On the Convex Behavior of Deep Neural Networks in Relation to the Layers' Width