DeepMind published early research exploring the use of federated learning techniques to train language models