Asynchronous Federated Learning for Geospatial Applications

Michael R. Sprague, Amir Jalalirad, Marco Scavuzzo, Catalin Capota, Moritz Neun, Lyman Do, Michael Kopp

Abstract: Federated learning is an emerging collaborative machine-learning paradigm for training models directly on edge devices. The data remains on the edge device and this method is robust under real-world edge data distributions. We present a new asynchronous federated-learning algorithm (‘asynchronous federated learning’) and study its convergence rate when distributed across many edge devices, with hard data constraints, relative to training the same model on a single device. We compare asynchronous federated learning to an existing synchronous method. We evaluate its robustness in real-world situations; for example, devices joining part-way through training or devices with heterogeneous compute resources. We then apply asynchronous federated learning to a challenging geospatial application, namely image-based geolocation using a state-of-the-art convolutional neural network. Our results lay the groundwork for deploying large-scale federated learning as a tool to automatically learn, and continually update, a machine-learned model that encodes location.