Abstract: In recent work, we’ve shown how to learn to fix images that have undergone complex, severe corruptions, entirely without clean training data. This “Noise2Noise" algorithm (ICML 2018) supervises the restoration model using other corrupted realizations of the same underlying clean (but unobserved) signal. In this talk, I’ll describe an extension into the self-supervised regime (NeurIPS 2019), where the training examples consist of just a single corrupted image, instead of a corrupted pair. While this restricts the class of corruptions we can learn to remove to those where the errors in pixels are uncorrelated, we demonstrate results on par with fully supervised learning with clean targets.
Speaker: Professor Jaakko Lehtinen
Affiliation: Aalto University and NVIDIA Research
Place of Seminar: Zoom