Few shot learning : an overview of methods, applications and benchmarks
Μάθηση με λίγα παραδείγματα : μια επισκόπηση μεθόδων, εφαρμογών και δοκιμαστικών συνόλων

Master Thesis
Author
Foukanelis, Christos Georgios
Φουκανέλης, Χρήστος Γεώργιος
Date
2025-11View/ Open
Keywords
Few shot learning ; Deep learning ; Neural networks ; AI ; Artificial Intelligence ; Meta learning ; Computer vision ; Signal processing ; Survey ; Τεχνητή νοημοσύνη ; Machine learning ; Green AI ; Efficient AI ; NLP ; Prototypical networks ; Transfer learning ; Domain adaptationAbstract
Few-Shot Learning (FSL) is an emerging research trend in the field of neural networks, as it enables algorithms and models to rapidly adapt to previously unseen data using only a limited number of examples. This efficiency is crucial, serving as a key enabler for Green AI and edge device computing within the Internet of Things, where algorithms must operate under limited computational and data resources. This thesis examines the fundamental principles of FSL and its early experimental developments, which were conducted primarily on computer vision tasks, before extending to the audio and natural language processing (NLP) domains and their most recent advancements. The architectures, methods, and datasets relevant to these domains are explained in detail. Furthermore, the applications of the explored methods are analyzed within each domain, illustrating their use in specific problems such as object detection. An implementation report-replicating the aforementioned experiments- presents experiments conducted on audio and image datasets using the most prominent few-shot learning architectures. Finally, experimental results from various works and benchmarks are presented, along with an exploration of the main challenges and potential future research directions in the field.


