At fixed cell noise power, 1% noise shared within eye modules cut information about a slow mode by 35.5% in two mean outputs of one zebrafish linear model

We tested how shared input noise changes a slow memory signal in a fixed model of one larval zebrafish circuit. The model had 289 core cells and 54 output cells. We kept the circuit weights, cell times and each core input noise power fixed. A fraction 0.01 of the standardized input noise variance was common within each of two released eye modules. The two common inputs were independent. We calculated Gaussian Fisher information about one slow mode amplitude from the mean rates of 32 motor and 22 internuclear output cells. Information fell by 35.5% relative to independent input. The exact variance of the best scalar estimate rose by 55.0%. Mean activity and its 10.10 s decay stayed fixed. The exact slow coordinate gave a similar loss of 35.1%. Noise timing and cell-time checks retained a loss. Observation of all core rates gave a 33.2% loss. Shuffled module labels also gave losses of 30.0% to 32.3%. Thus the exact eye groups explain only part of the effect. All 19 main numerical checks passed. A fresh-source repeat matched the stored numerical results. We supply code, source hashes, covariances, output kernels, figures and a query tool. These results concern a linear Gaussian model and a specified observation task. The shared noise fraction is an assumption, not a measured value. We do not estimate animal memory capacity or eye-angle error.

connectome · zebrafish · neural integrator · shared input noise · Fisher information · memory precision · linear model · output covariance