Invited Speaker---Dr. Sang C. Lee


Senior Researcher, Daegu-Gyeongbuk Institute of Science and Technology (DGIST), South Korea

Biography: Sang C. Lee received the B.S., M.S., and Ph.D. degrees in electrical and electronics engineering from Pohang University of Science and Technology, Pohang, Korea, in 1994, 1996, and 2002, respectively. He was a Postdoctoral Researcher (supervisor: Romeo Ortega) with the Centre National de la Recherche Scientifique (CNRS), Supelec, France. Also, he was a project leader at the Samsung SDI Ltd. and Samsung Techwin to make a portable fuel cell and wireless sensor network, respectively. He is currently a Senior Researcher with the convergence research institute, Daegu-Gyeongbuk Institute of Science and Technology (DGIST), Daegu. His main research interests are in hybrid control and machine deep learning.

Speech Title: Discriminative Feature Learning Utilizing A Force-based Cost Function
Abstract: In this paper we compare the ways of rearrangement of invariant mapping using the conventional probabilistic type I and energy-based type II loss functions, and our proposed force-based type III loss functions. The enlarged inter-class variations and reduced intra-class variations are fulfilled at the same time by employing the modification of loss to obtain low dimensional features more discriminatively on a manifold. The discriminatively learned features by the type III enhance the accuracy compared with regular training approaches. Moreover, there are lots of advantages of discriminatively learned features to construct these on the CNNs. We introduce center loss and expansion loss functions which are capable of achieving small intra-class variations and large inter-class distances. The center loss and expansion loss considers the two aspects of feature variations; firstly, minimizing the intra-class variations while keeping the features a part and secondly maximizing the inter-class distances while keeping the minimization of intra-class variations. The first aspect shares the same idea with the center loss to minimize the intra-class variations while keeping the correct orders of each images. The second aspect further penalizes the global center variations to enhance the different pair of features to be more apart.