CMStatistics 2018: Start Registration
View Submission - CMStatistics
B1156
Title: Distribution dependent learning with asymmetric label noise Authors:  Henry Reeve - University of Birmingham (United Kingdom) [presenting]
Ata Kaban - University of Birmingham (United Kingdom)
Abstract: Finite sample bounds are presented for a nearest neighbour based algorithm in the presence of unknown asymmetric label noise. Our first result shows that minimax optimal rates are attained whenever the regression function is Lipschitz continuous and the marginal density is uniformly bounded away from zero. In particular, fast rates may be attained whenever Tysbakov's noise condition holds. We then consider the more general non-compact setting in which the density may be arbitrarily close to zero within its support. In this setting, learning with label noise becomes more challenging and depends heavily upon the behaviour of the marginal distribution in neighbourhoods of the regression functions extrema.