Accelerating MCMC via Parallel Predictive Prefetching
This accelerates Bayesian inference for practitioners dealing with computationally intensive MCMC methods, though it appears incremental as it builds on existing parallelization ideas.
The paper tackles the problem of slow Markov chain Monte Carlo (MCMC) algorithms by introducing a parallel predictive prefetching framework that uses fast approximations to evaluate potential future steps concurrently, achieving near-linear speedup in burn-in phases without compromising exactness.
We present a general framework for accelerating a large class of widely used Markov chain Monte Carlo (MCMC) algorithms. Our approach exploits fast, iterative approximations to the target density to speculatively evaluate many potential future steps of the chain in parallel. The approach can accelerate computation of the target distribution of a Bayesian inference problem, without compromising exactness, by exploiting subsets of data. It takes advantage of whatever parallel resources are available, but produces results exactly equivalent to standard serial execution. In the initial burn-in phase of chain evaluation, it achieves speedup over serial evaluation that is close to linear in the number of available cores.