joint distribution adaptation via feature and model matching

It is usually supposed that the training (source domain) and test (target domain) data follow similar distributions and feature spaces in most pattern recognition tasks. however, in many real-world applications, particularly in visual recognition, this hypothesis has frequently been violated. thus, the trained classifier for the source domain performs poorly in the target domain. this problem is known as domain shift problem. domain adaptation and transfer learning are promising techniques towards an effective and robust classifier to tackle the shift problem. in this paper, a novel scheme is proposed for domain adaptation, named joint distribution adaptation via feature and model matching (jdafmm), in which feature transform and model matching are jointly optimized. by introducing regularization performed between the marginal and conditional distribution shifts across the domains, data drift can be successfully adapted as much as possible along with empirical risk minimization and rate of consistency maximization between manifold and prediction functions. extensive experiments were conducted to evaluate the performance of the proposed model against other machine learning and domain adaptation methods in three types of visual benchmark datasets. our experiments illustrated that our jdafmm significantly outperformed other baseline and state-of-the-art methods.