University of Oxford
Gareth Roberts (University of Warwick)
This talk will discuss ways of subverting the usual order of traditional Monte Carlo simulation algorithms to achieve huge computational advantages. I will give mainly examples, ranging from simple toy illustrations to infinite dimensional examples involving the simulation of stochastic processes.
ABC for Big Data
Paul Fearnhead (University of Lancaster)
Approximate Bayesian Computation (ABC) is a Bayesian method which uses the ability to simulate from a model to replace the need to be able to calculate likelihoods. It has received a lot of recent interest due to its promise of being able to fit complex models to data. However most current applications for ABC focus on relatively small data sets, and on models with relatively few (up to 5 or so) parameters. There are real issues with getting ABC to scale to big data applications, and to complex models which have many parameters.
This talk will introduce ABC, cover recent ideas on how to implement ABC effectively in practice, and consider the challenges of and opportunities for using ABC for big data applications.
Methods for big data in medical genomics
Modern measurement technologies and information streams are producing massive heterogeneous data sets, "big data", particularly in the areas of medical genomics and electronic health records (eHealth). This trend carries major challenges to traditional statistics both in terms of computation and in models that can handle the heterogeneity and variety of data. This talk will review the potential of emerging hardware developments such as general purpose graphic processing units (GPGPUs) for big data and the challenges in scaling flexible (nonlinear) Bayesian methods to real world problems arising in medical genomics and eHealth.
Composite Likelihood Methods: Advances and Challenges
David Firth (University of Warwick)
In this talk I will briefly describe what Composite Likelihoods are, and why they can be useful in connection with large and/or complex problems where the full likelihood is intractable. Recent advances and some current challenge areas will be sketched, along with some prominent current application areas.
For more detailed information a good place to start is the 2011 special issue of Statistica Sinica,
Adaptive Monte Carlo methods
Christophe Andrieu (University of Bristol)
This talk will discuss and review adaptive Monte Carlo methods, that is sampling numerical methods which tune themselves automatically in order to optimise performance.