Inference and learning in Bayesian Networks

Applications in fields such as medicine, image processing, bioinformatics, speech processing, etc, involve large-scale models in which thousands of random variables must be thoroughly linked. Probabilistic graphical models typically provide a rich platform where such problems can be approached. Most notably, probabilistic graphical models have the potential to provide compact and tractable representations of several real-world problems and applications. Graphical models also allow us to abstract out the conditional independence relationships between the variables from the details of their parametric forms.

Bayesian networks are probabilistic directed acyclic graphical models that encode probabilistic relationships among variables of interest. Bayesian networks (BNs) allow the resulting model to statistically handle situations where some data entries are missing (incomplete data), to predict consequences of intervention related to a certain problem domain by learning causal relationships, among other advantages. After an introduction to BNs, we move on to how BN structures can be learnt from the observed data of a problem or application, how parameters of the learnt BN can in turn be learnt, and then to how inference can be performed on BNs.

Tameem Adel is a Postdoctoral Fellow at the Machine Learning Lab at University of Amsterdam, led by Max Welling. He will be moving in three weeks' time to the Machine Learning group, School of Computer Science at University of Manchester, UK. He is currently involved in three projects, two in collaboration with Max Welling at the University of Amsterdam. The first project is focussed on ABC (Approximate Bayesian computation) algorithms, and the second is about classification of brain MRI scans using unsupervised representations. The third project is focussed on structure learning of Bayesian networks. Tameem got his PhD from University of Waterloo, Waterloo, ON, Canada.