Topic: Model geometry and estimation with incomplete data
By: Adrian Corduneanu (MIT), Tommi Jaakkola (MIT)
Dealing with limited or incomplete information is a recurring problem in large scale estimation tasks. Suitable constraints on the probability models permit a beneficial fusion of complete and incomplete information whereas under or overconstrained models may degrade as a result. We study the information geometry of probability models starting from simple overconstrained models and analyze when they can and cannot effectively exploit incomplete information in a classification context. We are also concurrently developing new estimation algorithms and criteria that are better suited for estimation tasks with predominantly incomplete data.
Questions and comments to the webmaster. This page is copyrighted by the Massachusetts Institute of Technology.