Prototypical Networks: Few-Shot Classification by Distance

Episode Setup

N-way (classes) 4

K-shot (support / class) 3

Cluster spread σ 1.00

Show decision regions

Live Readouts

Query accuracy

—

Queries correct

—

Support points

—

Avg dist → proto

—

Classes

How it works

Prototypical Networks (Snell, Swersky & Zemel, 2017) solve N-way K-shot classification by mapping every example into an embedding space f_θ(x) and representing each class by a single point — its prototype: the mean of its support embeddings.

c_k = (1/|S_k|) · Σ_(x_i,y_i)∈S_k  f_θ(x_i)

p(y=k | x) = softmax_k( −d(f_θ(x), c_k) )
           = exp(−d(f_θ(x),c_k)) / Σ_k' exp(−d(f_θ(x),c_k'))

ŷ = argmin_k  d(f_θ(x), c_k)      (squared Euclidean d)

The small filled spheres are the support set — the few labelled examples per class used to compute each prototype (the ◆ diamond markers).
The rings are the query set — unlabelled points classified purely by nearest-prototype distance. A green ring means the prediction matched the true class; red means it didn't.
The tinted floor is the induced Voronoi decision map: every location's fill colour is whichever prototype is closest, exactly the boundary a real embedding-space classifier would draw.
This is the same mechanism used for real few-shot image/audio classification — only the embedding function f_θ (normally a trained CNN) is replaced here with a directly-visualised 2D embedding so the geometry is legible.

Increasing K averages out noise and pulls prototypes toward the true cluster centre, which is why Prototypical Networks scale gracefully from 1-shot to many-shot regimes with no change to the algorithm itself.

● support example ○ query example (ring = correct/wrong) class prototype

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.