Middleware-enabled frugality for intelligent and distributed edge applications

In the ever-evolving landscape of computing, the integration of Artificial Intelligence has become pervasive, empowering applications with unprecedented capabilities. However, as the scope of AI expands, so do the challenges associated with resource constraints, particularly in the realm of intelligent and distributed Edge Applications. This thesis delves into the intersection of two concepts—Frugality and Serverless paradigms—aiming to revolutionize the deployment and functionality of intelligent agents on constrained devices. Distributed Computing in all its facets is discussed, addressing the main challenges related to device heterogeneity and seamless networking management. Experimental contributions in edge and industrial scenarios are presented, confirming the practical application of digital twin technology. In addition, a smart routing approach is novelly presented, showcasing the beneficial application of AI on Distributed Computing tasks. The exploration extends to intelligent and Distributed Edge Applications, where the fusion of AI and decentralized computing offers unlimited potential for swift responsiveness and restrained energy consumption. Understanding the importance of power efficiency in constrained environments, this thesis places a particular emphasis on the trade-off between performance and power consumption. Greener and more sustainable approaches to Distributed Learning are proposed, taking into account power consumption in Federated Learning round planning strategies. Then, we address the challenges related to data constrained scenarios. Privacy and Trust related issues are addressed in a Federated Learning setting proposing a novel contribution. Additionally, considering the scarcity of data and model complexity, we introduce an ensembling of weak autoregressor as a striking solution for traffic volume forecasting. During the exploration at the intersection of AI, frugality, and Serverless Computing, we then focus our attention on those cases where networking conditions are unstable and unreliable. Following the Split Computing paradigm we present a distilled encoder capable of challenging the performance of deeper neural networks, enabling realtime semantic compression on mobile devices.