A 3D point cloud that learns from experience.
2000 points form a sphere. Each point is connected to 12 nearest neighbors. Impulses travel the network as branching dendrites. When an impulse crosses a connection, it strengthens it (Hebbian learning). Connections that aren't used slowly fade (decay). Over time, the network develops pathways — traces of what happened to it.
- Stimulus A / B — inject impulses into opposite regions of the sphere. Hold for 10–15 seconds.
- Show paths — visualize the strongest connections. Gold = stimulus A region, blue = stimulus B.
- Reset weights — clear all learned connections, start fresh.
- Before stimulus — impulses wander randomly, all connections equal.
- During stimulus — activity concentrates in one region, connections strengthen.
- After stimulus — pathways remain. The network remembers where the activity was.
- Two stimuli — two distinct traces coexist in the same network.
| Concept | Implementation |
|---|---|
| Reservoir | 2000 random points on a sphere (radius 3.5, distortion ±0.7) |
| Sparse activation | Impulses select neighbors probabilistically, not exhaustively |
| Hebbian learning | Traversed edge weight += 0.12 (branches += 0.06) |
| Decay | All weights × 0.999985 per frame (~60 fps) |
| Neighbor graph | K=12 nearest neighbors, built once at init via random sampling |
One mechanism, governed by two numbers (learning rate and decay). The same principle as biological memory: what fires together wires together; what doesn't is forgotten.
Started as a quantum measurement visualization (2024) — three spheres of random points, destroyed and recreated on each visit. Impulses added March 2026 — dendrites emerged from geometry, no one programmed branching. Hebbian learning added 23.03.2026 — the network gained memory.
30 years of one question: where does order emerge from chaos?
- Ripple — same principle (reservoir + Hebbian + sparse activation) as a flat 2D grid with four experiments (eye, short-term memory, long-term memory, creature)
- newBrain — physical neuromorphic network on ESP32 with Hebbian learning. BlochSphere is the software mirror.
- Alive Pictures (2018) — the ancestor. 500 points pushing neighbors → islands, borders, emergence. No memory, no learning — just geometry.
https://alexvjack.github.io/BlochSphere/
Note: the public version may lag behind this repo. The live experiment lives here.