A Novel Transfer Metric Learning Approach Based on Multi-Group

Abstract

In recent years, transfer learning receives increasingly attention ranging from the communities of developmental robots, computer vision to artificial intelligence. In the research of transfer learning, knowledge should be transferred from the source domain to the target domain. The source domain is used to train a classifier while the target domain is for testing. Existing works consider the source domain as a whole, however, samples in the source domain might be extracted into different groups and the samples in the same group would have similar intrinsic attributes. In this work, we propose a novel transfer metric learning framework based on multi-group, called TMLMG. In TMLMG, based on each group both a Mahalanobis distance metric and a basic classifier are learned to make predictions. A weight matrix is used to describe the prediction capabilites of all the combinations of groups and Mahalanobis distance metrics. The weight matrix is initialized and optimized based on the labeled samples in the target domain. Experimental results on publicly available datasets of object recognition and handwriting recognition verify the effectiveness of our proposed TMLMG in knowledge transfer.