Transfer alignment network for blind unsupervised domain adaptation

摘要

How can we transfer the knowledge from a source domain to a target domain when each side cannot observe the data in the other side? Recent transfer learning methods show significant performance in classification tasks by leveraging both source and target data simultaneously at training time. However, leveraging both source and target data simultaneously is often impossible due to privacy reasons. In this paper, we define the problem of unsupervised domain adaptation under blind constraint, where each of the source and the target domains cannot observe the data in the other domain, but data from both domains are used for training. We propose TAN (Transfer Alignment Network for Blind Domain Adaptation), an effective method for the problem by aligning source and target domain features in the blind setting. TAN maps the target feature into source feature space so that the classifier learned from the labeled data in the source domain is readily used in the target domain. Extensive experiments show that TAN (1) provides the state-of-the-art accuracy for blind domain adaptation outperforming the standard supervised learning by up to 9.0% and (2) performs well regardless of the proportion of target domain data in the training data.