[Reading] Lecture 12 - A Comparison of Algorithms for Inference and Learning in Probabilistic Graphical Models

This paper demonstrates several inferences and learning algorithms by a simple computer vision application - occlusion. The graphical models that have been introduced in the paper includes the Bayesian Network, the Markov Random Field, and the Factor Graph. Since the occlusion application is simple, we could easily learn the pros and cons of each graphical model from the examples. Bayesian Network is good at indicating marginal independence while Markov Random Field is good at indicating conditional independence.

The rest of the paper discuss many different algorithms for inference, including exact and approximate inferences. Since the algorithms for exact inference is often intractable, we are interested more on the characteristics of the algorithms for approximate inferences. ICM reaches lower energy more quickly, but converges at a higher energy compared to other algorithms.

Reference :
B. J. Frey and N. Jojic, "A Comparison of Algorithms for Inference and Learning in Probabilistic Graphical Models," IEEE Trans. Pattern Analysis and Machine Intelligence, 27(9), pp. 1392-1416, September 2005.