flybrain-play
We took the complete wiring diagram of a male fruit fly's central nervous system, ran current through it, replayed the taste experiments of Shiu et al. on it, and asked it to ride a road bike.
- 166,700 neurons
- every synapse from male CNS v1.0
- leaky integrate-and-fire, 0.1 ms steps
- connectome never trained
The fly learns to ride
Starting from a blank decoder that cannot steer, the fly's whole brain learns to ride a Colnago V4Rs down a 7 m road at 20 km/h in side gusts, by trial and error: 1,776 attempts in 37 generations of 48, every one recorded. In generation 0, 36 of 48 are down within a second; by the last, 35 of 48 ride the full 15 s and none falls. Watch him fall, drift into the grass and set new records one attempt at a time, with his brain firing in the side panel. The connectome never changes; only the 90 numbers between his descending neurons and the handlebars do.
Watch him learn →Fly Brain Live
The firing of all 140,638 neurons with a known position, in 50 ms frames, while his sugar, water, bitter or Ir94e taste neurons, or his antennal Johnston's organ neurons, are driven as in Shiu et al. 2024. Watch whether the feeding motor neuron MN9 fires, and how bitter shuts it down.
Open the dashboard →More ride replays
What is real
Neurons, synapses and their signs come from the Janelia / Google male CNS v1.0 connectome; the neuron model is Shiu et al., Nature 2024, re-implemented by us and checked spike for spike against their code. How a bike's lean maps onto sensory neurons is modeled, and so is the bike decoder.
Where it ran
On a Slurm node with RTX A6000 GPUs. One whole-brain trial takes ~8 s per simulated second; batched on the GPU, 128 trials take 35 s. In the bike loop, 48 riders run at ~20 s of wall time per second of bike time.