Train / Fruit fly multisensory circuit (FlyWire)

The fly plays Pong

Can the real circuit turn the position of a moving ball into well timed paddle movements?

Loading 3D view…

Fly ·

Real wiringReadout from generation 0

Learning curve

press Train

No generations yet. Press Train: every generation adds a point for the best candidate, the population mean and the held out test.

Untrained (generation 1 mean)–

Best on training–

Held out, current readout–

Inside the circuit

live, current readout

readout neuronsfiring rateweight

All spikes, last 2 seconds

Saved runs

tick runs to overlay their held out curves

No saved runs yet. Train the real circuit and a control, save both, and compare their curves here.

What is being trained

The fly walks left and right behind a paddle at the end of a walled court. While the ball approaches, its offset from the paddle drives the LPLC1 neurons on that side of the fly, more strongly as the ball comes closer.

A readout of the 21 output neurons sets the walking speed. The score counts hits, minus misses, minus a little for being far from the ball.

Training uses a slower ball; the held out test a faster one with steeper angles.

Inspired by: DishBrain (Kagan et al. 2022), where cultured neurons played Pong.

How to read the result

Fixed: the wiring, the synapse counts, the transmitter signs and the neuron model (Shiu et al. 2024). Learned: 22 readout numbers.

A rising curve alone proves little: a trained readout can often make use of any network. The evidence is in the comparison. Train the real circuit, save it, train a rewired control, and overlay the two held out curves. If the real wiring learns faster or generalises better, the specific map carries information the controls lack.

Senses: Ball to the left of the paddle → LPLC1 left; Ball to the right → LPLC1 right. Readout neurons: DNa01 left, DNa01 right, DNa02 left, DNa02 right, DNa03 left, DNa03 right, DNb05 left, DNb05 right, DNa11 left, DNa11 right, DNp01 left, DNp01 right, MDN left, MDN right, DNp09 left, DNp09 right, DNg13 left, DNg13 right, DNp03 left, DNp03 right, MN9 (proboscis).

Training uses a 0.25 ms simulation step (the lab uses 0.1 ms); firing rates in these circuits differ by about 2 percent. Dataset: Subcircuit of FlyWire v783: inputs HSE, HSN, HSS, LPLC1, LC4, LPLC2, LC16, ORN_DM1, ORN_DA2, sugar/water, bitter; outputs DNa01, DNa02, DNa03, DNb05, DNa11, DNp01, MDN, DNp09, DNg13, DNp03, CB0701; paths with at least 40 synapses per step. About this dataset