Spiking neural network experiments
A small set of from-scratch spiking-neural-network experiments that run on a laptop CPU — no special hardware. Each one was written with its pass/fail bar declared before it ran, every run is kept (including the nulls and the failures), and the conclusions are reported as measured, not as hoped.
Setup
- snnTorch 1.0.0, PyTorch 2.14.1 (CPU), Brian2 2.10.1 — all open-source. Exact pins in
requirements.txt.
python -m venv .venv
.venv/Scripts/python.exe -m pip install -r requirements.txt # Windows
# or: .venv/bin/python -m pip install -r requirements.txt # Linux/macOS
Run any experiment with that interpreter, e.g. .venv/Scripts/python.exe exp1_mnist_snn.py.
Experiments
| File | What it shows | Result |
|---|---|---|
exp1_mnist_snn.py |
A 2-layer spiking network learns handwritten digits; information moves only as spikes over 25 time steps. | 9.0% before training (chance) → 91.2% after 2 short passes, in 19 s. 9.1% of hidden neuron-steps spiked. (exp1_spikes.png) |
exp2_brian2_neurons.py |
Ten biology-style LIF neurons with rising input: below the 15 mV threshold none fire; above it the rate climbs. | 0 spikes/s at 5–14 mV; 45→155 spikes/s at 17–32 mV. (exp2_neurons.png) |
exp3_stdp.py |
Unsupervised learning from spike timing (STDP). One output neuron listens to 100 inputs, 20 of which secretly fire together. | Bar (pattern group > 2× the rest) set before running. See the three runs below. (exp3_stdp.png) |
exp5_competing_neurons.py |
Three neurons compete (STDP + winner-take-all) to each claim one of three hidden patterns. | Not passed — see below. |
exp6_normalised.py |
The exp5 redesign: per-neuron weight normalisation (Diehl & Cook 2015). | Not passed — see below. |
exp3, every run kept
- Run 1 (
exp3_stdp_run1_null.py, 10 mV, 30 s): null — the output neuron never fired (a 20-spike burst gives ~6 mV), so STDP had nothing to learn from. - Run 2 (
exp3_stdp_run2_30s.py, 5 mV, 30 s): pattern group 0.58→0.84, rest 0.58→0.59 — ratio 1.42×. The 2× bar was not met. - Run 3 (
exp3_stdp.py, 120 s, declared beforehand as the only change): pattern 1.76, rest 0.56 — 3.1×. Met. The neuron found the hidden group from spike timing alone.
controls/ holds the exp3 control runs (sync vs async inputs, three seeds each) used to check
the effect was real and not an artifact.
exp5 / exp6, honest failure
Three designs were tried against the three-pattern competition; all three hit the same failure shape — a "dead" third neuron that never gets the early wins it needs, so one pattern stays unclaimed (exp5 run 1 and the +3 mV tweak, exp6 with weight normalisation). Each bar was committed before its run; tuning was stopped rather than searched until it passed. Unrun next ideas: present each pattern alone in early blocks, or give each neuron a different starting bias.
Method notes
Bars before runs, nulls and failures kept beside the successes, controls where an effect is claimed. The aim is a record you can check, not a highlight reel.
Part of one system — each piece usable on its own
These repositories are independent instruments that also compose into a single apparatus for building and honestly auditing trustworthy multi-intelligence systems. Any one can be used standalone; together they form the stack:
- Governed core — covenant: a self-governing multi-agent system with an ethics gate inside the transaction path.
- Observability & provenance — threefold (the L-lens layer over Jacobian Lens) and threefold-memory (its persistent, provenance-tracked memory): every action measured and recorded under one task identity.
- Measurement instruments — contradiction-probe and flinch-observability, with the session records flinch-record and grokflinch: preregistered, blind-scored tests of AI behaviour that embody the covenant's rule — investigate a signal, never convict on one.
- Neuromorphic research — spiking-experiments: from-scratch spiking-network experiments with the pass/fail bar declared before each run.
- Ledger surface — Sentinel-Witness: the trading / ledger dashboard.
This repository's place in it: the neuromorphic research lab behind the covenant's stdlib spiking reflexes.