Abstract: Compiling graphical models has emerged over the years as a major approach for inference. Historically, the motivation behind compilation is to split the inference process into two phases, offline and online, with the aim of pushing much of the computational work into the offline phase, while amortizing this work over many online queries. In modern applications, however, the value of model compilation has been quite influential in exploiting the local (parameteric) structure of models, which usually incurrs too much overhead to justify its exploitation by standard inference approaches (unless the local structure is extreme). Theoretical results on model compilation have also presented a unifying framework for classical inference algorithms, based on jointrees, variable elimination and conditioning.

This tutorial will cover the subject of model compilation from both theortical and practical perspectives, providing an exposition to state of the art algorithms, and discussing the impact it has had on scaling up exact inference to levels never attained before. The presented techniques will be illustrated in the context of graphical models that have very high treewidth, making them inaccessible to standard inference approaches. Empirical results from various domains - such as relational models, genetic linkage analysis, coding, and probablistic planning - will also be discussed.

Abstract: This tutorial gives an overview of rich representations for probabilistic reasoning. The first third of the tutorial gives the basics of logic, knowledge represenation and probability. We then overview first-order representations where the semantics is the grounding of the representation. This is then compared to procedural langauges. We then consider the problems of identity uncertainty and existence uncertainty. Finally we overview combinations of ontologies with uncertainty. Commonalities amongst formalisms is emphasized rather than the differences between them.

About the speaker: David Poole is a Professor of Computer Science at the University of British Columbia. He is known for his work on knowledge representation, default reasoning, assumption-based reasoning, diagnosis, reasoning under uncertainty, combining logic and probability, algorithms for probabilistic inference and representations for automated decision making. He is a co-author of an AI textbook, Computational Intelligence: A Logical Perspective (Oxford University Press, 1998), co-editor of the Proceedings of the Tenth Conference in Uncertainty in Artificial Intelligence (Morgan Kaufmann, 1994), is former associate editor and on the advisory board of the Journal of AI research, is secretary of the Association for uncertainty in AI, and is a Fellow of the American Association for Artificial Intelligence.

Abstract: Causal knowledge is essential in making decisions and plans, and causal inference is a central goal of most sciences. In this tutorial I will provide an introduction to recent developments in using graphical models in causal inference. First, I will introduce the concept of a causal manipulation of a variable X to a value, and contrast it with conditioning on a value of X. The task of inferring the effect of manipulating a variable to a particular value can be broken into two parts: inference from data and background knowledge to (partial knowledge of) causal graphical models, and from causal graphical models (or partial knowledge of causal graphical models) to the effects of manipulations. I will introduce and illustrate the basic algorithms for both kinds of inference and describe the basic assumptions underlying these the inference algorithms, and contrast the goals and methods of causal inference with the goals and methods of estimating conditional probabilities. I will also describe a number of important open problems. The material in this tutorial will serve as background for a second tutorial, Causal Inference and Graphical Models - II.

Abstract: In this second tutorial on causal inference and graphical models, I will address the following topics: 1) I will present algorithms for computing the effects of manipulations (or interventions, actions) from a combination of nonexperimental data and qualitative causal assumptions in the form of causal graphs; 2) I will discuss recent developments in inferring the constraints imposed on the probability distributions of the observed variables that are not conditional independence relations, in the case that causal graphs contain hidden variables. I will discuss both equality and inequality constraints; 3) I will introduce the concept and semantics of counterfactuals and illustrate its application to the problem of determining the causes of effects to calculate quantities such as the probability that a person would be alive had he not taken a drug given that the person took the drug and died.