Disentangled cyclic reconstruction for domain adaptation

The domain adaptation problem involves learning a unique classification or regres-sion model capable of performing on both a source and a target domain. Althoughthe labels for the source data are available during training, the labels in the targetdomain are unknown. An effective way to tackle this problem lies in extractinginsightful features invariant to the source and target domains. In this work, wepropose splitting the information for each domain into a task-related representa-tion and its complimentary context representation. We propose an original methodto disentangle these two representations in the single-domain supervised case. Wethen adapt this method to the unsupervised domain adaptation problem. In partic-ular, our method allows disentanglement in the target domain, despite the absenceof training labels. This enables the isolation of task-specific information fromboth domains and a projection into a common representation. The task-specificrepresentation allows efficient transfer of knowledge acquired from the source do-main to the target domain. We validate the proposed method on several classicaldomain adaptation benchmarks and illustrate the benefits of disentanglement fordomain adaptation.