About the speaker
Tom is a Professor of Materials Physics in the School of Chemical, Materials and Biological Engineering, University of Sheffield. His research group explores the use of nano-scale magnetism and spintronics (spin-electronics), as well as other types if material, to create new, energy-efficient computing devices, and to explore fundamental physics. His research is currently focused on developing new neuromorphic (brain-like) computing devices that will allow us to use artificial intelligence at a fraction of its current energy cost.
Abstract: Go Hard(ware network) or Go Home: Networking Spintronic Metamaterials for Neuromorphic Computation
Spintronic devices offer a promising route toward neuromorphic computing due to their inherent non-volatility, rich nonlinear and transient dynamics, and compatibility with conventional CMOS technologies. Current proposals for spintronic neural hardware typically fall into one of two camps. At one extreme are physical neural networks (PNNs), which attempt to reproduce the architecture of software multilayer perceptrons (MLPs) using large arrays of nominally identical synaptic and neuronal devices. At the other are physical reservoir computing (PRC) approaches, which treat complex spintronic systems as monolithic dynamical substrates that project inputs into high-dimensional spaces where simple linear readouts perform classification or regression. However, neither approach is without limitations. PNNs require the fabrication and interconnection of huge numbers of devices, a challenge when working with novel technologies. PRC is more easily implemented but lacks the expressivity of the massively parameterized networks used for state-of-the-art applications.In this talk, we illustrate alternatives to this false dichotomy, showing how networking modest numbers of complex spintronic systems can produce powerful computational architectures. Building on our research into the computational properties of magnetic metamaterials [1,2], we show that their computational capabilities are significantly enhanced when multiple devices are connected into modestly sized networks. We first present a machine-learning approach for constructing noise-aware, differentiable transient models of these devices, and demonstrate how this enables the connectivity of dynamic device networks to be trained to tackle challenging time-domain problems such as neuroprosthetic gesture recognition [3]. We then present new theoretical and experimental results showing how the spin–orbit torque ferromagnetic resonance spectra of these materials can act as powerful basis functions for physical Kolmogorov–Arnold Networks (KANs) [4]. This recently introduced neural network paradigm replaces the simple linear synaptic weights of MLPs with trainable nonlinear functions, offering substantial advantages in parameter efficiency and interpretability. Collectively, these results illustrate how networking spintronic devices can provide a scalable and expressive route toward physical neuromorphic computing.
Environmental Statement Modern Slavery Act Accessibility Disclaimer Terms & Conditions Privacy Policy Code of Conduct About IOP
© 2021 IOP All rights reserved.
The Institute is a charity registered in England and Wales (no. 293851) and Scotland (no. SC040092)