We are happy to announce a seminar by A. Prof John Ormerod and Jackson Zhou. We hope to see you there.
Any questions, please feel free to contact: secretary.nswbranch@statsoc.org.au.
Date: Wednesday, 19th June 2024
Time:
5.30pm - 6.15pm: Jackson Zhou
6.15pm - 6.30pm: Break (dietary requirements)
6.30pm - 7.30pm: A. Prof John Ormerod
7.30pm onwards - dinner with speaker (RSVP)
Venue: F10A.00.026.Law Building Annex.Law Annex Lecture Theatre 026, University of and via Zoom.
RSVP: Register at https://www.statsoc.org.au/event-5735130
Speaker: A. Prof John Ormerod (University Of Sydney)
Title: Moment propagation
Abstract:
We introduce and develop moment propagation for approximate Bayesian inference. This method can be viewed as a variance correction for mean field variational Bayes which tends to underestimate posterior variances. We show for some simple models (with two components) that moment propagation can be applied to recover the posterior distribution exactly. We then discuss how this idea can be extended to more complicated models with more than two components and a post-hoc correction that can be applied to any Gaussian based approximate Bayesian inference method. We demonstrate these ideas on a number of models where the moment propagation approximation of the marginal posterior distributions is nearly exact.
Biography:
Associate Professor John Ormerod completed is PhD in 2008 University of New South Wales supervised by Professor Matt Wand. He then held a postdoc at Wollongong before moving to the University of Sydney in 2010. In 2013 he was awarded the prestigious DECRA fellowship. His research covers approximate Bayesian inference methods, in particular variational Bayes where he has published over 30 papers establishing some of the first asymptotic theory in the area.
Speaker: Jackson Zhou (University Of Sydney)
Title: Scalable and Tractable Expectation Propagation for Mixed-Effects Regression
Abstract:
Divide-and-conquer has become an appealing paradigm in modern statistical inference, wherein the inference problem is split into multiple sub-problems which are solved separately before combining the results. One of the current state-of-the-art divide-and-conquer approaches for general Bayesian inference is expectation propagation (EP), which naturally splits the statistical inference across groups of data points. However, there are currently outstanding issues which limit the application of EP to mixed-effects models for correlated data in particular. One issue is that when using the standard implementation, the run time scales polynomially with respect to the number of groups. Another issue is that the updates to the approximations of the random-effects posterior factors are currently intractable when the random effects are multivariate. We aim to address both these issues for single-level univariate and multivariate random-effects structures in particular, as they are the most common. To address scalability, we outline scalable update, downdate, marginal recovery, and sampling schemes which take advantage of sparsity in the parameters of the EP approximation. To address tractability, we describe an approximate approach based on a moment propagation step which still allows for tractable EP-based inference.
@StatSocAus
