Research on Distribution Alignment and Semantic Consistency in the Adversarial Domain Adaptation

摘要

Domain adaptation is an effective method solving the learning tasks lack of labeled data. In recent years, the adversarial domain adaptation (ADA) has achieved attractive results in a series transfer learning tasks. ADA reduces the distribution discrepancy between the source and the target by extracting the domain invariant features. However, the lack of constraints on the transferable features leads to poor results even negative transfers. A novel ADA method is proposed to solve this problem which contains two main improvements: the conditional distribution alignment and the semantic consistency regularization. The experiment demonstrate that the proposed method has promising improvement in the classification accuracy on the benchmark dataset. The code and data can be downloaded from https://github.com/kiradiso/EADA.