Mind in the mirror
Surya Ganguli is helping pioneer a new science of intelligence—one that studies how brains perceive, how AI learns, and what each can reveal about the other.
)
Illustrations: Jonathan Chaves
“The human brain runs on 20 watts. For reference, our old lightbulbs ran on 100 watts,” says Surya Ganguli. “That means every one of us is dimmer than a lightbulb.”
Within that modest energy footprint, roughly 86 billion neurons fire in intricate networks, sending signals that allow us to recognize a face, follow a conversation, or navigate a crowded street. Training a state-of-the-art AI model, by contrast, can require millions of watts of power—along with vast data centers, enormous financial investment, and a growing environmental cost.
For Ganguli—an associate professor of applied physics and senior fellow at the Stanford Institute for Human-Centered Artificial Intelligence—that contrast reveals a basic trade-off: Faster or more precise answers usually require more energy. Much of modern AI has improved performance simply by scaling up—more data, more processing power, more electricity.
The brain, however, seems to solve this problem differently.
)
Surya Ganguli
The contrast becomes even clearer when we look at how learning happens. AI systems typically train on enormous datasets scraped from the internet, absorbing patterns from vast streams of text and images with little guidance about what matters most.
Ganguli sometimes sees a faint reflection of that process at home. His 4-year-old daughter occasionally uses words he has never taught her. When he asks what they mean, she shrugs: She doesn’t know—but she often uses the word correctly.
If we trained children the way we train AI systems, Ganguli notes, we would simply push them into the world and hope they picked up the right information. Instead, parents read stories, teachers design lessons, and environments are carefully structured over time.
How the brain achieves such flexibility, efficiency, and resilience without the staggering energy demands of today’s AI systems lies at the heart of Ganguli’s research—and points to a deeper question: how intelligence emerges at all.
)
Order from noise
If you zoom in far enough, the brain does not look that intelligent. A single neuron simply receives signals, performs a small calculation, and passes the result along. There is no thought, no imagination, no reasoning inside any one cell.
And yet, when billions of these simple units interact, perception, memory, imagination, and reasoning emerge.
Artificial neural networks follow the same basic principle. Their components—mathematical abstractions inspired by neurons—are simple on their own. But connect enough of them and train them at scale, and they can recognize patterns, generate language, and produce images.
“We can use AI as a lens aimed back at ourselves,” says Ganguli.
That instinct—to treat machines as mirrors—has shaped Ganguli’s thinking from the beginning.
As an undergraduate at MIT, Ganguli arrived intending to study artificial intelligence. But the field he encountered—focused on symbolic reasoning rather than neural networks—left him uneasy. During one lecture, he raised this question: If intelligence emerges from the brain, shouldn’t computer scientists try to reverse engineer it?
He still remembers his professor’s response: “He said, ‘Ignore the brain. It’ll just confuse you. We just need to figure out the software the brain is running.’”
Unsatisfied, Ganguli began drifting across disciplinary boundaries, taking any course that fascinated him. By the time he graduated, he had unintentionally triple majored—in mathematics, physics, and electrical engineering and computer science. He later turned to string theory for his doctorate, becoming less interested in building AI systems and more interested in discovering the principles that allow simple systems to give rise to complex behavior.
That perspective would lead to a breakthrough in 2015.
At the time, Ganguli was thinking about thermodynamics, not AI benchmarks. The second law describes one of the most reliable tendencies in nature: Heat spreads, structures decay, the universe drifts toward noise.
But what if a learning system could run that process in reverse?
Instead of watching structure collapse into randomness, a neural network might be trained to transform randomness into structure. The trick was to take real images, corrupt them step by step with increasing amounts of random noise, and train the network to reverse each step—learning, in effect, to reconstruct order from chaos. Beginning with a field of random static—like the snow on an untuned television—the system gradually learned to impose order on noise. With each step, faint outlines sharpened into recognizable forms until an image emerged.
That process became the foundation of diffusion models—a technique that would later power some of the most striking advances in generative AI, including DALL-E, Stable Diffusion, and Midjourney—tools now capable of generating photorealistic images and artwork from nothing more than a text description.
)
This diagram shows the main technical idea behind diffusion models. Take a bunch of training images, add increasing amounts of noise to them (diffusion) and train a neural to reverse this process (reverse diffusion). Courtesy: Surya Ganguli
In more recent work, Ganguli developed the first theory explaining how creativity arises in these systems. When diffusion models generate images by reassembling digital noise, they work patch by patch, without any awareness of how each piece will fit into the whole—like a worker tiling a mosaic who can only see the square directly in front of them. This narrow focus is precisely what makes the models creative: Because each patch is assembled independently, different parts of an image can draw from different sources, producing combinations that never existed in the training data. Rather than a mysterious accident, this creativity emerges from the way the models are built.
But if one thread of Ganguli’s research asks how minds—artificial or otherwise—generate something new, another asks how they stay grounded in reality. The two questions are, in a sense, mirror images: What allows a system to reach beyond what it knows, and what keeps it from losing its grip on reality? For Ganguli, they are not separate pursuits but two sides of the same inquiry into the nature of intelligence.
)
The brain’s balancing act
Ganguli wanted to see what would happen when the brain’s internal predictions grew louder than signals arriving from the outside world. To find out, he turned to a colleague: Karl Deisseroth, the Stanford neuroscientist who pioneered optogenetics—a technique that activates specific neurons with pulses of light.
This collaboration reflects why Ganguli came to Stanford in the first place—the rare opportunity to work simultaneously across departments as disparate as applied physics, neurobiology, and computer science.
“Often the most interesting discoveries happen when you’re not looking for them.You’re just probing a system to see how it behaves.”Surya Ganguli
The system he and Deisseroth were probing was the brain’s machinery for perception.
At every moment, the brain balances two streams of information. Bottom-up signals arrive from the outside world—from the eyes, ears, and other senses—carrying evidence about what is happening around us. Top-down signals originate within the brain itself: expectations, memories, and predictions about what we are likely to see or hear. When you read a sentence, your brain often predicts the next word before your eyes reach it. Those predictions help perception unfold quickly and efficiently. But the two streams must stay in balance. If internally generated signals grow strong enough, they can begin to crowd out sensory evidence—what you imagine or remember, in effect, would seem more real than the world outside (think The Matrix).
Working with rodents, the researchers used optogenetics to selectively activate neurons that send top-down signals into the cortex, while recording activity from tens of thousands of nearby neurons. As the data accumulated, the same small population of cells kept lighting up—regardless of which top-down neurons the researchers activated.
Looking more closely, they realized they were seeing something neuroscientists had not previously recognized: a distinct functional cell type, present in both rodent and human brains, that appears to stabilize the circuit. When internally generated signals begin to cascade through the network, these cells activate to dampen the activity—a neural fail-safe that keeps the system from spinning out of control.
The implications reach beyond basic neuroscience. Without that stabilizing influence, internally generated signals could begin to crowd out sensory evidence—a dynamic that may help explain how hallucinations arise. Deisseroth, who studies schizophrenia, found this cell type absent in patients with the disorder. If confirmed, the discovery could point toward an entirely new class of treatments for conditions that have long resisted them.
The connection to AI runs deeper than analogy. To analyze the recordings from tens of thousands of neurons, Ganguli’s team relied on the same computational tools used to study large neural networks in machines—techniques for finding structure in high-dimensional data that Ganguli had helped develop. At Stanford HAI, where Ganguli serves as associate director, this is precisely the kind of work the Science of Intelligence initiative was built to pursue—using AI and neuroscience together, each as a lens on the other.
While Ganguli sees ample room to improve the efficiency, accuracy, and explainability of generative AI, he is clear about which system presents the greater challenge. “The brain is still much more of a black box than AI,” he says.
“We really need to come to terms with how a complex distributed circuit learns and computes—whether that circuit is a brain or a machine.”
Karl Deisseroth is the D. H. Chen Professor and a professor of bioengineering and of psychiatry and behavioral sciences.
)
:focal(976x555:977x556))
:focal(740x1051:741x1052)){kind=spacer}
:focal(745x365:746x366))
)