ΠΡΟΣΚΛΗΣΗ ΣΕ ΔΗΜΟΣΙΑ ΠΑΡΟΥΣΙΑΣΗ ΜΕΤΑΠΤΥΧΙΑΚΗΣ ΔΙΠΛΩΜΑΤΙΚΗΣ ΕΡΓΑΣΙΑΣ

ΕΛΛΗΝΙΚΟ ΜΕΣΟΓΕΙΑΚΟ ΠΑΝΕΠΙΣΤΗΜΙΟ
ΤΜΗΜΑ ΗΛΕΚΤΡΟΛΟΓΩΝ ΜΗΧΑΝΙΚΩΝ ΚΑΙ ΜΗΧΑΝΙΚΩΝ ΥΠΟΛΟΓΙΣΤΩΝ

ΠΡΟΣΚΛΗΣΗ ΣΕ ΔΗΜΟΣΙΑ ΠΑΡΟΥΣΙΑΣΗ
ΜΕΤΑΠΤΥΧΙΑΚΗΣ ΔΙΠΛΩΜΑΤΙΚΗΣ ΕΡΓΑΣΙΑΣ

Exploitting compressed sensing to distributed machine learning του Σταματάκη Εμμανουήλ, μεταπτυχιακού φοιτητή του Προγράμματος Μεταπτυχιακών Σπουδών “Ενεργειακά Συστήματα” Επιβλέπων Καθηγητής: Σπυρίδων Παναγιωτάκης

H παρουσίαση θα πραγματοποιηθεί την Τετάρτη 6 Απριλίου 2022 στις 12:30π.μ https://vdc.hmu.gr/b/spy-zdo-all-zry.

ΠΕΡΙΛΗΨΗ ΤΗΣ ΕΡΓΑΣΙΑΣ
A range of contemporary applications, such as remote monitoring of crucial measurements, mechanical fault recognition, remote detection of structural strain in constructions, and many others, have become possible today thanks to the Internet of Things (IoT). Due to the fog-based design of these systems, devices are placed at the extreme edge, requiring data transmission to a central node. At the same time, the performance of these devices is limited both by power requirements for wireless transmission via the network and by their limited computational capabilities and features.
Ideally, there would be a way to compress the data to a great extent so that, upon receipt by the receiving node, it could be accurately reconstructed. This need is addressed by a new technique called compressive sensing. Its operation is based on the sparsity characteristic of most natural signals when represented in a specific basis. This technique allow reconstruction with far fewer points than traditional sampling techniques, such as Nyquist.
The intersection of the compressive sensing (CS) and machine learning (ML) domains has garnered significant research interest, combining the fundamental principles from both areas, and is the focus of the present master’s thesis. Through this study, the use of CS as a compression tool before transmission is explored, and the limits of its application in distributed ML systems are determined. More specifically, the impact of CS on the recognition capability of reconstructed signals by a trained ML model is examined. The study aims to achieve a high recognition rate for reconstructed signals at the network’s edge and extreme edge.

Μετάβαση στο περιεχόμενο