Adaptive neural networks object-oriented approach

Abstract

The paper presents a method how to configure neural network with an optimum number of neurones in the hidden layer(s). Based upon the fact that the configuration varies with the nature of objects to be recognised and can hardly be set optimally in advance which often results in over-estimation, the network is configured dynamically during self supervised learning process. Concept of design of the whole system is presented using an object oriented approach which enables easy construction of neurones and their link to existing configuration automatically during the learning. Type of the neural neural network considered consists of a single hidden layer and uses error back propagation method of learning.