#Number
TR-PDS-1995-005
#Title
A Parallel Algorithm for Optimal Margin Classifiers
#Author
Ross Baldick and Craig M. Chase
#Abstract
In a recent paper, Boser et al. describe an algorithm based
on quadratic programming for training a classifier that maximizes
the classification margin between dichotomous classes of patterns.
Boser et al.'s algorithm can treat classification problems
having parameter spaces of very high dimension because it solves the
problem in the dual space. If the margin in the original problem is
positive, that is, if there exists a hyperplane separating the two
classes, then the dual problem is a convex quadratic program. The
dual problem has dimension equal to the number of patterns.
Unfortunately, the number of patterns, while finite, is also very
large. Consequently, Boser et al.'s algorithm requires
considerable computational effort to execute. In this paper we
propose a parallel implementation of Boser et al.'s algorithm.
We demonstrate the implementation on an example database and analyze
its performance. We also indicate extensions to other problems
of similar structure.
#Bib
@TechReport{BC95,
author = "Ross Baldick and Craig. M. Chase",
title = "A Parallel Algorithm for Optimal Margin Classifiers",
institution = "Parallel and Distributed Systems Laboratory,
ECE Dept., University of Texas at Austin",
year = "1995",
number = "TR-PDS-1995-005",
note = "submitted to Symposium on Parallel and Distributed
Processing",
note = "available via ftp or WWW at maple.ece.utexas.edu
as technical report TR-PDS-1995-005"
}