A penalty control reduced the shared input gain from 1.49% to 0.69% in a fixed test of estimated motion polarity in 1,721 MICrONS neurons

We tested whether shared inputs predict which motion direction a neuron prefers when its orientation axis is known. We used 1,721 MICrONS neurons with manual registration and motion estimates from a digital twin. The full input graph retained 5,596,131 synapses from 3,123,824 source segments. Five outer folds withheld spatial cubes. Inner folds chose kernel weights and ridge penalties. Shared inputs reduced Brier error from 0.232049 to 0.228599, a 1.49% gain. A later control added only an identity kernel. It had no shared input information but reduced error to 0.230188 by changing the effective penalty. It removed 54% of the initial absolute gain. The remaining input gain was 0.69%, with a conditional 95% block range of -0.14% to +1.55%. That range holds fits fixed and is not an animal-level interval. Other designs gave mixed results: after the same control, gains were -0.30% in V1 only, 0.27% in neuron folds, 0.77% in a second spatial split, and 1.09% with each imaging session withheld. Removing 11 conflicting cell class labels gave 1.21%. The result permits small conditional information but does not establish a stable added gain in the main controlled test. Full inputs reduced error by 0.37% against a nucleus-only control, with a range across zero. The package gives a model with unchanged error when the ends of an axis are exchanged, a proof of the penalty identity, compact exact graph inputs, provenance, code, all fits and predictions, figures, artificial controls, and repeat steps. One mouse, model labels, source fragments, and fixed prediction ranges limit the result.

connectome · MICrONS · motion direction · orientation · shared inputs · kernel ridge regression · regularization · source coverage