Federated Bilevel Optimization

Federated bilevel optimization (FBO) addresses the challenge of training hierarchical machine learning models across decentralized datasets, aiming to improve efficiency and communication in federated learning settings. Current research focuses on developing communication-efficient algorithms, such as those employing momentum-based variance reduction and aggregated iterative differentiation, to overcome the computational and communication bottlenecks inherent in nested optimization structures. These advancements are crucial for enabling practical applications of FBO in areas like hyperparameter tuning, meta-learning, and robust federated learning with noisy labels, ultimately improving the scalability and performance of distributed machine learning systems.