Visual similarity gave no error decrease in a split-trial variance test of 75 mouse V1 connections

We tested whether visual response similarity improves prediction of repeated synaptic voltage variance after mean strength and recording controls. We used selected PV-positive and pyramidal cell connections from Znamenskiy et al., Figshare version 1. The main panel had 75 directed connections, 52 cell pairs and 19 mice. Each connection supplied 16 accepted trials. Odd and even groups gave separate calibration and test data. A fixed 2–20 ms voltage window avoided peak selection from test trials. Nested ridge fits excluded the whole test mouse. The control rule had mean squared log variance error 0.75875. Adding similarity gave 0.76134, a 0.34% error increase. The added slope was omitted in 15 of 19 outer tests. Full mouse resampling with repeated fitting and selection gave a central 95% error-decrease range of −14.0% to +7.9% in 199 draws. Alternate windows, trial groups and source filters gave no observed decrease. A later trace audit removed nine directions with early extra crossings or alignment fallback and retained the result. Independent source, variance and solver checks passed. A large artificial predictor effect was detected. The package gives 11,200 compact trial records, source hashes, code, all predictions, figures and repeat steps. This result applies to the stated prediction rules and selected slice recordings. It does not exclude a modest benefit or identify biological release variance.

functional connectivity · synaptic variability · visual cortex · paired recordings · parvalbumin · trial variance · prediction · model selection