Physical reservoir computing with dynamical electronics

Liang, Xiangpeng (2022) Physical reservoir computing with dynamical electronics. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2022liangphd.pdf] PDF
Download (96MB)

Abstract

Since the advent of data-driven society, mass information generated from human activity and the natural environment has been collected, stored, processed, and then dispersed under conventional von Neumann architecture. However, further scaling the computing capability in terms of speed and power efficiency has been significantly slowed down in recent years due to the fundamental limits of transistors. To meet the increasingly demanding requirement for data-intensive computation, neuromorphic computing is a promising field taking the inspiration from the human brain, an extremely efficient biological computer, to develop unconventional computing paradigms for artificial intelligence.

Reservoir computing, a recurrent neural network algorithm invented two decades ago, has received wide attention in the field of neuromorphic computing because of its unique recurrent dynamics and hardware-friendly implementation schemes. Under the concept of reservoir computing, hardware’s intrinsic physical behaviours can be explored as computing resources to keep the machine learning within the physical domain to improve processing efficiency, which is also known as physical reservoir computing.

This thesis focuses on modelling and implementing physical reservoir computing based on dynamical electronics, along with its applications with sensory signals. First, the fundamental of the reservoir computing algorithm is introduced. Second, based on the reservoir algorithm and its functionalities, two different architectures for physically implementing reservoir computing, delay-based reservoir and parallel devices, are investigated to perform temporal signal processing. Thirdly, an efficient implementation architecture, namely rotating neurons reservoir, is developed. This novel architecture is evaluated in both theoretical analysis and experiments. An electrical prototype of the rotating neurons reservoir exhibits unique advantages such as resource-efficient implementation and low power consumption. More importantly, the theory of rotating neurons reservoir is highly universal, indicating that a rotational object embedded with dynamical elements can act as a reservoir computer.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
Colleges/Schools: College of Science and Engineering > School of Engineering > James Watt Nanofabrication Centre
Supervisor's Name: Heidari, Professor Hadi
Date of Award: 2022
Depositing User: Theses Team
Unique ID: glathesis:2022-83134
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 22 Sep 2022 13:09
Last Modified: 23 Sep 2022 12:57
Thesis DOI: 10.5525/gla.thesis.83134
URI: https://theses.gla.ac.uk/id/eprint/83134

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year