Deep and Wide Tiny Machine Learning

Abstract

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