Scientists have long operated under the assumption that the remarkable abilities of the human brain are primarily due to the vast number of neurons and the extensive networks they form.
New research suggests the story begins much earlier. Researchers have discovered that a single human brain cell can perform computations complex enough to rival a small artificial neural network. The findings were published in the journal Proceedings of the National Academy of Sciences.
The human cell also proved to be a far more sophisticated information processor than comparable cells in rats. The results indicate that human intelligence depends not only on the number of neurons but also on the amount of computational operations each individual neuron performs before transmitting information to the next cell.
The work was led by Professor Idan Segev from the Hebrew University of Jerusalem (HUJI), whose team set out to address a question for which no proper evaluation metric had long existed.
Human cortical cells look noticeably different from those of other animals. No one had a standard way to convert this visible difference into a single number. Without it, the claim about the uniqueness of human neurons remained qualitative, easily asserted but difficult to verify.
So the team borrowed a method from the field of machine learning. For each real neuron, they created a detailed computer model of its shape and electrical behavior, then trained an artificial neural network to mimic the cell’s responses to incoming signals.
The idea behind this method is simple. A neuron receiving a stream of input data responds with a sequence of electrical impulses, and the artificial copy attempts to predict these impulses with millisecond precision. If the copy consistently fails, it means the cell is performing a computation too complex for a shallow neural network to capture.
The focus of this work was on pyramidal neurons—large, roughly triangular cells that make up most of the cerebral cortex and transmit the majority of signals.
The researchers modeled about two dozen such models, 12 from humans and 12 from rats, sampling from different layers of the cerebral cortex. On a new scale from zero to one, the average score for human cells was about 0.38, while for rat cells it was about 0.22. The difference was substantial and statistically significant, not the result of a few unusual cells.
The difference was especially striking in individual comparisons. For one pair of upper-layer cells, the human neuron performed more than twice as well as its rat counterpart, and reproducing it required a noticeably deeper artificial neural network. The study shows that a single human pyramidal cell behaves not so much like a simple switch but rather like a small multi-layered network.
“People often imagine a neuron as a simple switch that is either on or off. We show that a single human neuron is itself an extraordinarily complex computational device,” said Segev.
This methodology builds on earlier work by the same research group. In that project, it was found that copying a single rat neuron could create an artificial neural network with a depth of five to eight layers. In the previous study, the main complexity was attributed to one type of synapse, and the new index turns this idea into a general metric applicable across different species and layers.
Human brain cells differ in two key features. The first is dendrites, branched filaments that collect incoming signals from a neuron. Human dendrites are larger and have a more complex branching structure than those of other species. The total surface area of this branched tree turned out to be the strongest predictor of cell complexity, explaining about three-quarters of the variation.
Size was not the only factor. How this length was distributed also mattered. Cells that invested more effort in developing internal branched segments, rather than long straight trunks, turned out to be structurally more complex.
The second feature is synapses—junctions where one neuron transmits signals to another. Human synapses rely heavily on NMDA receptors, whose response is far from stable.
Below a certain threshold, they react only weakly. Once about 35 nearby inputs fire simultaneously, the response jumps sharply, creating a small local spike that amplifies the signal. Complexity is also unevenly distributed across the cerebral cortex. In rats, the most complex cells are in the deep layer 5, but in humans, the order is reversed, with upper-layer cells in layer 2/3 taking the lead.
An earlier study showed that these same human cells can solve a logical problem that was long thought to require an entire network of neurons.
The finding complicates the long-standing narrative about what makes the human brain exceptional. This observation has been known for a while. Ever since anatomist Santiago Ramón y Cajal drew cortical cells over a century ago and noted how large and intricate they are in humans, this complexity has been obvious but difficult to pin down numerically.
The quantitative analysis suggests that the sheer number of neurons tells only part of the human story. The rest depends on how much computation each cell performs on its own before passing the signal along. There is also a practical aspect. Modern artificial networks are built from primitive, switch-like elements stacked in enormous numbers.
The study suggests that much smaller networks could perform the same work if each unit carried more of the load, like a living neuron. Previously, there was no way to measure a neuron’s computational power and relate it to its physical structure.
This tool now exists and can be applied to other cell types, other species, and ultimately to how billions of these tiny processors collectively form a thinking brain.
About the Author
