AdaMatch: A Unified Approach to Semi-Supervised Learning and Domain Adaptation

We extend semi-supervised learning to the problem of domain adaptation to\nlearn significantly higher-accuracy models that train on one data distribution\nand test on a different one. With the goal of generality, we introduce\nAdaMatch, a method that unifies the tasks of unsupervised domain adaptation\n(UDA), semi-supervised learning (SSL), and semi-supervised domain adaptation\n(SSDA). In an extensive experimental study, we compare its behavior with\nrespective state-of-the-art techniques from SSL, SSDA, and UDA on vision\nclassification tasks. We find AdaMatch either matches or significantly exceeds\nthe state-of-the-art in each case using the same hyper-parameters regardless of\nthe dataset or task. For example, AdaMatch nearly doubles the accuracy compared\nto that of the prior state-of-the-art on the UDA task for DomainNet and even\nexceeds the accuracy of the prior state-of-the-art obtained with pre-training\nby 6.4% when AdaMatch is trained completely from scratch. Furthermore, by\nproviding AdaMatch with just one labeled example per class from the target\ndomain (i.e., the SSDA setting), we increase the target accuracy by an\nadditional 6.1%, and with 5 labeled examples, by 13.6%.\n