Comparison of Two Memristor Based Neural Network Learning Schemes for Crossbar Architecture

Abstract

This paper compares two neural network learning schemes in crossbar architecture, using memristive elements. Novel memristive crossbar architecture with dense synaptic connections suitable for online training was developed. Training algorithms and simulations of the two proposed learning schemes, winner adjustment training (WAT) and multiple adjustments training (MAT) are presented. Tests performed using MNIST handwritten character recognition benchmark dataset confirmed the functionality of proposed learning schemes. Proposed learning schemes were compared accounting for noise, device variations and multipath effects. The proposed learning schemes improve available neural network learning schemes.