PRUNING ARTIFICIAL NEURAL NETWORK USING BACK PROPAGATION TRAINING WITH APPLICATION TO PATTERN CLASSIFICATION

Abstract

Artificial Neural Network is a machine which models the way in which human brain perform its task of learning from new environment. Neural Network has been extensively used for task of learning patterns from training samples for purpose of classification. Pattern Classification involves mapping the given set of input features to two or more classes. Formerly, completely connected neural network with Back Propagation learning was used to predict membership of data instance to particular class. But same task can be done by using the neural network of smaller size and less complexity. This work aims to propose a new paradigm to prune an artificial neural network using error back propagation learning algorithm. In this work, neural network is trained partially and redundant weights are removed. There are two issues involved in the method. First is when the network should be pruned? Second is heuristic to measure importance of weights in network. In particular performance of pruned neural network is compared with its completely connected version using four different datasets and significant increase in learning speed is observed, while maintaining similar generalization ability.