A Deep Reinforcement Learning Perspective on Adaptive Federated Dropout

Federated learning is a distributed learning solution that achieves high-quality machine learning models while ensuring privacy and collaboration among various end devices. However, different kinds of end devices can lead to an unstable training process due to limited and dynamic communication and computation resources. Federated Dropout is an important technique for mitigating such resource bottlenecks by randomly removing a fixed percentage of activations of model components. However, most of the existing FL dropout methods cannot effectively utilize the distinguished characteristics of different clients or fail to automatically adapt to end device heterogeneity. In this paper, we thus propose Adaptive Federated Dropout with Reinforcement Learning (AFD-RL). AFD-RL employs reinforcement learning technology to adapt to the clients’ unbalanced computation and communication resources and provide a personalized dropout strategy for each client. By automatically selecting suitable model components for each client in each communication round, AFD-RL achieves better accuracy with fewer communication rounds. Extensive experiments validate the effectiveness of the proposed approach in improving training efficiency and inference accuracy.