An overview of Support Vector Machines and Kernel Methods

J. S. Marron
Department of Statistics
University of North Carolina

    There have been some exciting new methodologies developed for
addressing the classical statistical problem of classification (also known
as discrimination).  Most work has been done outside of the statistical
community, in a new branch of Computer Science called "Machine Learning".
This talk will present an overview, from an intuitive statistical
viewpoint, of two major ideas in this area, the Support Vector Machine and
Kernel Embedding methods.  Special attention will be paid to High
Dimension, Low Sample Size contexts.  Note that classical statistical
multivariate analysis is useless in HDLSS settings, because the first step
of "sphering the data" fails due to singularity of the covariance matrix. 
Performance of these methods will be illustrated in the context of
examples from medical image analysis, gene expression micro-array
analysis, and chemometrics, where HDLSS problems are endemic.