Utilizing Analog Circuits by Neural-Network based Multi-Layer-Perceptron

Abstract

This manuscript presents an artificial-neural-network based programmable-neuron for implementation of analog circuits with multi-layer-perceptron. The proposed programmable-neuron can estimate linear, hyperbolic, tangent and sigmoid functions which are used to activate the analog circuits. With this, a neural-network-designer can utilize maximum number of controller-bits to select an activation-function kind with no actual change. For this neuron, 0.18-µm CMOS-technology is used for simulations and demonstrates a good estimation in peak error with ideal sigmoid and hyperbolic tangent function by 7.3% and 29.34% respectively. To assess the usefulness of the neuron, a Multi-Layer-Perceptron-neural-network (MLP-NN) is used. The MLP-NN is trained to carry out XOR-logic gate for handling signals in frequency-range from 3mHz to 60MHz. The correctness of the proposed-neuron is over 99.9%. These results shows that there is a decrease of 49% in power consumption with related to previous works.