Prototypical Networks: Few-Shot Classification by Distance

Prototypical Networks tackle few-shot learning with a strikingly simple idea: embed every example, then represent each class by the mean of its few labelled ("support") embeddings — a single prototype point. A new ("query") example is classified by which prototype it lands closest to. This simulator renders that geometry directly in 3D: adjustable N-way / K-shot controls resample a fresh episode with its own class clusters, prototypes are recomputed live as the mean of the current support set, and every query point is judged by nearest-prototype distance against a shaded Voronoi decision map that shows exactly where each class's territory begins and ends.