There are no weights on a fixed grid here — the wire IS the weight. Every candidate edge carries a conductance gate: hot edges conduct, cold ones are dead "cold joints". As the network learns to classify, edges that carry useful signal thicken & heat; unused wires cool and are pruned. Watch the topology itself learn — colonies and pathways grow.
100% OURS · CANVAS 2D · OFFLINEReal learning: Hebbian credit + structural growth/prune — not a scripted animation
Reading the wire
hot / conducting edge
cold joint (near-dead)
exploratory new edge
input node (stimulus)
hidden node (pool)
output node (class)
Wire thickness & glow = conductance g. When g falls below the prune floor the wire dies and vanishes; the network keeps sprouting fresh cold candidates to explore.
The topology learning
Task accuracy—
Live edges0
Hot edges (g>.6)0
Pruned total0
Grown total0
Epoch0
Accuracy ■ · Live edges ■
accuracy (task)live-edge count
Idle. Press Grow to let the connectome wire itself.
Task the network must learn
8×
.16
.09
4
Each epoch: present every training pattern, propagate signal through live wires only, reward edges that carried correct signal (heat + thicken), cool the rest. Sub-floor wires are pruned; fresh cold candidates sprout. The shape is the learned program.