University of Thessaly (UTH)
Deep learning and its applications in wireless big data networks performance
Abstract
dc:descriptionThe success of Internet of Things (IoT) has led to a shift in data processing towards resource-limited, typically wireless, edge devices. This thesis examines the way cooperative neural network training among the aforementioned devices can exist, focusing not only on maintaining data autonomy, but also data security, without completely depleting devices’ resources. To achieve this goal, methods for reducing neural network parameters and training time are presented, such as dynamic pruning during the training phase, based on criteria inspired by network theory (e.g. scale free networks). Moreover, we introduce algorithms leveraging pre-trained neural network weights (transfer learning), accelerating convergence and improving model performance. Although power consumption is reduced due to decreased training time and memory requirements, the amount of data available on such a device is typically insufficient for the model to converge without encountering the phenomenon of overfitting. Even in scenarios where prior knowledge is utilized, the model may still train ineffectively because of heterogeneity between prior and newly acquired knowledge. Consequently, Decentralized (Distributed) Federated Learning method gained immense popularity. In this approach, devices exchange parameters rather than raw data, to train a shared global model collaboratively. By This way, the workload is distributed across devices, each training locally on its raw data and periodically sending updated weights for synchronization into aggregated weights. To prevent flooding or energy depletion—issues that can arise from all-to-all communication protocols—new methods are introduced to limit message exchange between nodes with similar data because of geographic proximity. However, in order to enable such communication, especially in wireless, resource-constrained environments (e.g., IoT), it is crucial to establish a suitable infrastructure (e.g., the minimal number of nodes capable of transferring all information across the network), enabling the adoption of the aforementioned learning strategy. Thus, new algorithms are introduced for constructing multi-layer backbone networks, along with new centrality measures for nodes, primarily aiming to identify the most "popular" nodes that lead to network formation using the fewest possible nodes (connected dominating sets - CDS).
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Fragkou, Evangelia
- Φράγκου, Ευαγγελία
Subjects
dc:subject × 18- Ανατροφοδοτούμενα νευρωνικά δίκτυα
- Ομόσπονδη κατανεμημένη μάθηση
- Επιτάχυνση εκπαίδευσης νευρωνικών δικτύων
- Δυναμική αραιοποίηση τοπολογίας
- Μεταφορά γνώσης
- Πολύ-επίπεδα κατανεμημένα δίκτυα κορμού
- Feed-forward neural networks
- Federated deep learning
- Neural network training acceleration
- Dynamic pruning
- Transfer learning
- Multi-layer distributed backbones
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και Τεχνολογίες, άλλοι τομείς
- Other Engineering and Technologies
- Engineering and Technology
- Engineering and Technologies, miscellaneous
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59113