Multi-Source Domain Adaptation with Sinkhorn Barycenter

We describe a multi-source and unsupervised domain adaptation method using Sinkhorn barycenters, which, given the labeled data in multi-source domains and unlabeled data in a target domain, uses the optimal transport Sinkhorn distance to measure gaps between data distributions in the source and target domains. For end-to-end classification learning, the feature extractor and classifier are simultaneously estimated on the basis of two criteria: the minimization of the Sinkhorn distance for the source and target domains and the minimization of the classification loss for the source domains. The first criterion is based on the assumptions that domain-invariant features would be captured in a latent feature space obtained by minimizing the Sinkhorn distance among all domains and that the space would be close to the Sinkhorn barycenter. Experiments on image classification using the Digit-Five dataset, which is comprised of digit datasets from five different domains, demonstrated that our method outperforms other state-of-the-art methods.