Test cell reuse cut pooled input error by 62% in seven zebrafish linear models with equal gains, known totals and matched training counts

We tested how cell pooling changes a connectome prediction test. We used released ribbon and membrane contact maps from seven zebrafish neuromasts. The main test excluded both cells of each same-direction pair together. It used 103 pairs from 60 cells. With known ribbon totals, equal cell and ribbon gains, and linear addition, contact area gave 68.5% less mean pooled share error than terminal dominance. The whole-animal bootstrap 95% interval for this decrease was 39.3% to 85.5%. We then added a descriptive control with matched training counts. Reuse of the other test cell reduced mean pooled error by 61.9%, from 0.22288 to 0.08488. The complete separate-cell loss distribution stayed fixed within every direction group. We prove this distribution equality by exchange of cell roles. We also prove two pooled-error factors for equal totals. These results show how input pooling can improve a validation score without better cell predictions. The scope is seven structural maps and the stated input model. Neural activity was not measured. The package preserves raw source workbooks, provenance, code, every split, predictions, proofs, figures and repeat instructions.

connectome · zebrafish · neuromast · validation · input pooling · ribbon synapse · membrane contact · cell groups