Generalization of an elliptic radial basis function neural network

Abstract

Radial Basis Function Neural Network (RBF NN) is one of the most considered neural networks with various data classification and function approximation applications. Here, we concern only the data classification applications. The classification correctness of an RBF network is defined by the data and the radius of the radial basis function. To cover a class of data, the data must be covered by the same neurons as much as possible without any data from the other classes. However, the correct classification of the training and testing data may not mean that the network can achieve its generalization. To overcome this, some statistical method for estimating the density distribution of the data must be applied to correctly adjust the center and radius of the RBF neuron. Bootstrap technique is considered and applied to estimate the center and size of each RBF neuron in this thesis. The experimental results show that this technique significantly increase the generalization.