B0625
Title: Error analysis of random subsampling methods for Bayesian inference
Authors: Han Cheng Lie - Universitaet Potsdam (Germany) [presenting]
Abstract: In Bayesian inference, one uses a prior probability measure to model the unknown object of interest before data collection. Given a data point, one updates the prior to the posterior, by integrating the data-dependent likelihood against the prior. For high-dimensional data in a Gaussian additive observation noise model, the cost of matrix-vector computations can make likelihood function evaluations very expensive. This motivates the need for dimension-reduction techniques. An error analysis of random subsampling methods is presented, using stability estimates for the posterior with respect to likelihood perturbations, and this analysis is applied to a subsampling method of a prior study that uses the Achlioptas distribution.
