Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
KFM: Fachverband Kristalline Festkörper und deren Mikrostruktur
KFM 23: Skyrmions 2 (joint session MA/KFM)
KFM 23.13: Vortrag
Donnerstag, 8. September 2022, 12:30–12:45, H37
Audio Recognition with Skyrmion Mixture Reservoirs — •Robin Msiska1, Jake Love1, Jonathan Leliaert2, Jeroen Mulkers2, George Bourianoff3, and Karin Everschor-Sitte1 — 1University of Duisburg-Essen, Duisburg, Germany — 2Ghent University, Ghent, Belgium — 3Senior Principle Engineer, Intel Corp. (Retired)
Physical reservoir computing is an information processing scheme that enables energy efficient temporal pattern recognition to be performed directly in physical matter [1]. Previously, random topological magnetic textures have been shown to have the characteristics necessary for efficient reservoir computing [2] and allowed for simple pattern recognition with two input channels [3].
We propose a skyrmion mixture reservoir computing model with multi-dimensional inputs. Through micro-magnetic simulations, we show that our implementation can solve audio classification tasks at the nanosecond timescale to a high degree of accuracy. Due to the quality of the results shown and the low power properties of magnetic texture reservoirs, we argue that skyrmion magnetic textures are a competitive substrate for reservoir computing.
Funding from the Emergent AI Centre (Carl-Zeiss-Stiftung), DFG (320163632), FWO-Vlaanderen and computer resources by VSC (Flemish Supercomputer Center) are gratefully acknowledged.
[1] G. Tanaka et al., Neural Networks 115, 100 (2019). [2] D. Prychynenko et al., Physical Review Applied 9, 014034 (2018) [3] D. Pinna et al., Phys. Rev. Applied 14, 054020 (2020)