Zero-shot learning

Zero-shot learning (ZSL) allows an AI model to classify entirely new, unseen data categories without explicit training, leveraging auxiliary information.

ZSL is a critical machine learning paradigm: it enables models to generalize to novel classes (e.g., classifying a 'zebra') without a single labeled training example for that class. The core mechanism involves mapping both seen and unseen classes into a shared semantic space, using auxiliary data like textual descriptions or attribute vectors. This knowledge transfer bypasses the need for massive, costly labeled datasets, which is a major bottleneck in traditional supervised learning. Key models like OpenAI's CLIP and Large Language Models (LLMs) such as GPT-4 utilize ZSL extensively for tasks ranging from image recognition to advanced Natural Language Processing. The most challenging variant, Generalized Zero-Shot Learning (GZSL), requires the model to correctly identify samples from both known and completely unseen classes at test time.

https://en.wikipedia.org/wiki/Zero-shot_learning