Advances in Scalable Bayesian Computation (14w5125)
Luke Bornn (Simon Fraser University)
Nando de Freitas (University of British Columbia)
Christian Robert (Université Paris-Dauphine)
Scott Schmidler (Duke University)
As the statistical models used to understand complex systems grow, the methods used to fit these models must scale accordingly. While advanced computational methods are being developed to fit these complex models, their speed and memory requirements often demand tremendous computational power via large clusters. This approach of relying on ever larger computational resources is quickly becoming unsustainable, particularly for practitioners for whom these resources are not available. As such, there is a large and growing need for statistically efficient methods which scale in terms of speed and memory while being straightforward to implement and communicate.
Several communities are currently working on methods for Bayesian computation which are scalable; however, interaction between these communities has been limited and as such significant cross-community learnings are being missed. As the communities grow and diverge, it is important to bridge the gap at this crucial stage in the development of scalable methods.
Currently, the exchanges between the Monte Carlo (including MCMC, SMC, and ABC), INLA, optimization, and other communities remain surprisingly limited. With each group having their own workshops and conferences, and little work being done to compare and educate between the various approaches, we believe this workshop would be an ideal place to:
(1) Gather people from different research communities and foster links between these communities. (2) Expose the various communities to the state of the art in scalable Bayesian computation methods, including MCMC, SMC, ABC, INLA, optimization, and other methods. (3) Classify the advantages, limitations, and possibilities from each class of methods, and equip all participants with the knowledge to bridge the existing gap in scalable Bayesian computation. (4) Create a venue to encourage innovation through the synthesis and development of new approaches by combining existing, currently disjoint, approaches to Bayesian computation.
In reaching out to statisticians, computer scientists, mathematicians, and others working in Bayesian computation as potential participants, the desire for such a workshop was overwhelming. To quote several of the contacted participants, “a fantastic and very timely conference topic,” “sounds incredibly interesting and highly pertinent,” “a particularly timely topic,” and “a great opportunity to see what others are doing.”