Dynamic search trajectory methods for neural network training

Abstract

Training multilayer feedforward neural networks corresponds to the global minimization of the network error function. To address this problem we utilize the Snyman and Fatti approach by considering a system of second order differential equations of the form, ẍ = -∇E(x), where x is the vector of network weights and ∇E is the gradient of the network error function E. Equilibrium points of the above system of differential equations correspond to optimizers of the network error function. The proposed approach is described and experimental results are discussed.