On the Complexity of Good Samples for Learning

In machine-learning, maximizing the sample margin can reduce the learning generalization-error. Thus samples on which the target function has a large margin (γ) convey more information so we expect fewer such samples. In this paper, we estimate the complexity of a class of sets of large-margin samples for a general learning problem over a finite domain. We obtain an explicit dependence of this complexity on γ and the sample size.