Unsupervised Multi-Class Domain Adaptation: Theory, Algorithms, and Practice.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised multi-class domain adaptation (multi-class UDA) algorithms often lack theoretical grounding.
• Existing methods' learning objectives are primarily empirically motivated.

Purpose of the Study:

• To develop a theoretical framework for multi-class UDA.
• To propose novel algorithms for improved domain adaptation performance.

Main Methods:

• Introduced Multi-Class Scoring Disagreement (MCSD) divergence to measure domain distance.
• Developed a new data-dependent, probably approximately correct (PAC) domain adaptation bound.
• Proposed Multi-class Domain-adversarial learning Networks (McDalNets) and Domain-Symmetric Networks (SymmNets).

Main Results:

• MCSD fully characterizes relations between multi-class scoring hypotheses.
• The theoretical bound naturally suggests adversarial learning objectives.
• SymmNets demonstrate efficacy across closed set, partial, and open set UDA settings.
• Empirical studies validate theoretical analysis and show SymmNets' superiority.

Conclusions:

• The proposed theoretical framework provides a foundation for multi-class UDA.
• SymmNets offer a robust and versatile solution for domain adaptation challenges.
• Publicly available code facilitates further research and application.