Using the Nonsmooth Analysis in a Learning Process

0355020 - ÚI 2011 RIV US eng C - Conference Paper (international conference)
Jiřina, Marcel - Jiřina jr., M.
Using the Nonsmooth Analysis in a Learning Process.
Proceeding of The 2010 IRAST International Congress on Computer Applications and Computational Science. -: IRAST, 2010 - (Chellappan, S.; Cheng, A.; Min, M.; Quang, V.; Ramalingam, P.; Yang, H.), s. 91-94. ISBN 978-981-08-6846-8.
[CACS 2010. 2010 IRAST International Congress on Computer Applications and Computational Science. Singapore (SG), 04.12.2010--06.12.2010]
R&D Projects: GA MŠMT(CZ) 1M0567
Institutional research plan: CEZ:AV0Z10300504
Keywords : nonsmooth analysis * pattern classification * multivariate systems * 1-NN classifier * weighted distances
Subject RIV: BB - Applied Statistics, Operational Research

We propose an unconventional updating algorithm for weighting features for a classification task with abrupt changes of the error function. Using a so-called nonsmooth analysis we prove the quadratic convergence of such a learning process. For the dynamic optimization of the step size we use approach similar to Runge’s method of a half step. Finally we demonstrate its classification abilities on artificial as well as on real-life classification tasks.
Permanent Link: http://hdl.handle.net/11104/0193878