TUI: A Conformal Uncertainty Indicator for Continual Test-Time Adaptation

Fan Lyu, Ziqi Shi, hanyu zhao, Fuyuan Hu, Zhang Zhang, Liang Wang

Keywords: Continual Test-Time Adaptation, Domain Adaptation, Conformal Prediction

TL;DR: We propose a simple uncertainty estimation method TUI for CTTA to measure the test uncertainty for each test prediction.

Abstract: Continual Test-Time Adaptation (CTTA) addresses the challenge of adapting models to sequentially changing domains during the testing phase. Since no ground truth labels are provided, existing CTTA methods rely on pseudo-labels for self-adaptation. However, CTTA is prone to error accumulation, where incorrect pseudo-labels can negatively impact subsequent model updates. Critically, during testing, a CTTA method can not detect its mistakes, which may then propagate through further adaptations. In this paper, we propose a simple uncertainty indicator called TUI for the CTTA task based on Conformal Prediction (CP), which generates a set of possible labels for each test sample, ensuring that the true label is included within this set with a given coverage probability. Specifically, since domain shifts can undermine the coverage of predictions, making uncertainty estimation less dependable, we propose compensating for coverage by dynamically measuring the domain difference between the target and source domains in continuously changing environments. Moreover, after estimating uncertainty, we separate reliable test pseudo-labels and use them to discriminatively enhance the adaptation process. Empirical results demonstrate that our algorithm effectively estimates the uncertainty for CTTA under a specified coverage probability and improves adaptation performance across various existing CTTA methods.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4045