A Versatile Framework for Unsupervised Domain Adaptation based on Instance Weighting.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) faces challenges with label shifts across domains.
• Existing methods struggle with diverse UDA settings like closed, partial, open, and universal domain adaptation.
• Discriminating common/private classes and measuring domain discrepancy are key difficulties.

Purpose of the Study:

• To propose a general and effective method for Unsupervised Domain Adaptation (UDA) across various settings.
• To address challenges posed by label shifts and domain-specific classes.
• To introduce Learning Instance Weighting for Unsupervised Domain Adaptation (LIWUDA).

Main Methods:

• Developed a weight network to assign instance probabilities for common classes.
• Designed Weighted Optimal Transport (WOT) for domain alignment using instance weights.
• Implemented Separate and Align (SA) loss and Intra-domain Optimal Transport (IOT) for instance separation/alignment and weight network guidance.

Main Results:

• The proposed LIWUDA method demonstrates effectiveness across four distinct UDA settings.
• Experimental evaluations on benchmark datasets validate the method's performance.
• LIWUDA successfully unifies the approach to diverse UDA challenges.

Conclusions:

• LIWUDA offers a unified and effective solution for complex Unsupervised Domain Adaptation problems.
• The method shows strong performance in handling label shifts and domain discrepancies.
• LIWUDA advances the field of domain adaptation by providing a versatile approach.