Large Margin Classifiers: Convex Loss, Low Noise, and Convergence Rates