Exploring deep learning for data-efficient and generalizable visual recognition

Abstract

This PhD research explores how artificial intelligence (AI) systems learn to recognize new image categories from a limited number of labelled samples. Unlike conventional image classification methods that rely on large training datasets, few-shot learning aims to enable models to learn effectively from scarce data and generalize to previously unseen categories.

The research investigates few-shot learning along three main directions: metric learning, feature enrichment, and knowledge transfer. Through a series of models and experiments, the work explores how AI can better distinguish visually similar categories, build richer and more informative representations from limited data, and make effective use of previously acquired knowledge. Four proposed methods—ProtoSimNet, AFPN, AMFF, and DSKT—demonstrate how these challenges can be addressed in complementary ways.

Overall, this research contributes to a better understanding of how image recognition systems can remain effective when labelled data are limited. By addressing limitations in similarity modelling, feature representation, and knowledge transfer, the work provides complementary strategies for improving learning under limited supervision. These findings contribute to the development of more data-efficient and adaptable AI systems, and may help extend image recognition to applications where collecting large-scale labelled datasets is difficult, costly, or impractical.