On the Storage Capabilities of Radial Basis Function Neural Networks

Pattern classification and function approximation have been found in many applications. The radial basis function network (RBFN) has shown a great promise in this sort of problems because of its faster learning capacity. Though RBFNs have storage properties similar to that ofHopfield networks, these properties have not been well explored so far. In this paper, an approach for analyzing the storage capacity of the RBFN is presented. An upper bound on cost function is found and the error over weighted input vectors is minimized by increasing the number of hidden units. The storage capacity is defined and the proposed method can be used to estimate the capacity in terms of the total probability density function by adding the partial information content associated with each class.