DCST: Dual Cross-Supervision for Transformer-based Unsupervised Domain Adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) seeks to apply models trained on labeled source data to unlabeled target data.
• Significant domain disparities between source and target data present a major challenge for UDA.
• Existing UDA methods struggle with noisy pseudo-labels and rigid domain alignment rules, especially in challenging domains.

Purpose of the Study:

• To propose a novel method, Dual Cross-Supervision Transformer (DCST), to improve UDA performance in challenging domains.
• To enhance cross-domain feature alignment and knowledge transfer despite significant domain gaps.
• To mitigate issues like pseudo-labeling noise, model collapse, and overfitting.

Main Methods:

• Integrating a vision transformer with a dual cross-supervision learning strategy for consistent learning across domains.
• Implementing domain-specific self-training and adaptive cross-domain feature alignment.
• Introducing a Domain Shift Filter to leverage source domain feature memory for smoother transitions and reduced risks.

Main Results:

• Achieved state-of-the-art classification performance on four benchmark datasets.
• Obtained high accuracy rates: 94.3% on Office-31, 86.0% on Office-Home, 89.3% on VisDA-2017, and 48.8% on DomainNet.
• Demonstrated effective knowledge transfer and improved model robustness in the presence of domain shift.

Conclusions:

• DCST significantly advances unsupervised domain adaptation, particularly for domains with substantial differences.
• The proposed Domain Shift Filter effectively handles noisy data and prevents model degradation.
• The method offers a robust solution for leveraging labeled data in unlabeled, challenging target domains.