Gatsby Computational Neuroscience Unit

UCL Gatsby Unit

Richard E. Turner

Wednesday 9th September 2015

Time: 4.00pm

Ground Floor Seminar Room

25 Howland Street, London, W1t 4JG

Gaussian Processes with non-parametric stationary covariance functions (and a little bit about scaling Expectation Propagation)

I'm going to start the presentation by briefly describing a problem with a popular approximate inference algorithm called expectation propagation (EP) and a simple way to fix it. In a nutshell, the problem is that EP requires huge amounts of memory to handle large models and large datasets. The fix replaces EP's local distributional approximation with a global one, but still employs EP's local updates. I'll show empirical evidence indicating that the new algorithm often performs almost as well as EP as well as some theory that shows it is closely related to stochastic variational inference.

In the main part of the presentation, I will switch tack and present very recent work that develops a Gaussian Process model that has a covariance function which is itself non-parametric. The model, which is essentially a non-parametric continuous-time generalisation of a moving-average process, is well suited to modelling possibly irregularly sampled time-series with complex spectra. I'll show that recent developments in Gaussian Process variational free-energy approximations allow analytic approximate inference. I'll wrap up by showing some preliminary results on synthetic and real-world signals.

Biography:
Richard E. Turner carried out his undergraduate degree at the University of Cambridge reading Natural Sciences and specialising in Physics. He then studied for his PhD in Computational Neuroscience and Machine Learning at the Gatsby Computational Neuroscience Unit, UCL. Following his PhD, he held an EPSRC Postdoctoral research fellowship which he spent at both the University of Cambridge and the Laboratory for Computational Vision, NYU, USA. He now holds a Lectureship in the Computational and Biological Learning Lab, Department of Engineering, University of Cambridge.