Friday 16 Oct 2020: Spintronic neural networks with radio-frequency connections
Julie Grollier - Unité Mixte de Physique CNRS/Thales, Université Paris-Saclay, 91767 Palaiseau, France
Spintronic oscillators are nanoscale devices realized with magnetic tunnel junctions which have the potential to be integrated by hundreds of millions in electronic chips. Their non-linear dynamical properties are rich and tunable, and can be leveraged to imitate different features of biological neurons. High performance pattern recognition was achieved through the coupled dynamics of the oscillators in small circuits. The transient dynamics of a single spintronic nano-oscillator has been used to implement reservoir computing, achieving state-of-the-art results on a simple spoken digit recognition task , . Four spintronic nano-oscillators have been trained to classify spoken vowels by phase locking their oscillations to the strong input signals produced by external microwave sources . Three spintronic nano-oscillators did bind temporal data through their mutual synchronization .
These demonstrations now need to be scaled to deep networks to establish their potential definitely. The neocortex, the seat of higher cognitive functions in the brain, has a hierarchical structure of six layers of neurons. Adopting such a layered structure in artificial neural networks was the key to their fantastic progress in the last ten years. Neuromorphic systems need to be scalable to deep networks to truly establish their promises.
A key asset of spintronic nano-oscillators towards this goal is their ability to emit radio-frequency (RF) signals. These oscillators indeed produce microwave voltages with varying amplitude and frequency in response to direct current inputs. They could therefore potentially communicate through radio-frequencies signals, allowing fully parallel operation with minimized wiring, at a speed seven orders of magnitude faster than the brain. But for this, it is necessary to devise radio-frequency synapses that can interconnect the oscillators.
In this talk, I will rapidly review recent results on neuromorphic computing with spintronic nano-oscillators. I will then describe how they can be interconnected layer-wise through RF spintronic nano-synapses, and present our recent simulation results of classification with these novel RF synapses.
 J. Torrejon et al., « Neuromorphic computing with nanoscale spintronic oscillators », Nature, vol. 547, no 7664, p. 428?431, juill. 2017, doi: 10.1038/nature23011.
 S. Tsunegi et al., « Physical reservoir computing based on spin torque oscillator with forced synchronization », Appl. Phys. Lett., vol. 114, no 16, p. 164101, avr. 2019, doi: 10.1063/1.5081797.
 M. Romera et al., « Vowel recognition with four coupled spin-torque nano-oscillators », Nature, vol. 563, no 7730, p. 230, nov. 2018, doi: 10.1038/s41586-018-0632-y.
 M. Romera et al., « Binding events through the mutual synchronization of spintronic nano-neurons », arXiv:2001.08044 [physics], janv. 2020