A VLSI-based Gaussian kernel mapper for real-time RBF neural networks

An analog VLSI circuit approach to a radial basis function (RBF) neural network is explored. For each of a number of reference pattern templates, the circuit calculates the Euclidean distance between that template and an unknown point, and maps each distance to a point on the Gaussian surface of that template. Then, these points may either be added in order to form the basis for an RBF approximator or laterally inhibited to form the basis for an RBF classifier. The circuitry for this network has been implemented in 2-micron CMOS technology, and will form the bases for truly parallel and simultaneous standalone neural networks that function in real time without intervention from conventional computers.