The deep neuron

For sixty years the picture was one-to-one: a neuron is a perceptron — sum your inputs, cross a threshold, fire. Then someone trained a deep network to imitate a single biological cortical cell, faithfully, spike for spike. It took five to eight layers. The dendrites — long treated as passive wires — were doing nonlinear computation all along. A real neuron may be a small deep net wearing a cell's clothing.

1943 McCulloch–Pitts: neuron = threshold unit → 1958 the perceptron → 2021 Beniaguev, Segev & London: one cortical neuron ≈ a 5–8 layer network

◔ EARLY · FRONTIER Beniaguev, Segev & London (2021, Neuron). Peer-reviewed and striking — a temporal deep net needed 5–8 layers to match one simulated cortical neuron's spikes. But the lesson for AI is open: does it mean we should build dendritic units, or just that biology is messy? The number depends on the cell type and the fidelity target. Real result, unsettled implications. This is a beginning, not a conclusion.

The gap, live — point neuron vs dendritic neuron

Both cells try to learn the same nonlinear rule (an XOR-like task — fire only when the two inputs disagree). On the left, the classic point neuron: every input flows straight to the soma, one weighted sum, one threshold — a single perceptron. On the right, a dendritic neuron: each branch first does its own nonlinear integration, then the soma combines the branches. That one extra step is a hidden layer — and it's the difference between failing and solving.

speed

left: point neuron (perceptron) parks below ~75% — one corner of the task is always wrong · right: dendritic neuron folds the space and reaches 100% · the 4 dots are the task's four input cases, ringed when the cell gets them right

The point neuron can draw only one straight cut through its inputs, so an XOR-shaped task always leaves a case stranded — exactly the wall from the perceptron's own story. Give each dendrite a nonlinearity and the cell can bend the boundary. Beniaguev's result says a real cortex neuron has far more than one such bend — enough to need 5–8 layers to copy. The humble perceptron isn't the neuron; it's one dendritic twig.

Why it's a beginning, said plainly

What's real: the demo's gap is exact (the point neuron provably can't do XOR; the two-stage one can — verified). And the published finding is real: a careful deep net needed several layers to mimic one simulated L5 cortical neuron's input-output, with the dendritic nonlinearities (NMDA spikes) as the reason depth was required.

What's unsettled: (1) "5–8 layers" is for one cell type at one fidelity — change either and the number moves; it's a characterization, not a constant. (2) This demo illustrates the principle (dendritic nonlinearity = hidden-layer power); it does not simulate a biophysical neuron — the real number comes from full compartmental models, not this page. (3) The big question — should AI replace its perceptrons with dendritic units, and would it help? — is wide open and actively debated. The finding reframes the neuron; what to do about it is the frontier.