Neuromorphic computing in nanomagnetic arrays
Artificial intelligence is increasingly ubiquitous across tech and broader society. While incredibly powerful, the energy demands of operating deep-learning networks on traditional von Neumann computers are spiralling unsustainably - limiting scalability and presenting a barrier to zero-carbon futures[1]. A huge reason for this is that existing computing architectures look nothing like the brain, and as a result struggle to efficiently run ‘neural network’ style computing.
Directly implementing machine-learning in complex physical systems is emerging as an attractive low-energy solution to this issue[2]. So-called ‘Neuromorphic Computing’[3] takes inspiration from the brain & migrates computing back to the complex physical systems which initially inspired AI[4]. Nanomagnetic arrays are ideal candidates for neuromorphic hardware. They passively store information, providing memory, and perform complex nonlinear processing via magnonics[5], their collective GHz dynamics. Remarkably, the maths powering modern software neural networks originate from theoretical frameworks developed by physicists in the 1970’s to describe strongly-interacting magnetic networks[6], with great synergy between the nanomagnets and neural network architectures. The early machine learning community adopted these frameworks (originally termed Hopfield networks[7]) and adapted & refined them into the AI of today. My team at Imperial College London (especially Dr. Kilian Stenning & Dr. Will Branford) recently engineered the world-first example of a functioning neuromorphic computer built from a specific nanomagnetic network[8] termed ‘Artificial Spin Ice’. In this talk I’ll tell you about this system, our recent progress[9] and new developments.
[1] David Patterson,et al. arXiv:2104.10350 (2022).
[2] Wright, L. G. et al. Nature 601, 549-+ (2022).
[3] Markovic, D. et al. Nat. Rev. Phys. 2, 499-510 (2020).
[4] Sherrington, D. et al. Phys. Rev. Lett. 35, 179
[5] Gartside, Jack C., et al. Nature Communications 12.1 (2021): 2488.
[6] Sherrington, David, and Scott Kirkpatrick. Physical review letters 35.26 (1975).
[7] Hopfield, John J. Proc. NAS 79.8 (1982): 2554-2558.
[8] Gartside, Jack C., et al. " Nature Nanotechnology 17.5 (2022): 460-469.
[9] Stenning, Kilian D., Gartside, Jack C., et al. arXiv:2211.06373 (2022).
Date and Time
Location
Hosts
Registration
- Date: 03 Mar 2023
- Time: 11:00 AM to 12:00 PM
- All times are (UTC-07:00) Mountain Time (US & Canada)
-
Add Event to Calendar
- 1420 Austin Bluffs Pkwy
- Colorado Springs, Colorado
- United States 80918
- Building: Osborne Center for Science and Engineering
- Room Number: A204
- Contact Event Host
- Co-sponsored by UCCS
Speakers
Jack Gartside of Imperial College London
Neuromorphic computing in nanomagnetic arrays
Artificial intelligence is increasingly ubiquitous across tech and broader society. While incredibly powerful, the energy demands of operating deep-learning networks on traditional von Neumann computers are spiralling unsustainably - limiting scalability and presenting a barrier to zero-carbon futures[1]. A huge reason for this is that existing computing architectures look nothing like the brain, and as a result struggle to efficiently run ‘neural network’ style computing.
Directly implementing machine-learning in complex physical systems is emerging as an attractive low-energy solution to this issue[2]. So-called ‘Neuromorphic Computing’[3] takes inspiration from the brain & migrates computing back to the complex physical systems which initially inspired AI[4]. Nanomagnetic arrays are ideal candidates for neuromorphic hardware. They passively store information, providing memory, and perform complex nonlinear processing via magnonics[5], their collective GHz dynamics. Remarkably, the maths powering modern software neural networks originate from theoretical frameworks developed by physicists in the 1970’s to describe strongly-interacting magnetic networks[6], with great synergy between the nanomagnets and neural network architectures. The early machine learning community adopted these frameworks (originally termed Hopfield networks[7]) and adapted & refined them into the AI of today. My team at Imperial College London (especially Dr. Kilian Stenning & Dr. Will Branford) recently engineered the world-first example of a functioning neuromorphic computer built from a specific nanomagnetic network[8] termed ‘Artificial Spin Ice’. In this talk I’ll tell you about this system, our recent progress[9] and new developments.
[1] David Patterson,et al. arXiv:2104.10350 (2022).
[2] Wright, L. G. et al. Nature 601, 549-+ (2022).
[3] Markovic, D. et al. Nat. Rev. Phys. 2, 499-510 (2020).
[4] Sherrington, D. et al. Phys. Rev. Lett. 35, 179
[5] Gartside, Jack C., et al. Nature Communications 12.1 (2021): 2488.
[6] Sherrington, David, and Scott Kirkpatrick. Physical review letters 35.26 (1975).
[7] Hopfield, John J. Proc. NAS 79.8 (1982): 2554-2558.
[8] Gartside, Jack C., et al. " Nature Nanotechnology 17.5 (2022): 460-469.
[9] Stenning, Kilian D., Gartside, Jack C., et al. arXiv:2211.06373 (2022).
Biography:
Jack C. Gartside is a Royal Academy of Engineering Research Fellow in Engineering Magnonic Metamaterials for Low-Energy Neuromorphic Computing. Their team is currently hiring with 2 funded Postdoctoral Researcher positions available & PhD studentships. Email j.carter-gartside13@imperial.ac.uk for info.
Email:
Address:United Kingdom