.

Weaknesses

(1) The paper lacks clarity and organization in presenting the proposed approach and the experimental results. For example, the mathematical derivation in Section 4 is difficult for readers to follow the flow of ideas. It is unclear to me how Eq. 4 is derived from Eq.3, and what is |·| refer to? (2) Ablation studies are comprehensive but messy. It is hard to follow the logic of paper writing. Please give an overview of ablation before the details of each ablation study for better understanding. (3) I am curious about the effect of combining GAIA-Z and GAIA-A while no relevant experiment or explanation about it. (4) The paper writing should be further improved. There are lots of mistakes. For example, the highlight in Table 2 is wrong (some bests and second bests are upside-down); In line 169 of Page 5, the range of c_i should be placed below the argmax. (5) Although the motivation is clear and the proposal is effective, the lack of clarity, implementation details and deficient writing quality are major weaknesses that impact the overall quality of the paper. I recommend you can polish this paper carefully and submit it to other conferences such as cvpr or iclr, which will be good work.

Questions

(1) I am curious about the effect of combining GAIA-Z and GAIA-A while no relevant experiment or explanation about it. (2) The paper writing should be further improved. There are lots of mistakes. For example, the highlight in Table 2 is wrong (some bests and second bests are upside-down); In line 169 of Page 5, the range of c_i should be placed below the argmax.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer Exp86/10 · confidence 4/52023-07-07

Summary

This paper presents an approach to OOD detection in deep neural networks. The authors propose a method based on analyzing the uncertainty that emerges when models attempt to rationalize their predictive decisions. The abnormalities are found by using two strategies: the zero-deflation abnormality that takes advantage of the observation that attribution gradients in OOD data have more zero values than in-distribution and channel-wise average abnormality that captures variations in the feature maps of OOD data compared to in-distribution data. The experiments are performed on ImageNet-1K and CIFAR benchmarks and include ablation studies to understand the impact of model capacities.

Strengths

The paper provides an interesting approach to OOD detection by leveraging the concept of attribution gradients. I find the two forms of gradient abnormalities for OOD detection very promising in approaching the problem. The paper is very well written and clear. The experiments use robust setups on well-known ImageNet-1K and CIFAR benchmarks and the ablation studies are interesting to validate the hypotheses and claims. I suggest experimenting with more datasets to validate the proposition in different contexts. OOD detection is an important problem in the field of deep learning and has numerous practical applications in enhancing the safety and reliability of deep neural network applications.

Weaknesses

A more direct comparison with other gradient-based methods would be beneficial. While the authors compare their method to GradNorm, it would be interesting to see how GAIA compares to other methods that also utilize gradient information for detection or other tasks. I think a more detailed explanation of the proposed abnormalities is needed, as it would be beneficial to have a more intuitive or visual explanation to aid in understanding these concepts. The authors could elaborate more in this direction. While ImageNet-1K and CIFAR are standard benchmarks, it would be important to see the method performs on different types of data, such as text or audio data or even different image datasets. The paper does not address the computational efficiency and scalability of the proposed method. I think a discussion on the proposed method's potential limitations and failure modes would be a valuable addition.

Questions

Could the authors provide a comparison with other gradient-based methods? Can the authors elaborate more or provide visual aids to help intuitively understand the proposed abnormalities? Can the method deal with other types of data (text, audio, etc.)? Could the authors comment on the computational efficiency and scalability of GAIA? The authors claim that no hyperparameters are required for their method. Can the authors clarify this point?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

It is an interesting paper, but a few areas could use more in-depth discussion on the scalability and computational efficiency of the method. While the authors show impressive performance on several benchmark datasets, they do not fully discuss the method's robustness to various forms of OOD shifts, which are common in real-world scenarios. While both method variants show superior performance compared to other methods, there's a noticeable difference between the two. GAIA-Z generally achieves a lower FPR95 and a higher AUROC than GAIA-A. Is GAIA-Z the more precise variant? Can you elaborate on this direction? The authors do not discuss where and why the GAIA-A and GAIA-Z methods fail. Failure case analysis is crucial for understanding the limitations of the methods and guiding future research.

Reviewer TjRi5/10 · confidence 4/52023-07-27

Summary

In this paper, the authors present an innovative perspective on quantifying the disparities between in-distribution (ID) and out-of-distribution (OOD) data. The authors observed that gradient-based attribution methods face challenges when assigning feature importance to OOD data, leading to significantly divergent explanation patterns. To address this issue, the authors investigate how attribution gradients contribute to uncertain explanation outcomes and introduce two forms of abnormalities for OOD detection: the zero-deflation abnormality and the channel-wise average abnormality. To overcome these challenges, they propose a new approach called GAIA (Gradient Abnormality Inspection and Aggregation), which is simple yet effective. Importantly, GAIA can be directly applied to pre-trained models without the need for further fine-tuning or additional training. The results demonstrate that GAIA outperforms existing approaches on commonly utilized benchmarks such as CIFAR and large-scale benchmarks like ImageNet. Specifically, on CIFAR benchmarks, GAIA reduces the average FPR95 by 26.75% on CIFAR10 and by 45.41% on CIFAR100 when compared to competing methods, highlighting its superiority in OOD detection.

Strengths

- The idea of the paper is clear, the writing is easy to follow, and provides theoretical support. - The proposed method GAIA is simple and effective on CIFAR benchmarks.

Weaknesses

But I'm more concerned about the effectiveness of the method: - In the comparison in Table 1 on large-scale benchmark ImageNet-1K, the proposed GAIA-A and GAIA-Z methods only compare the inferior version of Rankfeat (Block 4), but don't outperform the SOTA version of Rankfeat (Block 3 + 4)$[1]$. - There are two versions of GAIA: GAIA-A, and GAIA-Z, which have their own strengths and weaknesses on different benchmarks, but there is no guidance in the paper on which method to use in different benchmarks. - Important points of innovation and the bulk of proofs are related to GAIA-A, but GAIA-A does not outperform GAIA-Z on most benchmarks (e.g. CIFAR). This reduces the effectiveness of GAIA-A. - Typo: line 173, "with with". Reference: [1] Song Y, Sebe N, Wang W. Rankfeat: Rank-1 feature removal for out-of-distribution detection[J]. Advances in Neural Information Processing Systems, 2022, 35: 17885-17898.

Questions

1. What is the meaning of zero baseline output S_c ($0$) in Eq. (3)? 2. What are the inherent reasons for the difference in effectiveness between GAIA-A and GAIA-Z? For more please refer to the Weaknesses part. Open question: - Pre-training on one dataset and testing on another, yet the dataset itself is natural domain inconsistent, is there a real-world application for this? - What is the OOD sample in the real-world application? If we only have the dataset on a sunny day, but a bird on a rainy day, would that be considered as an OOD sample? Will the method GAIA in the paper recognize birds in rainy weather as OOD samples? Or only images that do not contain objects will be treated as OOD?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer TjRi2023-08-14

Response to Authors

The authors addressed my concerns to some extent. Could the authors provide a more professional analysis of why "GAIA-A has the advantage of collecting more anomalies in larger label spaces"? Not just from an experimental observation view.

Authorsrebuttal2023-08-15

Thanks for your comments

Thank you for taking the time to read our rebuttal and engaging in timely discussions with us. > Q: Could the authors provide a more professional analysis of why "GAIA-A has the advantage of collecting more anomalies in larger label spaces"? Not just from an experimental observation view. Of course! We put forward this viewpoint due to that GAIA-A has the ability to aggregate information from all predicted outputs. As mentioned in the rebuttal, GAIA-A aims to gather extreme anomaly values of weights (channel-wise average gradients) in the gradient-based attribution method to reflect uncertainty. Consider that the aggregation region has $L$ feature layers, and each layer has $K$ channels for ease of representation. The overall expectation of the abnormality $\mathbb{E}[\epsilon]$ can be represented as: $\begin{equation} \mathbb{E}[\epsilon] = \sqrt{\sum\limits_{l\in L} \sum\limits_{k\in K} \| \mathbb{E}[\epsilon|\textbf{A}^{kl}]\|^2} \end{equation}$ Then, we analyze the expectation of the abnormality on an individual k-channel, $\begin{equation} \mathbb{E}[\epsilon|\textbf{A}^{kl}] = \left\| \sum\limits_{i, j} \frac{\partial \Gamma(S_c(\textbf{A}^l))}{\partial \textbf{A}^{kl}_{ij}} \right\| = \left\| \Omega \right\| \end{equation}$ where $\Gamma(\cdot)$ represents a method of aggregating over label outputs, while $\Omega$ signifies the aggregation of attribution gradients. In our paper, we utilize the log-softmax aggregation approach. For the sake of simplicity in analysis, let us consider aggregation as a summation. Thus, the aggregation gradient $\Omega$ can be decomposed as follows: $\begin{equation} \Omega = \frac{\partial \sum\limits_{c\in C}S_c(\textbf{A}^l)}{\partial \textbf{A}^{kl}} = \sum\limits_{c\in C} \frac{\partial S_c(\textbf{A}^l)}{\partial \textbf{A}^{kl}} = \sum\limits_{c\in C} \omega_c \end{equation}$ where $w_c$ represents the weight corresponding to the attribution for the $c$-label output $y_c$. Therefore, the expectation of abnormality on this channel can be represented as the sum of the expectation of abnormality across the whole label space. $\begin{equation} \mathbb{E}[\epsilon|\textbf{A}^{kl}] = \sum\limits_{c\in C} \mathbb{E}[\epsilon|\textbf{A}^{kl}, y_c] \end{equation}$ In other words, since GAIA-A relies on collecting weight anomalies, a larger label space allows it to gather more anomalies, thereby better reflecting the model's uncertainty.

Area Chair S8dB2023-08-13

Discussion starts

Dear Reviewers, Thank you for reviewing this paper. Authors have provided their rebuttal. Would you please check it, and give your comments/rating based on the rebuttal letter and the comments from other reviewers? Best Regards AC

Reviewer Exp82023-08-16

Thank you for providing a comprehensive rebuttal and the effort to improve the paper. The visualizations incorporated offer a clearer understanding of the abnormalities. Looking at the images, I was also considering if a signal-to-noise analysis could provide some insights if the features considered are more related to the foreground than the background. Also, considering your last comment, I'm not sure I understand the problem of the proposed approach to deal with ViT. Given the growing popularity and effectiveness of transformer architectures, how to adapt or evolve the proposed to accommodate these new advancements?

Authorsrebuttal2023-08-17

We appreciate your insightful viewpoints and thoroughly enjoy the discussion. > Q1: Looking at the images, I was also considering if a signal-to-noise analysis could provide some insights if the features considered are more related to the foreground than the background. This is indeed an interesting perspective. We have given consideration to this issue and have come up with a few directions to explore. If we consider the attribution map on the feature map as a grayscale image, with attribution values as pixel values, features deemed to contribute to predictions (useful features) could be considered as signals. Under the assumption of the highly effective visual explanation method, these useful features, such as the foreground, should have high pixel values, making them meaningful signals. On the other hand, the background should possess low pixel values, representing insignificant noise. Therefore, we can use the signal-to-noise ratio (SNR) to quantify the uncertainty of the model's attribution explanations: $$\text{SNR} ( \text{db})= \frac{\text{Signal Power}}{\text{Noise Power}}$$ If the model makes a confident and correct decision, the SNR should be very high. However, this approach also presents some challenges: 1) How do we define the scope of the signal (foreground). We may need a threshold or algorithm to differentiate which regions on the attribution map are background and which are foreground. 2) When there are no in-distribution objects in the entire image (OOD), we should consider the entire image as noise. In our observations, this scenario can result in widely scattered and exceptionally high noise values. We need to design a more appropriate matrix to reflect SNR. **We consider that this approach could be highly beneficial for object-level OOD detection.** When an image contains multiple objects that need to be recognized, we can consider the regions within the bounding boxes as the signal (foreground) and those outside the boxes as the background. Then we can calculate the SNR on the attribution map to reflect the model's uncertainty about predictions within these regions. This method might be effective in detection scenarios with relatively stable environments, such as autonomous driving, industrial identification, etc. > Q2: Considering your last comment, I'm not sure I understand the problem of the proposed approach to deal with ViT. Given the growing popularity and effectiveness of transformer architectures, how to adapt or evolve the proposed to accommodate these new advancements? Given the current trends, providing visual explanations for large models has been a focal point for both the interpretability academic community and the engineering efforts in recent years. Honestly, this is also the research direction we are currently pursuing. Overall, our approach involves utilizing model-explained prediction uncertainty for OOD detection. For large models, our ongoing research primarily focuses on two directions: - Direction 1: **For transformer-based model.** Despite the differences between transformer-based architectures and the convolutional feature layers of CNNs, methods based on attribution gradients can still provide explanations for predictions made by models using transformer structures (such as ViT). For the transformer-based model, we consider the attribution gradients on the attention matrix. Many traditional CAM methods, including GradCAM, are initially proposed based on CNN structures, and therefore, their performance on transformer-based models might be suboptimal. **However, there is a line of improvements available now to enhance the gradient-based visual explanations on such models**. One prominent example is [1]. Building upon these enhancements, we are currently researching how to identify attribution gradient abnormality on the attention matrix to reflect the model's uncertainty. Furthermore, the Attention mechanism in transformer-based models can also offer directions for visual explanations. Researching uncertainty in this context can further enhance OOD detection. [1] Chefer, Hila, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. CVPR. 2021. - Direction 2: **For CNN-based backbone.** Although the mainstream of large models isn't purely CNN-based, many multimodal large models (like CLIP) and downstream tasks (such as object detection) still utilize CNNs as the backbone networks for visual feature extraction. How to reflect the model's uncertainty in visual feature extraction on these backbones is also a topic we are currently researching.

Reviewer Exp82023-08-18

I appreciate your answer, clarification, and additional edits. I think the review score is still consistent. Congratulations on the work.

Authorsrebuttal2023-08-20

Thanks for your positive feedback

Thanks for your positive feedback and insightful comments! We are delighted that our responses have been satisfactory, and we thoroughly enjoy the discussions with you!

Reviewer qAeX2023-08-19

Thanks for the response.

The authors have addressed my concerns, and I would like to raise my score to 5.

Authorsrebuttal2023-08-20

Thanks for raising the score

Thank you! We appreciate the updated score and we are glad that our clarifications have addressed your concerns. Thanks again for taking the time to review our paper and providing detailed comments.

Authorsrebuttal2023-08-20

Thanks for raising the score

Thank you for raising the score. We are delighted that our responses were able to address your concerns. Finally, we appreciate your careful review and engaging discussions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC