Alexander Gammerman

Conformal Prediction and Its Applications
This talk continues discussion of a new machine learning technique called Conformal Predictors and its applications in different fields. The technique is among the most accurate methods of machine learning and allows us to make reliable prediction in both batch and online modes of learning. Among its advantages are: conformal predictors can be used as set predictors, and we can control the number of erroneous predictions by selecting a suitable confidence level; unlike many conventional techniques it does not make any additional assumptions about the data beyond the exchangeability assumption; measures of confidence produced by conformal predictors are tailored to individual examples; it can be used in high-dimensional problems where the number of attributes greatly exceeds the number of objects. The technique will be illustrated with a number of applications in various fields, including an application to neuroimaging data for diagnostic and prognostic prediction in psychiatry. Its robustness will be illustrated by experiments comparing it to Bayesian methods.