Hybrid implementation of neural nets using switched resistor technique

Abstract

A simple hybrid implementation technique for realizing ANNs is presented. The technique is used for realizing a network of two layers in order to make a classification between two characters T and C independent of position, rotation, and scaling. The programmability of adaptive CMOS synaptic weights is achieved by employing the switched resistor (SR) technique. Due to the exponential nature of the bipolar transistors, the sigmoid function is represented by using bipolar transistors. So, the proposed neuron is fully compatible with BiCMOS technology. This implementation technique can be used for different applications.