Transfer Alignment Network for Double Blind Unsupervised Domain Adaptation

How can we transfer knowledge from a source domain to a target domain when each side cannot observe the data in the other side? The recent state-of-the-art deep architectures show significant performance in classification tasks which highly depend on a large number of training data. In order to resolve the dearth of abundant target labeled data, transfer learning and unsupervised learning leverage data from different sources and unlabeled data as training data, respectively. However, in some practical settings, transferring source data to target domain is restricted due to a privacy policy. In this paper, we define the problem of unsupervised domain adaptation under double blind constraint, where either the source or the target domain cannot observe the data in the other domain, but data from both domains are used for training. We propose TAN (Transfer Alignment Network for Double Blind Domain Adaptation), an effective method for the problem by aligning source and target domain features. 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 double blind domain adaptation, and 2) outperforms baselines regardless of the proportion of target domain data in the training data.