Hamiltonian Annealed Importance Sampling for partition function estimation
This work addresses the challenge of partition function estimation in probabilistic modeling, which is crucial for tasks like model comparison and likelihood evaluation, but it appears incremental as it extends an existing method (annealed importance sampling) with Hamiltonian dynamics.
The authors tackled the problem of estimating partition functions in probabilistic models by introducing Hamiltonian Annealed Importance Sampling, which uses Hamiltonian dynamics to accelerate normalization constant estimation. They demonstrated the method's effectiveness by computing log likelihoods for various image models, including linear generative models with Gaussian and Laplace priors, product of experts models, and others.
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.