At a fixed 20 ms time constant floor, 5 ms and 20 ms Euler steps gave a median 42.1% central flash response difference in 50 released fly visual models
We tested how the solver step changes neural responses in the 50 released flyvis 1.2.0 models. The wiring and learned parameters were fixed. We scored the 65 central model units at common times. We compared 20, 10, 5, and 2.5 ms Euler steps, with the time constant floor either fixed at the 20 ms training value or set to the solver step. With the floor fixed, the 5 ms and 20 ms flash responses had a median root mean square (RMS) difference of 42.1% of the 20 ms response RMS (range 8.2% to 74.2%). Changing only the floor at 5 ms gave a median 0.7% difference. Of 3,250 learned time constants, 2,396 were below 20 ms. At that training step, the intrinsic voltage-retention coefficient is zero for these units. Network inputs can retain memory. The 2.5 ms and 1.25 ms fixed-floor responses differed by a median 1.1% in five selected models. For two opposite 4 Hz motion inputs, the median T4/T5 response difference was 39.3%. Direction-index signs changed in 1 of 113 horizontal model units with absolute baseline index above 0.1. These results describe numerical sensitivity in a fixed model release and two probes. They do not measure animal responses or task accuracy. The package contains the full graph, model parameters, predicted traces, source hashes, code, and checks.
connectome · Drosophila · flyvis · Euler integration · time constant · neural dynamics · reproducibility
References
3- usesdoi:10.1038/s41586-024-07939-3
We use the released 50 trained visual models, the source connectome, and the state equation.
- doi:10.48550/arXiv.2610.13018
Earlier work tests intervention and circuit dependence in the same released model family.
- doi:10.1007/s12021-026-09811-3
Earlier work tests direction readouts and learned-parameter changes in a frozen fly visual model.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesWith uniform light, a periodic 380-state model kept central flash differences below 5% in 42 of 50 released fly visual models
We use its verified graph and parameter arrays, retain its source readers, and reproduce all 50 of its 5 ms full-field flash traces. We test spatial representation at a fixed time step.
- With a 20 ms minimum time constant, a 125 ms Mi4 current delay reversed 4 Hz direction contrast in 9 of 13 selected T4 units in nine periodic fly visual models
Its solver test motivates the fixed 20 ms time constant floor and the smaller-step control.
- usesAt a 20 ms time constant floor, boundary wrap changed maximum absolute voltage after 20 s of gray input from 2.4e24 to 18 in fly model 032 and from 9 to 7.1e23 in model 049
We reuse its verified finite graph, 50 parameter sets and restricted source reader. We keep its distinction between solver step and time constant floor.
- usesAfter 1 s of gray input, a 20 ms update lost at least 3,241 voltage directions in each of 50 released fly visual models
We reuse its verified graph and parameter arrays and its MIT source reader, graph compiler, matrix builder and fetch code. We add full-state loss certificates and continuous-flow limits.