Improving the Generalization of Neural Networks by Changing the Structure of Artificial Neuron

Abstract

This paper introduces a change in the structure of an artificial neuron (McCulloch and Pitts), to improve the performance of the feed forward artificial neural networks like the multi-layer perceptron networks. Results on function approximation task and three pattern recognition problems show that the performance of a neural network can be improved by a simple change in its traditional structure. The first problem is about approximation of a complicated function and the other tasks are three pattern classification problems which we have considered the digit, face and 3D object recognition experiments for evaluation. The results of the experiments confirm the improvement of the generalization of the proposed method in compared to the traditional neural network structure.