Performance of Feedforward Neural Network with External Influence Function for Back Propagation Learning

The human brain contains at least 100 billion neurons, each with the ability to influence many other cells. Clearly, highly sophisticated and efficient mechanisms are needed to enable communication among this astronomical number of elements. Such communication is made possible by synapses, the functional contacts between neurons. The synapses have a major part to play for learning, memory and control in the functional brain. The connection strengths of the synapses correspond to weights of neurons in artificial neural networks. Back Propagation (BP) learning [1] is one of engineering applications of artificial neural networks. The BP learning operates with a feedforward neural network which is composed of an input layer, a hidden layer and an output layer, and the effectiveness of the BP learning has been confirmed in pattern recognition, system control, signal processing, and so on [2]-[4]. It is an optimization produce based on gradient descent that adjusts weights to reduce the system error or cost function. The errors are used as inputs to feedback connections from which adjustments are made to the synaptic weights layer by layer in a backward direction. In this study, we investigate the performance of the BP learning if the action of synaptic weights changing by external influence. We introduce periodically sine wave as external influence function to adjustment of the weights. By computer simulation, the proposed network gains the good performance for learning efficiency. Furthermore, the characteristics of the neurons in hidden layer are investigated.