A versatile distributed MCMC algorithm for large scale inverse problems

For large scale inverse problems, inference can be tackled with distributed algorithms, dividing the task over multiple computing nodes or cores referred to as workers. Since random sampling methods yield not only estimates but also credibility intervals, we leverage data augmentations and MCMC algorithms to design a distributed sampler. In contrast with usual approaches relying on a client-server architecture, we propose a flexible distributed sampler relying on a Single Program Multiple Data implementation, in which all workers have a similar task. This distributed strategy allows the computing time and volume of communications to be reduced by separately handling blocks of data and parameters on different workers. Experiments on a large synthetic image inpainting problem illustrate the performance of the proposed approach to produce high quality estimates in a small amount of time.