Active Dynamic Weighting for multi-domain adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Multi-source unsupervised domain adaptation (MS-UDA) transfers knowledge from multiple labeled sources to an unlabeled target domain.
• Existing MS-UDA methods often overlook domain-specific discrepancies and struggle with complex cross-domain category boundaries.

Purpose of the Study:

• To propose a novel Active Dynamic Weighting (ADW) method for enhanced multi-source unsupervised domain adaptation.
• To address the challenges of global and local feature distribution discrepancies and unclear category boundaries across domains.

Main Methods:

• ADW employs a dynamic adjustment mechanism for feature alignment between source and target domains within training batches.
• A dynamic boundary loss function is introduced to focus on hard samples, improving decision boundary clarity.
• Active learning is integrated with importance sampling for efficient selection of target domain samples for annotation, refining category-level domain alignment.

Main Results:

• Experiments on benchmark datasets demonstrate the superiority of the proposed ADW method.
• The dynamic weighting and boundary loss effectively improve feature alignment and classification accuracy.
• The integration of active learning enhances domain adaptation efficiency and category-level understanding.

Conclusions:

• ADW offers a significant advancement in multi-source unsupervised domain adaptation.
• The method effectively handles domain discrepancies and improves model generalization.
• ADW provides a robust framework for leveraging multiple labeled domains to adapt to an unlabeled target domain.