Topic: Title: Some new approaches for performing efficient and reliable Bayesian inference on large datasets Authors:  Clare McGrory - University of Queensland (Australia) [presenting]
Abstract: Some new computational approaches for performing time efficient, practical and reliable Bayesian inference on large datasets will be presented. Massive datasets are becoming commonplace in modern statistical applications. While the ability to capture more information than we could before is exciting and enhances the potential to unlock even more interesting features of the data, the downside is that data storage as well as analysis can be challenging. We consider some important applications where these difficulties arise. A time-efficient approach for satellite image analysis will be described. The approach centres on the idea of carefully choosing a representative weighted subsample of the complete dataset in order to model the images in a fraction of the time. Hybrid algorithms combining more than one Bayesian inferential technique to create an approach which targets the posterior distribution more efficiently thereby saving on time while aiming to improve accuracy will also be explored. We will outline hybrid schemes for mixture model estimation and hidden Markov modelling, considering applications to genetics problems and time series data arising from ocean regime shift modelling.