Domain Adaptation Meets Disentangled Representation Learning and Style Transfer
share
In order to solve the unsupervised domain adaptation problem, some methods based on adversarial learning are proposed recently. These methods greatly attract people's eyes because of the better ability to learn the common representation space so that the feature distributions among many domains are ambiguous and non-discriminative. Although there are many discussions and results, the success of the methods implicitly funds on the assumption that the information of domains are fully transferrable. If the assumption is not satisfied, the influence of negative transfer may degrade domain adaptation. In this paper, we proposed to relieve the negative effects by not only adversarial learning but also disentangled representation learning, and style transfer. In detail, our architecture disentangles the learned features into common parts and specific parts. The common parts represent the transferrable feature space, whereas the specific parts characterize the unique style of each individual domain. Moreover, we proposed to exchange specific feature parts across domains for image style transfer. These designs allow us to introduce five types of novel training objectives to enhance domain adaptation and realize style transfer. In our experiments, we evaluated domain adaptation on two standard digit data sets. The results show that our architecture can be adaptive well to full transfer learning and partial transfer learning. As a side product, the trained network also demonstrates high potential to generate style-transferred images.
READ FULL TEXT