On Numerical Methods for Efficient Deep Neural Networks