On the Convergence of Overparameterized Problems: Inherent Properties of the Compositional Structure of Neural Networks