Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned\nfrom a labeled source dataset to solve similar tasks in a new unlabeled domain.\nPrior UDA methods typically require to access the source data when learning to\nadapt the model, making them risky and inefficient for decentralized private\ndata. This work tackles a practical setting where only a trained source model\nis available and investigates how we can effectively utilize such a model\nwithout source data to solve UDA problems. We propose a simple yet generic\nrepresentation learning framework, named \\emph{Source HypOthesis Transfer}\n(SHOT). SHOT freezes the classifier module (hypothesis) of the source model and\nlearns the target-specific feature extraction module by exploiting both\ninformation maximization and self-supervised pseudo-labeling to implicitly\nalign representations from the target domains to the source hypothesis. To\nverify its versatility, we evaluate SHOT in a variety of adaptation cases\nincluding closed-set, partial-set, and open-set domain adaptation. Experiments\nindicate that SHOT yields state-of-the-art results among multiple domain\nadaptation benchmarks.\n