Raghuraman Gopalan , AT&T Labs-Research, Middletown
Ruonan Li , Harvard University, Cambridge
Rama Chellappa , University of Maryland, College Park
With unconstrained data acquisition scenarios prevalent in recent times, the ability to handle changes in data distribution across training and testing is an important feature to possess for statistical learners. One way to approach this problem is through domain adaptation, and in this paper we primarily focus on the unsupervised scenario where the labeled source domain data is accompanied by unlabeled target domain data. Starting with linear generative subspace representation of domains, we first utilize the geometry of these subspaces, the Grassmann manifold, to obtain a geodesic path between the domains. We then sample points along the geodesic to obtain intermediate data representations, using which a discriminative classifier is learnt to estimate the target labels. We subsequently incorporate non-linear domain representations, and examine the applicability of our approach for semi-supervised adaptation where target domain is partially labeled, and multi-domain adaptation where there could be more than one domain in source and target. Finally, we supplement our technique with a domain-shift prior to encode the notion of physical relevance to the cross-domain representations, and a boosting analysis to obtain robustness to algorithmic parameter choices. We evaluate our approach for object recognition problems and report competitive results on Office and Bing datasets.
Grassmann manifold, Domain adaptation, Object recognition, Unsupervised
R. Chellappa, R. Li and R. Gopalan, "Unsupervised Adaptation Across Domain Shifts By Generating Intermediate Data Representations," in IEEE Transactions on Pattern Analysis & Machine Intelligence.