Joint visual response terms gave an uncertain 12% strength prediction gain in 52 mouse V1 inhibitory connections

We tested whether joint visual response effects add to strength prediction after separate feature effects are supplied. We used 52 detected PV-to-pyramidal connections from 18 mice in a public mouse V1 study. At each frame, we split calcium responses into a mean term, three separate feature terms, and joint terms. The split gives an exact sum for full response cosine. A fixed ridge model used cell distance and the first four response terms. Adding the joint term reduced mean squared log voltage error by 11.9% when each test mouse was excluded from fitting. The 95% whole-mouse bootstrap range ran from a 16.6% error increase to a 38.5% decrease. A secondary test of 40 reverse connections gave a 0.7% error increase, with a range across zero. Across all 93 cells, median odd/even trial cosine was 0.73 for separate feature effects and 0.14 for joint effects. These are descriptive repeat scores. All fixed model checks had gain ranges across zero. The result does not establish an added prediction benefit. The test concerns joint pattern agreement under restricted linear models, with the full response norms supplied to both models. It cannot separate unrestricted cell effects from pair effects. We preserve exact selected visual trials, response components, predictions, code, source checks, figures, and repeat instructions.

visual cortex · structure and function · inhibition · joint tuning · factorial decomposition · paired recordings · mouse holdout · repeat agreement