Sandwiching the marginal likelihood using bidirectional Monte Carlo

Abstract

Computing the marginal likelihood (ML) of a model requires marginalizing out all of the parameters and latent variables, a difficult high-dimensional summation or integration problem. To make matters worse, it is often hard to measure the accuracy of one's ML estimates. We present bidirectional Monte Carlo, a technique for obtaining accurate log-ML estimates on data simulated from a model. This method obtains stochastic lower bounds on the log-ML using annealed importance sampling or sequential Monte Carlo, and obtains stochastic upper bounds by running these same algorithms in reverse starting from an exact posterior sample. The true value can be sandwiched between these two stochastic bounds with high probability. Using the ground truth log-ML estimates obtained from our method, we quantitatively evaluate a wide variety of existing ML estimators on several latent variable models: clustering, a low rank approximation, and a binary attributes model. These experiments yield insights into how to accurately estimate marginal likelihoods.

References in corpus (2)

• Practical Bayesian Optimization of Machine Learning Algorithms
• Testing MCMC code

Cited by in corpus (18)

• Inference Suboptimality in Variational Autoencoders
• Differentiable Molecular Simulations for Control and Learning
• Adversarial Message Passing For Graphical Models
• Partition Functions from Rao-Blackwellized Tempered Sampling
• Normalizing Constant Estimation with Gaussianized Bridge Sampling
• Learning Deep Generative Models with Annealed Importance Sampling
• An Easy to Interpret Diagnostic for Approximate Inference: Symmetric Divergence Over Simulations
• AIDE: An algorithm for measuring the accuracy of probabilistic inference algorithms
• Importance Weighted Hierarchical Variational Inference
• Measuring the non-asymptotic convergence of sequential Monte Carlo samplers using probabilistic programming
• Filtering Variational Objectives
• Annealed Importance Sampling with q-Paths
• MCMC Variational Inference via Uncorrected Hamiltonian Annealing
• Quantifying the probable approximation error of probabilistic inference programs
• Stochastic Gradient Annealed Importance Sampling for Efficient Online Marginal Likelihood Estimation
• Stochastic Variational Inference via Upper Bound
• q-Paths: Generalizing the Geometric Annealing Path using Power Means
• Differentiable Annealed Importance Sampling and the Perils of Gradient Noise