Autograd enables us to push the limits of machine learning by automatically calculating gradients. This levitates us from manually deriving and implementing the gradients. This technique comes to its limits when trying to solve a Bi-Level optimization problem (hyperparameter, meta-learning). This paper proposes a new system aimed at solving this using existing autograd systems by leveraging a mapping function to define optimality.
Link to the paper: https://arxiv.org/abs/2105.15183