Learning Graphical Model Parameters with Approximate Marginal Inference
This provides a more robust alternative for graphical model learning in domains like imaging where exact models are unavailable.
The paper tackles the problem of learning graphical model parameters by maximizing marginal prediction accuracy instead of likelihood, addressing computational complexity and model mis-specification issues. Experiments on imaging problems show this approach outperforms likelihood-based methods when models are approximate.
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.