Domain consistency regularization for unsupervised multi-source domain adaptive classification

Deep learning-based multi-source unsupervised domain adaptation (MUDA) has been actively studied in recent years. Compared with single-source (SUDA), shift MUDA exists not only between the source and target domains but also among multiple domains. Most existing algorithms focus on extracting domain-invariant representations all whereas task-specific decision boundaries classes are largely neglected. In this paper, we propose an end-to-end trainable network that exploits Consistency Regularization for Multi-source Adaptive classification (CRMA). CRMA aligns distributions of each pair For domains, employ intra-domain consistency to regularize a domain-specific classifiers achieve alignment. addition, design inter-domain targets joint alignment To address different similarities domain, authorization strategy assigns authorities adaptively optimal pseudo label prediction self-training. Extensive experiments show tackles effectively under setup achieves superior consistently across datasets.