What is the difference between zero-shot learning and few-shot learning?

Zero-shot learning (ZSL) enables models to recognize or classify objects for which no labeled training examples were available during training. In contrast, few-shot learning (FSL) utilizes a very small number (e.g., 1-5) of labeled examples per new class to adapt the model.

Zero-shot learning primarily works by leveraging pre-existing semantic knowledge or auxiliary information (e.g., class attributes, textual descriptions, or embeddings) to transfer learned concepts to unseen classes. Few-shot learning focuses on rapid adaptation, often using techniques like meta-learning, metric learning, or fine-tuning, where the model learns effectively from minimal supervision. Both paradigms address data scarcity challenges beyond traditional supervised learning. ZSL relies entirely on side-information, while FSL depends critically on the quality and representativeness of its few examples.

Applications of ZSL are crucial when collecting labeled data for new classes is impossible or prohibitively expensive, such as recognizing rare species or novel object categories. FSL shines in scenarios requiring rapid deployment with minimal annotation effort, like niche product classification. Both provide significant business value by enabling AI systems to generalize efficiently to novel tasks and categories without massive retraining datasets. FSL is generally more practical where some examples can be provided.