The textbook neuron — add up the inputs, fire if you cross a line — was assumed on paper in 1943. Nobody ever measured it.
You have probably seen the picture: a circle with a few lines coming in, each input multiplied by a weight, all of it summed, and an output that fires once the sum crosses a threshold. The picture is so convenient that nearly every introduction to brains and to artificial intelligence still uses it. But it is an assumption, not a measurement. Real dendrites — the bushy branches a neuron receives on — are not obedient wires: they leak, they care where an input lands, and they fire little spikes of their own. This issue is about something genuinely subversive: a serious share of the "computation" happens inside a single cell, finished before it ever decides whether to spike.
First, the parts. The branches a neuron receives on are its dendrites. An upstream cell hands its signal over at a contact point on one of those branches (that contact is a synapse — synaptic transmission). The signal flows inward toward the cell body, and the cell body is where the decision to send out a spike gets made (neurons & action potentials).
The textbook diagram treats those branches as ideal cables: whatever arrives is delivered intact. In fact they leak everywhere. The membrane is not an insulator, so current bleeds out along the way and the signal shrinks as it travels. In the 1960s Wilfrid Rall turned this into equations, solving dendrites as segments of leaky pipe — an approach known as cable theory. The conclusion is blunt: a synapse out on a distal twig may deliver only a fraction of its push by the time it reaches the cell body.
That alone is interesting enough: position is itself a weight. Move the same upstream cell's contact a little closer to the cell body and its vote counts for more — and this really happens, since synapses grow, shift and get pruned. But if the story ended here, the neuron would still be computing a weighted sum, just with distance folded into the weights. The real turn comes next.
The dendritic membrane carries gates that can open by themselves, and the most important of them is the NMDA receptor. It has an odd temper: a magnesium ion plugs it shut, and the plug only pops out once the local stretch of membrane has already been lifted by other inputs. So it isn't "open when a signal arrives" — it's "open when a signal arrives and the neighbourhood is already busy". A part that demands two things at once is called a coincidence detector.
The consequence is dramatic. If several inputs land on the same branch within a few milliseconds, they first lift that little stretch of membrane; past a certain point the NMDA gates open together, current pours in, and the branch fires a local spike of its own — a dendritic spike — amplifying the total far beyond what the inputs would sum to separately. The same number of inputs scattered across different branches lift nothing: each branch adds its modest bit, and the total is close to a straight line.
So a neuron is not an adder. It is a set of semi-independent little adders — roughly one per branch — each making its own nonlinear judgement first, with the cell body summing those judgements afterwards. In 2003 Poirazi and Mel put it bluntly: viewed as an input–output device, a cortical pyramidal neuron behaves like a two-layer artificial network, not like one unit.
In artificial networks, "one layer" versus "two layers" is not a difference of size but of capability. One layer means: multiply the inputs by weights, add them up, pass the result through a simple nonlinearity — geometrically, that can only cut the space with a single straight boundary; curved boundaries are out of reach. Add a layer, so that several local nonlinear decisions happen before the summing, and curved — even disconnected — boundaries become possible. That is the real weight of Poirazi and Mel's claim: what you were calling one unit has already crossed that threshold. Counting biological neurons against artificial units breaks down right here.
First, what XOR is. Picture a lamp with two switches and this rule: flip one and the lamp comes on; flip both and it goes off again; flip neither and it stays off. Real houses have this — the pair of switches at either end of a staircase that each control the same light. What makes it awkward is that "more input, more output" fails: halfway is brightest, and pushing further makes it darker.
This left a scar on the history of AI. In 1969 Minsky and Papert wrote a book proving that the single-layer networks of the day (perceptrons) could not compute XOR. The result pressed an entire research direction flat for years, until multi-layer networks and ways to train them matured. "A single unit cannot compute XOR" became textbook fact.
In 2020 Gidon and colleagues put a knife into that verdict in Science. Recording from dendrites of pyramidal neurons in the upper layers (2/3) of human cortex — tissue removed during epilepsy and tumour surgery — they found a previously unseen kind of dendritic spike, driven by calcium. Its most counterintuitive property: the response is not monotonic. Push the stimulus past a certain strength and the response gets smaller. That is exactly the shape XOR needs, and the authors showed this single dendritic property suffices to implement it.
The limits matter: this was measured in slices, with an electrode delivering the stimulus directly. Whether a waking human brain actually uses the mechanism this way is not known. It demonstrates a ceiling on capability, not everyday practice.
A second example, with more behavioural evidence behind it, comes from Matthew Larkum's line of work. The big pyramidal cells of cortex have two separate input zones: a bushy set at the bottom receiving feedforward signals just arriving from the world, and one long branch reaching up into the topmost layer to collect feedback flowing back down from higher areas — context, expectation, where attention has settled. The two zones do not simply add. With input to one zone the cell fires mildly; with both arriving within tens of milliseconds it erupts into a dense burst. A single cell natively multiplies "the evidence now" by "what was expected" — which is the perception-as-inference story, implemented in one cell.
In 2021 Beniaguev, Segev and London ran a very clean experiment. They took a detailed biophysical model of a layer-5 cortical pyramidal cell (the kind where every membrane gate is written into the equations with measured parameters), fed it masses of random input, recorded exactly when it spiked, and asked: how deep does an artificial network have to be before it can imitate this cell's input-to-output map?
The answer was roughly five to eight layers. The control is even prettier: strip the NMDA receptors out of the model and change nothing else, and the required depth collapses back to about one. Nearly all of that depth is held up by NMDA's "only opens if the neighbourhood is busy" temper.
The number has its boundaries: five to eight is the fit for that cell type under those input statistics; another cell class or another input distribution would move it. Nor does it mean brains beat artificial networks. It says something plainer: the word "unit" does not refer to the same object on the two sides, so comparisons like "86 billion neurons versus N billion parameters" fail on units alone.
The artificial neuron has a clear lineage: in 1943 McCulloch and Pitts, wanting to show that neural activity could be described in logic, deliberately reduced a neuron to "weighted sum plus threshold". Perceptrons inherited that skeleton, and so do today's deep networks. It was a mathematical assumption from the start, never a measurement of biology — it has just been in service so long that many people take it for a fact.
Read it the other way and it gets interesting. Artificial networks build depth by stacking layers, and every layer computes all of its units on every pass. Brains fold part of that depth inside the unit — different branches of one cell compute separately, none of them knowing what the others are doing, and only one number comes out. The first is easy to train and to parallelise; the second saves wiring and energy. Neither copied the other; they are the economical solutions on two different kinds of hardware.
Step close enough to any "smallest unit" and it opens into another network. Neuroscience is not the only field to have learned this: