NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:8412
Title:Algorithmic Guarantees for Inverse Imaging with Untrained Network Priors


		
This paper describes new theory associated with using untrained neural networks to conduct regularization. The idea has been explored empirically before, but the theory provides new insight into why these ideas work. Some of the proofs are derivative of proofs in other contexts, and the key assumption (analogous to the RIP) is strong, especially since in the deep image prior makes this assumption much less interpretable than it would be in a sparse recovery problem.