Unsupervised Multitarget Domain Adaptation With Dictionary-Bridged Knowledge Exploitation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) typically focuses on single-source to single-target scenarios.
• Adapting models from a single labeled source to multiple unlabeled targets presents unique challenges due to inter-target domain relationships.
• Existing UDA methods are limited in addressing the complexities of multitarget domain adaptation.

Purpose of the Study:

• To propose a novel unsupervised multitarget domain adaptation method (DL-UMTDA).
• To effectively leverage knowledge from a single labeled source domain to multiple unlabeled target domains.
• To address the challenge of modeling relationships between source and multiple target domains, as well as among target domains themselves.

Main Methods:

• Dictionary learning-based unsupervised multitarget domain adaptation (DL-UMTDA).
• Construction of a common dictionary to correlate source and multitarget domains.
• Design of individual dictionaries to capture private knowledge within each target domain.
• Utilizing dictionary representation coefficients for knowledge transfer and relationship exploitation.
• Development of an alternating algorithm with theoretical convergence guarantee.

Main Results:

• DL-UMTDA effectively exploits correlations between source and target domains, and among target domains.
• The proposed method demonstrates superior performance compared to existing approaches.
• Validation on benchmark (Office + Caltech) and real-world datasets (AgeDB, Morph, CACD).

Conclusions:

• Dictionary learning provides a robust framework for unsupervised multitarget domain adaptation.
• DL-UMTDA successfully addresses the challenges of adapting a single source to multiple targets.
• The method offers a significant advancement in UDA, particularly for multitarget scenarios.