Learning to Prune Deep Neural Networks via Layer-wise Optimal Brain Surgeon