Dynamic bias alignment and discrimination enhancement for unsupervised domain adaptation

Unsupervised domain adaptation (UDA) aims to explore the knowledge of labeled source domain to help training the model of unlabeled target domain. By now, while most existing UDA approaches typically learn domain-invariant representations by directly matching the distributions across the domains, they pay less attention on respecting the cross-domain similarity and discrimination exploration. To address these issues, this article designs a kind of UDA with dynamic bias alignment and discrimination enhancement (UDA-DBADE). Specifically, in UDA-DBADE we define a dynamic balance factor by the ratio of the normalized cross-domain discrepancy to the discrimination, which decreases gradually in the process of UDA-DBADE. Afterward, we construct domain alignment with adversarial learning as well as distinguishable representations through advancing the discrepancy of multiple classifiers, and dynamically balance them with the defined dynamic factor. In this way, a larger weight is originally assigned on the domain alignment and then gradually on the discrimination enhancement in the learning process of UDA-DBADE. In addition, we further construct a bias matrix to characterize the discrimination alignment between the source and target domain samples. Compared to current state-of-the-art methods, UDA-DBADE achieves an average accuracy of 88.8% and 89.8% on Office-31 dataset and ImageCLEF-DA dataset, respectively. Finally, extensive experiments demonstrate that UDA-DBADE has an excellent performance.