Unsupervised domain adaptation with weak source domain labels via bidirectional subdomain alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) facilitates knowledge transfer between labeled source and unlabeled target domains.
• Real-world UDA is challenged by noisy or missing source domain labels, hindering performance.
• Existing methods often struggle with the unreliability of source data labels.

Purpose of the Study:

• To develop a novel framework for robust feature extraction in UDA, specifically addressing unreliable source labels.
• To enhance UDA performance by improving feature discriminability and domain alignment.
• To enable effective knowledge transfer even when source domain labels are imperfect.

Main Methods:

• Iterative pseudo-labeling and queue-based clustering to generate label-independent class centroids.
• Bidirectional Subdomain Alignment (BSA) for rectifying and enhancing features by matching subdomain distributions.
• A two-stage adversarial process for global feature alignment, optimizing via the Expectation-Maximization algorithm.

Main Results:

• Achieved significant improvements over state-of-the-art methods on benchmark datasets (Office-31, Office-Home, VisDA-2017).
• Reported average accuracies of 91.5% (Office-31), 76.6% (Office-Home), and 87.4% (VisDA-2017).
• Demonstrated consistent superiority in UDA tasks with both fully and weakly labeled source domains.

Conclusions:

• The proposed framework effectively extracts robust and discriminative features despite unreliable source labels.
• BSA and adversarial learning significantly enhance feature rectification and global alignment.
• The method offers a robust solution for UDA, outperforming existing approaches in challenging real-world scenarios.