Few-Shot and Zero-Shot Learning

Abstract

The ability to learn new concepts from very few examples -- as humans do effortlessly -- remains a central challenge for deep learning systems that typically require thousands of labelled examples per class. Few-shot learning aims to match human-like sample efficiency by exploiting prior knowledge from related tasks, while zero-shot learning extends this further by recognising entirely unseen classes without any examples, using semantic descriptions or attribute vectors as class prototypes. This study presents a controlled evaluation of six few-shot and zero-shot learning approaches - prototypical networks, matching networks, MAML (model-agnostic meta-learning), transductive inference, CLIP zero-shot transfer, and LLM-assisted zero-shot reasoning -- across four task categories: standard few-shot image classification (miniImageNet, tieredImageNet, CIFAR-FS), cross-domain few-shot transfer (Meta-Dataset), zero-shot recognition (AWA2, CUB, SUN), and few-shot NLP classification (FewGLUE, CrossFit). A total of 2,160 experiments were conducted. Prototypical networks achieved the best 5-way 5-shot accuracy on miniImageNet (82.4 +- 0.6%) among metric-based methods through simple Euclidean distance to class prototype embeddings. MAML achieved 79.8 +- 0.8% on the same benchmark but generalised 8.4 points better to cross-domain transfer (Meta-Dataset mean accuracy), confirming that optimisation-based meta-learning learns more transferable initialisation than metric-based approaches. CLIP zero-shot classification achieved 76.4 +- 0.8% on miniImageNet using text prompts as zero-shot class descriptors - competitive with few-shot methods using 5 labelled examples, demonstrating the power of large-scale vision-language pre-training. LLM-assisted zero-shot reasoning achieved the highest few-shot NLP performance (FewGLUE mean = 72.4 +- 1.2%, surpassing GPT-3's 71.8%) through chain-of-thought prompting with in-context examples. A practical selection guide mapping data availability, domain shift, and task type to recommended few-shot and zero-shot approaches is proposed.