Label Space-Induced Pseudo Label Refinement for Multi-Source Black-Box Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Unsupervised Domain Adaptation (UDA) typically requires source data/models, posing privacy and IP concerns.
• Black-box Domain Adaptation (BDA) uses API predictions for pseudo-labeling, but struggles with multi-source settings.
• Existing multi-source BDA methods lack effective pseudo-label generation strategies.

Purpose of the Study:

• To develop a novel training framework for multi-source black-box domain adaptation (MSBDA).
• To introduce a method that refines pseudo-labels by learning relationships among multiple source domains.
• To enable effective target model adaptation using only source API predictions.

Main Methods:

• Proposed Label Space-Induced Pseudo Label Refinement (LPR) framework for MSBDA.
• Introduced a Pseudo label Refinery Network (PRN) to learn inter-source domain relationships.
• Employed a dual-phase PRN: a warm-up for initial pseudo-labels and a refinement phase for improved accuracy.

Main Results:

• LPR effectively refines pseudo-labels by leveraging relationships between source domains.
• The dual-phase PRN successfully adapts the target model, mitigating noisy samples.
• Achieved competitive performance on four benchmark datasets in various domain adaptation settings.

Conclusions:

• LPR offers a robust solution for MSBDA, overcoming limitations of existing BDA methods.
• The framework demonstrates the efficacy of pseudo-label refinement through learned domain relationships.
• Provides theoretical support for the proposed mechanism, validating its effectiveness.