All-Spin Artificial Neural Network Based on Spin–Orbit Torque-Induced Magnetization Switching
Zhen Cao, Shuai Zhang, Jincheng Hou, Wei Duan, Long You
Abstract
A reliable design of all spin artificial neural networks based on spin–orbit torque (SOT) devices has been proposed and demonstrated in W/CoFeB/MgO heterostructures. In our scheme, a single device acts as a neuron with an rectified linear unit (ReLU) activation function. Besides, the synaptic function is also realized using the devices made of the same film structure as that used in neural devices, but with different film thicknesses. Furthermore, system-level simulations are performed to classify the MNIST database by exploiting SOT neurons and SOT synapses characteristics. The high recognition rate (91.01%) confirms the feasibility of our scheme.
Topics & Concepts
MNIST databaseArtificial neural networkActivation functionTorqueSpin (aerodynamics)Computer scienceMagnetizationControl theory (sociology)HeterojunctionFunction (biology)Orbit (dynamics)Materials scienceElectronic engineeringCondensed matter physicsPhysicsBiological systemOptoelectronicsTopology (electrical circuits)Artificial intelligenceElectrical engineeringEngineeringQuantum mechanicsEvolutionary biologyMagnetic fieldThermodynamicsAerospace engineeringBiologyControl (management)Magnetic properties of thin filmsAdvanced Memory and Neural ComputingFerroelectric and Negative Capacitance Devices