Another machine learning paper ("ancient", 2015) where being able to exactly reverse a computation was useful: https://arxiv.org/abs/1502.03492