CONVERGENCE PROPERTIES OF A CLASS OF LEARNING VECTOR QUANTIZATION ALGORITHMS

Authors
KOSMATOPOULOS EB CHRISTODOULOU MA
Citation
Eb. Kosmatopoulos et Ma. Christodoulou, CONVERGENCE PROPERTIES OF A CLASS OF LEARNING VECTOR QUANTIZATION ALGORITHMS, IEEE transactions on image processing, 5(2), 1996, pp. 361-368
Citations number
24
Categorie Soggetti
Engineering, Eletrical & Electronic
Journal title
IEEE transactions on image processing → ACNP
ISSN journal
10577149
Volume
5
Issue
2
Year of publication
1996
Pages
361 - 368
Database
ISI
SICI code
1057-7149(1996)5:2<361:CPOACO>2.0.ZU;2-Z
Abstract
In this paper, a mathematical analysis of a class of learning vector q uantization (LVQ) algorithms is presented, Using an appropriate time-c oordinate transformation, we show that the LVQ algorithms under consid eration can be transformed into linear time-varying stochastic differe nce equations. Using this fact, we apply stochastic Lyapunov stability arguments, and we prove that the LVQ algorithms under consideration d o indeed converge, provided that some appropriate conditions hold.