Bilevel Optimization (BO) is primarily used for solving complex problems in machine learning, such as hyperparameter optimization, meta-learning, and reinforcement learning5. It is particularly effective in scenarios where the solution to one problem depends on the solution to another, allowing for a hierarchical approach to optimization.
BO's interdependence affects its scalability as it creates a mutual dependency between the upper and lower levels of problems, introducing significant computational challenges, especially in large-scale problems. This interdependence hinders the efficient parallelization of BO, making it difficult to scale for massive data and high-dimensional problems.
The two main types of Bilevel Optimization (BO) methods discussed are: (a) approximate implicit differentiation (AID) methods, and (b) iterative differentiation (ITD) methods. Both approaches follow a two-loop manner and require significant computational costs for large-scale problems.