The paper introduces a decentralized learning approach called Federated Learning, which allows mobile devices to collectively train deep networks without centralizing the raw data. This method, termed *FederatedAveraging*, combines local stochastic gradient descent (SGD) on each client with a server that performs model averaging. The authors present extensive empirical evaluations on five different model architectures and four datasets, demonstrating that the approach is robust to unbalanced and non-IID data distributions and reduces communication costs by 10-100 times compared to synchronized stochastic gradient descent. The experiments cover image classification and language modeling tasks, showing that FederatedAveraging can achieve high accuracy with significantly fewer communication rounds. The paper also discusses the privacy benefits of Federated Learning and addresses challenges such as non-IID data and communication constraints.The paper introduces a decentralized learning approach called Federated Learning, which allows mobile devices to collectively train deep networks without centralizing the raw data. This method, termed *FederatedAveraging*, combines local stochastic gradient descent (SGD) on each client with a server that performs model averaging. The authors present extensive empirical evaluations on five different model architectures and four datasets, demonstrating that the approach is robust to unbalanced and non-IID data distributions and reduces communication costs by 10-100 times compared to synchronized stochastic gradient descent. The experiments cover image classification and language modeling tasks, showing that FederatedAveraging can achieve high accuracy with significantly fewer communication rounds. The paper also discusses the privacy benefits of Federated Learning and addresses challenges such as non-IID data and communication constraints.