AI Is Nothing Like a Brain, and That’s OK

The brain’s neuronal diversity and networked complexity is lost in artificial neural networks. But computational neuroscientists — experts on both brains and computers — say that’s OK. Although the two systems have diverged along separate evolutionary paths, computer scientists and neuroscientists still have much to learn by comparing them. Infusing biological strategies could improve the efficiency and effectiveness of artificial neural networks. The latter could, in turn, be a model to understand the human brain.

With AI, “we are in the process not of re-creating human biology,” said Thomas Naselaris, a neuroscientist at the University of Minnesota, but “of discovering new routes to intelligence.” And in doing so, the hope is that we’ll understand more of our own.

An Electronic ‘Brain’

In 1958, two years after the term “artificial intelligence” was coined at a math and computer science workshop at Dartmouth College, the U.S. Navy unveiled what The New York Times called an “electronic ‘brain’” that teaches itself.

This computer, known as the “perceptron,” wasn’t particularly advanced. It could read a simple code — a series of holes punched into a card — and predict whether the next hole would appear on the right or the left, represented by a binary output: 0 or 1. The perceptron made these calculations through a series of nodes, also called neurons. But they were neurons only in spirit.

“Neurons are so much more than just nodes: They’re living cells … with DNA and organelles and specialized structures,” said Mac Shine, a systems neurobiologist at the University of Sydney.

In your brain, 86 billion neurons chitchat with one another in complex networks. They communicate by tossing molecules called neurotransmitters into the spaces between cells and catching them with arms called dendrites. These molecules can shut down a neuron or spur it to activate, which triggers a sharp burst of electricity that flows down its long tail (axon). That then triggers branches (axon terminals) on the other end of the cell to send a new wave of molecules to the next neurons in the network.

All neurons — in the brain and beyond — share this basic mechanism, but practically every other feature varies among neuron types and even individual neurons of the same type. “No two neurons look alike; they’re completely different,” Shine said. Some excite other neurons; some inhibit them. Some are long — for instance, the axons of some touch neurons extend from the base of the spinal cord to the big toe — while others are relatively small, snug within the brain. They can carry meaning not only in binary code — whether they fire or not — but also in analog, in a symphony of signals with variable patterns, strengths and frequencies. They send different neurotransmitters to encode different messages; some communicate through electrical currents instead. And, of course, neurons serve vastly different functions: regulating emotions, moving food through the gut, storing and retrieving memories, moving your hands and so much more.

Even the simplest brains in the animal kingdom feature this complexity. For example, maggots have a neuron that signals when their body has stretched too far. It has 130 inputs and 200 outputs. “And that’s just to say, ‘Hey, I’ve stretched,’” Cobb said. The more a maggot’s body stretches, the more that neuron fires. In that way it contains a frequency code, he said. “It’s not saying, ‘Oh, I’m being stretched.’ It’s saying, ‘I’m being stretched [this] much.’” That’s analog, not binary.

In contrast, artificial neurons are “caricatures of biological neurons,” Poggio said. The perceptron’s neurons, for example, take in information, analyze it and output a final answer — 1 or 0. In that sense, like a biological neuron, they fire or they don’t. But that’s about as far as the similarities go.

What made the perceptron remarkable compared with prior computers was that it could learn. For the first time, a machine — made not of tissue but of wires and circuits — displayed an ability that only biology had claimed before. Its ability to learn was itself inspired by neuroscience. In 1949, the psychologist Donald Hebb had pointed out that neural pathways in the brain are strengthened when they’re used and weakened when they’re not, an idea often summarized as “neurons that fire together, wire together.” In other words, the brain learns, in large part, by adjusting the connections between its neurons.

In the perceptron, this meant assigning to each connection a “weight” — a number that determined the input’s significance to the final output. The neural network learned by adjusting its weights based on whether its predictions were accurate. In this way, the perceptron was “the first machine which is capable of having an original idea,” according to its creator, Frank Rosenblatt.

However, the perceptron’s artificial learning didn’t take off. At the time, computing power was too limited to advance machine learning beyond this simple model, and the funding to support its development wasn’t there. “The perceptron approach was neglected, and computer science itself [became] about programming,” Poggio said. Programming meant helicopter parenting. Algorithms didn’t learn; they were told what to do. As computer scientists focused on writing ever more detailed instructions and ignored systems that could teach themselves, an AI winter set in.

But neuroscience research didn’t slow; it accelerated. The following decades revealed how different parts of the human brain, especially the visual cortex, work — laying down more biological breadcrumbs that computer scientists would later pick up.

Networked Complexity

Around the time the perceptron was unveiled, the mathematician Oliver Selfridge proposed a framework for how the brain’s visual system works. In his metaphorical “Pandemonium model,” a network of screaming demons represented firing neurons: When a demon recognized a visual feature, it would scream (the neuron would fire). For example, when presented with the letter “A,” the first set of demons would identify the left stem of the letter, and then scream to the next set of demons, who were prepared to recognize a larger pattern, such as the left stem connecting to the right stem at the letter’s apex. If that was indeed what they saw, those demons would scream to the final demon, who — as you might guess at this point — screamed when presented with a complete “A.” In that way, screaming demons could process the input information and identify the letter “A.”

It turned out that this screaming in our heads was real. Soon after, the neuroscientists David Hubel and Torsten Wiesel, by probing the visual cortex of cats, found that the visual system sends signals through layers of neurons that respond to increasingly complex details, from lines to blurry shapes to full images, a discovery that would earn them the 1981 Nobel Prize in Physiology or Medicine.

Today, the visual system is one of the best understood networks in the human brain, and a simplified version of the model Hubel and Wiesel revealed is core to modern neural networks that identify objects in images. Such algorithms, called convolutional neural networks, start by detecting edges and simple shapes in a picture, and then move through layers of artificial neurons to identify larger shapes or patterns, such as ears or faces.

The brain is made up of many such networks, taking in and processing flows of information, interacting with one another in feedback loops and constantly shifting connections. These networks make it a superb multitasker, performing a dazzling array of functions across its 100 trillion connections. Information, in the form of photons or sound waves or scent molecules and more, is captured by sensory neurons and then analyzed by layers of other neurons. They trigger networks to rapidly fire sequences to pull up memories and put information in context. If the brain deems an experience worth keeping, it stores it as a memory.

Meanwhile, the brain is also talking to the gut to assess hunger, and interpreting sunlight cues to coordinate circadian time for almost every cell in your body. One second it is daydreaming, and the next it is telling your hand to brush a spider off your shoulder. It’s planning for tomorrow and recalling yesterday, and at the same time regulating temperature, blood pressure, inflammation and heartbeat. When we sleep, the brain switches to a different mode — assembling images as dreams, consolidating memories, doing a bit of housekeeping and directing many more activities we don’t understand.

These processes are all running across various scales, from single neurons to local networks to networks that span the entire brain and even the entire body. The ever-shifting dynamics of the nervous system are made possible by neuromodulators, a subset of neurotransmitters that are slower acting and diffused more broadly across brain regions. They are “the master switches in the brain,” according to Srikanth Ramaswamy, the head of the neural circuits lab at Newcastle University.

Neuromodulators are released from elaborate dendritic trees at the ends of some neurons, and they allow the brain to adapt to new situations over seconds to minutes. For example, cortisol release during stress primes the body for action. The system is finely tuned: Studies have shown that molecules released from different branches of the same tree can impact an animal’s behavior, such as whether a mouse runs or stops.

“You would have no idea where to put that in a neural network,” Shine said. “There is complexity hidden in neuroscience that is just inaccessible to modern neural networks because they’re constructed differently.”

Crucially, an artificial neural network is not made of physical connections like neurons in the brain. The network is abstract, residing in a world of math and calculations as algorithms that are programmed into silicon chips. It’s “basically just linear algebra,” plus some other nonlinear computations, said Mitchell Ostrow, a computational neuroscience graduate student at MIT.

To reach the complexity of even one biological neuron, a modern deep neural network requires between five and eight layers of nodes. But expanding artificial neural networks to more than two layers took decades. In deeper networks, it becomes much harder to figure out which weights the network should tweak to minimize the error in its predictions. In the 1980s, the computer scientist Paul Werbos came up with an innovation called backpropagation that solved this problem.

In 1986, Geoffrey Hinton — the so-called godfather of AI who was awarded the 2024 Nobel Prize in Physics for his work on machine learning — and his colleagues wrote an influential paper about how neural networks could be trained using backpropagation. This idea, which wasn’t directly based in neuroscience, would become key to deepening neural networks and improving their learning.

In the 1990s, computer scientists finally deepened neural networks to three layers. But it wasn’t until the 2010s, when computer scientists learned to structure their algorithms to run faster calculations simultaneously on smaller chips, that neural networks deepened to dozens and hundreds of layers.

These advances led to today’s powerful neural networks, which can surpass the human brain in certain tasks. They can be trained on billions of images or words that would be impossible for a human to analyze in a lifetime. They beat human world champions in games such as chess and Go. They can predict the structure of almost any known protein in the world with a high degree of accuracy. They can write a short story about McDonald’s in the style of Jane Austen.

However, while these abilities are impressive, the algorithms don’t really “know” things the way we do, Cobb said. “They do not understand anything.” They learn mainly by recognizing patterns in their training data; to do that, they typically need to be trained by an immense amount of it.

Meanwhile, even the simplest nervous systems in the animal kingdom have knowledge. “A maggot knows things about the outside world in a way that no computer does,” Cobb said. And one maggot is different from another maggot because each learns by interacting with and gaining information from its environment. We don’t know how to infuse machines with knowledge beyond feeding it a set of facts, he said.

Artificial neural networks are simpler and not as dynamic as the systems that give them their name. They work very well for what they’ve been designed to do, such as recognize objects or answer prompts. But they have “no way to reason” like a human brain does, Ramaswamy said. “I think adding biological detail would play a huge role in enabling this.”

And that is what he is trying to do.

Infusing Biology

Because the gears of biology, honed by evolution, have proven to work pretty well in the brain, some researchers think that artificial neural networks could be improved by returning to their inspiration and better mimicking some neurobiological features. Such brain-inspired computing, also known as neuromorphic computing, doesn’t require chemistry, Ramaswamy said. Rather, it’s possible to abstract the idea of molecules into algorithmic equivalents that work across circuits.

His team has found that infusing some diversity into the way artificial neurons behave makes neural networks work better. For example, in preliminary work published on the scientific preprint site arxiv.org in 2024, his team found that programming artificial neurons to fire at different rates improved how well their systems learned. Ramaswamy is also looking at network effects seen in biological nervous systems. This year, his team theorized that designing neural networks to include the kind of information that neuromodulators provide would improve their ability to learn continuously like the brain does.

Other researchers also aim to capture the complex dynamics that exist in populations of neurons — but instead of translating that information into an algorithm, they want to speak biology’s native language. The nuclear engineer Mattia Gazzola and his team at the University of Illinois Urbana-Champaign are attaching biological neurons to artificial components to improve the computer’s ability to process information. Biological networks can evolve new computational dynamics because they’re made of neurons that innately encode just the right amount of information, he said.

“We see this as a key to access those capabilities of creativity, curiosity, learning,” Gazzola said. “These are all hallmarks of biological intelligence that we find all over the place, not only just in humans … but even in simple creatures.”

Other neuromorphic efforts focus on improving artificial neural networks’ efficiency — a race the brain is winning by a landslide. The brain is “incredibly energy efficient,” Ramaswamy said. It operates on about 20 watts of power, roughly twice that of a typical LED lightbulb. “There is no way a neural network could ever operate with just [20] watts,” he said.

However, comparing the energy use of artificial and biological neural networks isn’t straightforward because of the way a brain is trained. Do we count hundreds of millions of years of evolution, or an individual’s lifetime of learning, as training? How much energy does that demand? “There’s more debate in the field about what the comparison is,” said Kozachkov, the IBM researcher. “Are we comparing apples to oranges?”

Despite the nuances, the energy use of artificial neural networks concerns many people. The current trend toward scaling — making the networks ever bigger to increase their computational power — will continue to accelerate their energy demand until more efficient chips and processes are invented. Adding too much biological detail into algorithms comes with a significant cost in computing power, energy and other resources. “We need to identify a sweet spot where we think that this level of detail is actually useful,” Ramaswamy said.

Still, for most computer scientists, the brain is not top of mind. For many neuroscientists, however, AI is.

Model Systems

Already, AI is influencing and, in some cases, accelerating the study of biology. Researchers can train neural networks on biological data, such as proteins and genomes, with the hope that computers can help design novel, functional biological products. But they can also use AI to study the brain itself.

“We’re entering a really exciting time,” said Jenelle Feather, a research fellow at the Flatiron Institute Center for Computational Neuroscience* who will soon lead a lab at Carnegie Mellon University. “We are now moving on to a place where we can actually take advantage of these models to learn something new about the brain.”

Where the brain was once a model for artificial neural networks, now neural networks are models for the brain. Despite their many differences, today’s AI algorithms are significantly more similar to parts of a real brain than any models we’ve had in the past, Feather said. And there are other benefits. Neuroscientists typically study animals to try to understand human neurobiology. But an algorithm doesn’t need to be fed, housed or kept alive. Plus, neural networks have a leg up on rodents on writing essays, solving math problems and playing chess.

A series of studies have suggested that despite having vastly different architectures, artificial and biological neural networks could behave in similar ways. In a preprint published in 2024 on biorxiv.org, Ev Fedorenko, who studies language and the brain at MIT, and her team found that the activity of different neural networks can mimic each other and the activity of the brain in what’s called “universality of representation.” In another study, they found that artificial neural networks and the human brain can process sentences in a similar way. The fact that they have “some similarity to the brain — that’s really exciting,” Fedorenko said.

But any similarities are hard to identify given that neural networks and the brain are close to “black boxes”: It’s hard to see from the outside how exactly a neural network or a brain work on the inside. In that sense, comparing the two is like “the myopic leading the myopic,” Naselaris said. Still, there’s much to gain and little to lose from comparing our two smartest networked systems, despite their many differences. “It will never not be interesting to compare humans to AI,” he said, similar to how we never tire of comparing animal and human minds.

But they’re not the same — and probably never will be. “[Neural networks] are now sufficiently different from the way that actual brains are in so many different ways that I think it’s actually more sensible to think of them as a really different information-processing object,” said Shine, the systems neurobiologist, “one that’s extremely interesting in its own right.”

If you ask ChatGPT-4, “Are you similar to a human brain?” it converts your text into numbers, processes them through hundreds of thousands of nodes — once, twice, many times. Each time, it predicts the best next word to say back to you.

Yes and no, it might write. Sometimes, just like the brain, it needs some feedback. “Do you think of me like a brain, or more like something else?”

Editor’s note: The Simons Foundation funds both the Flatiron Institute Center for Computational Neuroscience and this editorially independent publication. Simons Foundation funding decisions have no influence on our coverage.[back]