Reduced-Order Modeling of Deep Neural Networks