AI-Optimized Neuron Models¶

Nine novel neuron models designed for AI workloads, not biological simulation.

Modules: sc_neurocore.neurons.models.arcane_neuron, sc_neurocore.neurons.models.ai_optimized

ArcaneNeuron¶

Five coupled subsystems in a single coherent ODE:

FAST (tau=5ms): spike timing, immediate sensory processing
WORKING (tau=200ms): working memory via sustained activity
DEEP (tau=10s): long-term context accumulation, personality drift
GATE: learned attention over inputs, modulated by confidence
PREDICTOR: forward model of own future state, fires on surprise

The deep compartment accumulates identity: it changes only when the neuron encounters genuine novelty (prediction errors), not routine input. Confidence modulates the threshold (confident = lower threshold = faster responses) and the learning rate (uncertain = learn faster).

from sc_neurocore.neurons.models.arcane_neuron import ArcaneNeuron

neuron = ArcaneNeuron()
for t in range(1000):
    spike = neuron.step(current=0.5)

print(neuron.identity_state)      # deep compartment value
print(neuron.confidence)           # 1 - mean(novelty_history)
print(neuron.meta_learning_rate)   # lr_base * (1 + eta * novelty)

Rust implementation: ArcaneNeuron in network_runner.rs (included in the 81-model NetworkRunner).

AI-Optimized Models¶

Model	Equation Summary	Use Case
`MultiTimescaleNeuron`	Three-compartment (fast/medium/slow), slow compartment modulates threshold	Temporal context accumulation
`AttentionGatedNeuron`	Sigmoid gate with learned key/query weights	Selective input filtering
`PredictiveCodingNeuron`	Fires only on prediction errors	Novelty detection
`SelfReferentialNeuron`	Introspects spike history, auto-regulates dynamics	Stable autonomous firing
`CompositionalBindingNeuron`	Phase-coding with amplitude, in-phase = bound	Variable binding
`DifferentiableSurrogateNeuron`	Trainable alpha/beta/theta for surrogate gradients	SNN training
`ContinuousAttractorNeuron`	Ring attractor, Mexican-hat connectivity	Continuous working memory
`MetaPlasticNeuron`	Error-trace-driven meta-learning rate	Adaptive learning speed

All models share the standard step(current) -> int / reset() interface.

from sc_neurocore.neurons.models.ai_optimized import (
    MultiTimescaleNeuron,
    AttentionGatedNeuron,
    PredictiveCodingNeuron,
)

mtn = MultiTimescaleNeuron()
for t in range(200):
    spike = mtn.step(current=0.8)
print(f"Slow compartment: {mtn.v_slow:.4f}")