結果 : need of regularization in deep learning