12:00, 27 Sep 1995, WeH 7220 "What to do when you're feeling out of sorts: rankprop and multitask learning for pneumonia risk prediction" Rich Caruana Abstract There are 3 million cases of pneumonia each year in the U.S. About 10% of patients hospitalized for pneumonia die within 60 days. In this talk we present several approaches to training artificial neural nets to predict severity of pneumonia illness to aid in the decision to hospitalize patients or treat them as outpatients. Since the goal is to identify those patients at least risk, we develop a method called rankprop that outperforms traditional backprop in identifying low risk patients. It is applicable wherever the usual goal of function approximation can be relaxed so that learning to order data by the (possibly unknown) function values suffices. We also show that using multitask learning with rankprop further improves performance in this domain. This is joint work with Shumeet Baluja and Tom Mitchell.