Unsupervised Domain Adaptation Method Based on Relative Entropy Regularization and Measure Propagation

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) is crucial for applying models to new data distributions.
• Distributional discrepancies between source and target domains hinder model generalization.
• Existing UDA methods often struggle with significant domain shifts.

Purpose of the Study:

• To propose a novel UDA framework integrating information-theoretic principles.
• To mitigate distributional discrepancies between source and target domains effectively.
• To enhance both global feature alignment and semantic consistency.

Main Methods:

• Relative entropy regularization using Kullback-Leibler (KL) divergence to align label distributions.
• Measure propagation to transfer probability mass and create pseudo-measures for the target domain.
• A dual mechanism combining these components for robust adaptation.

Main Results:

• The proposed framework consistently outperforms State-of-the-Art methods on OfficeHome and DomainNet datasets.
• Superior performance is observed, especially in scenarios with significant domain shifts.
• Demonstrated robustness, scalability, and theoretical grounding of the UDA framework.

Conclusions:

• The novel UDA framework effectively reduces domain distributional discrepancies.
• The integration of information-theoretic principles offers a new perspective on domain adaptation.
• The method shows significant improvements in cross-domain generalization capabilities.