New Algorithm Improves Universal Domain Adaptation with Joint Distribution Alignment.

Unsupervised Domain Adaptation (UDA) is a well-established field in machine learning, typically assuming that the source and target domains share identical label spaces, with the primary challenge being sample distribution drift. However, real-world applications often present a more complex scenario: Universal Domain Adaptation (UniDA), where both sample distribution drift and differences in class spaces exist between domains. Current research rarely offers theoretical analysis for UniDA. This work addresses this gap by providing a theoretical upper bound for the generalization error in universal domain adaptation. Building upon this theoretical foundation, the authors propose a new UniDA algorithm called Joint Distribution Alignment for Universal Domain Adaptation (JAUA). JAUA tackles the problem by aligning the joint distributions of the source and target domains, minimizing their discrepancy using Chi-Square divergence. Additionally, it incorporates a progressive pseudo-labeling method to assign labels to unlabeled target samples, gradually improving the model's understanding of the target domain. Experimental results across six public image datasets demonstrate JAUA's superior capability in handling the complexities of the UniDA problem, outperforming existing methods.