Current Mode Analog Kohonen Neural Network

Abstract

In this paper we present Matlab analysis as well as CMOS implementation of an analog current mode Kohonen neural network (KNN). The presented KNN has been realized using several building blocks proposed earlier by the authors, such as: binary tree winner take all circuit, Euclidean distance calculation circuit, adaptive weights change mechanism. The example network contains four neurons, each of them having three weights. There are three input signals applied, which can be currents or voltages converted just at the beginning into currents. The network operates with the clock frequency of 20 MHz, dissipating 1.5 mW of power from 1.5 V supply voltage. For lower operation frequencies power dissipation can be reduced.