This website requires JavaScript.

Unsupervised domain adaptation by learning using privileged information

Adam BreitholtzAnton MatssonFredrik D. Johansson
Mar 2023
Successful unsupervised domain adaptation (UDA) is guaranteed only understrong assumptions such as covariate shift and overlap between input domains.The latter is often violated in high-dimensional applications such as imageclassification which, despite this challenge, continues to serve as inspirationand benchmark for algorithm development. In this work, we show that access toside information about examples from the source and target domains can helprelax these assumptions and increase sample efficiency in learning, at the costof collecting a richer variable set. We call this domain adaptation by learningusing privileged information (DALUPI). Tailored for this task, we propose asimple two-stage learning algorithm inspired by our analysis and a practicalend-to-end algorithm for multi-label image classification. In a suite ofexperiments, including an application to medical image analysis, we demonstratethat incorporating privileged information in learning can reduce errors indomain transfer compared to classical learning.