Multi-layer Neural Networks based on spintronic RF devices for the classification of RF signals

Arnaud DE RIZ

Radio-Frequency (RF) signals classification tasks such as identification of aerial vehicle using RF fingerprints, cancer identification and gesture sensing are ubiquitous. These tasks can be performed successfully by using Artificial Neural Networks. Today however, these networks are implemented together with a digitization of the input signals, using classical architecture hardware. Such systems use a great amount of energy to perform classification tasks, hence the urge of finding ways to implement these systems into specialized hardware that would be more energy efficient. Spintronic devices are good candidate for such implementations where digitization would be avoided.
In a recent work [1], it has been shown through simulation that a single layer neural network based on an assembly of spintronic resonators can be used for classifying analogue input RF signals encoding images of digits with an accuracy close to a software neural network. This spintronic neural network is based on the fact that the resonators convert the radio-frequency input signals into direct voltages through the spin-diode effect. Hence, each input signals are multiplied by a weight which depends on the resonators’ resonance frequency, thus performing a Multiply-and-Accumulate operation characteristic of the software neural networks.
To build spintronic neural networks capable of performing more complex classification tasks such as the ones previously cited, it is necessary to consider neural networks with several layer depth. In this work, we show through simulation that it is possible to extend the depth of the previous single layer neural network by using spintronic oscillators acting as activation functions. Both resonators and oscillators physical response to RF signals are based on the ideal auto-oscillator model [2] with parameters extracted from experimental devices. We consider the task of classifying RF signal fingerprints of drones [3] and we show that the spintronic neural network recognizes the signals with an accuracy almost equivalent to a software neural network. It is a step towards deep spintronic neural networks which could be a solution for fast, low-power radio-frequency classification applications.

[1] N. Leroux et al., “Radio-Frequency Multiply-and-Accumulate Operations with Spintronic Synapses,” Phys. Rev. Appl., vol. 15, no. 3, pp. 1–11, 2021, doi: 10.1103/PhysRevApplied.15.034067
[2] A. Slavin and V. Tiberkevich, “Nonlinear auto-oscillator theory of microwave generation by spin-polarized current,” IEEE Trans. Magn., vol. 45, no. 4, pp. 1875–1918, 2009, doi: 10.1109/TMAG.2008.2009935
[3] S. Basak, S. Rajendran, S. Pollin, and B. Scheers, “Combined RF-based drone detection and classification,” IEEE Trans. Cogn. Commun. Netw., pp. 0–11, 2021, doi: 10.1109/TCCN.2021.3099114