题目: Independent component analysis for non-independent sources and nonlinear mixtures 报告人：Kun Zhang
Postdoctoral fellow
Chinese University of Hong Kong 时间：6月17日，上午10：30-11：30 地点：蒙民伟楼404会议室 摘要：Independent component analysis (ICA) is a powerful statistical technique
for finding hidden and meaningful sources underlying the observations. Genera
lly it assumes that the sources are mutually independent and linearly mixed. T
his greatly limits the application of ICA for many real-world problems. In thi
s talk, I extend ICA to admit non-independent sources or nonlinear mixtures. F
irst, I consider the subband decomposition ICA (SDICA) model, which relaxes th
e independence assumption by assuming that sources are independent only in som
e frequency subbands, and show its separability. An adaptive method for SDICA
is then presented. The relationship of SDICA to overcomplete ICA is also discu
ssed. Next, I present the “minimal nonlinear distortion” (MND) principle to
tackle the inherent ill-posedness of nonlinear ICA. MND prefers the solution w
ith the mixing procedure close to linear. As an application, nonlinear ICA wit
h MND is used to discover the linear causal relations among major stocks in Ho
ng Kong. 简历：Zhang Kun is currently a postdoctoral fellow in Department of Computer S
cience and Engineering at the Chinese University of Hong Kong. He is also work
ing in Department of Mathematics, Hong Kong Baptist University, as a visiting
scholar. From 1996 to 2001, he studied for the Bachelor degree in Automation a
t University of Science and Technology of China. He received his Ph.D degree i
n Computer Science from the Chinese University of Hong Kong in 2005. He was t
he organizer of special session on independent component analysis at ICONIP 20
06, and has been serving as a reviewer for TSP, TNN, Quantitative Finance, Neu
rocomputing, etc. His research interests include multivariate statistics, spar
se coding, high-dimensional statistical learning, and computational finance. |
