Nonlinear random projection machines are efficient neural networks capable of classifying real- life data with lower computational demands compared to standard artificial neural networks. They are well-suited for hardware implementation using nonlinear devices, enabling the creation of low-power hardware neural networks. We implement such a network using vortex-based spin-torque oscillators (STVOs), magnetic tunnel junctions (MTJs) that transform input signals nonlinearly at low power. We identify three physical parameters affecting the STVO dynamics and the network’s performance during data classification. We demonstrate their impact on a simplified nonlinear separation task and optimize them using ultrafast data-driven simulations for image recognition on the MNIST dataset. This approach holds potential for further hyperparameter optimization in STVO-based hardware random projection machines, and for the efficient development of custom neural architectures tailored for neuromorphic data classification
Authors: Moureaux, Anatole ; de Wergifosse, Simon; Chopin, Chloe, Abreu Araujo, Flavio.
Also available at Zenodo: https://zenodo.org/records/14811395
https://multispinai.eu/wp-content/uploads/2025/02/Moureaux_SPIE.pdf