With one random contact, a fitted rule cut squared mean-area error by 48% in 63 connections from one fly larva and by 7.5% in the 32 with five or more contacts

We tested whether sparse contact area measurements improve connection weights based on synapse counts. We used 540 measured areas from 80 connections in one first-instar fly larva. The main test had 63 connections with at least two contacts. Each test excluded both connected neurons from all calibration. The loss was squared total-area error divided by contact count squared, averaged equally across connections. We compared count only, a rule that used a measured part and the training mean for the rest, and a fitted blend that also predicted the rest from the measurement. We computed exact mean errors for uniform sampling of one contact. The fitted rule cut error by 48.19% relative to count only. Using only the measured part cut error by 43.48% and supplied 90.24% of the full error decrease. Predicting the rest added 8.33%. In a fixed panel of 32 connections with five or more contacts, the fitted rule cut error by 7.49%. Predicting the rest raised error by 6.44% relative to the measured-part rule. Sample mean times count gave 5.22 times the count-only error in this panel. Four measurements cut error by 49.06%. These are exact finite-table test results under the stated sampling rule. They do not measure physiological weight accuracy or establish effects across animals. We provide source tables, code, all predictions, an exact loss formula, and repeat instructions.

connectome · synapse count · synaptic area · Drosophila · sparse measurement · shrinkage · sampling without replacement · structure and function