When Neural Networks Fail to Generalize? A Model Sensitivity Perspective