Proximally Regularized Multitask Learning over Networks

We consider a collection of intelligent agents with the objective of learning distinct, but related tasks. Classical algorithms for federated and decentralized learning employ consensus mechanisms to learn a single, common model across the different agents and tasks. In recent years, there has been increasing interest in multitask or meta-learning techniques to develop personalized learning algorithms over networks, though most studies focus on the federated setting with single fusion center. We develop a fully decentralized strategy for personalized learning over networks based on proximal regularization and explore relations with alternatives based on quadratic regularization or meta-learning. Numerical results demonstrate the benefit of the proposed approach.