Turing visiting researcher Jeremias Knoblauch, a third year PhD student within the Oxford-Warwick Statistics programme, is the first graduate student based in the UK and one of only 21 graduate students worldwide to receive the Facebook Fellowship award. This year there were more than 900 applicants.
The award consists of a stipend worth $84,000, a yearly trip to Facebook's global headquarters in California and additionally covers two years of university tuition fees.
Jeremias is based at the University of Warwick, where he is supervised by Dr Theo Damoulas (data-centric engineering project lead). Jeremias’ research forms part of the London Air Quality project led by Theo at the data-centric engineering programme, run in partnership with the Lloyd's Register Foundation, at The Alan Turing Institute. He officially joined the programme as a visiting researcher in September 2018 but had been working intimately within the London air quality project and data-centric engineering since September 2017 at the beginning of his PhD.
This collaboration led to two publications on Bayesian on-line changepoint detection which extended the method into space-time (publication presented at ICML 2018) and made it robust to outliers and misspecification (publication presented at NeurIPS 2018).
Both papers are highly relevant for addressing the real-world complications of inference in large scale and non-stationary data streams and have immediate applications in air pollution measurements on sensor networks. For example, they can help evaluating the effects of policy changes such as congestion charges.
Beyond Jeremias’ work on time series problems, he has become increasingly interested in understanding and generalizing the information-geometric basis of Bayesian inference.
Jeremias’ interests revolve around scalable inference methods for spatio-temporal data streams that can run in real time. Inference for complex dynamical systems generating high-dimensional structured data is typically complicated by non-stationarity, changepoints, model uncertainty, misspecification and outliers. While the analysis of real-world data streams almost always needs to address these complications, tackling them jointly leads standard likelihood-based learning rules to break down.
Jeremias works on alternative learning rules derived from generalised Bayes theorems which can solve this collection of problems jointly, efficiently and effortlessly.