A penalty-one target correction reduced Poisson contact-count scores in one selected MICrONS mouse

We tested target-cell rate corrections for contact counts in 275,986 selected MICrONS pairs from one mouse. Five folds kept each pair and its reverse together. A Poisson mean used anatomy, response similarity and source identity. The primary target correction, with penalty one, reduced mean test log score by 0.001804 natural-log units per pair. The shared-cell interval for the score change was -0.002344 to -0.001265. Of penalties 0.1, 1 and 10, only 10 improved the Poisson score. A negative binomial distribution with the same corrected means improved scores at all penalties. At penalty one its gain was 0.000757, with interval 0.000603 to 0.000911. Thus the tested primary correction did not explain the broader-count prediction gain. Sparse targets, conditional correction, training-based dispersion, fixed-fit uncertainty and one animal limit the result. This is not a test of a contact-formation mechanism.

MICrONS · Allen Institute · synapse count · target rate · predictive calibration