Institut des
NanoSciences de Paris
insp
Accueil > Evénements > Séminaires > Neuromorphic computing
insp
1.jpg

Séminaire « Magnétisme et physique du spin » de l’INSP

Neuromorphic computing with spin-torque nano-oscillators - Philippe Talatchian - Mardi 18 septembre 2018 à 11 h

Philippe Talatchian - Unité Mixte de Physique CNRS-Thalès

INSP - Sorbonne Université - 4 place Jussieu - 75005 Paris - Barre 22-32, 2e étage, salle 201

Abstract

Spin-torque nano-oscillators are non-linear, nano-scale, low power consumption, tunable magnetic microwave oscillators which are promising candidates for building large networks of coupled oscillators. Those can be used as building blocks for neuromorphic hardware which requires high density networks of neurons-like complex processing units coupled by tunable connections. Recently a first demonstration of neuromorphic computing with a single spin- torque nano-oscillator was established allowing spoken digit recognition with state of the art performances [1]. However, to realize more complex cognitive tasks, it is still necessary to demonstrate a very important property of a neural network : learning  an iterative process through which a neural network can be trained using an initial fraction of the inputs and then adjusting internal parameters to improve its recognition or classification performance. One difficulty is that training networks of coupled nano-oscillators requires tuning the coupling between them. Here, through the high frequency tunability of spin-torque nano-oscillators, we demonstrate experimentally the learning ability of coupled nano-oscillators to classify spoken vowels with a recognition rate of 88% [2]. To realize this classification task, we took inspiration from the synchronization of rhythmic activity of biological neurons and exploit the synchronization of spin-torque nano-oscillators to external microwave stimuli. According to our simulations, the high experimental recognition rates stem from the weak-coupling regime and the high tunability of spin-torque nano-oscillators. Finally, in order to realize more difficult cognitive tasks requiring large neural networks, we show numerically that arrays of hundreds of spin-torque nano-oscillators can be designed with the constraints of standard nano- fabrication techniques. These results open new paths towards highly energy efficient bio- inspired computing on-chip based on non-linear nano-devices that can adapt and learn.

Acknowledgements

This work was supported by the ERC grant bioSPINspired n°682955.

References

[1] J. Torrejon et al, Nature. Vol 547, page 428-431 (2017). [2] M. Romera, P. Talatchian et al, arXiv : 1711.02704 (2017).