Skip to content
BrainlessLab BrainlessLab BrainlessLab

SORN

Which populations are constructed?

SORN separates excitatory and inhibitory nodes. The parameter inhibitory_fraction is the inhibitory-to-excitatory ratio, despite its historical name. With its default 0.2, the excitatory count is round-to-even of count / 1.2, clamped to at least one; the remainder is inhibitory. A one-node model therefore has no inhibitory nodes.

from brainlesslab import CompositionSpec, NodeSpec, TaskSpec, profile, resolve


resolved = resolve(profile(CompositionSpec(
    node=NodeSpec("sorn", parameters={"eta_stdp": 0.001, "eta_ip": 0.001}),
    task=TaskSpec("delayed_cue"),
    count=32,
)))
print(resolved.targets[0].target.composition.node.kind)

This resolves SORN with explicit learning rates. The default excitatory connection probability is 10 / 166; it is a probability, not a fixed in-degree. Excitatory self-links are removed.

Read the four update stages

  1. Compute new excitatory and inhibitory spikes from the old populations. Excitatory drive combines excitatory recurrence, inhibitory subtraction and task input. Inhibitory drive receives old excitatory activity. Threshold equality fires.
  2. Update excitatory weights by STDP: eta_stdp * (new_destination * old_source - old_destination * new_source).
  3. Remove weights at or below zero and permanently clear their connectivity masks. Normalise nonzero excitatory incoming rows to their initial row-sum targets. Zero rows are skipped; pruned edges do not regrow.
  4. Shift excitatory thresholds by eta_ip * (new_spike - H_ip). This intrinsic update is not clipped.

Inhibitory weights and thresholds stay fixed. Task input weights stay fixed. The common model noise argument is unused by SORN.

Emitted activity contains excitatory followed by inhibitory spikes. Task effectors average selected excitatory spikes only; an empty selection returns zero. Matrices use destination/source orientation.

Freeze and reset deliberately

freeze_weights stops STDP and synaptic normalisation while intrinsic plasticity continues. freeze_plasticity also stops threshold adaptation. Setting eta_stdp=0 alone does not disable normalisation or intrinsic plasticity.

Full reset restores weights, pruned masks, thresholds, spikes and plasticity flags. Each independent trial starts from its declared construction, including initially empty rows. Current storage is dense.

Terminal window
uv run pytest python/tests/test_sorn.py

These tests establish equation conformance and runtime ownership for the declared implementation. They do not establish the authors’ code trajectories or biological validity. Continue with capacity probes or a paired ablation.