Maintain Plasticity in Long-timescale Continual Test-time Adaptation

2024-12-28

AI Summary

In continual test-time adaptation (CTTA), models suffer from progressive plasticity degradation during long-term online adaptation—especially under non-stationary label streams—leading to sustained performance decline. This work is the first to empirically and theoretically establish a strong correlation between plasticity decay and label-flip frequency. To address this, we propose Adaptive Scaling-and-Restoration (ASR): a label-flip-driven dynamic detection mechanism that triggers weight reinitialization, coupled with adaptive interval optimization to regulate update timing. ASR significantly improves long-term adaptation stability and accuracy across multiple CTTA benchmarks, effectively mitigating plasticity decay and enabling robust, persistent unsupervised online adaptation. Our core contributions are: (i) establishing a theoretical linkage between label dynamics and model plasticity, and (ii) introducing the first label-flip-aware reinitialization framework for CTTA.

Technology Category

• Machine Learning: Life-Long and Continual Learning
• Computer Vision: Adversarial Attacks & Robustness
• Cognitive Modeling & Cognitive Systems: Adaptive Behavior

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• User Modeling, Personalization and Recommendation: User modeling for targeted and personalized online advertising

Abstract

Continual test-time domain adaptation (CTTA) aims to adjust pre-trained source models to perform well over time across non-stationary target environments. While previous methods have made considerable efforts to optimize the adaptation process, a crucial question remains: can the model adapt to continually-changing environments with preserved plasticity over a long time? The plasticity refers to the model's capability to adjust predictions in response to non-stationary environments continually. In this work, we explore plasticity, this essential but often overlooked aspect of continual adaptation to facilitate more sustained adaptation in the long run. First, we observe that most CTTA methods experience a steady and consistent decline in plasticity during the long-timescale continual adaptation phase. Moreover, we find that the loss of plasticity is strongly associated with the change in label flip. Based on this correlation, we propose a simple yet effective policy, Adaptive Shrink-Restore (ASR), towards preserving the model's plasticity. In particular, ASR does the weight re-initialization by the adaptive intervals. The adaptive interval is determined based on the change in label flipping. Our method is validated on extensive CTTA benchmarks, achieving excellent performance.

Problem

• Continuous Testing Time Adaptation
• Plasticity Decay
• Label Shift

Innovation

• ASR strategy
• plasticity maintenance
• continuous task-agnostic adaptation