Loss landscape Characterization of Neural Networks without Over-Parametrization