14 prototypes × 20 noisy samples — same generator as the dropout version, but with dropout_prob = 0. Every sample has all 5 pools intact, so there's no missing-modality handling to obscure the result.
Species: pine, oak, rose, daisy, robin, canary, sunfish, salmon Kingdom generics: tree, flower, bird, fish (color “dulled”) Super-generics: plant (above tree+flower), animal (above bird+fish) — damped on every pool their children disagree on Pools: Color, Locomotion, Covering, BodyParts, Habitat (Locomotion + Habitat are simplex, sum to 1) Source: generated by scripts/generate_synthetic_12.js (no dropout variant). Downloads:samples + prototypes (.xlsx)
Hyperparameters
Training
Episodes
Select Episode
Learned Hubs
Probe Mode
Click input feature nodes to set partial values. The network will activate hubs by similarity and reconstruct the full feature pattern from the best-matching hub.