I am trying to understand few-shot learning. I understand the support sets and query sets and that the fundamental use of few-shot learning is to learn to classify (in computer vision classification tasks) with few labeled images. Generally, it learns the similarities and differences between images.
There are a few things I don't understand:
- What is the general train and test task paradigm in few-shot learning?
- Say I want to classify art paintings by artist (100 images per artist, 10 artists). Would it make sense to use few-shot learning for this specific problem? So, how to I know that a certain problem qualifies as for few-shot learning?
- When deploying this model in the real world, how is it used? Do you have to provide a support set along with the image you are trying to classify? What happends if the image is not part of the dataset?
I appreciate any feedback!