Unsupervised Learning & Reservoir Computing Leveraging Analog Spintronic Phenomena
Joseph S. Friedman · 2021
We have proposed three distinct spintronic neural network approaches that leverage analog spintronic phenomena: 1) Unsupervised learning systems with spin-transfer torque magnetoresistive random-access memory (STT-MRAM) in which analog behavior is produced by stochastic STT switching; 2) Unsupervised learning systems with three- and four-terminal MTJs in which analog behavior is produced by magnetic domain wall motion; and 3) Reservoir computing systems with irregular arrays of nanomagnets in which analog behavior is produced by frustrated nanomagnetism. All three spintronic neural network approaches exploit the hysteresis intrinsic to binary spintronic memory devices while providing analog behavior with significant advantages over analog memory devices.