A Pseudo-Label Guided Hybrid Approach for Unsupervised Domain Adaptation

Abstract

[Abstract]: Unsupervised domain adaptation focuses on reusing a model trained on a source domain in an unlabeled target domain. Two main approaches stand out in the literature: adversarial training for generating invariant features and minimizing the discrepancy between feature distributions. This paper presents a hybrid approach that combines these two methods with pseudo-labeling. The proposed loss function enhances the invariance between the features across domains and further defines the inter-class differences of the target set. Using a well-known dataset, Office31, we demonstrate that our proposal improves the overall performance, especially when the gap between domains is more significant.