Learning Graphical Model Parameters with Approximate Marginal Inference

TL;DR

This paper explores methods for training graphical models by directly optimizing marginal accuracy, improving robustness and performance in complex, approximate inference scenarios, especially in imaging applications.

Contribution

It introduces a marginalization-based learning approach that outperforms likelihood-based methods in challenging approximate inference tasks.

Findings

Marginalization-based learning yields better accuracy in complex models.

The approach is more robust to model mis-specification.

Experiments demonstrate improved performance on imaging problems.

Abstract

Likelihood based-learning of graphical models faces challenges of computational-complexity and robustness to model mis-specification. This paper studies methods that fit parameters directly to maximize a measure of the accuracy of predicted marginals, taking into account both model and inference approximations at training time. Experiments on imaging problems suggest marginalization-based learning performs better than likelihood-based approximations on difficult problems where the model being fit is approximate in nature.