Choisir la langue :

On Markov chain Monte Carlo for tall data

Markov chain Monte Carlo methods are often deemed far too computationally intensive to be of any practical use for big data applications, and in particular for inference on datasets containing a large number of individual data points, also known as tall datasets. In the case where the model assumes independence of the data, various approaches to scale up Metropolis-Hastings (MH) have been recently proposed in machine learning and statistics. These approaches can be grouped in two categories: subsampling-based algorithms and divide-and-conquer approaches. In this talk, I will give a tutorial introduction to subsampling-based approaches, which randomly subsample the dataset at each MH iteration. I will illustrate existing theoretical results (or lack thereof) with simple examples. This talk is based on the following paper, which is joint work with Arnaud Doucet and Chris Holmes, and is both a survey and a follow-up to work I presented at a Modal seminar in 2014. Of course I will assume everyone remembers perfectly what I said back then.

Tuesday, June 21, 2016 - 16:00 to 17:00
Inria Lille - Nord Europe, bâtiment A, salle plénière
Rémi Bardenet