FEATURE EXTRACTION VIA KERNEL WEIGHTED DISCRIMINANT ANALYSIS METHODS

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "FEATURE EXTRACTION VIA KERNEL WEIGHTED DISCRIMINANT ANALYSIS METHODS"

By

Mr. Guang Dai


Abstract

In recent years, as the kernel extension of linear discriminant analysis
(LDA), kernel discriminant analysis (KDA) has become one of popular and
powerful tools for dimensionality reduction and feature extraction in
fields of machine learning and pattern recognition, with superior
performance in many practical applications. In this thesis, I investigate
three adverse problems that can impair the performance of KDA in many
real-world applications, and these problems are referred as: the small
sample size problem, the outlier classes problem, and the heteroscedastic
problem. In order to overcome the adverse problems for KDA, in this
thesis, I also trend to propose three different weighted KDA methods:
kernel fractional-step discriminant analysis (KFDA), complete kernel
fraction-step discriminant analysis (CKFDA), and heteroscedastic kernel
weighted discriminant analysis (HKWDA). KFDA effectively overcomes the
adverse effects of outlier classes via a weighted Fisher criterion; CKFDA
not effectively solves the outlier classes problem via the weighted
discriminant criteria but only solves the small sample size problem via
calculating the discriminant information in two orthogonal subspaces;
HKWDA not effectively solves both the outlier classes problem and the
heteroscedastic problem via a weighted chernoff criterion, but also
considers the small sample size problem via simply calculating the crucial
discriminant information in the null subspace of the within-class scatter.
Extensive comparative studies performed on many face databases reveal the
effectiveness of three weighted KDA methods proposed in this thesis.


Date:				Wednesday, 24 October 2007

Time:				11:00a.m.-1:00p.m.

Venue:				Room 5564
				Lifts 27-28

Committee Members:		Prof. Lionel Ni (Supervisor)
				Dr. Nevin Zhang (Chairperson)
				Dr. Lei Chen


**** ALL are Welcome ****