NeurIPS 2020

On Correctness of Automatic Differentiation for Non-Differentiable Functions

Meta Review

The reviewers are unanimous in their appreciation of the theoretical contributions of the paper, while they (and the authors) also agree that the practical implications for applications built on autodiff systems are less well explored. As the authors say, this is an important first step, so is valuable to convey to the NeurIPS audience. The plots proposed by R3 are indeed helpful, so we welcome the authors' proposal to include them in the final copy.