Fast learning and efficient memory utilization with a prototype based neural classifier

摘要

A prototype based neural network classifier (NNC) is presented. Studies on prototype based classifiers show that they can rapidly learn compared to classifiers that use gradient descent methods R. P. Lippmam. IEEE Communications Magazine 27, 47–54 (1989). Studies also show that prototype based classifiers are capable of forming complex decision regions, and that they can resolve pattern classes separated by nonlinear boundaries D. L. Reilly et al., Biol. Cybernetics 45, 35–41 (1982). Nonetheless, there is no technique to help the designer set the initial firing threshold conditions of the prototype neurons, which is usually done by experimenting with different thresholds and choosing the one that yield the best results in a particular problem domain. A technique is described in this paper that allows a NNC to automatically choose the initial firing threshold conditions for its prototype neurons. The guiding principle is to try to create a prototype and set it with the maximum initial firing threshold condition such that it does not overlap prototypes of other classes and without violating any given constraints. Examples are provided to illustrate the capability of the NNC in continuous and binary pattern classification problems.