An experimental study: on reducing RBF input dimension by ICA and PCA

Experimentally investigates using independent component analysis (ICA) and principle component analysis (PCA) in the reduction of the input dimension of a radial basis function (RBF) network such that the net's complexity is reduced. The results have shown that a RBF network with ICA as an input pre-process has similar generalization ability to the one without pre-processing, but the former's performance converges much faster. In contrast, a PCA based RBF leads to a deteriorated result in both convergent speed and generalization ability.

1 Introduction

The Radial Basis Function (RBF) network has been extensively applied to a lot of applications such as pattern recognition, image processing, and time series forecasting because it allows the independent tuning of RBF network parameters and sufficiency of using one layer of neural network to establish input-output mapping^[12]. In general, the complexity of the RBF network increases when the net's input size increases. Moreover, the noise and the irrelevant components itn the inputs will degrade the generation performance of RBF networks as well.