Zero-shot learning is a machine learning paradigm in which a model is able to recognize or classify instances of classes that it has never seen in training, without any direct labeled examples of those unseen classes​.

This is typically achieved by leveraging auxiliary information that links seen classes to unseen ones, such as semantic descriptors, attributes, or embeddings (for example, word embeddings of class names or textual descriptions of the classes). During training, the model might learn to map inputs to a semantic space of attributes or descriptions. At test time, given a new class, the model uses the learned semantic relationships to infer the correct class label even with zero training examples of that class​.

For instance, if a zero-shot model knows about horses and stripes, it could identify a zebra by understanding that “zebra” shares attributes with known classes (like horse plus stripe pattern) despite never having seen a zebra image. Zero-shot learning expands a model’s scope to handle new categories on the fly and is valuable in scenarios where collecting training data for every possible class is impractical.