Matthew Willetts is a Visiting Researcher at The Alan Turing Institute.
Previously he was a Research Fellow at UCL Computer Science, working with Prof Brooks Paige on empirical and theoretical aspects of deep generative models.
He completed his DPhil at the Statistics Department at the University of Oxford, supervised by Prof Chris Holmes and Prof Stephen Roberts.
Prior to that he studied Physics at Cambridge (BA and MSci), specialising in astrophysics and relativity, graduating in 2014.
His work is on combining deep learning with probabilistic modelling, with particular interest in unsupervised learning with deep generative models. These are models that combine probabilistic structures with deep learning.
His research is on a broad range of questions within the deep generative modelling framework: Can we learn highly-structured models for particular tasks? Can we make them robust, to noisy or missing data? What about deliberate attempts to fool them into acting a certain way? With colleagues his research has been foundational in the study of the robustness of Variational Autoencoders to adversarial attack, analysing the problem both empirically and theoretically.
Most recently he has been working on non-linear independent components analysis, models for learning statistically-independent latent representations of high-dimensional data (eg images), as well as applying functional and harmonic analysis to deep learning models.
Previously he has worked with the UK biobank on their activity dataset of ~100,000 participants' accelerometry data, making models that can segment and classify participants' activity (sitting, walking, etc) given only a limited number of partially-labelled time series.