So-called “neural networks” in AI have hardly anything to do with actual neural networks.

Excellent new study shows how enormous the the gulf is between the simplified neurons in typical AI systems and actual neurons.

For decades, neurons were modeled as point units in AI/neuroscience.

But biological neurons have dendrites and nonlinear conductances.

What capabilities are lost?

With @DavidBeniaguev @noampnueli @Segev_Lab & @mikilon, we set out to find out (https://www.biorxiv.org/content/10.64898/2026.06.08.730984v1). (2/15)

What is the mechanism? Dendrites.

PCA of dendritic voltages shows that harder tasks recruit higher-dimensional voltage dynamics, more semi-independent dendritic branches doing local computations. It's a distributed, high-rank representation across the tree. (9/15)

Our solution: TwinProp, a digital-twin-based backprop algorithm that propagates gradients through a twin to optimize input weights.

1: Train a DNN on the neuron's I/O 2: Optimize synaptic weights via frozen DNN 3: Map optimized weights back to the detailed model & verify (4/15)

We reverse-engineered the XOR solution.

The neuron splits the problem across 2 dendritic subunits: the apical tuft computes ¬P₀ ∧ P₁ and the basal tree computes P₀ ∧ ¬P₁. Their outputs converge at the soma that computes (¬P₀ ∧ P₁) ∨ (P₀ ∧ ¬P₁) = P₀ ⊕ P₁. (12/15)

TL;DR: A single cortical pyramidal neuron is not a point; it's a powerful, noise-robust, general-purpose computational unit.

Dendrites are not just wires to help neurons connect; they are the substrate for complex nonlinear computation. (15/15)

preprint: https://www.biorxiv.org/content/10.64898/2026.06.08.730984v1

A perceptron can't solve XOR, but our single neuron solves 10-dimensional parity and arbitrary random Boolean functions (~99% on 4-bit tasks), demonstrating genuine high-order feature binding, capabilities often attributed to multilayer networks. (8/15)

If one neuron can solve tasks that require networks of point units, then point-neuron brain models miss a key source of computational power.

TwinProp enables digital twins with full dendritic complexity; with connectomics, this opens a path to multi-scale brain twins. (13/15)

Can a single L5 pyramidal cell classify images of cats and dogs?

We encoded images into spike trains via a biologically plausible retina→LGN→V1→V2 model, then trained the neuron using TwinProp.

Accuracy: LIF: ~55% L5PC: ~80% 7-layer DNN upper bound: ~83% (5/15)

If one biological neuron is this powerful, what happens when we build artificial networks from dendritic units instead of point neurons?

TwinProp opens the door to bio-inspired architectures that do more with fewer units and less energy. More soon with @DavidBeniaguev. (14/15)

How far can it go? We tested parity: a classic hard benchmark for nonlinear computation, where one bit flip changes the answer.

The L5PC solved XOR (d=2) perfectly, reached 99.4% on 4-bit parity, and 91.2% on 10-bit parity: 1,024 patterns requiring nonlinear integration. (7/15)

Until now, the field lacked a way to systematically ask what can a detailed neuron model compute.

These models can't be solved analytically, and with thousands of synaptic weights to tune, manual search is intractable.

How do we overcome these optimization challenges? (3/15)

NMDA receptors are key.

As task complexity increases, NMDA currents grow monotonically, and spatial recruitment spreads across the dendritic arbor, especially into the apical tuft. (10/15)

Can a single L5 pyramidal cell recognize spoken words?

We encoded sounds into spike trains via a biologically plausible cochlea→VCN model, and:

LIF: ~55% L5PC: ~71% 7-layer DNN upper bound: ~73%

One neuron solves natural vision and speech tasks at near-DNN accuracy. (6/15)

To pinpoint the mechanisms, we performed causal ablations on 4-bit parity:

Intact L5PC: 99.4% No active conductances: 78.1% Soma-only: 76.9% No NMDA: 73.8% LIF: 68.8%

The computational power depends on dendritic structure, active conductances, and NMDA nonlinearities. (11/15)

@IdoAizenbud @threadreaderapp please #unroll

@IdoAizenbud @CursiveCrow thought you might find this interesting

@IdoAizenbud single cortical neuron is equivalent to single feature/single direction from interp work? also, the whole work gives me like déjà vu of reading early interp work from A\, like so much correspondence!

@GaryMarcus "Neural networks" is a phrase direct from the same crowd that brought you "Persil washes whiter." Lawless American capitalism is captive to its own view of life as a marketplace of grift

@IdoAizenbud @arb8020 cool work! so does this change the effective “parameter count” of the human brain? is this even a well-defined concept? do you have any estimate?