A0383
Title: Fast expectation propagation for regression models
Authors: Jackson Zhou - The University of Sydney (Australia) [presenting]
John Ormerod - The University of Sydney (Australia)
Clara Grazian - University of Sydney (Australia)
Abstract: Expectation propagation (EP) is an approximate Bayesian inference (ABI) method that has seen widespread use across machine learning and statistics owing to its accuracy, speed, and ability to be distributed. However, it is often difficult to apply EP to models with complex likelihoods, where the EP updates do not have a tractable form and need to be calculated using methods such as numerical quadrature. These methods increase run time and reduce the appeal of EP as a fast, approximate method. It is demonstrated that EP can still be made fast for important models in this category. In particular, fast EP-based inference frameworks are developed for when there are (a) nuisance parameters and (b) correlated observations in Bayesian regression models. This is achieved using (a) dimension reduction techniques and analytic integral reductions and (b) a sparse representation of the EP approximation, a modification of the divergence being minimized, and a moment propagation-based update with custom matching statistics. Numerical experiments are conducted, and EP in these settings was shown to provide a good balance between accuracy and speed compared to competing ABI methods.
