Constructive Neural Network: A Framework

Abstract

In this paper, two techniques for construction of feedforward neural network are being reviewed: pruning neural network algorithms and constructive neural network algorithms. In pruning method, training starts with a larger than required network and subsequently delete the redundant hidden nodes and redundant weights till there is a satisfactory solution. In the constructive method, training of the network starts with minimum structure and then according to some predefined rule some more layers of neurons are added. A number of major issues are discussed that can be considered while constructing a constructive neural network i.e. how to select network architecture, network growing strategy, weight freezing, optimization technique, activation function and stoppage criteria