Fitted positive input weights gave no clear block prediction gain at penalty 10 in 169 selected MICrONS targets

We tested whether fitted nonnegative input weights predict omitted entries of a target cell's mean movie response in one MICrONS mouse. We used 169 targets with at least five measured anatomical inputs and a ridge penalty of 10. In five block folds, mean scaled error was 2.27770; matched input sets gave 2.25550. The matched-set gain was -0.02221, with a descriptive target bootstrap interval of -0.11342 to 0.06765. The training-mean error was 2.04157. Interleaved folds gave positive-model error 1.18900 and matched error 1.21884. We also tested signed weights, equal pooling and penalties 1 and 100. Sparse inputs, shared stimuli and cells, one mouse, coarse matching and trial-mean data limit the test. It does not measure causal transmission or transfer across animals. Original code, aggregate results, a figure, hashes and external retrieval instructions support reproduction. External records remain outside the bundle.

MICrONS · Allen Institute · structure and function · positive regression · response prediction · matched input sets