Matching Networks: Attention-Weighted Few-Shot Classification

Episode

Ways (classes) 4

Shots per class 3

Drag the gold query marker across the embedding plane. Every support point pulls on it — the pull each one exerts is its attention weight. Drag empty space to pan, scroll to zoom.

Attention

Sharpness β 1.0

Prediction

Class A

confidence 0%

correct

Attn. entropy

0.00 bits

True class

How it works

Matching Networks classify a query x̂ without any gradient update by attending over every labelled support example (xᵢ, yᵢ) at once:

ŷ = Σᵢ a(x̂, xᵢ) · yᵢ
a(x̂, xᵢ) = softmax over i of −β·‖f(x̂) − g(xᵢ)‖²

f and g are learned embedding functions (in the real model, a CNN or an LSTM that reads the whole support set for context). Here the plane you see is the embedding space directly, so the attention kernel is a Gaussian over Euclidean distance — geometrically the same idea as the cosine-similarity attention Matching Networks learn, just laid out flat where you can watch it.

Sharpness β — scales the exponent. Low β spreads attention across many support points (soft, blurry vote); high β collapses it onto the single nearest neighbour (hard 1-NN).
Ways / Shots — N-way K-shot episode size: N classes, K labelled examples each, drawn fresh with "New Episode".
Attention entropy — H = −Σᵢ aᵢ log₂ aᵢ. Near 0 when one support point dominates the vote; near log₂(N·K) when the vote is split evenly.

Drag the query away from its home cluster toward a rival class and watch the predicted label flip once the rival's support points out-vote the true class — that flip traces the model's decision boundary directly.