Back to search

Harokopio University

Adaptations of rehearsal methods for online continual learning

Abstract

dc:description

The proliferation of Internet of Things (IoT) technologies and the emergence of edge-cloud computing architectures have enabled the deployment of intelligent systems that support and enhance various aspects of daily life, including healthcare, transportation, and automation. These systems rely heavily on deep learning models to perform complex tasks; however, the training and continual updating of such models demand substantial computational resources and storage capabilities. This introduces challenges related to cost, scalability, and sustainability, particularly in dynamic and resource-constrained environments such as those encountered in IoT. A critical limitation of deep learning in this context is catastrophic forgetting—the tendency of a model to lose previously acquired knowledge when trained on new data. Continual Learning (CL) has emerged as a promising solution to this issue by enabling models to learn incrementally over time, akin to human learning. However, conventional CL approaches often assume balanced data distributions and well-defined task boundaries, which rarely hold true in real-world IoT scenarios. Online Continual Learning (OCL) addresses this by processing data streams in a single pass, making it more suitable for deployment in environments where data is non-stationary and memory is limited. Among OCL strategies, rehearsal-based methods have gained prominence due to their simplicity, adaptability, and model-agnostic nature. These methods maintain a small buffer of past data, which is periodically replayed alongside new samples to reinforce prior knowledge. This thesis focuses on optimizing state-of-the-art rehearsal techniques for IoT applications, emphasizing resource efficiency and adaptability. To further enhance the robustness and reliability of continual learning systems, this work integrates two critical mechanisms: drift detection and out-of-distribution (OoD) detection. Drift detection involves identifying shifts in the data distribution—commonly referred to as concept drift—caused by changes in user behavior, sensor degradation, or environmental variation.OoD detection, in contrast, seeks to identify inputs that diverge significantly from the training distribution. Detecting such anomalies is essential for ensuring safety and dependability in autonomous and semi-autonomous IoT systems. The objective of this research is to establish a resource-aware, continually learning machine learning framework tailored to the demands of IoT environments. By combining efficient rehearsal-based continual learning with robust mechanisms for drift and OoD detection, the proposed framework aims to facilitate dependable and adaptive model deployment across a wide range of real-world applications. Together, these components—continual learning, drift detection, and OoD detection—form a holistic, resilient, and low-resource machine learning system that is capable of operating autonomously in dynamic environments. The resulting framework not only enhances the reliability and adaptability of intelligent systems but also promotes sustainable AI deployment in the growing field of IoT, where computational efficiency and robustness are paramount.

Degree

thesis:*
Grantor dc:publisher
Harokopio University
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Davalas, Charalampos
  • Δαβαλάς, Χαράλαμπος

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60504

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Davalas, Charalampos; Δαβαλάς, Χαράλαμπος. Adaptations of rehearsal methods for online continual learning. Harokopio University, 2025. http://hdl.handle.net/10442/hedi/60504