Bayesian and Markov networks have been widely successful for

learning and reasoning in domains with uncertainty, but each has

limitations.  Dependency networks are an alternative graphical

representation with more flexibility than Bayesian networks and more

efficient learning methods than Markov networks.  The disadvantages of

dependency networks are that they may represent inconsistent

probability distributions and few inference algorithms are applicable.

In this talk, I will show how we can improve the utility of dependency

networks with new learning and inference algorithms.  First, I will

show how mean field inference can be a faster alternative to Gibbs

sampling, even in inconsistent dependency networks.  Second, I will

show how dependency networks can be used to learn better Markov

networks in less time, compared to several state-of-the-art methods.

Finally, I will introduce a new method for directly converting an

inconsistent dependency network into a consistent Markov network.

 

Based on joint work with Jesse Davis and Arash Shamaei.