In this project, I trained a Reservoir Computing model for speech recognition on the Spoken MNIST Dataset. The model consists of feature extraction filters, followed by a masking procedure. Then, the resulting signal is applied to a Simulated Reservoir based on the dynamics of a novel nanomagnetic device, the Spin Torque Nano Oscillator, which is an emerging candidate for state-of-the-art neuromorphic computing. The signal response of the magnetic reservoir is then sampled and measured to obtain the responses of virtual neurons, a process called time multiplexing. With these neurons' response, a linear readout is performed, achieving an accuracy of 88.6 percent, similar to those of more computationally and energetically complex models such as an MLP with approximately 300,000,000 parameters (88% accuracy), as opposed to the reservoir network with 30,000. This showcases the potential for edge AI of the Spin Torque Nano Oscillator as novel neuromorphic AI hardware.
This project is part of my research thesis for my BSc in Physics Degree.
Copyright © [2024] [David Rojas, Universidad Catolica del Norte]. All rights reserved