RDumb: A simple approach that questions our progress in continual test-time adaptation

Summary

Test-time adaptation techniques seek to adapt a model to new unlabelled samples without access to the original training data. However, TTA approaches are typically not evaluated for long runtimes. This work proposes a new benchmark Continuously Changing Corruptions (CCC), which is a stream of corrupted ImageNet images with gradual changes and varying degrees of difficulty. Previous approaches fail on (CCC) when compared to the pretrained model as well as a simple new baseline method dubbed RDumb. RDumb is shown to achieve high performance across several experimental settings.

Strengths

Strengths: - The overall message of the paper is easy to follow and the experiments are illustrative of a serious shortcoming in the evaluation of TTA methods. - The baseline is simple and satisfies all the TTA assumptions, and should be considered as a baseline in future works in this area - Experiments are very extensive and thorough

Weaknesses

Weaknesses: - There are a few issues related to clarity that should be addressed in the next version of the paper. See Questions - I would encourage the authors to add error bars to their tables, so that it is clear that the differences are statistically significant

Questions

Questions: 1) The mathematical notation in Eq. 1 is a bit difficult to parse. In particular, it is unclear what is meant by $cos(y_t, \bar{y}_t)$. Is this cosine similarity or cosine distance? I would encourage the authors to clarify this and also provide a clear description of the equation in the text. 2) Can the authors precisely define "collapse"? This term is used but is never defined. 3) In Figures 3 and 4, if RDumb is being reset to the pretrained model weights, then why do the blue and purple lines never intersect? 4) It would be interesting to test CLIP image encoders [1] that are naturally robust and have high zero-shot accuracy on OOD samples. In future work, I would be curious to see how such models behave on the proposed CCC when TTA techniques are applied. References: - Learning Transferable Visual Models From Natural Language Supervision, Radford et al. 2021

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

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Limitations are addressed.

Summary

This paper proposes a new benchmark for continual test-time adaptation (CTTA): CCC (Continually Changing Corruptions). Experiments for existing state-of-the-art approaches using their proposed benchmark demonstrate that even a non-adapting model performs better than existing approaches for this benchmark. Further, they propose a simple approach as a baseline, named $\textit{RDumb}$, which periodically resets the model to its pre-trained state, which performs better than the SOTA approaches.

Strengths

1. This paper proposes a more challenging benchmark for CTTA. 2. Combining a pair of corruptions is a novel idea. 3. Experimental results demonstrate that several state-of-the-art approaches perform poorly on this benchmark. 4. A simple periodic reset mechanism with weighted entropy (as in ETA) adaptation loss outperforms the SOTA CTTA approaches.

Weaknesses

1. In CoTTA (Wang et al., CVPR 2022), there is a setting of gradually changing the corruptions (CIFAR10-to-CIFAR10C gradual). So the idea of gradually changing corruption is not entirely new, even though CCC has a combination of two corruptions. 2. This paper does not analyze the reasons for the performance degradation of the SOTA TTA approaches, such as overfitting, error accumulation, catastrophic forgetting, or interference, and does not provide any insights or suggestions on mitigating or preventing it. 3. Experimental results on CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C classification tasks, along with their CCC variants, are missing. Minor Comment: Line 62: use `` for starting quotes before RDumb

Questions

1. In CoTTA (Wang et al., CVPR 2022), there is a setting of gradually changing the corruptions (CIFAR10-to-CIFAR10C gradual). CoTTA, as well as PETAL [1], reports performance improvement in this gradual setting, so it would be interesting to see how they would work on the CCC variant of CIFAR10-to-CIFAR10C. Do the authors have any intuition behind this? 2. As the authors point out, the ImagenetC dataset is large and other benchmarks are much smaller; the dataset being smaller can also make the task more challenging, since there would be lesser data to learn from for a domain shift. It would be interesting to see how the performance is for CCC variants of CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. 3. Also, with regards to the semantic segmentation task, do the authors have some idea as to how this dataset/benchmark building mechanism would be applicable? *References* 1. Brahma, Dhanajit, and Piyush Rai. "A Probabilistic Framework for Lifelong Test-Time Adaptation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

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

1. Since the state-of-the-art approaches are being evaluated on a new benchmark dataset, it is very important to tune their hyperparameters optimally. 2. The details regarding hyperparameter tuning for various SOTA approaches are missing.

Summary

The paper newly introduces a new benchmark dataset for Continual Test-Time Adaptation (CTTA) named Continually Changing Corruptions (CCC), and suggests a simple technique - repeatedly initialize the learned model weights during CTTA. CCC is composed of interpolated corruption data and its data scale varies at each timestep. They can control the difficulty of CCC data by changing the degree of interpolation of two different corruptions.

Strengths

The paper suggests a new large-scale continual test-time adaptation benchmark dataset, considering corruption interpolation, scale variation, and repetition.

Weaknesses

The suggested method is too naive and not attractive in view of knowledge transfer. Periodically resetting the trainable weights to the initial state indicates simply discarding obtained knowledge and adaptive representation regardless of their importance and relevancy. Even though this remedy outperforms baselines, there is a lack of quantitative and qualitative analysis of why it happens and of the reason RDumb behaves differently from baselines. Additionally, the decision of resetting iteration is heuristic. The suggested idea is surprising but I strongly recommend more rigorous analyses and validation of the suggested method and CCC dataset since they are counterintuitive in some sense.

Questions

Why does RDumb consistently outperform a pre-trained model during TTA? Shouldn't its performance reach the pre-trained model's performance when re-setting the model weights to the initially pre-trained ones?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Please see the weakness.

Summary

The paper proposes a new benchmark (dubbed CCC) for test-time adaptation, which generalizes previous corrupted imagenet benchmarks. Specifically, the CCC gradually draws a sequence of corruptions (e.g. gaussian noise or motion blur), and gradually interpolates between two consecutive corruptions, creating a stream without hard boundaries, akin to many realistic settings. By considering long sequences of corruption pairs, CCC yields a stream longer than previous benchmarks, which sheds new light on how previous methods fair in such settings. Indeed, it is shown that given a long enough stream, methods collapse to a performance worse than a non-adapted model. The authors then propose a new baseline, Rdumb, which resets to the default parameters of the pretrained model at regular, fixed intervals. Over both CCC and previous benchmarks, it is shown that Rdumb, despite its simplicity, works well.

Strengths

1. The proposed benchmark is well-designed. I liked the use of calibration with a pretrained model to generate different stream difficulties, and the fact that this overall is stream generator can is flexible and can accommodate smooth transitions. 2. The proposed baseline, Rdumb, is extremely simple and should be considered as a standard baseline for TTA. 3. Overall the paper is well written and easy to follow.

Weaknesses

1. The idea of resetting the model to its original pretrained state in order to combat accumulating degradation in TTA is not new (as correctly stated in the paper). Indeed, it was shown in [45] that resetting after each task boundary helps the network. For a fair comparison with COTTA, how would such method perform if the probability of resetting the weights was determined by a similar cross-validation technique as used in Section 6 ? 2. COTTA also has some experiments with gradually changing corruptions (see Table 3). it would be good for this to be mentioned in the paper. minor comments : 1. regarding figure 2, I would use the color red as the most severe corruptions, and yellow as the least severe. 2. Optimal resetting interval : I think it should be Rdumb, and not Gdumb (lines 189-190)

Questions

1. My understanding of TTA's use-cases is mostly for online settings; the deployed model receives data for which to output predictions, and this data may encounter distribution shifts. In such a setting, how would one determine the frequency of when to reset the model ?

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

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Answer to Rebuttal

Thank you for the rebuttal. After reading the other reviewers' comments and your rebuttal, - I agree that lacking Technical complexity is not a negative attribute in itself; simple methods that work well should be prioritized. - I agree that an analysis as to *why* existing methods suffer from such performance degradation would a good addition to the paper. I thank the authors for the additional experiments, and making some progress towards answering this. While my expertise is not directly in continual test time adaptation, I do believe that well executed papers which revisit the current state of a field of research (and its evaluation protocols), as well as propose simple methods which perform well, can be of value to the research community. I believe that my previous score and confidence reflect my overall assessment of the paper, and will therefore leave as-is. Thank you.

Some of the concerns addressed with additional experiments

Dear authors, I want to thank the authors for responding to my comments and questions. Some of the questions asked have been addressed. The experiment with toy 2d Gaussian binary classification is interesting. However, the smaller benchmark dataset, as in CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C, can also make the task more challenging since there would be less data to learn from for a domain shift. This point has not been addressed. Experimental results on these benchmarks would make the case stronger for the proposed approach. Based on the response, I do not have any other queries at this point in time. Thank you.

Response to Rebuttal

Thank you for the clear rebuttal, all of my concerns are addressed so I have raised my score. I would recommend the authors to address why the lines do not intersect in the next version of the paper.

Discussion Phase Summary

Dear AC, Overall, we enjoyed a constructive discussion phase and the reviewers found our work meaningful, thorough, easy to understand, and found our contribution to be significant. The current scores are: 7,7,5,3 (and we are waiting for replies from the latter two). During the discussion phase, we did the following: - **Added analysis on collapse in a toy setting**: Reviewers BezD and h5ar asked for additional analysis/discussion on the causes for collapse. We now added a theoretical analysis that addresses both settings of model collapse and stable behavior and links this to the signal-to-noise ratio in the data. We confirm that the learning dynamics match between small scale (simulated 2D data) and large scale (CCC scale) experiments. The analysis sheds light on the causes for collapse during adaptation, indicating why RDumb (full model resetting), CoTTA (gradual model resetting) and EATA (regularization-based) outperform standard entropy minimization techniques. - **CIFAR10-C/CIFAR100-C experiments**: Reviewers h5ar and xRL5 asked for additional results on CIFAR scale. While we presented arguments of important shortcomings of CIFAR vs. our proposed CCC benchmark, we also ran these results. We now show that RDumb also significantly outperforms CoTTA in a comparable experimental setup on CIFAR, further corroborating our larger-scale experimental results. While the reviewer did not reply back yet, we hope that these new results address any remaining doubts about RDumb’s effectiveness. - **Additional CoTTA experiments**: Reviewer h5ar asked us to run CoTTA with more hyperparameters. We show that CoTTA still fails after an extensive hyperparameter search. The only reviewer, BezD, that currently leans towards rejection didn’t participate at all in the discussion. Their main reason for rejection was that our method was allegedly “too simple” - but that was our intention from the get-go: showing that an extremely simple method performs just as well or *better* than the more complex SOTA ones.. We believe that the importance of Occam’s Razor principle is still pertinent. Reviewers h5ar (“simple methods that work well should be prioritized”) , xRL5 (“A simple periodic reset mechanism with weighted entropy … outperforms the SOTA CTTA approaches” [under strengths]), and Gw5H (“The baseline is simple and satisfies all the TTA assumptions, and should be considered as a baseline in future works in this area” [under strengths]) agreed with us that the simplicity of our method is an *advantage*. Thank you for taking this into consideration.