Deep and Wide Tiny Machine Learning

Abstract In the last decades, on one hand, Deep Learning (DL) has become state of art in several domains, e.g., image classification, object detection, and natural language processing. On other pervasive technologies—Internet Things (IoT) units, embedded systems, Micro-Controller Units (MCUs)—ask for intelligent processing mechanisms as close possible to data generation. Nevertheless, memory, computational, energy requirements characterizing DL models are three or more orders magnitude larger than corresponding computation, capabilities devices. This work aims at introducing a methodology address this issue enable particular, by defining Tiny Machine (TML) solutions, i.e., machine deep learning that take into account constraints target device. The proposed addresses problem different levels. first approach, devices inference-based TML solutions approximation techniques, model runs device but was trained elsewhere. Then, introduces on-device TML. Finally, third approach develops Wide split distribute over connected heterogeneous