BIAS: Bridging Inactive and Active Samples for active source free domain adaptation

Document Type

Article

Publication Title

Knowledge-Based Systems

Abstract

As a practical domain adaptation scenario for data protection, source free domain adaptation has suffered from a performance bottleneck. Active source free domain adaptation (ASFDA) could alleviate this problem by exploring a few actively labeled target samples. Due to active selection bias in ASFDA, the active samples are hard samples as they are challenging to recognize by the source model, and the inactive samples are composed of well-learned not-hard samples and unexplored mild-hard samples. Thus the distributions of inactive and active samples are usually different. We provide a rigorous theoretical analysis, revealing that reducing their distribution difference is critical for model generalization. However, the previous DA alignment methods cannot reduce such distribution discrepancy due to severe class imbalance and quantity imbalance between the inactive and active samples. In this work, we propose Bridging Inactive and Active Samples (BIAS) to implicitly align their distributions by utilizing their distribution characteristics. By constructing and learning the intermediate samples between the reliable not-hard and hard samples obtained by our novel reciprocal active selection, BIAS can effectively align the distributions of inactive and active samples and learn unexplored mild-hard samples thoroughly. Extensive experiments verify that BIAS can be integrated into existing ASFDA methods and improve their performance significantly, with over 4% accuracy improvements in some transfer tasks.

DOI

10.1016/j.knosys.2023.111151

Publication Date

1-25-2024

Keywords

Active learning, Mix-up, Source free domain adaptation

Comments

IR conditions: non-described

Link to Full Text

Share

COinS