Learning with an unreliable teacher

摘要

In nonparametric pattern recognition problems one has to “learn” a good decision rule from a long training sequence, that is, independent pairs of observations and corresponding labels. In many practical situations there may be errors among the labels of the training sequence, that is, “the teacher may sometimes lie”. In this paper we investigate the behavior of two widely used methods in this situation under very general conditions. One of these methods is based on the maximization of the estimated a posteriori probabilities, the other is the nearest neighbor classification.