GoTTA be Diverse: Rethinking Memory Policies for Test-Time Adaptation
Researchers have developed a new approach to test-time adaptation (TTA) for machine learning models, focusing on how memory policies influence adaptation to distribution shifts. Their work introduces Guided Observational Test-Time Adaptation (GOTTA), which prioritizes intra-class diversity in memory selection alongside class balance. This diversity-aware memory management proves most effective under limited memory and challenging non-i.i.d. data streams, enhancing adaptation robustness. AI
IMPACT Enhances model robustness to distribution shifts, particularly under memory constraints.