Implementation of a Binary Neural Network on a Passive Array of Magnetic Tunnel Junctions

Abstract

Magnetic tunnel junctions (MTJs) provide an attractive platform for implementing neural networks because of their simplicity, non-volatility, and scalability. However, in hardware realizations, device variations, write errors, and parasitic resistance degrade performance. To quantify such effects, we perform inference experiments on a 2-layer perceptron constructed from a 15 x 15 passive array of MTJs, examining classification accuracy and write fidelity. Despite imperfections, we achieve median accuracy of 95.3% with proper tuning of network parameters. The success of this tuning process shows that new metrics are needed to characterize and optimize networks reproduced in mixed signal hardware.