Research On Architecture Selection For Single-hidden Layer Feed-forward Neural Networks

Neural Network is a power data modeling algorithm,it has been applied to classificationproblems and regression prolems.Reasearch on architecture selection of neural networks isone of the most important problems,suitable architecture is very significant to theperformance of neural networks for specific problems.This paper proposed two methods forarchitecture selection of Single-hidden Layer Feed-forward Neural Networks.The first methoddefined the sensitivity as the significance of hidden nodes.The algorithm train a net with largenumber of hidden nodes,then order the hidden nodes by the significance and remove the leastimportant node one by one until the pre-defined stop condition.Like the first method,thesecond one measure the relevancy between hidden nodes and classification,then prue thehidden nodes with low relevancy.The two proposed algorithms work without retraining thenetwork and they can get network with compact architecture and good generalization.