For decades, neuroscientists assumed that the brain's shared neural noise — the way millions of neurons fluctuate together — must eventually cap how much information a neural population can carry, raising the uncomfortable question of why evolution would sustain billions of seemingly redundant cells. A collaborative team from Kyoto University, Harvard, and UCLA has now challenged that assumption, finding in the visual cortices of mice that correlated noise slows information growth but does not halt it. The brain's extravagance, it seems, may be a form of wisdom — a system still gaining ground,
Brain's Information Capacity Not Limited by Shared Neural Noise, Study Finds
Shared noise slows growth, but does not bring it to a halt
Mark
So the basic problem is that individual neurons are unreliable. They fire differently each time you show them the same thing. How does the brain get around that?
Mimi
It uses a crowd. If you have enough neurons all encoding the same information, the noise in one gets averaged out by the signal in the others. It's called population coding.
Luke
Right, but the noise isn't random across neurons—that's the wrinkle. The neurons tend to fluctuate together. So their noise is correlated.
Mimi
Exactly. And for thirty years, people thought that shared noise would eventually cap how much information the population could carry. Like there's a ceiling you hit.
Mark
Why would that be a ceiling? Why wouldn't you just keep adding neurons and keep improving?
Luke
Because if the noise pattern overlaps with the signal pattern, adding more neurons just adds more of the same problem. You're not getting new information; you're amplifying the noise along with the signal.
Mimi
That's what everyone expected. But this team looked at actual recordings from mouse visual cortex—tens of thousands of neurons—and found that the noise does slow information growth, but it doesn't stop it.
Mark
How did they figure that out from the data?
Mimi
They used power laws. Mathematical relationships that describe how things scale. They found that even as populations got bigger, the scaling laws held their shape. That let them predict what would happen beyond what they could directly measure.
Luke
The key finding is that the signal spreads across less noisy activity patterns too. So even though some noise aligns with the signal, the signal isn't trapped by it.
Mark
So the brain's billions of neurons aren't wasteful?
Mimi
Apparently not. The growth just keeps going, even with the noise.
Luke
Though I'd note—this is five mice, all in visual cortex. Whether this holds across other brain regions or other species is still an open question.
The Pulse
- For thirty years, a foundational assumption held that shared neural noise would eventually strangle information capacity, making billions of extra neurons pointless — that assumption has now been directly challenged.
- Analyzing between 18,000 and 21,000 neurons across five mice, researchers found that information kept growing with population size rather than plateauing as correlated noise models predicted.
- The team's breakthrough came from identifying two power-law scaling relationships hidden within the noise structure, allowing them to mathematically project information growth far beyond what they could directly measure.
- The findings reframe correlated neural noise from a hard ceiling into a friction — something that slows the brain's information accumulation without ever fully stopping it.
- Beyond biology, the team's scaling framework applies to any noisy computing system, offering a new lens for designing artificial architectures that must wrestle with the same fundamental tension between signal and shared interference.
For decades, neuroscientists assumed that the brain's shared neural noise — the way millions of neurons fluctuate together — must eventually cap how much information a neural population can carry, raising the uncomfortable question of why evolution would sustain billions of seemingly redundant cells. A collaborative team from Kyoto University, Harvard, and UCLA has now challenged that assumption, finding in the visual cortices of mice that correlated noise slows information growth but does not halt it. The brain's extravagance, it seems, may be a form of wisdom — a system still gaining ground, neuron by neuron, against the tide of its own uncertainty.
Each neuron in the brain fires unreliably — show it the same image twice and it responds differently each time. The brain's answer to this problem is population coding: spread information across many neurons so that what one cannot convey reliably, a crowd can. But this solution carried a troubling asterisk. Neurons do not fluctuate independently; many rise and fall together in what researchers call noise correlation. If that shared noise aligns too closely with the actual signal, theory predicted it would impose a ceiling on information — a hard limit beyond which adding more neurons would accomplish nothing. The question haunted the field: if such a ceiling exists, why does the brain maintain billions of neurons at all?
A team from Kyoto University, Harvard, and UCLA decided to test whether that ceiling was real. They analyzed recordings from roughly 18,000 to 21,000 neurons in the primary visual cortex of five mice viewing subtly varied visual patterns, asking how much information larger and larger neural populations could extract about those differences.
What they found overturned three decades of expectation. By decomposing the noise into distinct patterns and identifying two power-law scaling relationships — one describing how noise strengths were distributed, the other capturing how closely each noise pattern aligned with the signal — the team could project information growth well beyond the populations they had directly observed. In every mouse, the math pointed to the same conclusion: correlated noise slows the rate at which information grows as neurons are added, but it does not stop that growth. Saturation is not inevitable.
The implications extend beyond neuroscience. The researchers also developed a general theory of information scaling in linearly readable systems — a framework applicable to any noisy computing architecture facing the same question of whether shared interference eventually locks performance in place. In the brain, at least, the answer is no. The abundance of neurons may not be extravagance. It may be the point.
The brain's billions of neurons do not work in isolation. Each one fires unreliably—shown the same image twice, a neuron's response will differ from trial to trial, its signal buried in noise. Yet the brain solves this problem by spreading information across many neurons at once, a strategy called population coding. What one neuron cannot reliably convey, a crowd of them can.
But there is a catch that has troubled neuroscientists for decades. Neurons do not fluctuate independently. Many of them rise and fall together, creating what researchers call noise correlation—a shared variability that can overlap with the actual signal carrying information about what the eye is seeing. If this shared noise aligns too closely with the signal itself, it could theoretically cap how much information a population of neurons can transmit, no matter how many you add. The puzzle became urgent: if shared fluctuations impose a ceiling on information, why does the brain bother maintaining billions of neurons?
A team spanning Kyoto University, Harvard, and UCLA set out to test whether this ceiling actually exists. They analyzed recordings from roughly 18,000 to 21,000 neurons in the primary visual cortex of five mice, each viewing subtly different visual patterns. The researchers then asked a straightforward question: as they looked at larger and larger groups of neurons, how much information about those visual differences could they extract?
What they discovered was unexpected. By randomly sampling different-sized populations and mathematically separating the noise in each into distinct patterns, the team identified two power laws—mathematical relationships that held their shape even when adjusted for population size. One law described how noise strengths were distributed. The other captured how closely each noise pattern aligned with the actual signal. These scaling laws, combined with the constraints of their sampling method, let the researchers predict how information would grow even beyond the populations they had directly observed.
In all five mice, the power-law exponents they calculated predicted the same thing: noise correlations would not create an upper limit on information capacity. Yes, the strongest noise components tended to align more closely with the signal. But the signal itself spread across less variable activity patterns too. This meant that shared noise slows the rate at which information grows as you add more neurons—but it does not stop that growth.
Hideaki Shimazaki of Kyoto University noted the significance of overturning three decades of expectation. The field had assumed that shared neural fluctuations would eventually cause information to saturate, reaching a plateau beyond which more neurons added nothing. The new results suggest this saturation is not inevitable. The brain's abundance of neurons may not be wasteful after all.
The implications reach beyond neuroscience. The team also developed a general theory of how information scales in systems that can be read linearly—a framework that applies to any noisy computing system facing the same question: does adding more components continue to improve performance, or does shared noise eventually lock you in place? The answer, at least in the brain, appears to be that growth continues.
Notable Quotes
Why does the brain have so many neurons if shared fluctuations impose a ceiling on information?— S. Amin Moosavi, UCLA
For three decades, shared neural fluctuations were widely expected to make information saturate. Our results show that this is not inevitable.— Hideaki Shimazaki, Kyoto University
Want the full story?
Read the original at News-Medical ↗