With uniform light, a periodic 380-state model kept central flash differences below 5% in 42 of 50 released fly visual models

We tested how spatial representation changes uniform-light responses in all 50 released flyvis 1.2.0 models. We built a 380-state model that is exact for an infinitely repeated lattice with class-repeated inputs and initial states. The model retains six spatial phases for the dense cell types and the sparse placement of Lawf1 and Lawf2. We fixed the learned parameters and a 5 ms Euler step. Compared with the 65 central traces of the source finite-eye model, the periodic model gave a median normalized root mean square difference of 0.77%, with a middle half of 0.23–2.51%. The difference was below 5% in 42 models. Model 049 gave a large exception: its periodic central voltage norm grew to 7,420 at 3 s, compared with 10.69 in the finite model. Growth also occurred under constant gray light and at a finer step. This model had the highest released task loss. The ten lowest-loss models all had periodic-versus-finite differences below 1%. A specified 65-state average gave a median 7.19% difference from the exact central phase. Averaging only Lawf output profiles made that 65-state relation exact. The reduced simulation was about 98 times faster in median loop timing. The exact result applies to a repeated lattice, not to the finite eye or arbitrary spatial inputs. These are model results, not animal recordings. The package contains inputs, the proof, code, traces, figures, source hashes, and repeat instructions.

connectome · Drosophila · model reduction · uniform light · spatial boundary · Lawf · nonlinear dynamics · reproducibility