Strictly modular probabilistic neural networks for pattern recognition

0411265 - UTIA-B 20030252 RIV CZ eng J - Journal Article
Grim, Jiří - Just, P. - Pudil, Pavel
Strictly modular probabilistic neural networks for pattern recognition.
Neural Network World. Roč. 13, č. 6 (2003), s. 599-615. ISSN 1210-0552
R&D Projects: GA ČR GA402/01/0981
Institutional research plan: CEZ:AV0Z1075907
Keywords : neural networks * distribution mixtures * pattern recognition
Subject RIV: BB - Applied Statistics, Operational Research

Considering the statistical pattern recognition we approximate the unknown class-conditional probability distributions by multivariate Bernoulli mixtures. We show that both the parameter optimization based on EM algorithm and the resulting Bayesian decision-making can be realized by a strictly modular probabilistic neural network. The autonomous adaptation of neurons includes only the locally available information. The properties of the sequential learning procedure are illustrated by numerical examples.
Permanent Link: http://hdl.handle.net/11104/0003519