- TL;DR: A novel unsupervised domain adaptation paradigm - performing adaptation without accessing the source data ('source-free') and without any assumption about the source-target category-gap ('universal').
- Abstract: There is a strong incentive to develop versatile learning techniques that can transfer the knowledge of class-separability from a labeled source domain to an unlabeled target domain in the presence of a domain-shift. Existing domain adaptation (DA) approaches are not equipped for practical DA scenarios as a result of their reliance on the knowledge of source-target label-set relationship (e.g. Closed-set, Open-set or Partial DA). Furthermore, almost all the prior unsupervised DA works require coexistence of source and target samples even during deployment, making them unsuitable for incremental, real-time adaptation. Devoid of such highly impractical assumptions, we propose a novel two-stage learning process. Initially, in the procurement-stage, the objective is to equip the model for future source-free deployment, assuming no prior knowledge of the upcoming category-gap and domain-shift. To achieve this, we enhance the model’s ability to reject out-of-source distribution samples by leveraging the available source data, in a novel generative classifier framework. Subsequently, in the deployment-stage, the objective is to design a unified adaptation algorithm capable of operating across a wide range of category-gaps, with no access to the previously seen source samples. To achieve this, in contrast to the usage of complex adversarial training regimes, we define a simple yet effective source-free adaptation objective by utilizing a novel instance-level weighing mechanism, named as Source Similarity Metric (SSM). A thorough evaluation shows the practical usability of the proposed learning framework with superior DA performance even over state-of-the-art source-dependent approaches.
- Keywords: unsupervised domain adaptation, knowledge transfer, source-free adaptation