A Quasi-Newton Approach to Nonsmooth Convex Optimization Problems in Machine Learning