Multi-state MRAM cells for hardware neuromorphic computing

TL;DR

This paper demonstrates that serially connected magnetic tunnel junctions (MTJs) can be used to create multi-state memory cells for hardware neuromorphic computing, enabling quantized weights and neural network implementation with comparable accuracy to software methods.

Contribution

It introduces a novel hardware neural network architecture using multi-state MTJ cells, with a behavioral model and a circuit for programming quantized weights.

Findings

Multi-state MTJ cells can store quantized neural network weights.

The hardware neural network achieves recognition accuracy comparable to software.

The proposed design integrates MTJ cells with CMOS components for neuromorphic computing.

Abstract

Magnetic tunnel junctions (MTJ) have been successfully applied in various sensing application and digital information storage technologies. Currently, a number of new potential applications of MTJs are being actively studied, including high-frequency electronics, energy harvesting or random number generators. Recently, MTJs have been also proposed in designs of a new platforms for unconventional or bio-inspired computing. In the present work, it is shown that serially connected MTJs forming a multi-state memory cell can be used in a hardware implementation of a neural computing device. A behavioral model of the multi-cell is proposed based on the experimentally determined MTJ parameters. The main purpose of the mutli-cell is the formation of the quantized weights of the network, which can be programmed using the proposed electronic circuit. Mutli-cells are connected to CMOS-based summing amplifier and sigmoid function generator, forming an artificial neuron. The operation of the designed network is tested using a recognition of the hand-written digits in 20x20 pixel matrix and shows detection ratio comparable to the software algorithm, using the weight stored in a multi-cell consisting of four MTJs or more.