Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.684291
Title: Algorithms and architectures for MCMC acceleration in FPGAs
Author: Mingas, Grigorios
ISNI:       0000 0004 5920 7399
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2015
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Markov Chain Monte Carlo (MCMC) is a family of stochastic algorithms which are used to draw random samples from arbitrary probability distributions. This task is necessary to solve a variety of problems in Bayesian modelling, e.g. prediction and model comparison, making MCMC a fundamental tool in modern statistics. Nevertheless, due to the increasing complexity of Bayesian models, the explosion in the amount of data they need to handle and the computational intensity of many MCMC algorithms, performing MCMC-based inference is often impractical in real applications. This thesis tackles this computational problem by proposing Field Programmable Gate Array (FPGA) architectures for accelerating MCMC and by designing novel MCMC algorithms and optimization methodologies which are tailored for FPGA implementation. The contributions of this work include: 1) An FPGA architecture for the Population-based MCMC algorithm, along with two modified versions of the algorithm which use custom arithmetic precision in large parts of the implementation without introducing error in the output. Mapping the two modified versions to an FPGA allows for more parallel modules to be instantiated in the same chip area. 2) An FPGA architecture for the Particle MCMC algorithm, along with a novel algorithm which combines Particle MCMC and Population-based MCMC to tackle multi-modal distributions. A proposed FPGA architecture for the new algorithm achieves higher datapath utilization than the Particle MCMC architecture. 3) A generic method to optimize the arithmetic precision of any MCMC algorithm that is implemented on FPGAs. The method selects the minimum precision among a given set of precisions, while guaranteeing a user-defined bound on the output error. By applying the above techniques to large-scale Bayesian problems, it is shown that significant speedups (one or two orders of magnitude) are possible compared to state-of-the-art MCMC algorithms implemented on CPUs and GPUs, opening the way for handling complex statistical analyses in the era of ubiquitous, ever-increasing data.
Supervisor: Bouganis, Christos-Savvas Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.684291  DOI: Not available
Share: