Near Sensor Artificial Intelligence on IoT Devices for Smart Cities

The IoT is in a continuous evolution thanks to new technologies that open the doors to various applications. While the structure of the IoT network remains the same over the years, specifically composed of a server, gateways, and nodes, their tasks change according to new challenges: the use of multimedia information and the large amount of data created by millions of devices forces the system to move from the cloud-centric approach to the thing-centric approach, where nodes partially process the information. Computing at the sensor node level solves well-known problems like scalability and privacy concerns. However, this study’s primary focus is on the impact that bringing the computation at the edge has on energy: continuous transmission of multimedia data drains the battery, and processing information on the node reduces the amount of data transferred to event-based alerts. Nevertheless, most of the foundational services for IoT applications are provided by AI. Due to this class of algorithms’ complexity, they are always delegated to GPUs or devices with an energy budget that is orders of magnitude more than an IoT node, which should be energy-neutral and powered only by energy harvesters. Enabling AI on IoT nodes is a challenging task. From the software side, this work explores the most recent compression techniques for NN, enabling the reduction of state-of-the-art networks to make them fit in microcontroller systems. From the hardware side, this thesis focuses on hardware selection. It compares the AI algorithms’ efficiency running on both well-established microcontrollers and state-of-the-art processors. An additional contribution towards energy-efficient AI is the exploration of hardware for acquisition and pre-processing of sound data, analyzing the data’s quality for further classification. Moreover, the combination of software and hardware co-design is the key point of this thesis to bring AI to the very edge of the IoT network.