In some domains, the number of data (especially of some classes) is sometimes limited. Thus it would be best if our algorithm converged to a near-correct answer even from a small amount of data.
