Conformal predictors for reliable pattern recognition

Abstract

The talk reviews a modern machine learning technique called Conformal Predictors. The approach has been motivated by algorithmic notion of randomness and allows us to make reliable predictions with valid measures of confidence for individual examples. The developed technique guarantees that the overall accuracy can be controlled by a required confidence level. Unlike many conventional techniques the approach does not make any additional assumption about the data beyond the i.i.d. assumption: the examples are independent and identically distributed. The way to test this assumption is described. The talk also outlines some generalisations of Conformal Predictors and their applications to many different fields including medicine, cheminformatics, information security, environment, plasma physics, home security and others.