FEW-SHOT AND ZERO-SHOT LEARNING: TEACHING MODELS TO LEARN FROM VERY FEW EXAMPLES
Despite the remarkable progress achieved by deep learning over the past decade, the majority of high-performing models continue to depend heavily on large quantities of high-quality labeled data. This reliance poses a substantial barrier in many real-world applications where obtaining labeled examples is costly, time-consuming, or practically infeasible. Domains such as medical imaging, specialized industrial inspection, rare disease diagnosis, and low-resource languages often lack sufficient annotated datasets to train conventional supervised models effectively. Few-shot learning focuses on scenarios where only a handful of labeled examples-sometimes only one-are available for each target class. Zero-shot learning, by contrast, addresses an even more challenging setting in which models must recognize categories for which no labeled examples are provided during training. Instead of direct supervision, ZSL systems rely on auxiliary semantic information-such as natural language descriptions, attribute vectors, or large language models-to bridge the gap between seen and unseen classes. This paper provides a comprehensive comparative study of modern few-shot and zero-shot learning techniques, focusing on their conceptual foundations, methodological distinctions, and practical applications.
Keywords: Few-Shot Learning, Zero-Shot Learning, Meta-Learning, Prototypical Networks, Prompt-Based Learning, Transfer Learning, Low-Data Regime, Deep Learning.
Views (17)



