Consistent Assistant Domains Transformer for Source-Free Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Source-free domain adaptation (SFDA) faces challenges due to inaccessible source data, hindering the extraction of deterministic invariable features.
• Existing SFDA methods often struggle with hard samples and domain bias by focusing on target domain features resembling the source domain.

Purpose of the Study:

• To propose a novel method, Consistent Assistant Domains Transformer (CADTrans), for effective SFDA.
• To address limitations in feature representation diversity and improve robustness against domain bias and hard samples in SFDA.

Main Methods:

• Developed an assistant domain module within CADTrans to generate diversified feature representations from aggregated global attentions.
• Employed multiple consistent strategies using assistant and target domains to obtain invariable feature representations for distinguishing easy and hard samples.
• Introduced a conditional multi-kernel max mean discrepancy (CMK-MMD) strategy to align hard samples with easy samples by differentiating between intra-class and inter-class samples.

Main Results:

• CADTrans demonstrates significant performance improvements across various benchmarks, including Office-31, Office-Home, VISDA-C, and DomainNet-126.
• The proposed methods effectively construct invariable feature representations with domain consistency.
• The approach successfully distinguishes and aligns hard samples, mitigating issues of domain bias.

Conclusions:

• CADTrans offers a robust solution for SFDA by enhancing feature consistency and diversity.
• The method shows superior performance compared to existing approaches on standard SFDA benchmarks.
• The developed techniques provide a promising direction for future research in domain adaptation without source data.