Identity Uncertainty

Hanna Pasula (Berkeley)

Abstract

Identity uncertainty is uncertainty about the set of objects responsible for a set of observations. This form of uncertainty can exist in many Artificial Intelligence applications, from database cleanup (where we wish to merge co-referring records) to multi-target tracking (where we group together the observations referring to each actual object.) In this talk, I describe a general approach for representing and reasoning about identity uncertainty, and apply it to two domains: citation clustering, and freeway surveillance. Our approach is based on a relational probabilistic representation, and so, unlike simple clustering, it permits us to reason about the relationships between objects, such as papers and their authors. Since exact inference is intractable, we have developed a family of approximate inference methods based on Markov chain Monte Carlo. The performance of these methods is comparable to that of current algorithms.

Back to the Main Page