A constructive algorithm for neural networks that generalize

Abstract

APL functions were designed to describe a constructive algorithm that synthesizes a neural network while optimizing its ability to generalize. Algorithms are implemented in programs to discover networks of binary weights that assign unfamiliar, high-dimension binary patterns to their most similar classes. Constructive algorithms that create networks are important for the design of classifiers based on array-processors made from fast two-level circuits. APL is an effective tool for the exposition of a constructive algorithm that can discover a minimal neural network. APL experts can benefit from being introduced to this interesting application and demonstration of the language's potential for describing array-based software or hardware. For constructing networks, algorithms for target switching and minimizing the overlap in the separation of training patterns have been used. Typically, prototypes of constructive algorithms as commonly implemented with scalar-based procedural languages require hundreds of program statements. Array-based formulations of similar constructive algorithms with functional style programming languages like APL and J are completed with a few brief functions.