Weights initialization methods for MLP neural networks

Abstract

This paper describes the weights initialization methods for two-layer feedforward neural networks. The choice of method influences of a convergence and a minimal final value of cost function for a neural network training process. The neural networks application in different domain to be expected, that their users will be acquired the applicable neural network models. Hence it is important prune away all kinds of an uncertainties while the choice of neural network structure, the learning algorithm in context of other adjustable parameters as well as preparing of suitable learning and testing data set for the neural network learning.