Spintronics for image recognition: performance benchmarking via ultrafast data-driven simulations

TL;DR

The paper tackles energy and scalability concerns in AI by presenting a time-multiplexed echo-state network (ESN) that uses a single vortex-based spin-torque oscillator (STVO) as the nonlinear reservoir, with STVO dynamics simulated via the data-driven Thiele equation approach (DD-TEA) to achieve ultrafast, hardware-friendly modeling. The method processes images by PCA-based dimensionality reduction, random feature mapping, and sequential STVO-driven nonlinear transformation, followed by linear readout trained with the Moore-Penrose pseudoinverse. It reports state-of-the-art reservoir-computing performance on MNIST (≈98.1% accuracy) and reasonable results on EMNIST-letters and Fashion-MNIST, with STVO nonlinearity effectively matching conventional nonlinearities like ReLU and Sigmoid when the reservoir has enough learnable parameters. The DD-TEA framework enables extensive hyperparameter sweeps and supports the design of deeper architectures for improved accuracy, potentially enabling energy-efficient neuromorphic image recognition and time-series tasks on spintronic hardware.

Abstract

We present a demonstration of image classification using an echo-state network (ESN) relying on a single simulated spintronic nanostructure known as the vortex-based spin-torque oscillator (STVO) delayed in time. We employ an ultrafast data-driven simulation framework called the data-driven Thiele equation approach (DD-TEA) to simulate the STVO dynamics. This allows us to avoid the challenges associated with repeated experimental manipulation of such a nanostructured system. We showcase the versatility of our solution by successfully applying it to solve classification challenges with the MNIST, EMNIST-letters and Fashion MNIST datasets. Through our simulations, we determine that within an ESN with numerous learnable parameters the results obtained using the STVO dynamics as an activation function are comparable to the ones obtained with other conventional nonlinear activation functions like the reLU and the sigmoid. While achieving state-of-the-art accuracy levels on the MNIST dataset, our model's performance on EMNIST-letters and Fashion MNIST is lower due to the relative simplicity of the system architecture and the increased complexity of the tasks. We expect that the DD-TEA framework will enable the exploration of deeper architectures, ultimately leading to improved classification accuracy.

Figures (4)

• Figure 1: (a) Radio-frequency oscillations of the vortex core reduced position $s(t)$ are triggered by (b) the injection of an amplitude-modulated signal $J_\text{in}(t)$ in the STVO. A bias current density $J_\text{dc}$ is added to the input signal in order to trigger oscillations of the vortex core. (c) The non-linearity of $s(t)$ can be retrieved experimentally by recording the envelope $\tilde{V}(t)$ of the voltage $V_\text{ac}(t)$ across the oscillator (best seen in color).
• Figure 2: ESN using a single STVO delayed in time for image recognition. Images $\mathbf{x}$ are first denoised using PCA components $\mathbf{C}$ (blue layer). The resulting vectors $\mathbf{x}'$ are then encoded in a higher-dimension space using a random mask $\mathbf{M}$ (orange layer). The encoded data $\mathbf{x}"$ are transformed into the input signal $\mathbf{J}$ and fed to a STVO delayed in time for nonlinear transformation. The reduced position of the vortex core $\mathbf{s}$ is simulated using the DD-TEA. The final output $\mathbf{y}$ is obtained using the weights $\mathbf{W}$ learned by linear regression during the training phase (green layer). The category $\hat{t}$ is obtained with the argmax function.
• Figure 3: Accuracy and NRMSE of our time-multiplexed ESN on the MNIST dataset for an increasing number of nodes. The nonlinear transfer function of the reservoir nodes is set with the STVO dynamics and the reLU and sigmoid functions. Linear nodes are also tested using the identity activation.
• Figure 4: Accuracy of the STVO-based ESN for the MNIST, EMNIST-letters and FMNIST datasets. The dashed lines represent the random choice accuracy levels linked to the number of categories in each dataset ($100\%/26 = 3.85\%$ and $100\%/10 = 10\%$).