Federated and Tiny Machine Learning for Edge Computing in IoT

Abstract

With the emergence of IoT and microcontrollers in general, as well as ad- vancements in machine learning processes, the desire to continuously au- tomate and process the world on smaller and smaller mediums has grown with the shrinking of computational power. Known libraries exist for devel- oping ML models for inference on devices, however, the act of decentralizing the entire training process to devices is still in its infancy. As such this task seeks to address the potential of deploying such machine learning capa- bility on microcontrollers and explore what improvements can be gained by leveraging federated learning to have small devices cooperate in creat- ing a large enough, and sufficiently accurate mode for different use cases. The thesis seeks to test that primarily by running on-device LSTM model training on PM10.0 data gathered by NILU, however briefly explores the potentials of on-device training of DNNs, and CNNs for image classifica- tion specifically. An assessment was made to determine the necessities in parameters for gaining a favourable model outcome.