On-device training and pruning for energy saving and continuous learning in resource-constrained MCUs

Tiny Machine Learning has gained significant research interest in the Internet of Things and edge computing fields. In this direction, continuous on-device learning represents an emerging research topic that aims to improve performance and minimize concept drift and privacy violations related to the need to transmit on-field data to a remote server for periodic retraining and refinement of the model. However, the resource consumption resulting from the intensive and prolonged use of the neural network for inference activity, can represent a strong limitation in the use of tiny devices, particularly because of the stringent constraints on memory and power consumption.Therefore, this paper presents the implementation of a novel on-device pruning procedure, designed to reduce the model’s latency and power consumption, without the need to send data to a central server. It runs during the incremental on-board training process, directly on memory-constrained devices based on MCUs without relying on custom hardware. An experimental evaluation on real devices is provided to support machine learning developers in choosing what can be the best compromise between model performance and energy consumption.