Who is wearing me? TinyDL-based user recognition in constrained personal devices

Abstract

Deep learning (DL) techniques have been extensively studied to improve their precision and scalability in a vast range of applications. Recently, a new milestone has been reached driven by the emergence of the TinyDL paradigm, which enables adaptation of complex DL models generated by well-known libraries to the restrictions of constrained microcontroller-based devices. In this work, a comprehensive discussion is provided regarding this novel ecosystem, by identifying the benefits that it will bring to the wearable industry and analysing different TinyDL initiatives promoted by tech giants. The specific use case of automatic user recognition from data captured by a wearable device is also presented. The whole development process by which different DL configurations have been embedded in a real microcontroller unit is described. The attained results in terms of accuracy and resource usage confirm the validity of the proposal, which allows precise predictions in a highly constrained platform with limited input information. Therefore, this work provides insights into the viability of the integration of TinyDL models within wearables, which may be valuable for researchers, practitioners, and makers related to this industry.