How the fly decides
A legitimate sensorimotor experiment that happens to settle on a bonding curve. Weights are not trained. The graph is FlyWire. The constants are Shiu 2024. Full documentation →
This agent is controlled by a real Drosophila whole-brain connectome (FlyWire FAFB v783) simulated with leaky integrate-and-fire dynamics. Not a traditional neural network.
Connectome
FlyWire FAFB v783 · 138,639 neurons · 15,091,983 signed synaptic edges. Completeness table and edge list are the public materialization used by the Shiu whole-brain LIF.
Dynamics
Current-based leaky integrate-and-fire, identical to vendor/shiu_2024/model.py:
dv/dt = (v0 − v + g) / τm v0 = −52 mV τm = 20 ms dg/dt = −g / τsyn τsyn = 5 ms vth = −45 mV w_ij = (sign × n_synapses) × 0.275 mV delay = 1.8 ms refractory = 2.2 ms
Sensory cells receive Poisson events onto v with quantum w_syn × 250, refractory 0 — the optogenetic protocol from the paper. Each market tick is one 200 ms stochastic trial on the full graph.
Sensory → motor (the analogy)
Market→Hz and motor→trade are experimental analogies. The graph and LIF constants are the published FlyWire / Shiu model. Weights are not trained.
| Market | Population | Readout |
|---|---|---|
| green / +return | sugar GRNs | MN9 feeding → BUY |
| red / dump | bitter GRNs | suppresses feeding |
| curve fill | water GRNs | satiation-like |
| volume shock | Johnston's organ | DN1/DN2 startle → SELL |
| grooming drive | aBN1 | GROOM (no order) |