Learning Time Reduction Algorithms of Multi Layered Neural Network for Pattern Classification Problems

Abstract

Back propagation learning rule (BP rule) is used as an effective learning algorithm for multi layered neural network (NN). However, BP rule has such a basic problem that the learning speed becomes very slow. This paper proposes two learning time reduction algorithms of NN for pattern classification problems. We consider the reasons why the learning speed becomes very slow and the back propagation errors for the each learning pattern and output unit don't change at the same speed. An error function is introduced to improve the unbalance of the learning speed among patterns. By using the error function, a step correction algorithm of convergence condition is proposed, in which the convergence condition is switched from the loose condition to the strict one to stop the unbalance of learning speed among the output units and the learning patterns. An efficient learning algorithm is also proposed, in which inefficient correction procedures of weights are omitted in order to save the learning time of NN. From the simulation results for pattern classification problems, it is confirmed that the proposed algorithms are superior to BP rule and other learning algorithms with respect to the reduction of learning time.