Rademacher Complexities and Bounding the Excess Risk in Active Learning

摘要

Sequential algorithms of active learning based on the estimation of thelevel sets of the empirical risk are discussed in the paper. LocalizedRademacher complexities are used in the algorithms to estimatethe sample sizesneeded to achieve the required accuracy of learning in an adaptive way.Probabilisticbounds on the number of active examples have been proved and severalapplications to binary classification problems are considered. color="gray">