Connecting convex energy-based inference and optimal transport for domain adaptation

The connection of optimal transport and neural networks finds a rich application in machine learning problems. In this paper, we propose a simple algorithm for the mutual improvement of optimal transport and energy-based models for the semi-supervised domain adaptation. Having target and source domain samples we use convex energy-based inference to create a new domain that is class-wise cyclical monotone to the target domain and its samples contains features from the source domain. Mapping from target to such domain can be solved by optimal transport much more successfully. We study the performance of our approach by benchmarking it on a range of optimal transport methods and showed that in our settings optimal transport can achieve much higher results.