CMStatistics 2017: Start Registration
View Submission - CMStatistics
B0937
Title: Fast and scalable Bayesian spatial 3D priors for brain imaging Authors:  Per Siden - Linköping University (Sweden) [presenting]
Mattias Villani - Stockholm University (Sweden)
Abstract: Gaussian Markov random field (GMRF) priors will be considered, which have been successfully used in many large scale spatial problems, thanks to the sparsity induced in the precision matrices. However, most commonly used inference methods rely on the sparse Cholesky factorization, which is not feasible for problems of very large size, as those arising in spatial whole-brain modeling of task-related fMRI data, normally containing hundreds of thousands of data points and parameters. We instead develop fast and scalable inference algorithms utilizing among other preconditioned conjugate gradient methods, which are used both for sampling (MCMC) and stochastic optimization (VB and MAP). The methods are applied to fMRI data using a model which also has a non-trivial temporal component, and show to be both faster and more accurate than previous methods.