07-25, 16:15–16:45 (US/Eastern), Room 349
Stheno.jl is a probabilistic programming framework specifically designed for constructing probabilistic models based around Gaussian processes. Come to this talk to find out what that means, why you should care, and how you can use it with Flux.jl and Turing.jl to do cool things.
Gaussian processes (GPs) are probabilistic models for nonlinear functions that are flexible, easy to interpret, and enable the modeller to straightforwardly encode high-level assumptions about the properties of the function in question. In short, they're a really useful component of the probabilistic modelling tool box.
Implementations to date have not made possible to fully exploit the interpretability of GPs, making it harder than necessary encode prior knowledge and interpret results. Based on the ideas in our recently proposed GP probabilistic programming framework, we have developed Stheno.jl to provide an implementation that is straightforward for the user interested in applying GPs to their problem to use, while remaining hackable for experts and researchers.
This talk will provide an intuitive introduction GPs using Stheno.jl. We'll then show how Stheno.jl can be used solve extensions of a classical non-linear regression problem and explore structure in the solution, how it can be used in conjunction with Turing.jl to embed GPs as a component in a larger non-Gaussian probabilistic programme, and to explore how it can be combined with Flux.jl to hardness the complementary strengths of deep learning and probabilistic modelling.
Will is a PhD student in the Machine Learning Group at the University of Cambridge, supervised by Rich Turner. He's generally interested in probabilistic modelling and (approximate) inference, and is particularly fond of Gaussian processes (GPs). His work on GPs includes approximate inference for scaling to large problems, their use in both multi-output regression and the ensembling of climate models, and most recently on how best to exploit their unique properties in a probabilistic programming framework.