Model semantics (Brian2)

dotFly reproduces the leaky integrate-and-fire model of Shiu et al. 2024 (philshiu/Drosophila_brain_model, model.py) exactly. "Exactly" is checked against Brian2 with golden fixtures: on a 2,001-neuron / 160 k-synapse sugar circuit every spike step is identical and membrane traces agree to 1e-11 mV over 3,000 steps.

The model

Per neuron: dv/dt = (v0 − v + g)/τm, dg/dt = −g/τs, spike when v > v_th, reset v = v_rst, g = 0, refractory rfc. Parameters: v0 = v_rst = −52 mV, v_th = −45 mV, τm = 20 ms, τs = 5 ms, rfc = 2.2 ms, delay 1.8 ms, w_syn = 0.275 mV per synaptic contact, dt = 0.1 ms. A synapse adds w_syn × sign(NT) × contacts to g of the target, 18 steps after the source spike. Poisson inputs add 68.75 mV to v (so every accepted event fires the target) and set the target's rfc = 0.

What has to be right, and is tested

1. Exact integration, not Euler — the closed form per step: a = e^{−dt/τm}, c = e^{−dt/τs}, b = τs/(τs−τm)·(c − a), v' = v0 + a(v − v0) + b·g, g' = c·g. Two FMAs and a multiply per neuron.
2. unless refractory guards every write to v and g except the reset: synaptic deliveries and Poisson events that arrive while a neuron is refractory (spike step through t + R − 1) are discarded, not accumulated. A port that accumulates them diverges within ~100 ms. Refractoriness is an integer countdown (Brian2 compares timesteps with floor(t/dt + 1e-3)).
3. Reset zeroes g.
4. Schedule order per step: state update → threshold → synapses (deliveries and Poisson) → reset. A delivery arriving in the step a neuron spikes is wiped by the reset.
5. Homogeneous delay of 18 steps: a ring of 19 spike lists, no per-synapse delay.
6. Poisson targets have no refractory period and an event landing in a spike step is discarded, so a continuously driven target fires at most every other step.
7. Silencing zeroes only outgoing weights — the neuron still integrates and spikes.
8. Sign policy lives in the data: ACh +1, GABA/Glu −1, DA/OA/5-HT +1; MaleCNS histamine and unknown transmitters are dropped (recorded in the checkpoint provenance).
9. Precision: the reference backend is float64 and matches Brian2 to rounding; the fast SIMD backend is float32 and is validated statistically (per-neuron spike counts identical on the full v630 sugar run, ≥ 99 % of golden spike events) and is bit-identical across thread counts.

A note on scale

One synaptic contact raises v by at most ≈ 0.1575 × 0.275 mV; a single presynaptic spike needs ~162 contacts to fire a resting neuron on its own. Firing is convergence. This is also why the model has no tonic activity: nothing fires without input, and recurrent loops either die or — when the convergence is high enough, as in the Kenyon cells — run away. See Findings.