A submicron analog neural network with an adjustable-level output unit

Abstract

A submicron feedforward analog neural network is described. This network uses submicron Gilbert multipliers for its synapses and a novel circuit based on the current-comparator circuit for its neuron. The XOR problem is solved by this network to demonstrate the capability of implementing multi-layer networks. The network is designed in a 0.5 /spl mu/m technology. HSPICE simulation shows the validity of the operation of the network.