When Branch-Local Shunting Helps: A Gain-Load-Alignment Principle for Dendritic E/I Networks

What the math gives to ML

The paper gives a useful, testable criterion for when divisive or shunting interactions should outperform additive nonnegative readouts: the divisor must be local to the signal pathway, reliably estimate a nuisance gain, and suppress gain variation more than it suppresses task signal or adds denominator noise. The controlled model makes shared multiplicative variability an explicit rank-one covariance mode, suggesting that normalization pools should be designed from nuisance covariance rather than applied globally. A practical transfer is a branch-local nonnegative divisive layer with covariance-initialized pooling supports, trained alongside an additive bypass and regularized against denominator unreliability. The key prediction is a larger advantage under multiplicative gain corruption and small-data regimes, with the advantage disappearing when normalization supports are shuffled or noisy.