With modern technologies moving towards the internet of things where seemingly every financial, private, commercial and medical transaction being carried out by portable and intelligent devices; Machine Learning has found its way into every smart device and application possible. However, Machine Learning cannot be used on the edge directly due to the limited capabilities of small and battery-powered modules. Therefore, this thesis aims to provide light-weight automated Machine Learning models which are applied on a standard edge device, the Raspberry Pi, where one framework aims to limit parameter tuning while automating feature extraction and a second which can perform Machine Learning classification on the edge traditionally, and can be used additionally for image-based explainable Artificial Intelligence. Also, a commercial Artificial Intelligence software have been ported to work in a client/server setups on the Raspberry Pi board where it was incorporated in all of the Machine Learning frameworks which will be presented in this thesis. This dissertation also introduces multiple algorithms that can convert images into Time-series for classification and explainability but also introduces novel Time-series feature extraction algorithms that are applied to biomedical data while introducing the concept of the Activation Engine, which is a post-processing block that tunes Neural Networks without the need of particular experience in Machine Leaning. Also, a tree-based method for multiclass classification has been introduced which outperforms the One-to-Many approach while being less complex that the One-to-One method.\par The results presented in this thesis exhibit high accuracy when compared with the literature, while remaining efficient in terms of power consumption and the time of inference. Additionally the concepts, methods or algorithms that were introduced are particularly novel technically, where they include: • Feature extraction of professionally annotated, and poorly annotated time-series. • The introduction of the Activation Engine post-processing block. • A model for global image explainability with inference on the edge. • A tree-based algorithm for multiclass classification.
Artificial Intelligence for the Edge Computing Paradigm.
DAHER, ALI
2023-03-17
Abstract
With modern technologies moving towards the internet of things where seemingly every financial, private, commercial and medical transaction being carried out by portable and intelligent devices; Machine Learning has found its way into every smart device and application possible. However, Machine Learning cannot be used on the edge directly due to the limited capabilities of small and battery-powered modules. Therefore, this thesis aims to provide light-weight automated Machine Learning models which are applied on a standard edge device, the Raspberry Pi, where one framework aims to limit parameter tuning while automating feature extraction and a second which can perform Machine Learning classification on the edge traditionally, and can be used additionally for image-based explainable Artificial Intelligence. Also, a commercial Artificial Intelligence software have been ported to work in a client/server setups on the Raspberry Pi board where it was incorporated in all of the Machine Learning frameworks which will be presented in this thesis. This dissertation also introduces multiple algorithms that can convert images into Time-series for classification and explainability but also introduces novel Time-series feature extraction algorithms that are applied to biomedical data while introducing the concept of the Activation Engine, which is a post-processing block that tunes Neural Networks without the need of particular experience in Machine Leaning. Also, a tree-based method for multiclass classification has been introduced which outperforms the One-to-Many approach while being less complex that the One-to-One method.\par The results presented in this thesis exhibit high accuracy when compared with the literature, while remaining efficient in terms of power consumption and the time of inference. Additionally the concepts, methods or algorithms that were introduced are particularly novel technically, where they include: • Feature extraction of professionally annotated, and poorly annotated time-series. • The introduction of the Activation Engine post-processing block. • A model for global image explainability with inference on the edge. • A tree-based algorithm for multiclass classification.File | Dimensione | Formato | |
---|---|---|---|
phdunige_4693284.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Dimensione
3.24 MB
Formato
Adobe PDF
|
3.24 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.