Techniques and Applications for Efficient Low-power TinyML

Tiny machine learning (tinyML) is becoming popular in Internet of Things (IoT) systems to add intelligence to end nodes with limited resources. Nowadays, there are tens of billions of connected devices that exchange data wirelessly, making dense IoT networks. However, the data involved in interconnected devices is increasing their memory footprint, making the transmission of raw data over low-power wide-area networks a challenging and expensive step. TinyML implements in-situ data processing, ensuring the efficiency and reliability of IoT systems without overloading communication channels. Developing tinyML systems is complex because it involves the implementation of the traditional ML algorithm and then its optimization and compression to ensure successful deployment in resource-constrained devices. However, ML algorithms range from simple systems (e.g., 1-layer feed-forward neural networks), to the most complex ones, such as deep neural networks (DNNs) with tens of hidden layers and millions of parameters. The optimization of DNNs is an active research area as it presents many challenges to deploy them in IoT end-nodes efficiently. This dissertation presents a general framework aiming at developing tinyML systems. It investigates different ML algorithms and embedded platforms to validate the correct operation of the proposed framework. Furthermore, it selects different use cases to motivate and demonstrate the effectiveness of the proposed solution for developing tinyML algorithms in IoT systems. The use cases consist of real-world applications providing actual techniques and methods to implement tinyML algorithms in constrained devices successfully. Furthermore, this thesis provides clear evidence of the benefits of tinyML considering energy efficiency, reliability, and maintenance. Finally, it improves the capability of standard tinyML systems with on-device learning techniques. In this way, it is possible to obtain tinyML systems which follow the trend of the environment, learning new patterns and reducing maintenance operations.