Abstract
dc:descriptionThis Ph.D. thesis was conducted in the Intelligent Systems Laboratory of the Department of Computer Science and Biomedical Informatics at the University of Thessaly. The focus of the thesis is to investigate the new field of Tiny Machine Learning, also known as TinyML. Machine Learning is an expanding field that is actively researched in both academic and industrial settings. The technology is employed in diverse domains, such as Internet of Things and Data Science. Moreover, Deep Learning, a subset of Machine Learning, is a highly utilized and prominent technology. The feasibility of this has been achieved by hardware breakthroughs and the capacity to extract vast quantities of data. The basic structure of Deep Learning closely mirrors the arrangement of neural networks found in the human brain. Researchers have redirected their focus towards compression approaches and optimization strategies to improve the overall efficiency of Deep Learning networks without sacrificing precision and performance due to the substantial computing load associated with these networks. The extensive adoption of both Machine Learning and Deep Learning techniques across several sectors enables the extraction of crucial insights from large volumes of data and the retrieval of vital information. In a typical IoT framework, data is gathered, interpreted, analyzed, and stored to monitor variables and forecast or anticipate events. Networked devices, such as sensors and actuators, generate data which is subsequently sent to the cloud for analysis and storage. IoT gateways are employed to establish a connection between the network of sensors and the cloud network. The previously stated optimization and compression approaches enabled the combination of embedded systems and Artificial Intelligence, resulting in the emergence of a new field called Edge AI. Edge AI is a nascent technology that can effectively tackle the difficulties faced by IoT devices. The gadgets in issue now have the capability to gather, analyze, and process data on the device itself in real-time, eliminating the need to send data to external entities. The potential to transform the architecture of smart applications and services lies in enabling machine intelligence at the network's edge, where data is generated. This can lead to the development of apps that possess enhanced safety, heightened privacy, and are unaffected by issues related to latency and bandwidth restrictions. These applications can operate on cost-effective and energy-conserving hardware, providing advantages to diverse sectors including healthcare, smart homes, agriculture, and automotive. TinyML is a fusion of hardware and software designed to enable the integration of Machine Learning models and neural networks on resource-constrained hardware, allowing for local processing. One significant advantage of the TinyML method is its capability to do real-time analysis of collected data, even in scenarios with constraints in latency and bandwidth. Furthermore, Internet of Things devices that integrate TinyML on their Microcontroller Units necessitate reduced dependence on cloud services. This results in cost and power efficiency, as well as enhanced security and privacy of data. The aforementioned benefits have led to the development of multiple applications utilizing TinyML in various industries, including healthcare, automotive, agricultural, security, and the industrial sector. Furthermore, the increasing prevalence and swift growth of TinyML led to the emergence of numerous development boards, frameworks, libraries, and other toolkits that are fully compatible with the technology under investigation. This thesis explores optimization strategies that can be used to generate models and neural networks suitable for constrained hardware. Moreover, a wide range of applications were developed across different sectors. Finally, a novel platform for profiling development boards used in TinyML technology has been developed.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsoukas, Vasileios
- Τσούκας, Βασίλειος
Subjects
dc:subject × 16- Μηχανική μάθηση
- Υλικό
- Ενσωματωμένα συστήματα
- Αλγόριθμοι τεχνητής νοημοσύνης
- Συμπίεση νευρωνικών δικτύων
- Machine learning
- Hardware
- Embedded systems
- Artificial intelligence
- Artificial intelligence optimization methods
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58714