Research
SnnGoL: a spiking network on a Game-of-Life grid
Can a spiking neural network on a 2-D cellular grid, trained only by local, Hebbian-style wiring rules and no backpropagation, learn to continue English text?
How it works
- The grid. 256×256 = 65,536 leaky integrate-and-fire cells. Each tick a cell sums input from its 8 neighbours and up to 32 learned long-range wires, leaks, fires if over its threshold, and goes refractory. Adaptive thresholds keep activity near 0.6% of cells.
- Words are codes, not places. Every word is a set of 40 cells chosen from its embedding, half private to the word and half shared with similar words.
- Training. For each position in a Wikipedia article: inject the last few words, run a few ticks, score every candidate word against the activity, then wire the context towards the true next word and depress wires into every word that outranked it.
- On the GPU. It runs through ILGPU; the chat service runs it in a Docker container on the same machine that serves this site.
Where it stands
It has learned word order, what a context is about, and many Wikipedia phrases, and it stops at plausible sentence ends. It does not yet write fluent sentences: replies are chains of short phrases, recalled and stitched together. Every result, dead end and retraction is kept in the project's findings log.
The chat
The chat service puts the latest pinned release in front of visitors. Each reply comes with a per-word trace, so you can see which recalled article and which wires produced each word. Conversations are kept only in memory and forgotten an hour after their last message.