David Dunson-Duke University

Tuesday, January 19, 2016

12:30pm-1:30pm

Computer Science, Room 105

*Lunch will be provided*

Title: “Probabilistic inference from complex and high-dimensional data” 

Abstract: There is a well-known increasing trend in complexity and size of data being collected across fields. Although many approaches have been developed for analyzing such data, there is a clear lack of methods that are robust, computationally scalable, and provide realistic uncertainty quantification.  We develop broad new tools for probabilistic inference from big and complex data along several threads, including a new framework for robust Bayesian inferences using coarsening, and new algorithms for scalable computation relying on (i) breaking data into subsets, estimating posterior probability measures for each subset and combining via Wasserstein barycenters; and (ii) approximate Markov chain Monte Carlo algorithms, which replace expensive transition kernels with approximations. We discuss theoretical results on robustness, computational complexity, and statistical accuracy and illustrate the methods with several applications.

If you would like to be added to the CSML listerv, please email capizzi@princeton.edu.