In this talk, we explore a duality between federated learning and subspace correction, which are concepts from two very different fields. Federated learning is a paradigm of supervised machine learning in which data is decentralized into a number of clients and each client updates a local correction of a global model independently via the local data. Subspace correction is an abstraction of general iterative algorithms such as multigrid and domain decomposition methods for solving scientific problems numerically. Based on the duality between federated learning and subspace correction, we propose a novel federated learning algorithm called DualFL (Dualized Federated Learning). DualFL is the first federated learning algorithm that achieves communication acceleration, even when the cost function is either nonsmooth or non-strongly convex.