Learning in the Presence of Inaccurate Information

Abstract

In this paper we discuss the effects of errors in input data on recursion theoretic learning. We consider three types of inaccuracy in input data depending on the presence of extra data (noise), missing data (incompleteness) or both (imperfection). We show that for function learning incompleteness harms strictly more than noise. However for language learning, identification from incomplete text and identification from noisy text are incomparable. We also prove hierarchies based on the number of inaccuracies present in the input.