Controlling the Skyrmion Density and Size for Quantized Convolutional Neural Networks

TL;DR

This paper explores how magnetic skyrmions can be controlled in thin films to develop energy-efficient, quantized neural network hardware, demonstrating high accuracy on CIFAR10 with skyrmion-based synapses.

Contribution

It presents experimental and micromagnetic studies on skyrmion control, proposes a skyrmion-based synaptic device, and demonstrates neural network recognition performance using skyrmion devices.

Findings

Skyrmion density and size depend on magnetic field, Dzyaloshinskii Moriya interaction, and anisotropy.

The skyrmion-based device exhibits high linearity and uniformity in resistance states.

Achieved 87% recognition accuracy on CIFAR10 with skyrmion synapses.

Abstract

Skyrmion devices show energy efficient and high integration data storage and computing capabilities. Herein, we present the results of experimental and micromagnetic investigations of the creation and stability of magnetic skyrmions in the Ta/IrMn/CoFeB/MgO thin film system. We investigate the magnetic-field dependence of the skyrmion density and size using polar magneto optical Kerr effect MOKE microscopy supported by a micromagnetic study. The evolution of the topological charge with time under a magnetic field is investigated, and the transformation dynamics are explained. Furthermore, considering the voltage control of these skyrmion devices, we evaluate the dependence of the skyrmion size and density on the Dzyaloshinskii Moriya interaction and the magnetic anisotropy. We furthermore propose a skyrmion based synaptic device based on the results of the MOKE and micromagnetic investigations. We demonstrate the spin-orbit torque controlled discrete topological resistance states with high linearity and uniformity in the device. The discrete nature of the topological resistance makes it a good candidate to realize hardware implementation of weight quantization in a quantized neural network (QNN). The neural network is trained and tested on the CIFAR10 dataset, where the devices act as synapses to achieve a recognition accuracy of 87%, which is comparable to the result of ideal software-based methods.