Approximate radial basis function neural networks (RBFNN) to learn smooth relations from noisy data

In this paper a novel RBFNN scheme is presented introducing the idea of strongly delocalized neural receptive fields. Based on delocalization, a robust deterministic annealing procedure is proposed for determining the RBF centers. It is shown that highly overlapped receptive fields exhibit good noise rejection capability. A real world sensor application of this RBFNN is described. By approximating the exact gaussian fields with a suitable radial function (RF), the forward step of the RBFNN can be efficiently implemented in a single digital chip for real time sensor applications.