Independent noise did not make input sums more reliable in 55 of 130 selected Allen median-response models

We tested a conditional model of measured inputs to the same neuron in Allen Synaptic Physiology r2.1. Each first-pulse median was assigned as a model mean. Responses added linearly with independent noise. The sum had lower squared signal-to-noise ratio than the best single input in 11 of 29 mouse VisP excitatory groups, 38 of 89 mouse inhibitory groups, and 6 of 12 human TCx excitatory groups. Ten losses remain under a conservative box that lets every assigned mean and standard deviation vary by a factor of 1.10. A 50 microvolt floor left 38 losses in 114 groups. Noise subtraction left 33 losses in 83 groups. At least 50 first-pulse responses per edge left only 14 groups. We give exact correlation thresholds and an ideal positive-weight reference. These are selected model calculations, not measurements of simultaneous input reliability. Median-to-mean error, unknown covariance, measurement noise, best-input selection, and small slice groups limit the result. Code, aggregate evidence, input hashes, and retrieval steps support reproduction. External records are not bundled.

Allen Institute · synaptic convergence · transmission noise · linear integration · signal-to-noise ratio · conditional model