Commonsense-2009: Invited Speakers
Commonsense-2009 features invited talks by three leading researchers. The abstracts of their talks are below.
Anthony G. Cohen, University of Leeds, "Acquiring Commonsense Knowledge from Perceptual Observations," Monday, June 1, 11:15-12:15.
Abstract: Crucial to the ultimate attainment of the goal of building an autonomous cognitive agent is endowing the agent with an ability to perceive, understand, formulate hypotheses and act based on the agent's perceptions. I will discuss work undertaken at Leeds in pursuit of this goal. A key focus of our work is to integrate quantitative and qualitative modes of representation and to learn as much as possible from observation of the world, and thus to acquire high level symbolic models. As one example of our approach, I will show how by characterising video sequences using a qualitative spatio-temporal relational descriptions, event classes can be mined, and in turn how a taxonomy of functional object categories can be induced from these event descriptions.
Ernest Davis, New York University, "Commonsense Reasoning about Chemistry Experiments: Ontology and Representation," Tuesday, June 2, 11:00-12:00.
Abstract: How should matter be conceptualized to best support commonsense reasoning about simple physics and chemistry experiments? To address the question, we consider a sheaf of eleven benchmark physical concepts, rules, and scenarios: Part/whole relations among bodies of matter; additivity of mass; motion of a rigid solid object; continuous motion of fluids; fixed mass proportions and spatial continuity at chemical reactions; conservation of mass at chemical reactions; gasses in a container attaining equilibrium; the ideal gas law and the law of partial pressures; liquid at rest in an open container; carrying liquid in an open container; the constant availability of oxygen for reactions in an atmosphere; and surface passivization of metals. We then present a number of different ontologies and representational schemes: the model of atoms and molecules with statistical mechanics; models of spatio-temporal fields, with either points, regions, or histories; models of continuous moving material in terms of chunks of matter, with or without point particles; and a hybrid theory that combines atoms and molecules, chunks of matter, and continuous fields using each where appropriate. We evaluate each of the representational schemes in terms of the ease of representing the benchmark problems and other features. Overall, the field model with histories and hybrid model seem to be best, though neither is unproblematic. We conclude by discussing the major challenges for extending this work.
Sheila McIlraith, University of Toronto, "Diagnosis Revisited," Wednesday, June 3, 11:00-12:00.
Abstract: In 1987, Ray Reiter proposed a logical characterization of diagnosis from first principles that has had significant influence on the study of diagnostic problem solving. Together with de Kleer and Mackworth he extended this characterization in 1992. Since that time there have been several attempts to build on his fundamental work by creating a characterization of diagnosis of dynamical systems. In this talk, we revisit Reiter's original work on diagnosis, as well as more recent work on diagnosis of dynamical systems. We discuss potential shortcomings of this more recent work, and propose a more general formulation of dynamical diagnosis together with some associated computational machinery.