Percolating atomic switch networks for neuromorphic computing (2020)
Type of ContentTheses / Dissertations
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury
AuthorsMallinson, Joshua B.show all
Conventional computer technology is ubiquitous in modern society. Initially used only to perform mathematical calculations, computers are now integrated with most daily human activity. As the demand on computers shifts towards more and more complex tasks, the limitations of conventional computing architectures are being realized. Conversely, many such tasks can easily be carried out by the human brain, which is capable of abstraction, recognition, prediction and decision making with extremely low power consumption.
The vast capabilities of the brain have motivated the development of novel neuromorphic computing architectures which can naturally facilitate brain-like computation. Several such architectures are being developed internationally, but few aim to mimic the complex connectivity and critical dynamics of the cortex; features which are central to the brain’s computational power.
This thesis investigates percolating atomic switch networks (PASNs) as candidates for novel neuromorphic architectures. It is demonstrated that PASNs exhibit many of the same emergent phenomena as the brain, including scale-free temporal dynamics and long-range temporal correlations. It is further shown that the critical avalanche dynamics observed in neuronal firing patterns is also present in the sequences of switching activity exhibited by PASNs. Scaling exponent relations are used to definitively demonstrate that, like the brain, PASNs are critical systems.
Reservoir computing (RC) is a novel brain-inspired computing paradigm that utilizes temporal correlations to perform classification and prediction tasks. The nonlinear dynamical response of PASNs to electrical stimulation makes them particularly suitable candidates for physical reservoirs in neuromorphic RC systems. PASNs are thus utilized as the dynamical nodes in a time-delay RC framework. Through numerical simulations and experiments, it is demonstrated that PASNs can perform waveform classification and time-series prediction tasks, marking the first ever implementation of a PASN as a neuromorphic device.