Deep neural networks are more powerful than shallow networks in approximating complex functions that are compositional in nature, such as language and vision.