Towards Online Machine Learning Libraries for Embedded Systems

In the context of Smart Buildings, data arrive continually from sensors. Learning about those data is challenging if the option is to realize it at the edge. Under this hypothesis, learning may occur on resource-constrained embedded devices like ESP32. This paper shares our experience developing an online machine learning (ML) library for such devices for learning at the edge. The paper aims to discuss the difficulties in this task, both from a technology point of view and from an algorithmic one. We specifically cover the case of online clustering for the Arduino and Micropython ecosystems. We place our study in the context of Smart Buildings. It lends well to the challenges of online machine learning because of the nature and temporality of the data. Our dedicated use case exemplifies all the challenging tasks of building an online machine-learning ecosystem for lowpower IoT, from the algorithmic, benchmarking, and toolkit perspectives, including power measurements. The contribution of this paper is that we simultaneously seek to draw up an overview of the domain and prove that it is possible to realize online learning based on low-tech equipment (ESP32 board).