Dual Domain-Attribute Learning Framework With Asynchronous Adapters for Continual Test-Time Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual test-time domain adaptation (CTTA) addresses model deployment in dynamic, non-stationary environments by adapting pre-trained models to evolving unlabeled target domains.
• Current CTTA methods often synchronize domain-specific (DS) and domain-invariant (DI) information learning, leading to error accumulation and catastrophic forgetting due to cross-domain discrepancies.

Purpose of the Study:

• To propose a novel dual domain-attribute learning framework inspired by the brain's Attribute Memory Model (AMM).
• To enable models to learn new knowledge without forgetting previous information during continual adaptation.

Main Methods:

• Decomposition of CTTA into an event-based learning system (ELS) for DS style and a knowledge-based learning system (KLS) for DI structure.
• ELS actively builds adapters for new domains based on distribution shifts, while KLS uses a shared adapter for general knowledge.
• Implementation of a gradient conflict solver to manage conflicts between past and current DI knowledge for interference-free adaptation.

Main Results:

• The proposed framework demonstrated superior performance on both segmentation and classification CTTA tasks.
• Outperformed state-of-the-art approaches across four benchmark datasets.
• Successfully addressed error accumulation and catastrophic forgetting issues inherent in previous CTTA methods.

Conclusions:

• The dual domain-attribute learning framework offers an effective solution for CTTA by separating and managing DS and DI knowledge asynchronously.
• The independent asynchronous update mechanism, inspired by neuroscience, enhances model adaptability and knowledge retention in evolving domains.
• This approach represents a significant advancement in handling non-stationary environments for machine learning models.