SoTTA: Robust Test-Time Adaptation on Noisy Data Streams

Summary

The authors point out that model may suffer from non-interest samples while TTA. Existing TTA methods are not robust to these samples. To address these issues, the authors proposed a methods called SoTTA with two key components, input-wise robustness via high-confidence uniform-class sampling and parameter-wise robustness via entropy-sharpness minimization.

Strengths

The authors focus on TTA in the wild and point out that TTA with non-interest would lead to performance degradation.

Weaknesses

The proposed method seems to be not novel. A) The proposed “Input-wise robustness via high-confidence uniform-class sampling” method excludes low-confidence samples in TTA. Such samples filtration strategies have been explored in ETA [Efficient Test-Time Model Adaptation without Forgetting]. B) The proposed “Parameter-wise robustness via entropy-sharpness minimization” method introduces entropy-sharpness minimization, which have been explored in SAR [Towards Stable Test-time Adaptation in Dynamic Wild World]. It would be better for the authors to clarify the differences between the proposed methods and the TTA methods I mentioned above. In the experimental results, I found the proposed always outperforms SAR. But in my understanding, the main components of these two methods are almost the same. Both of them include low-confidence sample filtration strategy and entropy-sharpness minimization. So I have no idea why the proposed method can yield much better results.

Questions

See comments above.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

See comments above.

Summary

This paper studies a practical problem of test-time adaptation where non-interest testing samples may appear and mislead the adaptation. This problem is quite serious in practical applications, and the problem setting is relatively novel. To address this problem, the authors propose the SoTTA method, which solves this problem in two aspects. For input-wise robustness, SoTTA filters out the non-interest samples with a high-confidence uniform-class memory buffer. For parameter-wise robustness, entropy-sharpness minimization is adopted to ensure that the landscape is smooth during the adaptation. The proposed method is evaluated on two benchmark datasets under one standard setting and four robust settings with different types of non-interest samples. The results show that SoTTA gives state-of-the-art performance compared to existing TTA methods.

Strengths

1. This paper studies a novel problem setting of test-time adaptation, where non-interest testing samples may appear and mislead the adaptation. This is a relatively novel and practical problem in real applications. 2. The overall idea of this paper makes sense. This paper tackles the harmful impact of non-interest with a filter and robust loss function. The experiments in the main paper also show that the proposed method gives state-of-the-art performance.

Weaknesses

1. First, this paper ignores one common TTA benchmark, i.e., CIFAR100. Additionally, the results of ImageNet-C have been hidden in the appendix. These key results should be presented in the main paper with comprehensive analysis. 2. One TTA algorithm, EATA [1], adopts a similar sample selection strategy during adaptation. This method should be taken into consideration in the experiment parts because it has the potential to handle the problem of non-interest samples. 3. This paper does not provide any discussion about the selection and sensitivity of threshold C0. Intuitively, the optimal C0 depends on both in-distribution data and out-of-distribution data, and it is not easy to decide in real applications. 4. There is more related work that can be discussed to further improve this paper. For test-time adaptation, the above-mentioned EATA method [1] and recent advanced TTA methods [2-4] about robustness and evaluation should be considered. For out-of-distribution detection, methods [5-6] that can be efficiently optimized in an unsupervised manner should be taken into consideration. **Reference** [1] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, Mingkui Tan: Efficient Test-Time Model Adaptation without Forgetting. ICML 2022: 16888-16905 [2] Hao Zhao, Yuejiang Liu, Alexandre Alahi, Tao Lin: On Pitfalls of Test-Time Adaptation. ICML 2023 [3] Zhi Zhou, Lan-Zhe Guo, Lin-Han Jia, Dingchu Zhang, Yu-Feng Li: ODS: Test-Time Adaptation in the Presence of Open-World Data Shift. ICML 2023 [4] Tong Wu, Feiran Jia, Xiangyu Qi, Jiachen T. Wang, Vikash Sehwag, Saeed Mahloujifar, Prateek Mittal: Uncovering Adversarial Risks of Test-Time Adaptation. ICML 2023 [5] Zhi Zhou, Lan-Zhe Guo, Zhanzhan Cheng, Yu-Feng Li, Shiliang Pu: STEP: Out-of-Distribution Detection in the Presence of Limited In-Distribution Labeled Data. NeurIPS 2021: 29168-29180 [6] Jiangpeng He, Fengqing Zhu: Out-Of-Distribution Detection In Unsupervised Continual Learning. CVPR Workshops 2022: 3849-3854

Questions

1. The experiment shows that RoTTA outperforms existing TTA methods even in standard settings. Can the authors explain why this phenomenon occurs? Intuitively, the techniques mentioned in this article are all aimed at addressing the problem of non-interesting samples. What factors enable them to achieve better performance when non-interesting samples do not exist? 2. Can the authors provide some sensitivity analysis regarding the threshold C0? It is very important for the actual effect of filtering out non-interesting samples. An adaptive scheme for setting C0 would also be acceptable. 3. One naive way to solve the proposed setting is to adopt an out-of-distribution process before test-time adaptation. How does this naive method work in the proposed setting? **[IMPORTANT]** I would like to raise my score if you successfully address my concerns in the Weakness and Question sections. Overall, this paper is interesting and I appreciate it. However, it is currently slightly below the level of acceptance.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

1 poor

Presentation

3 good

Contribution

2 fair

Limitations

The authors have properly discussed the limitations.

Summary

This article presents a new Test-Time Adaptation (TTA) scenario, wherein the model is adapted to noisy test streams. To address the challenges posed by this scenario, the paper introduces the Screening-out Test-Time Adaptation (SoTTA) algorithm, which leverages input-wise and parameter-wise robustness. The effectiveness of SoTTA is demonstrated through extensive comparison experiments conducted on TTA benchmarks.

Strengths

1. This work proposes a novel test time adaptation setup, which considers the noisy data (non-interest) in real world application during test phase, whose motivation is convincing. 2. The significant performance gain shows SoTTA addresses the challenges under noisy test stream well.

Weaknesses

1. The initial component of the method is inspired by the memory bank concept introduced in RoTTA, albeit in a simplified form. The second component, pioneered by SAR[1], represents a novel approach within the TTA domain. However, it is worth noting that the methodology section may lack originality. 2. The sensitivity analysis and selection criterion for key hyperparameters, specifically $m,C_0$ , are not included in the current study. 3. The motivation of this paper stems primarily from experimental findings, and it would be advantageous to analyze the motivation or methods from a theoretical perspective. 4. The paper contains several typographical errors, and it would be valuable to thoroughly revise it. [1] Towards stable test-time adaptation in dynamic wild world

Questions

1. What would be the outcome if CSTU and ESM were combined? Would the performance surpass that of SoTTA? 2. It is acknowledged that in real-world scenarios, the non-interest scenes mentioned in the experimental section may occur simultaneously. Have any relevant experiments been conducted to address this?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Contribution

2 fair

Limitations

Actually, the problem setting is interesting to some extent, however, the techniques in this paper are similar to several compared methods, which may limit the novelty of this paper.

Summary

This paper proposes screening-out test-time adaptation which is claimed robust to non-interest samples. It filters out the impact of non-interest samples with a high-confidence uniform-class sampling. It proposes entropy-sharpness minimization to deal with large gradients. Experiments are completed on CIFAR10-C and ImageNet-C.

Strengths

1. The proposed problem (screening-out test-time adaptation) is novel and sharp. 2. This paper has a good presentation. For example, Fig.1 and Fig.2 are very clear. 3. Entropy-sharpness minimization seems reasonable as discussed in Section 3.2. 4. Experiments are completed on both small and large datasets including ImageNet-C. As shown in Tab.1, the improvement is good enough.

Weaknesses

1. The problem is interesting but the proposed two methods are straightforward. I am a little worried about the technical novelty. 2. The design of input-wise robustness is hard coding. Although some samples are bad for optimization, they may be essential for the perception. 3. Only classification is completed. However, segmentation is also very important in autonomous driving scenarios.

Questions

1. Can it also work on segmentation and detection? Please discuss.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Please refer to "weakness".

Further Comments

Thank you for the authors' response. The authors highlight two main distinctions: *(1) characteristics of the samples filtered out and (2) memory management strategies.* However, in their detailed response, I struggled to grasp the exact disparity in memory management. Moreover, I'm seeking a clear understanding of the technical contrasts between the proposed method and ETA/SAR. For instance, while ETA performs operation "A", your SoTTA performs operation "B". In your scenario, "B" proves more effective than "A," leading to improved accuracy. The differences presented in the current version of the response are somewhat general. I recognize that your method's motivation varies from ETA/SAR, yet I'm specifically interested in perceiving technical distinctions.

Official Comment by Reviewer gePD

Thanks for the detailed response, and some of my concerns have been addressed. Indeed, the test time setting on Noisy Data Streams is interesting and challenging. However, I still believe the techniques in this paper are similar to several SOTA TTA methods, and I will keep my score.