Stabilizing Machine Learning Prediction of Dynamics: Noise and Noise-inspired Regularization