Neural models

BrainlessLab currently implements two dense neural models. Both receive task inputs, emit binary node activity and read selected activity into task effectors. Neither is trained by a task-labelled loss.

• Falandays adapts synaptic weights and activity targets.
• SORN combines excitatory/inhibitory activity, STDP, synaptic normalisation and intrinsic plasticity.

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

for kind in ("falandays", "sorn"):
    resolved = resolve(profile(CompositionSpec(
        node=NodeSpec(kind), task=TaskSpec("delayed_cue"), count=32,
    )))
    print(resolved.targets[0].target.composition.node.kind)

Both configurations should resolve. Availability means the software can construct and step these models; it does not establish equivalent resource use, author-code trajectory parity or task competence.

Choose one model, declare its parameters and test a task protocol with independent blocks. Compare frozen modes only after reading the model’s plasticity rules. See performance for measurement of execution costs.