Towards a theory of model complexity in deep neural networks