Parallel Cross-Validation of Artificial Neural Networks

Abstract

In this paper, a simple parallel cross-validation method for estimating the optimal number of hidden units for multi-layered feedforward neural network models trained by the generalized delta learning rule is presented. A neural network model is considered optimal when an ideal compromise between modelling accuracy (bias) and generalizing ability (variance) is found. The problem of training by means of gradient descent based methods, on data containing local minima is adressed. The cross-validation program has been parallelized to operate in a local area computer network. Development and execution of the parallel application was aided by the HYDRA