An Unlearning-Enhanced General Framework for Test-Time Adaptation

Zhenyi Wang, Chenxi Liu, Yihan Wu, Heng Huang

Keywords: Test time adaptation

Abstract: Test-time Adaptation (TTA) aims to mitigate performance degradation caused by distribution shifts during testing time. While various TTA approaches exist, such as entropy minimization, pseudo-labeling, weight-space regularization and Bayesian methods, a generalized optimization framework for TTA is currently absent. To address this gap, we present a general framework for TTA. This framework provides a conceptual basis for understanding existing methods as specific instances within a broader optimization framework, and facilitates the development of new TTA methods. Additionally, our proposed framework brings attention to limitations in existing approaches by unveiling an implicit assumption that all source domain knowledge is universally beneficial for adapting to the target domain. In reality, only a portion of the source domain knowledge is useful due to potential large distribution discrepancies between the source and target domains. Based on this insight, we build upon our general framework and derive a novel method named Unlearning-enhanced test-time adaptation (Lana). Specifically, it adaptively unlearns irrelevant source domain knowledge and then adapts to the target test domain. Through thorough theoretical analysis and empirical results, we showcase the effectiveness of our proposed method in enhancing TTA performance. This work contributes not only a broader understanding of TTA through a general framework but also a novel practical solution, Lana, derived from our general framework, offering a foundation for further advancements in addressing distribution shifts during testing in machine learning models.

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

Submission Number: 12560