Abstract
Traditional machine learning models depend heavily on large labeled datasets for achieving good performance. However, in many real-world situations, collecting large amounts of labeled data is expensive, time consuming, or even impossible. This limitation motivated the development of Zero-Shot Learning (ZSL) and Few-Shot Learning (FSL) techniques, which aim to recognize new classes with little or no training examples. These approaches try to mimic the human ability of learning new concepts from very limited experience. Zero-shot learning relies on auxiliary information such as semantic attributes or textual descriptions to identify unseen classes, while few-shot learning leverages a small number of examples with advanced training strategies like meta-learning and metric learning. This paper presents a detailed review of the principles, techniques, architectures, and applications of ZSL and FSL. Recent advancements using deep learning, transformers, and large language models are also discussed. Challenges, evaluation methods, and future research directions are highlighted.
Keywords: Zero-Shot Learning, Few-Shot Learning, Meta-Learning, Transfer Learning, Metric Learning, Deep Learning, Semantic Embedding
Full Issue
| View or download the full issue | PDF 98-109 |