Logic-Oriented Model Of Artificial Neural Networks

This paper addresses the problem of logical interpretation of neural nodes, which are basically computational elements. We will look at generic models of neurons viewed as computing units carrying out logic operations on input signals. The mapping of the inherent logical background of the problem at hand, given in the form of an artificial neural network (ANN), onto the logic-oriented network can be realized by choosing appropriate logic-based neurons. Introducing ''a simple type'' of logical function in the form NofM, we concluded that it is possible to model ANN behavior with these functions without ad hoc approximations. During the modeling process, it is more important to take into account the range and relations between weight factors in ANN than their absolute values. Based on this analysis, we developed an algorithm for logical interpretation of a neuron's computational model based on the generalized logical function NofM. Also, we present in this paper the methodology and some experimental results concerning logical interpretation of ANN. Experiments provide a better understanding of ANN behavior, particularly multiple-layered perceptrons (MLPs). and the meaning of the related concepts: robustness, redundancy, pruning, etc. (C) Elsevier Science Inc. 1997.