MIT Department of Electrical Engineering & Computer Science

Statistical Inference Under Limited Resources

Tommi S. Jaakkola
University of California, Santa Cruz

Thursday, April 23, 1998
4:00 PM (refreshments 3:45)
NE43-941
EECS Special Seminar

Abstract

Graphical models provide a unified framework for (model based) statistical decision making. These models are interpretable graphical representations of probabilistic dependencies/independencies and accompany well worked-out algorithms for carrying out inferences. They are increasingly being adopted in major application areas, from medical diagnosis to software products. The main limitation of these models, however, is their rather high demand on computational resources. I will illustrate a general methodology for balancing the accuracy of inference with the available computational resources in the context of a large scale problem in medical diagnosis. Graphical models as statistical models may be also primarily limited by their accuracy. In such cases an application of discriminative (or model free) techniques typically leads to better performance. Ideally, however, we would like to combine the strengths of these rather complementary approaches. I will show how graphical models and discriminative techniques can be combined in a general way, and illustrate the effectiveness of such techniques in the context of DNA and protein sequence analysis.