PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

A fast and efficient toolbox for SVMs in Matlab
Gaëlle Loosli, Stéphane Canu, Alain Rakotomamonjy, Alex Smola and S V N Vishwanathan
In: NIPS'04, 13-18 Dec 2004, Vancouver, Canada.

Abstract

Technology: The presented Toolbox is an efficient implementation in MATLAB of the SimpleSVM algorithm. It provides the exact solution to the SVM problem. The extended version of the toolbox provides the treatment of invariances and multiclass. Both are done with the SimpleSVM algorithm as a baseline. Briefly, the main point of SimpleSVM is to add points one by one to the solution. We take advantage of this to treat different problems: Invariances Unlike the previous published methods to take invariances into account, the presented algorithm achieves the task with no pre-processing and only one run of the algorithm. When a point is to be added to the solution, we take either itself or one of its transformation if it improves the solution. Doing so we avoid a preprocessing phase and we divide a big optimization problem into small ones.

EPrint Type:Conference or Workshop Item (Other)
Additional Information:Toolbox webpage : http://asi.insa-rouen.fr/~gloosli/
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
ID Code:348
Deposited By:Gaëlle Loosli
Deposited On:16 December 2004