An improved neural processing element using pulse coded weights

Abstract

A technique is presented to allow both inhibition and excitation weights to take place on the same input lines of pulse coded neural processing element (NPE). This is achieved by using different ranges of pulse duty cycles for excitation and inhibition. The pulse coded NPE contains blocks of neural type cells (NTC), a summation, a nonlinear threshold logic, and a learning block for adaptive weight feedback. Modifications on the neural-type cell are presented to achieve this as well as wider oscillation ranges. Simulation results verify the functions for each block. The results allow for adapting weights through sign changes leading to more flexibility in design and less power consumption. Along with its functional block-based structure, the NPE, as an artificial neuron, can be applied in many different configurations of neural networks. A layout for a configuration for a winner-take-all network was carried out and sent to MOSIS for IC fabrication. >