Reconstruction from very few sampling measurements has recently received a huge boost in performance using supervised deep learning methods. However, while they perform extremely well on data satisfying the conditions they were trained on, their performance deteriorates significantly once these conditions are not satisfied. In this talk, we will introduce some unsupervised deep learning schemes combined with classical iterative procedure for highly under-sampling MRI reconstruction, from denoising autoencoder to score-based generative model. Integrating the learned deep prior knowledge into classical model-based reconstruction, comparable performance can be achieved under various sampling patterns and acceleration factors.