Prototypical Networks

Prototypical networks are a metric-based approach to few-shot learning introduced by Snell, Swersky, and Zemel in 2017. The core idea is to learn an embedding space in which each class is represented by a prototype — the mean of its support examples in that space — and classification of a query point is performed by computing its distance to the nearest prototype. No gradient descent is needed at test time; the entire learning has been compressed into the embedding function during meta-training. The elegance of prototypical networks lies in their reduction of classification to a nearest-neighbor problem in a learned metric space. They are particularly effective when the task distribution exhibits clear cluster structure, and they connect few-shot learning to classical methods like k-means and kernel density estimation. The limitation is their assumption that class distributions are spherical and unimodal in the embedding space — an assumption that fails for complex hierarchical or multi-modal concepts.