VLSI design of radial functions hardware generator for neural computations

Abstract

Multidimensional radial functions (RFs) are widely used in several neural network schemes and may have interesting applications also in fuzzy logic based systems. Unfortunately their classical look-up table hardware implementation needs an external board that does not allow high speed real world applications. In this paper we introduce a new approach in which approximate RFs are completely generated in a very efficient way by fixed-point digital operators. The VLSI design of our general purpose RF hardware generator is discussed in detail, showing that it can be integrated in a wider neural system with little effort. Many RF generators (up to 100) can be integrated on a single chip giving rise to a computational system which is two orders of magnitude faster than classical look-up table implementations. The RF generator architecture has been used to implement the forward step of a radial basis function neural network in a single chip for a real world sensor application. The VLSI design of a network with 30 neurons and a cycle time of a few microseconds is investigated.