V8 Inner Models Engine

Modular spiking neural network engine. Izhikevich neurons, 8 synapse types, composable building blocks, JSON brain recipes. NumPy-optimized.

The idea: neurons and synapses are self-contained parts with intrinsic behavior. The engine just ticks. Topology creates function.

What's here

engine/
  neurons/     5 Izhikevich types (RS, FS, IB, CH, LTS)
  paths/       8 synapse types (fixed, plastic, facilitating, depressing,
               gated, gap_junction, reward_plastic, developmental)
  runner.py    Brain class - tick(), deliver_reward()
  loader.py    Load brain DB
  recorder.py  Spike/weight recording
  encoder.py   Signal encoder (audio/tone -> neuron currents)

blocks/
  components/       11 circuit components (JSON)
  sub_components/   21 sub-circuit components (JSON)

recipes/           JSON brain recipes
schema.py          SQLite brain DB schema
brain_generator.py Recipe -> DB compiler
build_arena_brain.py  Arena brain + body/sensor map population
run_arena.py       Full brain-body-environment loop
arena.py           2D arena (Gaussian food gradients, circular boundary)
worm_body.py       24-segment rigid chain body (anisotropic drag, proprioception)

Quick start

Requires: Python 3.10+, NumPy

# Generate a brain from recipe
py brain_generator.py --recipe recipes/arena_v1.json

# Populate body/sensor maps for arena
py build_arena_brain.py --brain brains/zoo/arena_v1_s42.db

# Run chemotaxis (brain + body + environment)
py run_arena.py --brain arena_v1_s42.db --ticks 30000

# Run developmental (blank slate) brain
py brain_generator.py --recipe recipes/arena_dev_v1.json
py build_arena_brain.py --brain brains/zoo/arena_dev_v1_s42.db
py run_arena.py --brain arena_dev_v1_s42.db --ticks 30000

Building blocks

Each component is a JSON file defining neurons, internal synapses, and terminals (named I/O ports). Meaning comes from wiring, not labels.

Example: binary_decision is a generic 2-option circuit. Two bistable CH rings, cross-kill FS neurons, RS integrators with STDP. Wire mechanosensory to sensory_a_in/sensory_b_in and it decides forward/backward. Wire emotional_state to modulator_a_in/modulator_b_in and emotion biases the choice.

Example: sensory_bank is 3 neurons (2 RS + gap junction for noise reduction, 1 RS output with depressing adaptation). Modality-agnostic. What it senses depends on what you wire to channel_in.

Example: working_memory_cell is a bistable CH ring + RS trigger + FS kill + 2 RS association neurons with fast STDP (lr=0.5). Sets a 1-bit memory. Whatever was active when the memory set can later recall it.

Synapse types

Type	Behavior
fixed	Static weight. Reflexes, structural connections.
plastic	STDP. Weight changes based on pre/post timing.
facilitating	Gets stronger with repeated use. Short-term potentiation.
depressing	Gets weaker with repeated use. Adaptation/habituation.
gated	Only transmits when modulator neuron is active. Attention.
gap_junction	Electrical coupling. Continuous, bidirectional.
reward_plastic	3-factor: pre x post x reward. Reinforcement learning.
developmental	STDP + Fisher Information pruning. Blank slate.

Recipes

JSON files that declare components, counts, layer positions, and inter-layer projections with synapse types. The brain_generator places components in 3D space and wires them by proximity.

arena_v1.json: Hardwired animal brain. ~8K neurons, 21 component types, 5 layers. Fixed sensory-decision-motor spine. Reward_plastic on emotion->decision. Works immediately (CI=+0.43 in 30K ticks).

arena_dev_v1.json: Blank slate brain. Same components but inter-layer connections are developmental synapses. Only the reflex spine (sensory->decision->motor) is fixed. Everything else starts dense and gets pruned by Fisher Information during a 10K-tick critical period. Needs extended experience to develop.

Neurons

All 5 types use the same Izhikevich model with different parameters:

Type	Role	Key property
RS	Standard excitatory	Steady spiking
FS	Inhibitory interneuron	Fast, precise
IB	Motor, sustainers	Initial burst then steady
CH	Oscillators, bistable rings	Rhythmic bursting
LTS	Suppressor inhibitory	Low threshold

Performance

8070N brain @ 82 t/s with full body physics (24-segment worm, proprioception, arena). 50K brain @ 49-209 t/s (pure brain, no body).

The developmental approach

Human brains overproduce synapses ~2x (peaking around age 2), then prune through adolescence. The topology that survives is shaped by experience.

The developmental synapse type implements this:

1. Start with dense, random connections (overproduction)
2. Track Fisher Information: coincidence_rate = (pre-post coincidences) / (source fires)
3. Every eval_interval ticks during critical period: prune synapses where coincidence rate < threshold
4. Surviving synapses continue as regular STDP plastic
5. After critical period: topology is set, no more pruning

This is the blank slate. The animal brain (arena_v1) works from birth because pathways are pre-wired. The developmental brain (arena_dev_v1) must discover its own wiring through experience.

Design principles

1. Architecture > quantity. Topology matters more than neuron count.
2. Memory density predicts learning. Dense WMC cross-wiring = more STDP surface.
3. Each part owns its own rules. The engine doesn't know what "emotional" means.
4. Meaning from wiring, not labels. binary_decision doesn't know it's deciding forward/backward.
5. Projection:crosswire ratio > 1.5 needed for learning at scale.