Data Selection Scheme for Energy Efficient Supervised Learning at IoT Nodes

In this letter, we consider a system model comprised of an Internet-of-Things (IoT) node connected wirelessly to a cloud server. The IoT node is assumed to generate data by sensing its environment and make inferences from the data. To this end, the IoT node can rely on its on-device neural network and make inference locally, which incurs small energy cost but a relatively inaccurate inference, or it can wirelessly transmit the data sample to the cloud so that the cloud makes the inference and feeds it back to the IoT node, which incurs a large energy cost but a more precise inference. For this system model, we propose a scheme that the IoT device can employ to select the data samples that would likely lead to inaccurate inferences if processed locally so that those data samples are transmitted to the cloud. Thereby, the overall inference precision of the system is significantly improved for a given energy cost compared to the case when the inference is always made locally at the IoT device.