Source-specific slopes gave no clear primary gain at penalty 10 for new targets in one MICrONS mouse

We tested whether source-specific movie-similarity slopes improve synaptic connection prediction for new target neurons in one MICrONS mouse. The source table has 295,371 positive-opportunity directions and 8,128 connections. Five identity groups gave 235,797 test directions, 6,523 connections, 148 sources and 11,028 targets. Each test target was absent from training in either cell role. All test sources had other training opportunities. Models controlled source rates, geometry, region and layer pairs, scan, and response ceilings. At the primary source-deviation penalty of 10, log loss fell from 0.1109813 to 0.1109708. The gain was 0.0094%, with a conditional 95% interval from -0.0837% to 0.1026%. Penalty 100 gave a small positive gain of 0.0813%; penalty 1 gave an uncertain negative gain. A shared nonlinear and area-specific reference left no clear source-slope gain. Spatial and quality checks at penalty 10 also gave intervals across zero. The result depends on the penalty and model family. Intervals hold fitted predictions fixed and omit new-animal and model-training variation. One mouse, selected arbors, imperfect opportunity measures, and response error limit inference. Code, exact compact input, predictions, coefficients, figures, provenance, and replay checks are included.

MICrONS · Allen Institute · structure and function · connection prediction · source-specific slopes · neuron exclusion · conditional uncertainty