MURI Research
Topic: Efficient use of partially annotated information
By: M. Szummer (MIT), T. Jaakkola (MIT)
Efficient use of partially annotated/labeled data involves extracting structure from large unlabeled set and combining this information with limited labeled examples. A typical albeit unstated assumption in this context associates separable clusters in the unlabeled set with unique but unknown labels. When this assumption is valid, labeled examples are needed only to the extent that they can facilitate labeling of the clusters. We capture and formalize this intuition in terms of Markov random walks, determine the appropriate time scales for the processes to guarantee unambiguous classification of examples, and analyze the abstract problem structure that permits efficient exploitation of partial information.

Questions and comments to the webmaster. This page is copyrighted by the Massachusetts Institute of Technology.