Semantic Network Models

Interactive visualizations of classic and modern connectionist models for concept learning, semantic memory, and unsupervised category formation.

Supervised

IAC Visualizer

Interactive Activation and Competition network — localist semantic memory with bidirectional pool inhibition.

Rumelhart Model

Backprop-trained concept-property semantic memory learning distributed representations.

Unsupervised — Single ART

Concept Learning (Default)

Vigilance-based clustering on a preset dataset of fruits, blocks, rocks, stars, etc.

Concept Learning (CSV)

Same model, but paste your own CSV data — supports categorical or numeric pool-prefixed columns.

Hierarchy Emergence ART New

Flat ART with noisy samples + random sensory dropout. A general "fish" hub and species-level hubs emerge without forcing any hierarchy.

Hierarchy Emergence (Synthetic 12, Clean) New

Generated dataset of 8 species + 4 generic superordinates (tree, flower, bird, fish), no dropout. The cleanest demonstration that flat ART can recover all 12 prototypes.

Hierarchy Emergence (Synthetic 12, Dropout)

Same dataset but with random sensory dropout — pools randomly zeroed out per sample. Demonstrates the model's robustness to missing modalities.

Hierarchy Emergence (Corrupted + Decay) New

14 prototypes with 2 deliberately corrupted samples per concept mixed in. Templates decay every training step — orphan hubs recruited by garbage samples wither away while the 14 real hubs stay refreshed.

Unsupervised — Hierarchical

Forced Hierarchical ART New

Three parallel ART layers at increasing vigilance — coarse super-categories, mid-level categories, and fine-grained specific concepts. Hierarchy imposed by design.