A Dual Optimization View to Empirical Risk Minimization with f-Divergence Regularization