This work is the first studying quantization for graph neural networks. We show that it is possible to train networks that use low-precision integer arithmetic at inference time, resulting in substantially lower memory usage, while causing little or no loss in accuracy in many cases.

In this paper we argue that activity recognition using inertial sensors, such as accelerometers, has stagnated. We make the case that imaging sensors should be the new default for activity recognition thanks to recent developments in the hardware and low-power machine learning field. Published at HotMobile ‘20.

Explored the viability of capturing internal body sounds with a wearable device for medical applications. This project involved the construction of wearable device (with specialised acoustic circuitry) which was used to collect a dataset; the experimental procedure was designed to evaluate performance during challenging user activities and acoustic conditions. Algorithms for continuous heart monitoring and asthma symptom detection are described and evaluated. It was shown that the algorithms could be run in real-time on plausible wearable hardware.

Supervised by Professor Cecilia Mascolo. Awarded “Highly Commended” dissertation prize by the University. Work from dissertation published at WellComp workshop at UbiComp 2020, winning best paper.

Conceptually similar to the lost device tracking features included by Apple in iOS 13 (2019) and used by world governments for contact tracing. The dissertation describes a novel technique for Bluetooth 5 that enables privacy-preserving human-proximity detection. Supervised by Dr Robert Harle. Some results from dissertation included in a paper at PerCom 2018.