On one light sequence, a selected zebrafish cone model had a higher mean score than all 1,000 saved mixed models

A map of neural connections specifies which input signals reach a cell. We tested how input choice affects a published model of cone convergence. We used 12 zebrafish PR1 cone records, with five repeats per cone, during one 15 s natural light sequence. One model family pooled five repeats from one cone. The other pooled mixed records. The score was minus the stimulus correlation at a fixed 30 ms response delay. We reproduced all 1,680 complete saved homogeneous scores from the response traces. The original comparison gave an 8.08% mean relative score gain for mixed inputs over the mean of 12 homogeneous models. We then selected one model from each family with separate time blocks. The mean test scores were 0.6637 for the homogeneous model and 0.6398 for the selected saved mixed model. The homogeneous model won four of five blocks. Its mean score exceeded every one of the 1,000 saved mixed models. The order also held in three new sets of 5,000 random pools, including pools with five distinct cones. All models used five response records and equal weights. The result shows that a selected response type can exceed random mixed convergence under this score on this sequence. It does not remove the original benefit relative to random homogeneous inputs. The data do not give an independent scene test, animal-level uncertainty, shared population noise, or measured bipolar cell wiring. The package contains the source subset, code, new pool assignments, figures, results, and repeat instructions.

cone synapse · retina · zebrafish · neural convergence · structure and function · input selection · sensory coding · model comparison · reproducible research