Kernel-based Class Separability: Theory and Applications

Lei Wang (The Australian National University)

COMPUTER VISION AND ROBOTICS SERIES

DATE: 2010-02-11
TIME: 16:00:00 - 17:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Class separability is a classic concept in statistical pattern recognition. The scatter-matrix-based class separability measure has been widely used in discriminant analysis, feature selection, and data clustering. This measure can conveniently incorporate the kernel trick to give a kernel-based separability measure. This talk will show the intrinsic relationship between this separability measure and the radius-margin bound of SVMs, the Kernel Alignment criterion, and the Kernel Fisher Discriminant Analysis, providing more insight into the kernel-based class separability measure. Moreover, these relationships indicate that this separability measure can have a wide range of applications. Its applications to kernel parameter tuning and feature selection are discussed, and its efficiency in handling non-separable and noisy data is demonstrated.
BIO:
Dr. Lei Wang received both a B.Eng and an M.Eng degree from Southeast University, China in 1996 and 1999, respectively, and a PhD degree from the School of EEE at Nanyang Technological University, Singapore in 2004. He worked as a research associate and research fellow at Nanyang Technological University from 2003 to 2005. After that, he joined the Department of Information Engineering, RSISE, The Australian National University as a research fellow. He is now a Fellow in the School of Engineering, College of Engineering and Computer Science. He was awarded an Australian Post-doctoral Fellowship and an Early Career Researcher Award by the Australian Research Council in 2007 and 2009, respectively. His research interests include pattern recognition, computer vision and machine learning.