What is a "shot" in machine learning?

Question

I keep on hearing this term "shot" used in machine learning.

Is a "shot" well-defined?

From what I can tell, "shot" is a synonym for "example". Most machine learning systems seem to be "multi-shot" meaning you have a huge dataset that has many different examples of different categories. However, for a system to have "one-shot" capabilities means that it is able to predict the category of something given exactly one example. Similarly, "few-shot" applications seem to only need a few examples in order to perform some function with the input. And "zero-shot" learning seems to be making predictions without any examples during training.

Is a shot just an example?

Given the evidence above, it seems like this is the case, but it also seems like it's a little bit more nuanced, something like a shot is a post-training example when 0 examples were given during training. But I'm not sure if this is right, thus the question.

Answer 1

A shot is a single example available for machine learning. So "one-shot" means you're given just the one example for each class.

Answer 2

I asked this question before starting my PhD in natural language processing, and I think I have a slightly better grasp of the term now.

When discussing prompting large language models (LLMs), a shot refers to an example given in the prompt to the model, rather than an example it was trained on. For example, 0-shot prompting is when you just ask an LLM a question; single-shot prompting is when you first show how one question should be answered and then ask a similar question and let the LLM answer it, all within one prompt.

I imagine that this terminology was borrowed from earlier training techniques, where, as the previous answer mentions, shots were merely training examples. But I do think the language has changed to differentiate between the two when talking about LLMs. For LLMs, shots are specifically given within a single prompt, so a single example may have multiple shots within it.