Few-shot learning (FSL) can be considered as a kind of meta-learning problem where the model learns how to learn to solve different problems.
FSL tasks are referred to as N-way K-shots, where N corresponds to the number of examples in each training classes and K is the number of separate training tasks for the model meta-training. A test time, the model will only see N examples of each of the classes it has to learn. One specific case if one-shot learning, where a model only sees a single image before being tested.
A typical task for few-shot learning is image classification on multiple sets of classes. For example, the mini-ImageNet Dataset (Vinyals et al. 2016) is an image dataset with 64 meta-training, 16 meta-validation and 20 meta-testing classes. To evaluate a FSL model on mini-ImageNet, one trains the model on the meta-training classes and tests on the meta-testing ones.
- Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, Daan Wierstra. . "Matching Networks for One Shot Learning". In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, edited by Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, 3630–38. https://proceedings.neurips.cc/paper/2016/hash/90e1357833654983612fb05e3ec9148c-Abstract.html.