Back Propagation with Randomized Cost Function for Training Neural Networks

A novel method to improve both the generalization and convergence performance of the back propagation algorithm (BP) by using multiple cost functions with a randomizing scheme is proposed in this paper. Under certain conditions, the randomized technique will converge to the global minimum with probability one. Experimental results on benchmark Encoder-Decoder problems and the NC2 classification problem show that the method is effective in enhancing BP’s convergence and generalization performance.