Unsupervised domain adaptation via feature alignment and redundancy reduction

Gauthier Miralles, Pietro Gori, Yanis Kahil, Vincent Jugnon, Loic Le Folgoc

Keywords: Unsupervised domain adaptation, deep learning, feature alignment, redundancy reduction

Abstract:

Unsupervised domain adaptation (UDA) addresses the challenge of transferring knowledge from a labeled source domain to an unlabeled target domain for the same task under data distribution shifts. Current approaches rely on strong hypotheses about the type of domain shift, or task to perform.
We propose FARR, a novel UDA method with 3 main contributions: (i) a new feature-alignment strategy based on redundancy reduction, that is task-adaptable and agnostic to the type of domain shift; (ii) a theoretical proof that our formulation effectively aligns source and target features; (iii) a comprehensive empirical evaluation across classification and segmentation tasks, using seven public and two private datasets covering diverse domain shifts. Our results show that FARR consistently outperforms existing feature-alignment methods, while remaining competitive with state-of-the-art UDA approaches across tasks and datasets.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 11184