Researchers at NVIDIA, MIT, and Aalto University have applied "basic statistical reasoning to signal re-construction by machine learning – learning to map corrupted observations to clean signals – with a simple and powerful conclusion: it is possible to learn to restore images by only looking at corrupted examples, at performance at and sometimes exceeding training using clean data, without explicit image priors or likelihood models of the corruption." "NVIDIA says there are several real-world applications for this AI, as sometimes there simply aren’t any actual “clean” examples to train AI with. Examples include low-light photography in certain fields (e.g. astrophotography) in which low-quality images are all we have." (petapixel)
"Basic statistical reasoning to signal reconstruction by machine learning – learning to map corrupted observations to clean signals" (paper)
50,000 photos used to train the model
Their results "show that a single model learns photographic noise removal, denoising synthetic Monte Carlo images, and reconstruc- tion of undersampled MRI scans – all corrupted by different processes – based on noisy data only." (paper)