DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised\n Domain Adaptation

In computer vision, one is often confronted with problems of domain shifts,\nwhich occur when one applies a classifier trained on a source dataset to target\ndata sharing similar characteristics (e.g. same classes), but also different\nlatent data structures (e.g. different acquisition conditions). In such a\nsituation, the model will perform poorly on the new data, since the classifier\nis specialized to recognize visual cues specific to the source domain. In this\nwork we explore a solution, named DeepJDOT, to tackle this problem: through a\nmeasure of discrepancy on joint deep representations/labels based on optimal\ntransport, we not only learn new data representations aligned between the\nsource and target domain, but also simultaneously preserve the discriminative\ninformation used by the classifier. We applied DeepJDOT to a series of visual\nrecognition tasks, where it compares favorably against state-of-the-art deep\ndomain adaptation methods.\n