Focusing on Difficult Directions for Learning HMC Trajectory Lengths
share
Hamiltonian Monte Carlo (HMC) is a premier Markov Chain Monte Carlo (MCMC) algorithm for continuous target distributions. Its full potential can only be unleashed when its problem-dependent hyperparameters are tuned well. The adaptation of one such hyperparameter, trajectory length (τ), has been closely examined by many research programs with the No-U-Turn Sampler (NUTS) coming out as the preferred method in 2011. A decade later, the evolving hardware profile has lead to the proliferation of personal and cloud based SIMD hardware in the form of Graphics and Tensor Processing Units (GPUs, TPUs) which are hostile to certain algorithmic details of NUTS. This has opened up a hole in the MCMC toolkit for an algorithm that can learn τ while maintaining good hardware utilization. In this work we build on recent advances along this direction and introduce SNAPER-HMC, a SIMD-accelerator-friendly adaptive-MCMC scheme for learning τ. The algorithm maximizes an upper bound on per-gradient effective sample size along an estimated principal component. We empirically show that SNAPER-HMC is stable when combined with mass-matrix adaptation, and is tolerant of certain pathological target distribution covariance spectra while providing excellent long and short run sampling efficiency. We provide a complete implementation for continuous multi-chain adaptive HMC combining trajectory learning with standard step-size and mass-matrix adaptation in one turnkey inference package.
READ FULL TEXT