Using this reference#

Use this page when setting a CLI flag, reading a saved configuration, or checking a neuron update. The legacy SNNSIM biophysical model uses ms for time, mV for voltage, nF for capacitance, μS for conductance, nA for current, and Hz for rates. A --dt 1 means 1 ms, not 1 s. Names are not uniformly suffixed: sim_ms, ref_ms_E, tau_gaba, and dt all represent milliseconds.

The values below describe library defaults or explicitly labelled examples, not a universal experiment recipe. CLI defaults, loaded configurations, and committed experiment recipes can differ. Graph bundles declare their own units and parameters; see exp105 — Networks, signals, and parameters.

Quantities#

COBA / PING biophysical constants#

These are the default constants in tools/snnsim/models.py used by exp100 — COBANet. They are model choices, not universal biological constants. Capacitance and leak satisfy ${𝜏}_{𝑚}=\frac{{𝐶}_{𝑚}}{{𝑔}_{𝐿}}$. Here ${𝜏}_{𝑚}$ is the passive membrane time constant, ${𝐶}_{𝑚}$ the capacitance, and ${𝑔}_{𝐿}$ the leak conductance.

Synapse time constants: ${𝜏}_{\text{AMPA}}=2$ ms (excitation), ${𝜏}_{\text{GABA}}=9$ ms (inhibition). These are decay times, not a sufficient timestep-stability criterion. Check timestep sensitivity for the intended input, recurrence, and measurement. --tau-gaba overrides the inhibitory decay; --train-leak allows bounded, per-neuron membrane time constants instead of fixed leak defaults.

Internal consistency#

The chosen units are self-consistent — no conversion factors appear in the integration code. Two equations carry the whole system.

The membrane time constant is ${𝜏}_{𝑚}={𝐶}_{𝑚}/{𝑔}_{𝐿}$. With ${𝐶}_{𝑚}$ in nF and ${𝑔}_{𝐿}$ in μS,

so ${𝐶}_{𝑚}=1$ nF and ${𝑔}_{𝐿}=0.05$ μS give ${𝜏}_{𝑚}=20$ ms directly.

The LIF voltage update is $d{𝑉}_{𝑚}=\left(\mathrm{\Delta }{𝑡}_{\text{sim}}/{𝐶}_{𝑚}\right)\left(-{𝑔}_{𝐿}\left({𝑉}_{𝑚}-{𝐸}_{𝐿}\right)+{𝐼}_{\text{ext}}\right)$. With $\mathrm{\Delta }{𝑡}_{\text{sim}}$ in ms, ${𝐶}_{𝑚}$ in nF, ${𝑉}_{𝑚}$ and ${𝐸}_{𝐿}$ in mV, ${𝑔}_{𝐿}$ in μS, and ${𝐼}_{\text{ext}}$ in nA,

because ms·nA / nF = mV exactly.

Conductance-current products share the same ledger: $𝑔\left({𝑉}_{𝑚}-𝐸\right)$ is μS × mV = nA, so synaptic currents fold into ${𝐼}_{\text{ext}}$ alongside any direct input current without a scale factor.

Why not SI?#

Pure SI (F, S, V, A, s) forces every value to a large negative exponent — ${𝐶}_{𝑚}={10}^{-9}$ F, ${𝑔}_{𝐿}=5\times {10}^{-8}$ S, $\mathrm{\Delta }{𝑡}_{\text{sim}}=2.5\times {10}^{-4}$ s. The neuroscience convention (ms, mV, nF, μS, nA) keeps every typical value between ${10}^{-3}$ and ${10}^2$, which makes numerical debugging and human intuition faster. Conversion to SI remains available as an independent check; the equations above show why no extra scale factor is needed in these units.

Checking a configuration#

1. Separate defaults from overrides. Read the saved config.json and the recipe that supplied it. For CLI inference, explicit flags override inherited values; omitted fields fall back to defaults.
2. Convert rates once. With integration timestep $\mathrm{\Delta }{𝑡}_{\text{sim}}$ in ms and input rate ${𝑟}_{\text{input}}$ in Hz, a Bernoulli spike encoder uses event probability ${𝑝}_{\text{event}}={𝑟}_{\text{input}}\mathrm{\Delta }\frac{{𝑡}_{\text{sim}}}{1000}$ per step. Here ${𝑝}_{\text{event}}$ is dimensionless. Check that the probability is meaningful for the chosen rate and timestep.
3. Check duration and counters. The legacy path uses int(t_ms / dt) simulation steps. Avoid assuming a non-integral duration is preserved exactly. Refractory times are also discretised to steps.
4. Check the stored weights. Initialization means are on a summed-coupling scale; individual stored edges are fan-in normalised. Readout weights can use direct initialization instead. See exp006 — Training.

exp003 — SNNSIM API Reference · exp100 — COBANet