Progressive Decision Boundary Shifting for Unsupervised Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) addresses domain shift between labeled source and unlabeled target domains.
• Current UDA methods using biclassifiers and self-training struggle with error accumulation from semantically uncertain instances.

Purpose of the Study:

• To develop a novel algorithm for unsupervised domain adaptation that mitigates error accumulation in semantically uncertain target instances.
• To improve task-specific generalization in target domains by enhancing instance discriminability and category-level alignment.

Main Methods:

• Designed a progressive decision boundary shifting algorithm to model and leverage semantic uncertainty.
• Introduced uncertainty decoupling using contrastive learning to extract discriminative information from low-uncertainty instances.
• Minimized predictive entropy for high-uncertainty instances to reduce prediction confidence.

Main Results:

• The proposed method effectively models semantic uncertainty in target domain data.
• Achieved improved discriminability and category-level alignment compared to existing UDA techniques.
• Demonstrated superior performance over state-of-the-art UDA methods on three benchmark datasets.

Conclusions:

• The progressive decision boundary shifting algorithm offers a robust solution for unsupervised domain adaptation.
• Addressing semantic uncertainty is crucial for preventing discriminability degradation and category misalignment in UDA.
• The proposed approach significantly advances the capabilities of UDA models in real-world applications.