What is zero-shot learning?

Zero-shot learning describes a model’s ability to solve a task for which it has not seen a single training example. Instead of examples, it receives a description of the target categories, for instance as a list of possible answers. With few-shot learning, by contrast, a handful of examples are provided; with classic training, many hundreds are.

This is made possible by the shared meaning space that large models build up during pre-training. An unknown category gets connected to already-learned concepts through its description, allowing the model to compute the similarity. For language models, an instruction in the prompt is enough — listing the allowed categories, each with a brief explanatory phrase. Accuracy depends on how clearly these descriptions are worded.

Zero-shot learning pays off where a new classification is needed and no labeled data exists yet. That ranges from opening up a new subject area, to short-notice special analyses, to pre-sorting a dataset that will later produce training data. For specialist terms from a narrow domain, the hit rate drops noticeably.

The advantage over training a model from scratch lies in the starting point. A new category is ready to use within minutes, since only the description needs to be added and no examples need to be annotated. Changes to the classification cost correspondingly little.

Quality varies between categories more than with a trained model. That’s why a sample of a few hundred manually checked cases is used to measure how well each individual category performs before the method goes into regular operation.