Genetic selection and neural modeling for designing pattern classifiers

This work shows how genetic algorithms and neural modeling provide powerful new tools for the design of trainable pattern classifiers--including neural classifier--based on statistical decision theory and cluster analysis and describes applications of these techniques to an adaptive detector of abnormal tissue in mammograms and a detector of straight lines and edges in noisy aerial images.

Abstract

We show how genetic algorithms and neural modeling provide powerful new tools for the design of trainable pattern classifiers--including neural classifiers--based on statistical decision theory and cluster analysis. We describe two major uses of genetic algorithms for designing linear and piecewise linear classifiers: (a) the selection of features from an initially large set, and (b) the global optimization of these classifiers, thereby avoiding local performance maxima. Neural modeling involves the replacement of step decision functions by differentiable decision functions. By applying neural modeling to genetically designed piecewise linear classifiers, we obtain globally optimized neural classifiers in which the number of neurons strike a good balance between classification accuracy and generalizing ability. We describe applications of these techniques to an adaptive detector of abnormal tissue in mammograms and a detector of straight lines and edges in noisy aerial images.