USI - Email
Score matching with conditional exponential families for likelihood-free inference
Host: Prof. Ernst-Jan Camiel Wit
USI, Lugano Campus, room SI-008, Informatics building
University of Oxford, UK
Likelihood-Free Inference (LFI) is concerned with performing Bayesian inference for stochastic simulator models for which the likelihood is not accessible. Standard methods (e.g. Approximate Bayesian Computation, ABC) generate simulations in order to make up for the missing likelihood;usually, simulations are adaptively tailored to the value of the observation. We consider instead a framework in which a large set of simulations is generated independently on the observations; then, an approximate likelihood is fit to the data. Specifically, we use a conditional exponential family parametrized by two neural networks as approximate likelihood, and we fit that using Score Matching. The obtained likelihood can be used for performing inference on multiple observations at a minor cost with respect to standard LFI methods. Further, our method also provides summary statistics that can be used in ABC.
Lorenzo Pacchiardi (University of Oxford) grew up near Torino, Italy, and he obtained a bachelor's degree in physical engineering from the Politecnico di Torino. Afterwards, he moved to an MSc in Physics of Complex Systems awarded by Politecnico di Torino and Université Paris-Sud, France. He is now a PhD student in Statistics at the University of Oxford, working under the supervision of prof. Geoff Nicholls (Oxford) and prof. Ritabrata Dutta (Uni. Warwick). His research is focused on likelihood-free inference methods and the use of machine learning tools therein. He is also interested in applying likelihood-free inference methods to the setting of Numerical Weather Prediction.
Click here for the MS Teams link.
INF Newsletter Archive