Regularizing Deep Neural Networks by Noise: Its Interpretation and Optimization