Estimation of Class Probability through Adversarial Training for Partial Domain Adaptation

Abstract

Domain adaptation is an approach to transfer knowledge from a certain source domain to another target domain. Several studies are recently reported on partial domain adaptation (PDA), where the class set of the source domain is larger than that of the target domain as a more realistic setting, but then the source domain specific classes make the adaptation more difficult. Most existing methods for PDA give small weights to the source domain specific classes to prevent the target data from being matched. The present paper proposes a PDA method which introduces a novel mechanism that gives additional weights to an individual target data by estimating the probability that the data belongs to each source class. The estimation is given by multiple discriminators that measure the distance between the data distribution of each source class and the entire target data distribution through adversarial training against a data encoder. Computer experiments using two handwritten digit datasets as two domains show that the proposed method achieves more stable and accurate domain adaptation compared with state-of-the-art existing methods for PDA.