Spintronic Neuromorphic Hardware Using Domain Wall Based Neurons and Quantized Synapses

TL;DR

This paper demonstrates a spintronic neuromorphic hardware model using domain wall dynamics to emulate neurons and synapses, achieving high accuracy on MNIST datasets with quantized weights.

Contribution

It introduces a novel spintronic device-based approach to emulate neural and synaptic functions, including quantized synapses for low-memory neural networks.

Findings

Achieved ~97% accuracy on MNIST with full-precision weights.

Achieved ~95% accuracy on MNIST with quantized weights.

Demonstrated controlled domain wall pinning as stable synaptic weights.

Abstract

In this work, we simulate the functionality of artificial neuron and synapse using spin-orbit torque-based spintronic devices and implemented a fully connected artificial neural netwrok (ANN). These neuro-synaptic devices are emulated using transverse domain wall dynamics in a rectangular magnetic nanotrack comprised of heavy metal/ferromagnet (HM/FM) heterostructures. The ReLU activation function of the neuron has been mimicked using the domain wall motion induced by a 3 ns current pulse. The synapse has been modelled using current-induced domain wall (DW) dynamics through a corrugated HM/FM nanotrack under the influence of a 10 ns current pulse with varying current density. The semicircular corrugations are in the form of notches, which are symmetrically located on both sides of the nanotrack. By applying 10 ns current pulses of varying densities, we achieve controlled DW pinning, revealing a step-like motion caused by temporary pauses at each pinning center. The electrical conductance of the pinned DW across various pinning points, act as stable synaptic weights for our ANN. Furthermore, we observe a threshold-dependent delay effect where each depinning event is influenced by previous ones, successfully mimicking synaptic memory and adaptability in neuromorphic systems. The fully connected ANN has been modeled using the conventional float32 synaptic weights for the MNIST and Fashion MNIST datasets with an accuracy of ~97 % and ~86 %, respectively, which serves as a test bed of our neuromorphic simulations. With the aim of implementing a sparse and low memory footprint ANN, we quantize the trained synaptic weights into discrete quantized level and tested the network, which demonstrate an accuracy of ~95% and ~62%, for the MNIST and Fashion-MNIST dataset, respectively.