Abstract: Wild-used gradient-based Multi-Objective Bi-Level Optimization (MOBLO) algorithms rely on solving numerous approximation subproblems with high accuracy, which significantly impacts numerical efficiency in terms of time and memory complexity. To address this issue, we propose a simpler and more efficient gradient-based algorithm for MOBLO, called g-MOBA, which has fewer hyperparameters to tune. We validate the theoretical soundness of the approach by achieving the desired Pareto stationarity, and numerical experiments on few-shot learning applications confirm its practical efficiency.
