Distributed Newton Methods for Deep Neural Networks