Alex of UMD/LPS

Physical reservoir computing has been recently proposed as an analog computational technique that exploits device nonlinearity and crosstalk instead of minimizing them, as in digital logic. As opposed to conventional neural networks, the reservoir in a reservoir computer is untrained and transforms low-dimensional input information into a significantly-higher-dimensional representation, ideally transforming a non-linear temporal classification problem into a linearly separable problem in the high-dimensional reservoir state space.  A trained linear output layer is therefore used to convert the reservoir state into the classification result.

We proposed and analyzed the dynamics of a reservoir of tightly packed nanomagnets, passively relaxing in the presence of their neighbors' stray magnetic fields.  Micromagnetic simulations demonstrate good expressivity on non-linear, temporal tasks, and resource estimations of the electrical MTJ readout indicate a combined area, energy, and delay decrease of 10,000,000x compared with CMOS, necessitating further study.

We anticipate testing the concept using a commercial MRAM array.  We have developed a streamlined high-speed FPGA-based interface for the MRAM, enabling full control of the MRAM signals.  However, as MRAM MTJs are designed for minimal crosstalk, it is expected that delayering and deposition techniques will also be needed to increase interactions between devices.  We hope that with such techniques, we can demonstrate both the functionality and efficacy of nanomagnet reservoir computing for next-generation neuromorphic applications.

Edwards, A.J., Bhattacharya, D., Zhou, P. et al. Passive frustrated nanomagnet reservoir computing. Commun Phys 6, 215 (2023). https://doi.org/10.1038/s42005-023-01324-8

Biography:

Alexander Edwards received his Ph.D. in Computer Engineering in 2024 from the Erik Jonsson School of Engineering and Computer Science at the University of Texas at Dallas, and his dissertation was entitled "Stochasticity and Spintronics for Bio-inspired Computing and Hardware Security".  During his Ph.D., he received several honors including a Chateaubriand research fellowship at Universite Paris-Saclay in France and a DOE Office of Science Graduate Student Research award to support work at Sandia National Laboratories.  Alex is presently a post-doc with the University of Maryland at the Laboratory for Physical Sciences, and his research interests include beyond-CMOS computing, spintronic devices, and hardware security.