Efficient Processing of Deep Neural Networks