Traditional machine learning methods mainly focused on the modeling on the deterministic knowledge. While in real application cases, machine learning methods easily occur robust issues for data noises. While such robust issue is closely related to the proper specification of loss function used in the machine learning model. In this talk, I will make a discussion on how to approximately learn a proper loss function from data under MAP/MLE framework. Such modeling strategy has been performed on various application problems, like online background subtraction from dynamic videos and CT image reconstruction.