Accelerating markov random field inference using molecular optical gibbs sampling units

摘要: The increasing use of probabilistic algorithms from statistics and machine learning for data analytics presents new challenges opportunities the design computing systems. One important class is Markov Chain Monte Carlo (MCMC) sampling, which can be used on a wide variety applications in Bayesian Inference. However, this iterative algorithm inefficient practice today's processors, especially problems with high dimensionality complex structure. source inefficiency generating samples parameterized probability distributions.This paper seeks to address sampling approach support that leverages native randomness Resonance Energy Transfer (RET) networks construct RET-based units (RSU). Although RSUs designed applications, we focus specific described as Random Field Our proposed RSU uses RET network implement molecular-scale optical Gibbs unit (RSU-G) integrated into processor / GPU specialized functional or organized discrete accelerator. We experimentally demonstrate fundamental operation an using macro-scale hardware prototype. Emulation-based evaluation two computer vision HD images reveal augmented provides speedups over 3 16. Analytic shows accelerator limited by 336 GB/s DRAM produces 21 54 versus implementations.