The use of deep learning is to construct architectures that can converge towards a function which generalizes efficiently rather than relying on piecewise linear functions which are limited in their scope.