SETTLE: SEquential Training-based Transfer LEarning in heterogeneous wireless sensor networks

The application of machine learning (ML) in Wire-less Sensor Networks (WSNs) has gained significant attention due to the enormous amount of data being collected by sensor nodes and the introduction of ML solutions tailored for low-power devices. To overcome the limitations in computation and communication resources required for training ML models, Transfer Learning (TL) has been proposed. However, in het-erogeneous WSN settings in which different nodes are equipped with different types of sensors and execute different tasks, the application of TL solutions is not straightforward. In this paper, we presents SETTLE that enables the application of TL solutions in such heterogeneous WSN settings. Specifically, the proposed algorithm operates in such a way that a subset of the neural network layers is specialized to represent the input data, while the remaining layers are specialized to execute the desired ML task. Experimental results assess the efficiency and effectiveness of SETTLE.