Target-area calibration reduced a feature-related transfer loss in one selected MICrONS mouse

We tested whether a separate target-area sample can reduce a wiring prediction loss after transfer from primary visual cortex (V1) to pooled higher visual areas (HVA). We used signed model predictions from a prior campaign study and public MICrONS v1 data from one mouse. Calibration used 14,215 pairs. Evaluation used 52,444 pairs from separate neurons. Adding model feature similarity to a geometry model gave a mean log-loss gain of -0.0008388 nats per pair before calibration and -0.0000719 after calibration. Positive gain means lower error. The increase in gain was 0.0007670, with a 95 percent neuron-weight interval from 0.0000464 to 0.0014061. The calibrated gain interval included zero. All five fixed neuron splits gave an increase in gain. Thus a target-area intercept adjustment reduced much of the tested transfer loss, but did not establish a feature advantage. The intervals hold source-trained models fixed. One mouse, selected cells, pooled area labels, and uncertain network dependence limit transfer. Code, input designs, model coefficients, split records, uncertainty draws, source hashes, and a figure support reproduction.

MICrONS · structure and function · prediction transfer · calibration · visual cortex