Compact and Computationally Efficient Representation of Deep Neural Networks