Safaai et al.'s DendriNet reveals when dendritic shunting outperforms additive integration

A new paper on arXiv by Safaai, Richards, Khoshnevis, and Sabatini tackles a long-standing question in computational neuroscience: when does the shunting inhibition seen in biological dendrites actually improve neural population readout over simple additive integration? The authors introduce DendriNet, a trainable framework that systematically varies integration rules, morphology, synaptic allocation, and dendritic nonlinearities. They show that while any shunting readout can be locally linearized into an additive model, performance beyond that limit follows a gain-load-alignment principle. Shunting only helps when a reliable divisor (inhibitory signal) suppresses gain that is aligned with the task-relevant signal more than it attenuates the signal or adds denominator variability.

Through simulated hierarchies and real mouse V1 recordings, the team found that deep shunting networks can outperform linear and fitted-linear controls in designed hierarchies, but flexible nonlinear predictors (like deep networks) overtake them with enough training labels. The same support-and-reliability interaction appears in frozen-feature normalization. In V1 data, the shunting-over-additive decoder gap is largest for narrow readouts, reverses under strong private noise at the widest readout, and varies across behavioral states like running. Crucially, neither dendritic depth nor shunting is intrinsically advantageous—the benefit depends entirely on the alignment of nuisance estimates with task-relevant signals. This work provides a principled framework for when biologically realistic dendritic computation should matter for building better neural network architectures.