Particle MCMC algorithms and architectures for accelerating inference in state-space models

Area of Science:

• Computational Statistics
• Bioinformatics
• Hardware Acceleration

Background:

• Particle Markov Chain Monte Carlo (pMCMC) is crucial for Bayesian inference in State-Space Models (SSMs).
• High computational costs and poor multi-modal posterior performance limit pMCMC's scalability with large datasets.
• Existing methods struggle with complex SSMs in scientific applications.

Purpose of the Study:

• To enhance pMCMC efficiency for multi-modal posteriors using multiple chains (ppMCMC).
• To develop custom parallel hardware architectures on FPGAs for pMCMC and ppMCMC.
• To enable previously intractable large-scale SSM-based data analyses.

Main Methods:

• Proposed a novel parallel pMCMC algorithm (ppMCMC) utilizing multiple Markov chains.
• Designed and implemented custom parallel hardware architectures on Field Programmable Gate Arrays (FPGAs).
• Evaluated the algorithm and architectures using a large-scale genetics case study.

Main Results:

• ppMCMC demonstrated 1.96x higher sampling efficiency than pMCMC on CPUs.
• FPGA implementations achieved significant speedups: pMCMC (12.1x CPU, 10.1x GPU) and ppMCMC (34.9x CPU, 41.8x GPU).
• FPGA architectures offered substantial energy efficiency gains (pMCMC: 53x, ppMCMC: 173x).

Conclusions:

• The ppMCMC algorithm and FPGA architectures effectively address computational challenges in SSM analysis.
• This work significantly advances the feasibility of large-scale, complex data analysis in fields like genetics.
• The developed methods pave the way for more efficient and accessible Bayesian inference in scientific research.