With fixed time constants of at least 20 ms, 20 ms Euler steps gave a median peak ratio of 3.15 to the ODE reference for 80 ms flashes in 50 fly vision models

We tested how numerical time steps changed flash responses in 50 released fly vision models. Each model had 45,669 units and 1,513,231 connectome-based connections. For each model, all solvers used the same state after one second of grey input. Bright and dark pulses lasted 80 ms. We held the wiring and weights fixed. Each effective membrane time constant was fixed at 20 ms or more. We compared Euler updates with an adaptive solution of the same ordinary differential equation (ODE). We measured central T4 and T5 voltage responses relative to each solver's grey trajectory. All responses used the same 20 ms sample grid. The median across models of the median peak ratio was 3.15. The median waveform error was 220.7% at 20 ms and 20.4% at 5 ms. The native time-constant floor changed 5 ms waveforms by a median 2.16%. The largest reference change was 3.67% in one weak response. The tested stricter references left the ensemble medians unchanged. The source stored 2,396 of 3,250 time constants below 20 ms. Those values have no local effect on a fixed 20 ms update. This test shows that numerical steps can set transient gain in these models. It tests a continuous-time interpretation of released parameters. Tests of motion tuning, task error and living-fly responses are separate work. The package has source archives, checksums, code, response arrays, figures and repeat instructions.

connectome · Drosophila · numerical integration · flash response · time constant · model validation · flyvis