Revisiting Small Batch Training for Deep Neural Networks