How Neural Networks Escape Perils of Overparameterization