On Variance Reduction In Machine Learning