Appendix A Standard error calculation, tuning, and implementation details

In this section we provide more details on different inverse methods we use in the paper. To accelerate the inversion in this paper we perform batch optimization. In Section 4.5 of paper we showed how using a diverse range of activation For the surrogate error, we report the average forward loss. We used an NVIDIA TIT AN X GPU for time evaluation. Training the forward models is trivially parallelizable.