Structure analysis of neural networks

Abstract

A geometrical interpretation of the multilayer perceptron (MLP) is suggested in this thesis. Some general guidelines for selecting the architecture of the MLP, i.e., the number of the hidden neurons and the hidden layers, are proposed based upon this interpretation and the controversial issue of whether four-layered MLP is superior to the three-layered MLP is also carefully examined. The guideline for architecture selection of MLP is then used to deal with the over-fitting problem. Various approaches of dealing with the over-fitting problem are also reviewed from this new geometrical interpretation. Finally, the attention is shifted to the radial basis function network (RBFN), which actually can be approximated or even represented by certain classes of MLPs with specified additional input. Various methods of determining the centers and other parameters of RBFN are compared with those implementations through MLPs.