Distribution-Matching Embedding for Visual Domain Adaptation

摘要

Domain-invariant representations are key to addressing thedomain shift problem where the training and test examples followdifferent distributions. Existing techniques that have attemptedto match the distributions of the source and target domainstypically compare these distributions in the original featurespace. This space, however, may not be directly suitable forsuch a comparison, since some of the features may have beendistorted by the domain shift, or may be domain specific. Inthis paper, we introduce a Distribution-Matching Embeddingapproach: An unsupervised domain adaptation method thatovercomes this issue by mapping the data to a latent space wherethe distance between the empirical distributions of the sourceand target examples is minimized. In other words, we seek toextract the information that is invariant across the source andtarget data. In particular, we study two different distances tocompare the source and target distributions: the Maximum MeanDiscrepancy and the Hellinger distance. Furthermore, we showthat our approach allows us to learn either a linear embedding,or a nonlinear one. We demonstrate the benefits of our approachon the tasks of visual object recognition, text categorization,and WiFi localization. color="gray">