tinyCPG

A closed-loop spiking model of the two-legged rat spinal central pattern generator (sCPG) in which the sensory synapses driving the locomotor rhythm are not hand-tuned, but self-organise through spike-timing-dependent plasticity (STDP).

The model asks two questions: how does a spinal circuit tune its own synaptic weights to produce a stable walking rhythm, and how well does that self-tuned gait survive when sensory feedback is progressively withdrawn — the central manipulation in spinal-cord-injury (SCI) rehabilitation.

What the model contains

Each leg has an extensor and a flexor rhythm generator (RG-E, RG-F) coupled by asymmetric reciprocal inhibition, driving motoneuron pools and Hill-like muscle proxies whose force and length close the loop back through Ia afferents. Left and right legs are coupled by commissural inhibition. Three projections are plastic — cutaneous CUT→RG-E and the proprioceptive Ia-E→RG-E and Ia-F→RG-F — while the brainstem drive is held tonic.

Extensions over the canonical computational sCPG lineage (Rybak 2006 → Shevtsova 2015 → Danner 2017 → Zhang 2022):

1. STDP self-organisation of the descending and afferent synapses
2. a closed-loop Ia pathway into the reciprocal-inhibition core
3. explicit motoneuron pools and muscle proxies
4. externally paced heel→toe cutaneous drive during stance

Experimental grid

Everything is reported on a 5 × 5 matrix: five locomotion modes (slow / medium=plantar=baseline / fast walk, toe stepping, air stepping) crossed with five STDP learning rates (λ = 10⁻² … 10⁻⁶).

Main findings: the plastic weights converge to the same set-point regardless of initialisation, leg or walking speed; gait quality peaks at an intermediate learning rate rather than the fastest; and holding the cutaneous drive intact under unloading rescues stepping that otherwise collapses — a model of the epidural-stimulation effect in SCI.

Repository layout

Path	Contents
cpg_2legs_fast.py	The model — neurons, connectivity, simulation loop, HDF5 export
run*.sh	SLURM array scripts for the production sweeps (MN5)
debug.sh	Fast local single-config run (~30 s)
scripts/	Figure/analysis generators feeding paper/figures/
scripts/legacy/	Superseded generators, kept for reference
paper/	LaTeX manuscript (main.tex, sections/, figures/)
results/	Simulation output, one dated folder per production run
validation/	Literature-validation notes and data requests

Quick start

pip install -r requirements.txt     # see notes there re: NEST/NEURON
./debug.sh                          # fast local run -> results/debug.h5
python3 scripts/cpg_plot_from_hdf5.py --in results/debug.h5 --save-prefix debug

Production runs are submitted as SLURM arrays (sbatch run_sensory_stdp.sh); see MN5_RUN.md for the full upload → submit → retrieve → plot workflow, and CLAUDE.md for model internals and the tuning-knob reference.

Implementations

The reduced model runs in NEST with Izhikevich neurons — cheap enough to make the 5 × 5 sweeps tractable. A conductance-based Hodgkin–Huxley implementation of the same circuit, used as a biophysical cross-check, lives in the companion repository memCPG/CPG_STDP/py.

Status

The manuscript is a work in progress: Introduction, Methods, Results and Discussion are drafted; the Abstract is not yet written. Figures are generated from the committed scripts and are reproducible from the data in results/.

License

MIT