Disentangled Representation Learning for Domain Adaptation and Style Transfer
share
In order to solve unsupervised domain adaptation problem, recent methods focus on the use of adversarial learning to learn the common representation among domains. Although many designs are proposed, they seem to ignore the negative influence of domain-specific characteristics in transferring process. Besides, they also tend to obliterate these characteristics when extracted, although they are useful for other tasks and somehow help preserve the data. Take into account these issues, in this paper, we want to design a novel domain-adaptation architecture which disentangles learned features into multiple parts to answer the questions: what features to transfer across domains and what to preserve within domains for other tasks. Towards this, besides jointly matching domain distributions in both image-level and feature-level, we offer new idea on feature exchange across domains combining with a novel feed-back loss and a semantic consistency loss to not only enhance the transfer-ability of learned common feature but also preserve data and semantic information during exchange process. By performing domain adaptation on two standard digit data sets, we show that our architecture can solve not only the full transfer problem but also partial transfer problem efficiently. The translated image results also demonstrate the potential of our architecture in image style transfer application.
READ FULL TEXT