Hamiltonian Annealed Importance Sampling for partition function estimation

Abstract

We introduce an extension to annealed importance sampling that uses Hamiltonian dynamics to rapidly estimate normalization constants. We demonstrate this method by computing log likelihoods in directed and undirected probabilistic image models. We compare the performance of linear generative models with both Gaussian and Laplace priors, product of experts models with Laplace and Student's t experts, the mc-RBM, and a bilinear generative model. We provide code to compare additional models.

References in corpus (3)

• MCMC using Hamiltonian dynamics
• Equilibrium free energy differences from nonequilibrium measurements: a master equation approach
• Thermodynamics of natural images

Cited by in corpus (10)

• Modified Hamiltonian Monte Carlo for Bayesian inference
• Hamiltonian Monte Carlo Without Detailed Balance
• Generalizing Hamiltonian Monte Carlo with Neural Networks
• Ergodic Inference: Accelerate Convergence by Optimisation
• MCMC Variational Inference via Uncorrected Hamiltonian Annealing
• Scalable Approximate Inference and Some Applications
• Beyond Brightness Constancy: Learning Noise Models for Optical Flow
• A Markov Jump Process for More Efficient Hamiltonian Monte Carlo
• Differentiable Annealed Importance Sampling and the Perils of Gradient Noise
• Variational Optimization of Annealing Schedules