On-Device Training of Machine Learning Models on Microcontrollers with Federated Learning

Recent progress in machine learning frameworks has made it possible to now perform inference with models using cheap, tiny microcontrollers. Training of machine learning models for these tiny devices, however, is typically done separately on powerful computers. This way, the training process has abundant CPU and memory resources to process large stored datasets. In this work, we explore a different approach: training the machine learning model directly on the microcontroller and extending the training process with federated learning. We implement this approach for a keyword spotting task. We conduct experiments with real devices to characterize the learning behavior and resource consumption for different hyperparameters and federated learning configurations. We observed that in the case of training locally with fewer data, more frequent federated learning rounds more quickly reduced the training loss but involved a cost of higher bandwidth usage and longer training time. Our results indicate that, depending on the specific application, there is a need to determine the trade-off between the requirements and the resource usage of the system.