Convex Formulation of Overparameterized Deep Neural Networks