All Spin Artificial Neural Networks Based on Compound Spintronic Synapse and Neuron

Area of Science:

• * Spintronics and Neuromorphic Computing
• * Materials Science and Engineering
• * Non-volatile Memory Technologies

Background:

• * Resistive RAM (RRAM) shows promise for artificial synaptic devices but faces challenges in stability and controllability for multilevel states.
• * Spintronic devices, like magnetic tunnel junctions (MTJs), utilize electron spin for low-power neuromorphic computing, though typically binary.
• * Existing artificial synaptic devices struggle with stable multilevel resistance states crucial for complex neural network functions.

Purpose of the Study:

• * To propose and demonstrate a novel compound spintronic synapse (CSS) device capable of stable, controllable multilevel resistance states.
• * To develop a compound spintronic neuron (CSN) circuit with a multi-step transfer function.
• * To construct and evaluate an All Spin Artificial Neural Network (ASANN) for handwritten digit recognition.

Main Methods:

• * Fabrication and characterization of compound spintronic devices comprising multiple vertically stacked MTJs.
• * Interfacial and materials engineering to achieve designable and stable multiple resistance states.
• * Design and simulation of a compound spintronic neuron (CSN) circuit and an All Spin Artificial Neural Network (ASANN).
• * System-level simulations using the MNIST database for performance evaluation.

Main Results:

• * Demonstrated that the proposed compound spintronic device (CSS) achieves designable and stable multiple resistance states.
• * Successfully implemented a compound spintronic neuron (CSN) circuit exhibiting a multi-step transfer function.
• * Constructed an All Spin Artificial Neural Network (ASANN) that shows promising performance in handwritten digit recognition simulations.
• * Investigated the impact of CSS/CSN resolution and device variations on ASANN system performance.

Conclusions:

• * The compound spintronic synapse (CSS) offers a viable solution for stable multilevel resistance states, overcoming limitations of RRAM.
• * The developed compound spintronic neuron (CSN) and ASANN architecture show potential for efficient and high-performance neuromorphic computing.
• * Further research into device resolution and variation is crucial for optimizing ASANN performance in practical applications.