Radial Basis Function Network Pruning by Sensitivity Analysis

Daming Shi, Junbin Gao, Daniel S. Yeung, Fei Chen

Abstract: Radial basis function (RBF) neural networks have been extensively used for classification and regression due to the fact that they can provide fast linear algorithms to approximate any regular function. The most critical issue in the construction of an RBF network for a given task is to determine the total number of radial basis functions, their centers and widths. Conventional methods of training an RBF network are to specify the radial basis function centers by searching for the optimal cluster centers of the training examples. This paper proposes a novel learning algorithm for construction of radial basis function by sensitive vectors (SenV), to which the output is the most sensitive. Our experiments are conducted on four benchmark datasets, and the results show that our proposed SenV-RBF classifier outperforms conventional RBFs and achieves the same level of accuracy as support vector machine.