Revisiting Entropy Minimization for Long-Sequence Continual Test-Time Adaptation


WeiQin Chuah (Royal Melbourne Institute of Technology), Ruwan Tennakoon (Royal Melbourne Institute of Technology), Alireza Bab-Hadiashar (Royal Melbourne Institute of Technology)
The 35th British Machine Vision Conference

Abstract

Deep neural networks (DNNs) have revolutionized a wide range of tasks, yet they remain vulnerable to distributional shifts between training and test data, which is a common occurrence in real-world scenarios. These shifts often unfold continually over time, posing significant challenges for test-time adaptation. A critical issue we identify in this setting is \textbf{noisy label memorization}, where the model gradually overfits to its own erroneous predictions (pseudo-labels), leading to significant performance degradation. This problem is particularly severe in \textbf{long-sequence} continual test-time adaptation, where the model must adapt continuously in non-stationary environments. To address this, we propose a novel methodology that strategically constrains weight updates to prevent noisy labels from degrading the model, resulting in significant performance improvements, particularly in long-sequence continual test-time adaptation (CoTTA) scenarios. Extensive experiments on benchmarks such as CIFAR10-C, CIFAR100-C, ImageNet-C, CCC, and DomainNet demonstrate that our method not only outperforms state-of-the-art CoTTA techniques but also ensures sustained model reliability and robustness over time. Furthermore, our method exhibits greater stability across varying sample sizes and batch sizes, with reduced sensitivity to model selection—all achieved without the need for source data calibration.

Citation

@inproceedings{Chuah_2025_BMVC,
author    = {WeiQin Chuah and Ruwan Tennakoon and Alireza Bab-Hadiashar},
title     = {Revisiting Entropy Minimization for Long-Sequence Continual Test-Time Adaptation},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_259/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection