What if an AI model could correctly label something it has never seen before? That is the promise of Zero Shot Learning (ZSL), which lets systems recognize new classes beyond their original training data by reasoning from descriptions instead of examples.

ZSL moves away from the idea that every class needs labelled examples. Instead, it uses auxiliary information such as text descriptions, semantic attributes, or relationships between classes. The model learns how objects connect to the terms used to describe them, so it can infer what an unseen class should look like from its description alone. With rich metadata and clear semantic attributes, ZSL can generalize from “seen” to “unseen” categories.

ZSL is only as strong as the quality of its descriptions, and weak or vague metadata can seriously hurt performance. Measuring how well it generalizes is also tricky, often needing metrics that consider both seen and unseen classes. Even so, ZSL marks an important shift toward more generalized AI that does not rely entirely on exhaustive labelling, especially when paired with few-shot learning to boost accuracy and flexibility.