In this section, we'll be looking at one-shot learning (k = 1) and its generalization few-shot learning (k > 1). In this case, the support set is not empty and we have one or more labeled samples of each class. This is an advantage over the zero-shot scenario because we can rely on labeled samples from the same domain instead of using a mapping from the labeled samples of another domain. Therefore, we have a single encoder f and no need for additional mapping.
An example of a one-shot learning task is a company's facial recognition system. This system should be able to recognize the identity of an employee based on a single photo. It should be possible to add new employees with a single photo as well. Let's note that in ...
