Zero-shot learning (ZSL)

Zero-shot learning is a machine learning approach in which a model can recognize or classify objects, categories, or data it has never seen during training. Unlike traditional supervised learning, which requires labeled datasets for every class, ZSL allows models to generalize from known classes to unseen ones by leveraging shared semantic information.

Key components of zero-shot learning include

• attribute-based methods, which describe classes using human-interpretable attributes, like color, shape, or size;
• embedding-based methods, which map data and class descriptions into a shared semantic space where similarity can be measured; and
• knowledge transfer from pre-trained models, which reuses learned representations to reduce the need for labeled data.

So, in practice, modern ZSL systems often rely on pre-trained foundation models—such as large language models (LLMs) or vision language models (VLMs)—that already understand many concepts and relationships, allowing them to recognize new categories without additional task-specific training.