The evolution of Machine Learning (ML) has transitioned from centralized, cloud-based models to leveraging the entire Cloud-Edge continuum, enabling scalable and efficient learning in pervasive computing environments. These environments are characterized by heterogeneous devices streaming inherently temporal data, operating within dynamic, constrained hardware and network resources, as well as privacy constraints. In this thesis, we provide methodological and technological contributions to enable an effective and efficient use of ML in this setting. We propose the first methodological framework for continuously and efficiently learning from spatially distributed sources of temporal data by design. Our proposal matches the efficiency requirement basing our framework on Echo State Networks (ESNs) to learn from temporal data, proving its effectiveness in terms of efficiency-efficacy trade-off on on-board anomaly detection tasks within aerospace applications. We enhance its compliance to human-centric, pervasive applications by proposing an adaptation mechanism fostering fairness of the model. We tackle the spatial dimension with Federated Learning by equipping our framework with a communication-efficient, single-round learning mechanism, specifically tailored for ESNs and suitable for diverse optimization schemes. Ultimately, we make the framework suitable for learning continuously via Continual Learning, by integrating replay mechanisms to adapt the federation of models from continuous streams of non-stationary data. From the technological standpoint, we propose a framework for easily developing and deploying to production intelligent applications in real-world settings. Overall, we aim to provide a new perspective on ML as a paradigm that must account for the conditions of the environments where it operates.
Learning in Pervasive Environments
DE CARO, VALERIO
2024
Abstract
The evolution of Machine Learning (ML) has transitioned from centralized, cloud-based models to leveraging the entire Cloud-Edge continuum, enabling scalable and efficient learning in pervasive computing environments. These environments are characterized by heterogeneous devices streaming inherently temporal data, operating within dynamic, constrained hardware and network resources, as well as privacy constraints. In this thesis, we provide methodological and technological contributions to enable an effective and efficient use of ML in this setting. We propose the first methodological framework for continuously and efficiently learning from spatially distributed sources of temporal data by design. Our proposal matches the efficiency requirement basing our framework on Echo State Networks (ESNs) to learn from temporal data, proving its effectiveness in terms of efficiency-efficacy trade-off on on-board anomaly detection tasks within aerospace applications. We enhance its compliance to human-centric, pervasive applications by proposing an adaptation mechanism fostering fairness of the model. We tackle the spatial dimension with Federated Learning by equipping our framework with a communication-efficient, single-round learning mechanism, specifically tailored for ESNs and suitable for diverse optimization schemes. Ultimately, we make the framework suitable for learning continuously via Continual Learning, by integrating replay mechanisms to adapt the federation of models from continuous streams of non-stationary data. From the technological standpoint, we propose a framework for easily developing and deploying to production intelligent applications in real-world settings. Overall, we aim to provide a new perspective on ML as a paradigm that must account for the conditions of the environments where it operates.File | Dimensione | Formato | |
---|---|---|---|
phd_thesis_DeCaro.pdf
accesso aperto
Dimensione
11.39 MB
Formato
Adobe PDF
|
11.39 MB | Adobe PDF | Visualizza/Apri |
relazione_phd_DeCaro.pdf
non disponibili
Dimensione
66.43 kB
Formato
Adobe PDF
|
66.43 kB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/216212
URN:NBN:IT:UNIPI-216212