The proliferation of Internet of Things (IoT) devices has intensified the demand for energy efficient solutions supporting on-device and distributed learning applications. This research presents a circumscribed comparative analysis of radio-enabled ultra-low power IoT devices, specifically focusing on their suitability for computation heavy use cases. Our analysis centers on middle-end IoT devices that serve as a vital interface between the Electronics and Machine Learning communities. The evaluation encompasses a diverse range of IoT hardware equipped with integrated radios. We established functional datasheet-based criteria completed with accessibility and community-wise criteria to study and offer valuable insights into each selected node's performance tradeoffs, strengths, and weaknesses. This study provides crucial guidance for TinyML practitioners seeking to make informed device selections for their applications.

The integration of Artificial Intelligence (AI) and Ambient Intelligence (AmI) has emerged as a promising approach to creating responsive and contextually aware environments. AmI creates contextually aware environments by seamlessly integrating intelligent technologies, while AI develops algorithms for autonomous learning and decision-making. However, embedding AI within AmI environments faces challenges due to limited resources and energy constraints. While recent research on embedded AI has primarily focused on specific tasks of AmI, our goal is to develop a comprehensive framework encompassing all the necessary components for practical use cases. Through this endeavor, we aim to explore power-aware designs and distributed learning as fundamental approaches to address limited computational resources, energy constraints, and dynamic context variations challenges.

Integrating Artificial Intelligence (AI) into embedded systems is critical for the development of viable Ambient Intelligence (AmI). However, the energy requirements of current AI computations are not compatible with the limited resources of AmI devices and global sustainability goals. To address this issue, we propose using Federated Learning across multiple intermittent embedded systems to distribute data gathering, storage, and load balancing, thus enabling the system to assess and adapt appropriately to the environment. We initially focus on relevant real-world data acquisition to check the minimum data required by a local node in a federation. Our preliminary results demonstrate a direct correlation between signal decimation, a decrease in both training time and energy usage, and a collapsing threshold limit in accuracy.

The increase in autonomy of ambient intelligence devices has led to the evolution of Internet of Things (IoT) towards the notion of Edge Computing and is driving Wireless Sensor Networks (WSN) (considered as the Low End Devices of IoT) towards total energy autonomy by recovering ambient energy. While the majority of current Machine Learning applications are still centralized in Clouds, we could take advantage of those frugal developments to propose systems that take into account their hardware and energy limitations for collaborative and adaptive distributed learning.