Core Concepts
Diffusive Gibbs Sampling (DiGS) offers an innovative approach to sampling from multi-modal distributions by bridging disconnected modes using Gaussian convolution.
Abstract
The content introduces Diffusive Gibbs Sampling (DiGS) as a method to address the inadequate mixing of conventional Markov Chain Monte Carlo methods for multi-modal distributions. DiGS leverages diffusion models and Gaussian convolution to improve mode coverage in sampling tasks, demonstrating superior results compared to traditional methods like parallel tempering. The article discusses the application of DiGS in various domains, including mixtures of Gaussians, Bayesian neural networks, and molecular dynamics. It also provides detailed insights into score-based MCMC methods, convolution-based techniques, initialization strategies, hyperparameter selection, and multi-level noise scheduling.
Introduction
Generating samples from complex unnormalized probability distributions is crucial.
Goal: Draw independent samples from the target distribution and estimate expectations.
Score-Based MCMC Methods
Unadjusted Langevin Algorithm (ULA) follows a transition rule based on a discrete-time Langevin SDE.
Metropolis-adjusted Langevin Algorithm (MALA) corrects bias using the Metropolis-Hasting algorithm.
Convolution-Based Method
Gaussian convolution bridges disconnected modes in distributions effectively.
Convolved distribution exhibits better connectivity between modes than the original distribution.
Diffusive Gibbs Sampling
DiGS uses a Gibbs sampler to sample from joint distribution p(x, ˜x).
Alternately draws samples from conditional distributions p(˜x|x) and p(x|˜x).
Comparison to Related Methods
Contrasts with tempering-based sampling, score-based diffusion models, proximal MCMC methods like HMC.
Empirical Evaluation
Evaluates DiGS on synthetic problems like MoG-40, Bayesian neural networks, and real-world applications like molecular dynamics.
Conclusion
DiGS offers significant improvements in sampling multi-modal distributions efficiently and accurately.
Stats
Our approach exhibits a better mixing property for sampling multi-modal distributions than state-of-the-art methods such as parallel tempering.
Quotes
"Our approach exhibits a better mixing property for sampling multi-modal distributions than state-of-the-art methods such as parallel tempering."