Monte Carlo and reconfigurable computing in Bayesian inference

Stanford University, Stanford CA

Investigators

Abstract

The primary aim of this project is to develop the concepts, methods and algorithms to extract information about the nature of a joint distribution in high dimensional space from samples generated by Monte Carlo algorithms. The approach will be based on the equi-energy sampling approach recently developed by the principle investigator's group under prior NSF support. The secondary aim of this project is to implement Monte Carlo sampling methods based on radically new hardware and computation model, such as those based on field programmable gate arrays. Some of the Monte Carlo methods developed in the primary aim will be implemented using these new architectures, in order to enhance our ability to solve hard inference problem by Monte Carlo computation. This project will provide interdisciplinary training of next generation scientists working at the interface of statistics, computation and biology. By developing methods to study the shape, topology and entropy for the posterior distribution, this research will provide fundamental tools for Bayesian inference. As such, it will have impact on numerous application areas ranging from computational biology to economic analysis. Beyond Bayesian statistics, this research will also have impact on other scientific areas that utilize Monte Carlo sampling to study a distribution, e.g. in equilibrium statistical physics where one is interested in understanding the energy landscape associated with a Boltzmann distribution. This project will provide interdisciplinary training of next generation scientists working at the interface of statistics, computation and biology.