Smart Split: Leveraging TinyML and Split Computing for Efficient Edge AI

The rapid advancement of Internet of Things (IoT) devices requires innovative approaches to implement machine learning (ML) in resource-constrained environments. This paper explores the integration of Tiny Machine Learning (TinyML) with split computing, focusing on classification using an ESP32 microcontroller connected to a Raspberry Pi edge server. We conduct a series of experiments to measure the time required for image capture and classification, rather than focusing solely on model accuracy. Our findings indicate that while local processing on the ESP32 is limited by its computational capabilities, the split computing approach significantly reduces the processing time by leveraging the Raspberry Pi's computing resources. We then highlight the benefits of our approach considering a dynamic scenario, in which networking changes hence the possibilities to perform split computing vary over time.