BEGIN:VCALENDAR
VERSION:2.0
METHOD:PUBLISH
CALSCALE:GREGORIAN
PRODID:-//WordPress - MECv6.4.2//EN
X-ORIGINAL-URL:https://www.lps.u-psud.fr/
X-WR-CALNAME:Laboratoire de physique des Solides
X-WR-CALDESC:
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-PUBLISHED-TTL:PT1H
X-MS-OLK-FORCEINSPECTOROPEN:TRUE
BEGIN:VEVENT
CLASS:PUBLIC
DTSTART;TZID=Europe/Paris:20251202T133000
DTEND;TZID=Europe/Paris:20251202T180000
DTSTAMP:20251118T094200
UID:MEC-0a7d83f084ec258aefd128569dda03d7@lps.u-psud.fr
CREATED:20251118
LAST-MODIFIED:20251118
PRIORITY:5
TRANSP:OPAQUE
SUMMARY:Julie Grollier : Spintronic neural networks
DESCRIPTION:\nSpintronics introduces intrinsic memory, wave-based dynamics, and nonlinearity into neuromorphic computing, offering a promising route toward energy-efficient, brain-inspired hardware. In recent years, it has been demonstrated that spintronic devices such as magnetic tunnel junctions can implement both synapses and neurons at the nanoscale with high accuracy and reliable behavior [1]. However, a major challenge remains: interconnecting these nano-objects to build scalable neural networks. Neural architectures require not only a massive number of synapses, from millions to billions depending on the task, but also very high connectivity that often exceeds ten thousand synapses per neuron.\n\n\n\nAfter introducing the current state of the art in neuromorphic computing, I will discuss spintronic implementations and their key challenges. I will then show that operating in the frequency domain to establish links between synapses and neurons, instead of relying solely on spatial wiring, offers a promising strategy for creating large-scale and densely connected networks.\n\n\n\nI will present the concept of frequency synapses and frequency-domain neural networks, followed by our first experimental demonstrations using networks of magnetic tunnel junctions [2] and spin diodes [3]. I will show that these networks can already perform tasks such as drone, digit, or clothing recognition when pre-trained, and that they can be reconfigured in a single step through their unique frequency selectivity. Finally, I will discuss our initial results on training these networks directly on chip, which opens the way toward edge-learning spintronic systems.\n\n\n\n[1]       J. Grollier, D. Querlioz, K. Y. Camsari, K. Everschor-Sitte, S. Fukami, et M. D. Stiles, « Neuromorphic spintronics », Nat. Electron., p. 1‑11, mars 2020\n\n\n\n[2]       A. Ross et al., « Multilayer spintronic neural networks with radiofrequency connections », Nat. Nanotechnol., vol. 18, no 11, p. 1273‑1280, nov. 2023\n\n\n\n[3]       E. Plouet, H. Singh, P. Sethi, F. A. Mizrahi, D. Sanz-Hernández, et J. Grollier, « Convolutions with Radio-Frequency Spin-Diodes », in 2024 IEEE International Electron Devices Meeting (IEDM), déc. 2024\n\n\n\na, Schematic of the spintronic frequency-domain network, where MTJs act as neurons and synapses. b, Spin-diode response of a single synaptic junction. c, Synaptic multiplication, showing the dependence of output voltage on input RF power and weight. d, Response of two synaptic junctions connected in series, illustrating combined resonances. e, Neuron emission power and frequency as a function of drive current.ive current.\n\n\n\nBiography\n\n\n\nJulie Grollier is a Research Director at Laboratoire Albert Fert where she leads the Neuromorphic Physics group (https://www.neurophysics.cnrs-thales.fr/people/julie-grollier/) Her research explores how concepts from physics can inspire new kinds of computing hardware, especially brain-inspired, energy-efficient systems based on spintronic and ferroelectric devices. She is a Fellow of the APS and a recipient of the CNRS Silver Medal and the Irène Joliot-Curie Prize, among other distinctions.\n\n\n\n\n
URL:https://www.lps.u-psud.fr/en/events/julie-grollier-spintronic-neural-networks/
CATEGORIES:General Seminars,Seminars
LOCATION:amphi Blandin
ATTACH;FMTTYPE=image/jpeg:https://www.lps.u-psud.fr/wp-content/uploads/2025/11/julie.jpg
END:VEVENT
END:VCALENDAR