The reviewers unanimously recommended to accept this work --- congratulations! In your revision, please address all requests for filling in missing derivations/details (see reviewer 1's comments and others). Of course address all promises from the rebuttal phase as well. After the review process, this paper sparked some discussion on whether it really belongs at a machine learning conference, since it is quite close to the "adjoint method" for simulation and examples 4.1-4.2 are not technically involved from a machine learning perspective. If possible in the revision, any additional experiments/applications/examples that can be added that justify inclusion at a learning conference (as opposed to simulation/graphics) would help.