Improving Generalization of Deep Neural Networks by Optimum Shifting