Abstract
dc:description.abstractAs the exponential scaling of computing systems predicted by Moore’s law begins to slow, in part due to reaching fundamental physical limitations in the continued miniaturisation of transistors, the development of novel unconventional computing technologies with improved scaling laws is becoming increasingly important. Unconventional computing systems use substrates other than silicon transistors to encode and process information. As an umbrella term, it encompasses fields such as analogue, physical and neuromorphic computing, which focus on computation through continuous variables, complex physical processes, and artificial neurons respectively. While there exist a wide range of physical processes and dynamics which one could imagine exploiting to process information, the key challenge is in designing scalable systems which can be easily programmed to solve specific tasks. The success of existing computing technologies relies on the strong predictions which can be made about their behaviour. When using alternative physical processes for computing, one must work against noise, instabilities, and unmeasurable internal dynamics, which can limit the determinism and predictability of a system. In the following work, we present a series of results on optimising physical computing systems with these factors in mind, where we focus primarily on optical substrates due to their natural potential for energy efficiency, speed and parallelism. We approach this in two ways, first considering the design of high-level system architectures suitable for computing, and secondly considering ways of programming and optimising these systems to solve specific tasks. In the case of the former, we introduce a new physical computing architecture which uses quantum resources to improve scaling over an equivalent classical counterpart, and which is realisable with currently available technologies. In the latter, we apply meta-learning and reinforcement learning techniques to develop new optimisation strategies for training physical neural networks in situ in a scalable way. Throughout, we analyse the criteria necessary for efficiency and scalability, while also considering ways in which we can ensure the resulting systems are accessible and sustainable. Beyond these examples, we discuss the broader motivations and requirements for the practical adoption of unconventional and neuromorphic computing systems, and the potential impact they could have on the scaling of future computing technologies, including the efforts towards realising artificial general intelligence and artificial consciousness.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Neill, Oliver D.
Subjects
dc:subject × 1Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.84943