Title: Bayesian inference with R-INLA: The road ahead
Authors: Haavard Rue - KAUST (Saudi Arabia) [presenting]
Abstract: Recent methodology progress is discussed, which concerns INLA and its R-package R-INLA. First, the use of the variational form of a Bayes theorem is discussed. This result frames the variational inference scheme methodologically within approximate Bayesian inference, and allows us to do a highly accurate correction to improve the current estimates. We will show how to do a low-rank mean and variance correction within the R-INLA framework. Secondly, we will discuss our effort to improve the parallel performance for HPC, using both OpenMPand SIP. This includes new algorithms for improving numerical gradients, a parallel line-search algorithm in the BFGS optimisation, and a new dense-matrix re-implementation of INLA.