2006 ApproximateInfTechForIdentUnc

Subject Headings: Identity Uncertainty Problem, Approximate Inference Algorithm.

Notes

Many interesting tasks, such as vehicle tracking, data association, and mapping, involve reasoning about the objects present in a domain. However, the observations on which this reasoning is to be based frequently fail to explicitly describe these objects' identities, properties, or even their number, and may in addition be noisy or nondeterministic. When this is the case, identifying the set of objects present becomes an important aspect of the whole task.
My talk will discuss how this task can be handled using methods that add relational elements to probabilistic representations; specifically, to directed and undirected graphical models. A recurring problem with such graphical models, ones that express uncertainty over the set of objects in existence as well as over their properties, is that they are highly connected, which makes exact inference and learning highly intractable. Fortunately, many of the connections become irrelevant given a specific set of objects, so the problem can be overcome with the help of approximate techniques based on sampling or stochastic search over the set of objects. I will describe such techniques, and explain how they can be applied to citation matching and topological map construction. In both cases, I will demonstrate that the ability to reason about the properties of the objects responsible for the observations (papers and authors, or locations) can improve a system's ability to identify these objects.

we want to reason about the objects, their properties, their existence, their numbers.

,

	Author	volume	Date Value	title	type	journal	titleUrl	doi	note	year
2006 ApproximateInfTechForIdentUnc	Hanna Pasula			Approximate Inference Techniques for Identity Uncertainty			http://www.researchchannel.org/prog/displayevent.aspx?fID=569&rID=5176			2006