This PhD studentship has been developed by the University of Liverpool in partnership with the Defence, Science and Technology Laboratory (DSTL) of the Ministry of Defence of the United Kingdom.
The focus of this project is on scalable machine learning of time-series models. Such machine learning necessarily involves assessing candidate models by processing data using each of many candidate models.
This assessment is challenging because optimal processing of a time-series (e.g., using a Kalman smoother in the context of linear Gaussian models) involves iterative consideration of the data in time order, making it hard to capitalise on parallel computational resources. Conversely, techniques that are readily parallelised (e.g., belief propagation) can fail to provide an accurate assessment of a model’s efficacy, particularly in contexts where the phenomenology that the model is trying to describe generates artefacts in the data that are only visible over long timescales. This project will investigate hybrid approaches that combine the ability to capitalise on parallel resources with near-optimal processing.
DSTL have a number of specific use cases (e.g., related to both cyber and physical surveillance) that will help focus the research comprising the PhD. The aim is to use the diversity of individual use cases to exemplify and demonstrate the generic utility of the research. One exemplar use case involves using historic GP and hospital admissions data to learn the parameters of a partially-observed non-linear epidemiological model for flu. Another exemplar use case involves learning the patterns-of-life associated with benign access to MoD’s intranet systems with a view to detecting anomalous activity that might be indicative of a cyber-attack. In the context of the use cases, DSTL will provide, for example, data, benchmark algorithms and metrics for comparison.