Leveraging Spintronic Nonlinear Random Projections for Handwritten Digit Recognition
(2024) Intermag 2024 — Location: Rio de Janeiro, Brazil (5.May.2024)
Files
No attached file found for this publication.
Details
- Authors
de Wergifosse, Simon 
Chopin, Chloé
- Abstract
- Nonlinear random projections are a powerful tool to efficiently classify real-life data while requiring much less computational resources than conventional artificial neural networks. We showcase the implementation of an echo-state network (ESN) based on a single spin-torque vortex oscillator (STVO) delayed in time. This network achieves accuracy of over 98% on the MNIST handwritten digit recognition task. Ultrafast data-driven simulations based on the Thiele equation approach are used to show that the performance of our STVO-based network is equivalent to that of conventional software implementations. We demonstrate how hardware neural networks based on STVOs can be studied for specific tasks through data-driven simulation, hence speeding up the development of such hardware intelligent systems.
- Affiliations
Citations
Moureaux, A., de Wergifosse, S., Chopin, C., & Abreu Araujo, F. (2024). Leveraging Spintronic Nonlinear Random Projections for Handwritten Digit Recognition. Intermag 2024, Rio de Janeiro, Brazil. https://hdl.handle.net/2078.5/232184