Two-Way Traffic
Artificial intelligence began as an attempt to copy the brain, and neuroscience has since borrowed back: an algorithm built for machines turned out to describe dopamine, and AI now maps brains and gives voices to people who lost theirs. Eighty years of traffic in both directions.
The words give it away. Artificial intelligence talks of neurons, networks and learning because it began as an attempt to imitate the brain. For decades the influence seemed to run one way, from biology to engineering. It no longer does. Ideas invented to train machines have become theories of how real neurons work; machines trained for practical tasks have become the best available models of parts of the brain; and AI has become an instrument that neuroscience can no longer do without. This essay follows the traffic in both directions, and ends where the two fields part company. It builds on the comparison of models and brains in How a Language Model Learns.
Copying the neuron, then the visual cortex
The first artificial neuron was a theory of the real one. In 1943 the neurophysiologist Warren McCulloch and the logician Walter Pitts described a neuron as a unit that fires when its combined input passes a threshold: “At any instant a neuron has some threshold, which excitation must exceed to initiate an impulse.” Networks of such units, they showed, could compute logical functions.[1] Six years later the psychologist Donald Hebb proposed how such networks might learn: when one cell “repeatedly or persistently takes part in firing” another, “some growth process or metabolic change takes place” so that the connection strengthens.[2] The slogan “cells that fire together wire together”, often attributed to him, came decades later. Frank Rosenblatt’s perceptron of 1958, the first machine that learned from examples, was presented explicitly as a model of “information storage and organization in the brain”.[3]
The most consequential borrowing came from vision. In 1959 and 1962 David Hubel and Torsten Wiesel recorded single neurons in the visual cortex of cats and found a hierarchy: cells responding to edges at particular angles, feeding cells that responded to more complex patterns regardless of their exact position. The work won them the Nobel Prize in 1981.[4] In 1980 Kunihiko Fukushima built the “neocognitron”, a network whose structure, he wrote, became “similar to the hierarchy model of the visual nervous system” of Hubel and Wiesel.[5] In 1989 Yann LeCun and colleagues trained a network of that layered kind with backpropagation to read handwritten postal codes.[6] When a much larger descendant won the ImageNet image-recognition competition in 2012, with an error rate of 15.3 per cent against 26.2 per cent for the runner-up, the modern era of AI began — on a design traced back to the cat’s visual cortex.[7]
Not every key idea came from biology. Backpropagation, the method that trains almost all networks, was described in 1986 as a procedure “for networks of neurone-like units”, but nobody has shown that the brain learns this way.[8] Whether it does something similar remains an open question, discussed below.
An algorithm found inside the brain
The clearest case of influence in the other direction concerns learning from reward. In 1988 the computer scientist Richard Sutton proposed “temporal-difference” learning: rather than waiting for the final outcome, a system improves its predictions using “the difference between temporally successive predictions”. It was an engineering method for teaching machines to predict and act.[9]
Neuroscientists recording from dopamine neurons in monkeys had noticed something puzzling. These cells fired when an animal received an unexpected reward; but once a cue reliably predicted the reward, they fired at the cue instead, and when an expected reward failed to arrive, their activity dropped below its usual level. In 1997 Wolfram Schultz, Peter Dayan and Read Montague showed that this is exactly the “prediction error” at the heart of temporal-difference learning: dopamine neurons appear to signal “changes or errors in the predictions of future salient and rewarding events” (Figure 2).[10] An idea from computer science became one of the most influential theories in neuroscience, with consequences for the study of motivation and addiction.
It happened again in 2020. AI researchers had found that machines learn better if they predict not just the average reward but its whole range of possible values. Researchers at DeepMind and Harvard reasoned that if the brain did the same, individual dopamine neurons should differ, some systematically optimistic and others pessimistic. Re-analysing recordings from mice, they found exactly that.[11] The machine had generated a prediction about the brain, and the brain confirmed it.
Traffic in this area ran both ways at once. DeepMind’s 2015 system that learned to play Atari games used “a biologically inspired mechanism termed experience replay”, storing past experiences and replaying them during training; the authors noted that the hippocampus does something similar, with “the time-compressed reactivation of recently experienced trajectories during offline periods”, for instance in sleep.[12] Demis Hassabis and colleagues argued in 2017 that “better understanding biological brains could play a vital role in building intelligent machines”.[13]
Machines trained for tasks end up resembling brains
A second kind of influence began around 2014: networks trained for practical tasks turned out to be good models of brain areas they were never designed to imitate. Daniel Yamins, James DiCarlo and colleagues trained networks only to recognise objects in images and then compared their internal activity with recordings from the monkey’s inferior temporal cortex, a late stage of visual processing. “Even though we did not constrain this model to match neural data”, they wrote, its top layer predicted the neurons’ responses better than earlier models built for the purpose.[14] A public benchmark, Brain-Score, now ranks AI models by how closely they match brain data.[15]
The same happened with space. The brain has “grid cells”, which fire in a regular hexagonal pattern as an animal moves and help it keep track of where it is. In 2018 two groups independently trained networks to navigate, and grid-like units appeared in them without being designed in.[16][17] One interpretation is that grid cells are what any efficient navigation system tends to develop.
Then language. Several studies found that the internal activity of language models predicts how the human brain responds to the same sentences, and that the similarity “primarily depends on their ability to predict words from context”.[18][19] Recording from electrodes placed on the brains of nine patients as they listened to a 30-minute podcast, Ariel Goldstein and colleagues found that people, like models, are “engaged in continuous next-word prediction before word onset”.[20]
These findings need care. That a model predicts brain activity shows that the two represent some things in similar ways, not that they compute the same way: a model can match a brain area’s responses while working by entirely different means, and the brains studied are those of monkeys, mice or a few patients. Still, the method has changed neuroscience. Instead of guessing what a brain area does and building a model by hand, researchers now train networks on plausible tasks and test which ones the brain resembles.
Mapping wiring, decoding speech
The most practical influence is the simplest: AI has become an instrument. Mapping a brain’s wiring means tracing every neuron through millions of microscope images, far beyond human effort. In 2024 the FlyWire consortium published the complete wiring diagram of an adult fruit fly’s brain, “5 × 107 chemical synapses between 139,255 neurons”. A neural network segmented the cells automatically; people then spent the equivalent of about 33 years correcting it.[21] The same year, Google and Harvard researchers reconstructed a cubic millimetre of human cortex, from tissue removed during surgery: “about 57,000 cells, about 230 millimeters of blood vessels, and about 150 million synapses”, in 1.4 petabytes of images.[22]
The most moving use is giving speech back. Implants that record brain activity, combined with AI decoders that turn it into words, have advanced quickly. In 2023 a person with ALS spoke through one at 62 words per minute, “3.4 times as fast as the previous record”; a second team, recording from the surface of the brain of a person with severe paralysis, reached a median of 78 words per minute, with a digital avatar that spoke for them.[23][24] In 2024 a man with ALS used an implant that “sustained 97.5% accuracy over a period of 8.4 months” at about 32 words a minute.[25]
The everyday work of laboratories has changed too. A tool called DeepLabCut, released in 2018, tracks the movements of animals in video after people label only about 200 frames, replacing weeks of manual scoring.[26] None of this makes AI a theory of the brain; it makes it the microscope and the stethoscope of a new kind of neuroscience.
The brain does not learn like a transformer
For all the borrowing, today’s most powerful AI works in ways the brain almost certainly does not. Language models learn by backpropagation, read trillions of words, attend to all of a text at once, and stop learning when training ends. The brain learns continuously from a lifetime of far less data, through a body, on about 20 watts. Whether the brain does anything like backpropagation is still argued: a 2020 review by Timothy Lillicrap, Geoffrey Hinton and colleagues proposed that feedback connections in the cortex might produce the differences in activity needed to approximate it, but there is no confirmed mechanism.[27]
The differences are now pushing some researchers back towards the brain. Chips that imitate spiking neurons promise far greater efficiency: IBM’s TrueNorth, with a million artificial spiking neurons, processed video using 63 milliwatts in 2014, although such chips have not displaced conventional hardware for large models.[28] Learning without forgetting, learning from few examples and learning through action — the subjects of The Body Problem — are where AI is furthest from biology, and where many expect the next ideas to come from.
A relationship, not a copy
The record supports three conclusions. Neuroscience gave AI its founding ideas — the model neuron, learning by strengthening connections, the layered visual system — but not the method that trains it, and modern AI owes as much to engineering, data and computing power as to biology. AI has given neuroscience more than tools: the reward prediction error is a theory that came from computer science and survived contact with real neurons, and task-trained networks have become standard models of the brain. And the two have diverged where it matters most for both: efficiency, continual learning and the body.
The most useful lesson may be methodological. Each field now uses the other as a test. If a network trained only to see or to navigate develops what the brain has, that suggests the brain’s solution is not an accident; if the brain manages something no network can, that marks where AI is missing an idea. Eighty years after McCulloch and Pitts, the traffic is still running both ways.
This piece was written collaboratively with Claude Opus 5.5 (Anthropic): human specification, editorial direction and critical review; machine research and drafting. Every citation, figure and quotation was checked against the original paper, its abstract or the publisher’s record on 3 October 2026. Two common claims were corrected in checking: the slogan “cells that fire together wire together” is not in Hebb, and the TrueNorth chip’s reported power was 63 milliwatts. Figure 1 dates are those of the cited publications; Figure 2 is a schematic. The cover is computed by scripts/two_way_traffic_cover.py.
Authored by: Luis Matos Ferreira — Physicist, Developer, Writer
- How a Language Model Learns — what is inside a model, and how it compares with a brain.
- The Body Problem — what changes when AI controls robots, and whether intelligence needs a body.
- How an AI Agent Works — the series glossary.
- W. S. McCulloch and W. Pitts, “A logical calculus of the ideas immanent in nervous activity”, Bulletin of Mathematical Biophysics 5, 1943, doi.org.
- D. O. Hebb, The Organization of Behavior, Wiley, 1949, chapter 4, p. 62.
- F. Rosenblatt, “The perceptron: a probabilistic model for information storage and organization in the brain”, Psychological Review 65(6), 1958, doi.org.
- D. H. Hubel and T. N. Wiesel, “Receptive fields of single neurones in the cat’s striate cortex”, Journal of Physiology 148, 1959; and “Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex”, 160, 1962, doi.org.
- K. Fukushima, “Neocognitron”, Biological Cybernetics 36, 1980, doi.org.
- Y. LeCun et al., “Backpropagation Applied to Handwritten Zip Code Recognition”, Neural Computation 1(4), 1989, doi.org.
- A. Krizhevsky, I. Sutskever and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks”, NIPS 2012, nips.cc.
- D. E. Rumelhart, G. E. Hinton and R. J. Williams, “Learning representations by back-propagating errors”, Nature 323, 1986, doi.org.
- R. S. Sutton, “Learning to predict by the methods of temporal differences”, Machine Learning 3, 1988, doi.org.
- W. Schultz, P. Dayan and P. R. Montague, “A neural substrate of prediction and reward”, Science 275, 1997, doi.org.
- W. Dabney et al., “A distributional code for value in dopamine-based reinforcement learning”, Nature 577, 2020, doi.org.
- V. Mnih et al., “Human-level control through deep reinforcement learning”, Nature 518, 2015, doi.org.
- D. Hassabis, D. Kumaran, C. Summerfield and M. Botvinick, “Neuroscience-Inspired Artificial Intelligence”, Neuron 95, 2017, doi.org.
- D. L. K. Yamins et al., “Performance-optimized hierarchical models predict neural responses in higher visual cortex”, PNAS 111, 2014, doi.org.
- M. Schrimpf et al., “Brain-Score: Which Artificial Neural Network for Object Recognition is most Brain-Like?”, bioRxiv, 2018, doi.org.
- A. Banino et al., “Vector-based navigation using grid-like representations in artificial agents”, Nature 557, 2018, doi.org.
- C. J. Cueva and X.-X. Wei, “Emergence of grid-like representations by training recurrent neural networks to perform spatial localization”, ICLR 2018, arxiv.org.
- C. Caucheteux and J.-R. King, “Brains and algorithms partially converge in natural language processing”, Communications Biology 5, 2022, doi.org.
- M. Schrimpf et al., “The neural architecture of language: Integrative modeling converges on predictive processing”, PNAS 118(45), 2021, doi.org.
- A. Goldstein et al., “Shared computational principles for language processing in humans and deep language models”, Nature Neuroscience 25, 2022, doi.org.
- S. Dorkenwald et al., “Neuronal wiring diagram of an adult brain”, Nature 634, 2024, doi.org.
- A. Shapson-Coe et al., “A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution”, Science 384, 2024, doi.org.
- F. R. Willett et al., “A high-performance speech neuroprosthesis”, Nature 620, 2023, doi.org.
- S. L. Metzger et al., “A high-performance neuroprosthesis for speech decoding and avatar control”, Nature 620, 2023, doi.org.
- N. S. Card et al., “An Accurate and Rapidly Calibrating Speech Neuroprosthesis”, New England Journal of Medicine 391, 2024, doi.org.
- A. Mathis et al., “DeepLabCut: markerless pose estimation of user-defined body parts with deep learning”, Nature Neuroscience 21, 2018, doi.org.
- T. P. Lillicrap, A. Santoro, L. Marris, C. J. Akerman and G. Hinton, “Backpropagation and the brain”, Nature Reviews Neuroscience 21, 2020, doi.org.
- P. A. Merolla et al., “A million spiking-neuron integrated circuit with a scalable communication network and interface”, Science 345, 2014, doi.org.
Comentários
Enviar um comentário