🧲 3D Matching Networks: Attention-Weighted Few-Shot Classification, Softmax Temperature and the Nearest-Neighbour Limit
A Blender-built stage where a support set of labelled tokens answers a query beacon through attention rods whose thickness is the softmax weight: sweep the temperature from the nearest-neighbour limit to uniform voting, change way, shot and noise, and compare episodic accuracy with the Bayes-optimal classifier.
⚙ Under the hood
A Blender-built stage where a support set of labelled tokens answers a query beacon through attention rods whose thickness is the softmax weight: sweep the temperature from the nearest-neighbour limit to uniform voting, change way, shot and noise, and compare episodic accuracy with the Bayes-optimal classifier.
Matching NetworksFew-Shot LearningAttentionSoftmax TemperatureNearest NeighbourEpisodic TrainingMetric LearningBayes Optimal Classifier
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install