Domain Adaptation with Multiple Latent Domains

Judith Hoffman, Kate Saenko1, Brian Kulis2 and Trevor Darrell

Defense Advanced Research Projects Agency

Domain adaptation is becoming increasingly popular as a research topic in object recognition. Recent methods successfully learn cross-domain transforms to map points between source and target domains, yet are either restricted to a single training domain, or assume that the separation into source domains is known a priori. We introduce a method for multi-domain adaptation based on learning nonlinear cross-domain transforms, that can be applied with both known or unknown (latent) domain labels. For the latent case, we develop a method based on a novel version of constrained clustering. Unlike many existing constrained clustering algorithms, ours can be shown to provably converge locally while satisfying all constraints.

1UC Berkeley, ICSI
2UC Berkeley, ICSI