How Q-PNA Works
Q-PNA replaces continuous embedding spaces (ℝⁿ) with Bruhat-Tits tree geometry. Features are mapped to tree leaves via a learned linear transformation. Classification happens by walking the tree — every decision is traceable by construction.
This explorer demonstrates the architecture: a 4-ary tree with 4 class regions. Click Classify to simulate a feature vector being mapped to a leaf, then trace the decision path from root to leaf. Green = correct class. Yellow = wrong class but traceable.
Controls
--
Correct/Total
--
Aggregate Accuracy
--
Tree Depth Traversed
--
Accuracy Level
Decision Trace
Click Classify or click a leaf in the tree to trace the decision path from root to leaf.
Bruhat-Tits Decision Tree
Yellow regions = class boundaries. Click any leaf to trace its path. Green highlight = selected leaf.