Lei Xu
Dept. of Computer Science and Engineering, The Chinese Univ. of Hong Kong

Abstract
Independent component analysis (ICA) has been widely studied for a decade with many applications. Still it is a hot topic in the current literature. Actually, ICA is a typical example of learning independence subspace. This talk makes an overview on over two decade advances on learning independence subspaces from a unified perspective. These independence subspaces involve (a) PCA, orthogonal subspace, and factor analysis; (b) ICA, LMSER learning, and independent factor models, as well as (c) their modular and temporal extensions. The subspaces are systematically examined not only from the natures of spanning bases, the distributions of variables on the bases, and the ways of mapping samples into the subspaces, as well as on how dependence between samples are considered, but also in an integrated bidirectional approach as a special case of Bayesian Ying Yang (BYY) harmony learning. Also, new results are presented, including a mathematical proof of one-bit-matching ICA theorem in a strong sense and adaptive algorithms that are able to learn subspaces with their dimensions deterimined atomatically during learning.

Speaker
Lei Xu is a chair professor of Chinese Univ Hong Kong. He completed his Ph.D thesis at Tsinghua Univ. by the end of 1986, then joined Dept.Math, Peking Univ in 1987 first as a postdoc and then an exceptionally promoted associate professor in 1988. During 1989-93, he worked at several universities in Finland, Canada and USA, including Harvard and MIT. He joined CUHK in 1993 as senior lecturer, became professor in 1996 and took the current chair professor in 2002. Prof. Xu has served or is serving as associate editor for several international journals, including Neural Networks, IEEE Trans. on Neural Networks, as a governor of International Neural Network Society(01-03), the chair of Computational Finance Technical Committee of IEEE Neural Networks Society(01-03), and a past president of Asian-Pacific Neural Networks Assembly. Prof. Xu has made several well-cited contributions on adaptive PCA and independence learning, classifier combination and mixture model based learning, rival penalized competition, and Bayesian Ying-Yang learning system and theory. Also, he and Oja’s invention on Randomized Hough Transform has a wide impact in the field of pattern recognition. He has given many keynote/plenary/invited/tutorial talks in international conferences. He has received several Chinese national prestigious academic awards (including 1993 National Nature Science Award) and international awards (including 1995 INNS Leadership Award). Prof Xu is an IEEE Fellow and a Fellow of International Association for Pattern Recognition, and a member of European Academy of Sciences.