DCSE: Domain Confusion with Self-Ensembling for Unsupervised Domain Adaptation

Data collection and annotation are time-consuming in machine learning, especially for large-scale problems. A com-mon approach for these problems is to transfer knowledge from the related labeled domain to the target one. There are two pop-ular ways to achieve this: adversariallearning and self-training. In this paper, we first analyze the training instability problem and feature-conditional distribution misalignment problem in adversariallearning process. Then, inspired by domain confusion and self-ensembling methods, we propose a combined model for unsupervised domain adaptation, Domain Confusion with Self-Ensembling (DCSE), to learn joint invariant representations of features and classes. Experimental results show that the proposed DCSE achieves better performance than the existing methods on various unsupervised domain adaptation benchmarks.