a IBM Semea, Scientific and Technical Solution Center, Rome, Italy
b Department of Anatomy, UCL, London WC1E 6BT, United Kingdom
Abstract:
The usability of the constructive neural algorithms as pattern classifiers is investigated. It is pointed out that the unboundedness of the decision regions formed by most neural recognizers leads to substantial limitations of the generalization capabilities of these nets. We specify a constructive neural recognizer that forms bounded decision regions, and report the results of this algorithm on a series of benchmark problems that resemble the usual pattern recognition problems.