Modification of Feed Forward Process and Activation Function in Back-Propagation

Abstract

Research on neural networks has grown significantly over the past decade, with valuable contributions made from many different academic disciplines. While there are currently many different types of neural network models, Back-propagation is the most popular neural network model. However, the input vectors in the Back-propagation neural network model usually need to be normalized and the normalization methods affect the prediction accuracy. In this study, a new method is proposed in which an additional feed-forward process was included in the Back propagation model and a sigmoid activation function was modified, in order to overcome the input vector normalization problem. The experimental results showed that the proposed approach might produce a better training and prediction accuracy than the most current common approach using input vector normalization and that it has the potential to improve performance in machine vision applications.