In this talk, I will introduce various unsupervised risk estimators for inverse problems. These estimators can handle the noise in the measurement data and tackle incomplete or heavily ill-conditioned forward operators. I will show how we can leverage these estimators for accurate uncertainty quantification and fully unsupervised learning in various imaging inverse problems, such as computed tomography, magnetic resonance imaging and image inpainting. The resulting methods rely only on noisy and incomplete measurements (no ground truth references), thus paving the way for knowledge discovery in imaging using (deep) learning.