A General Class of Layered, Trainable, Threshold Logic Networks for Pattern Classification

摘要

The system developed uses a three-layer network of TLE's, only one layer of which is made up of adaptive elements. Each network is modular, permitting individual primary subnet modules to be altered without affecting the partition realized by the remainder of the network. Training involves a modified error-correction rule for weight adjustment, and appropriate addition or deletion of decision elements. Heuristic procedures are developed to implement these features in the form of an adaptive program. The training process identifies the boundaries of populated regions of patternspace. Each net module defines a convex solution region described by a small set of hyperplanes. The nearest boundary classifying rule is introduced to categorize patterns lying outside populated regions. This classifying rule is shown to approach the accuracy of the nearest neighbor rule. (Author)