A Novel Stochastic Sensitivity Definition Of Radial Basis Function Neural Network And Its Applications

This Paper mainly concerns the stochastic sensitivity definition for radial basis function neural network and its application in the sense of function approximation. Each input is modeled by a random variable. Given that the inputs of the network are independent and normal distributed, a novel stochastic sensitivity definition for radial basis function neural network is provided, following the corresponding computation formula. This sensitivity could be served as a measure of features' importance. An order could be obtained by sorting the magnitudes of the features' sensitivity. This sensitivity definition has relationship to the distribution of the inputs; furthermore, the order of features' sensitivity will be independent of the input perturbation. There are two application of this sensitivity: one is feature selection, the other is to control the magnitude of input perturbation to keep the robust of the network. For classification problems, an initial form of the sensitivity definition is given. Computer simulations are compose of there parts. Firstly, we demonstrate applying sensitivity definition to feature selection. Secondly, the application of sensitivity definition to control the input perturbation is illustrated. Last one is to verify the sensitivity definition for pattern classification problems. Experiments results show that the sensitivity definition can effectively delete the redundant feature and give a guideline to keep the robust of the network. For classification problems, the sensitivity definition is reasonable.