Few-Shot Learning

Programs that learn independently from data normally need very many examples. Anyone wanting to teach such a program to recognize cats in photos would classically show it thousands of cat pictures. Few-shot learning is the counter-model: the program is given only a handful of examples, often two to five, and afterward still manages to solve the task usably. “Shot” here is the English word for a single attempt or example case. If there is exactly one example, this is called one-shot; if there are none, it’s called zero-shot.

Why this matters

Examples are expensive. Someone has to label images, sort texts, or write out answers in advance. For many tasks, large collections of such ready-made labeled data simply don’t exist. A doctor might have thirty cases for a rare disease, not thirty thousand.

That’s why few-shot learning often decides whether an AI project is worthwhile at all. If a few examples are enough, a small company can try out in a week what used to be a year-long project. The use of today’s chat programs also works on this principle: you describe the task, show two patterns, and get suitable results.

How it works

The trick lies in prior knowledge. A large language model — that is, a program trained on huge amounts of text to predict words — has in the process learned a great deal of general structure. It knows sentence construction, factual relationships, and typical task formats. The few examples therefore don’t need to teach it anything new. They only show which of its existing abilities is currently required.

You can compare this to an experienced translator. If a client shows him three example sentences in the desired style, the translator immediately understands what matters. He isn’t learning a new language; he’s adapting to a request. A beginner, on the other hand, couldn’t do much with just three sentences.

Technically, there are two approaches. With few-shot prompting, you simply write the examples into the request. The model itself doesn’t change; it only reacts to the text supplied along with it. With fine-tuning, by contrast, the model’s internal numerical values are slightly adjusted using the few examples. The first approach is faster and costs nothing; the second lasts longer and has a stronger effect on very specific tasks.

Where you encounter the term

In everyday life, people use few-shot learning without naming it. Anyone who shows a chat program two examples of a desired summary is doing exactly that. Text templates in email programs and automatic sorting of support requests also often work according to this pattern.

The term has appeared regularly in trade news since 2020. At that time, the model GPT-3 showed that a single large system could handle many tasks with few examples, without being specifically trained for each one. This is considered a turning point for the industry. When companies today talk about their models' zero-shot or few-shot capabilities, they mean exactly this flexibility.

For investors, this point is economically relevant. The fewer examples needed, the less effort lies between a model and a sellable product. At the same time, there is a limitation: few examples mean little control. The result can tip over if the examples are poorly chosen or if a task really requires specialized knowledge that was missing during training.