A neuron receives thousands of signals on its branching processes — the dendrites. Some signals push the cell toward firing (excitation), others hold it back (inhibition). For a long time, it was thought that the neuron simply adds up all the 'fors' and 'againsts.' But in real dendrites, inhibition works differently: it can not subtract but divide the excitatory signal. This is like a volume knob that attenuates sound proportionally, rather than reducing it by a fixed amount.
This mechanism is called shunting inhibition. It occurs when an inhibitory input opens channels that make the membrane 'leaky' to current right at that spot. As a result, the excitatory signal is weakened not by a constant but by a factor. The authors built DendriNet — a trainable model that can switch between summation and division on different branches of the dendritic tree. They tested whether such division helps a neuronal population read out encoded information better.
An important detail emerged. Division pays off when the inhibitory signal is reliable and targets that part of the excitation that varies with noise, not with the signal of interest. If disturbances are strong or inhibition is unstable, the simple sum starts to work better. In experiments on mouse visual cortex recordings, the advantage of division over summation was noticeable only with narrow readout of the signal. With broad readout, it disappeared or even turned into a loss. So dendritic architecture is not always an advantage per se, but a tool that is good in the skilled hands of reliable data.
as written by the authors arXiv:2607.24990
Biological neurons combine excitatory and inhibitory (E/I) activity on branched dendrites through shunting, in which inhibition divisively attenuates excitation. Whether this improves population readout over additive E/I integration of the same nonnegative inputs remains unclear. We introduce DendriNet, a trainable framework that varies integration rule, morphology, synaptic allocation, divisor locality, and dendritic nonlinearities. For population codes with multiplicative gain, a local linearization of any realizable shunting readout yields a decision direction within the positive additive E/I cone; matching the additive optimum requires a positive self-consistent shunting realization. Every scalar shunting threshold also has an exact affine additive realization. Beyond this local limit, performance follows a gain-load-alignment principle: branch-local shunting helps when a reliable divisor suppresses signal-aligned gain more than it attenuates signal or adds denominator variability. Passive additive trees flatten to linear readouts, whereas shunting trees compose local divisors. In a designed hierarchy, deep shunting outperforms tangent and fitted-linear controls, but flexible nonlinear predictors overtake it with enough labels. Support shuffling reverses the linear comparisons, sensor corruption reverses the fitted-linear comparison, and resource-matched activated training shows no consistent depth benefit. The same support and reliability interaction appears in frozen-feature normalization. Across three mouse V1 sessions, the shunting-over-additive decoder gap is largest for narrow readouts, reverses under strong private noise at the widest readout, and varies across running states. Morphology can determine where reliable nuisance estimates meet task-relevant signals, but neither depth nor shunting is intrinsically advantageous.
The same inhibitory neuron can affect its neighbors differently depending on which part of the dendrite it 'lands' on: closer to the cell body (simple subtraction, further out on the branch) division.