Improving the convergence of back-propagation learning with second-order methods