Fig. 1: Nanomagnetic reservoirs.
From: Neuromorphic overparameterisation and few-shot learning in multilayer physical neural networks
a, b Square macrospin-only artificial spin ice (MS). c, d Width-modified square artificial spin-vortex ice (WM). e, f Disorderded pinwheel artificial spin-vortex ice (PW). Bars vary from fully disconnected to partially connected. a, c, e Scanning electron micrographs (SEM). All scale bars correspond to 1 μm. b, d, f Ferromagnetic resonance spectroscopy (FMR) spectral evolution from a sinusoidal field-series input. Scale bar represents amplitude of FMR signal dP/dH (arb. units). g Reservoir computing schematic. Data are applied via magnetic field loops (+Happ then −Happ), which leads to collective switching dynamics in the array. FMR output spectra measured at −Happ is used as computational output. h Memory-capacity (MC) and nonlinearity (NL) of the reservoirs. Variations in sample design produce a diverse set of metrics. i Mean-squared error (MSE) when predicting future values of the Mackey-Glass time-series. High memory-capacity and low nonlinearity (WM) gives best performance. Shading is the standard deviation of the prediction of 10 feature selection trials. j Attempted predictions of t+7 of the Mackey-Glass equation. No single reservoir performs well. k Transforming a sine-wave to cos(2t)sin(3t). PW has 31.6 × lower MSE than MS. Shading represents the residual of the prediction.
