Summary:
Challenging a long-held assumption in sensory neuroscience, researchers have shown that correlated fluctuations, often called noise correlations, do not impose a strict upper limit on the amount of sensory information large neural populations can encode. By analyzing recordings from more than 20,000 neurons in the primary visual cortex (V1) across mammalian models, the team discovered scale-invariant power-law relationships that allow information capacity to continue growing as population size increases.
Key Facts:
- Overturning a 30-Year Hypothesis: For decades, theoretical neuroscience held that shared trial-to-trial fluctuations—noise correlations—would inevitably create a hard ceiling on how much sensory information a population of neurons can represent.
- Discovery of Invariant Power Laws: Analysis of large-scale visual cortex recordings revealed two scale-invariant power laws governing how noise strength is distributed across activity modes and how those noise patterns align with sensory signals.
- Information Continues to Scale: Although shared noise slows the accumulation of information as neurons are added, the sensory signal projects into many quieter dimensions, preventing information from saturating.
Source: Kyoto University
The Puzzle of Unreliable Neurons
The brain builds an internal representation of the sensory world using the combined activity of billions of cortical neurons. Yet individual neurons are notoriously variable: present the same visual stimulus repeatedly and a single neuron’s firing will vary widely from trial to trial.
To overcome this inherent variability, the nervous system relies on population coding—distributing sensory information across many neurons so that averaging and collective patterns can reveal reliable signals.
Still, neurons do not act independently. Their membrane potentials and spike times commonly fluctuate together. For about thirty years, many computational models argued that if those shared fluctuations align with stimulus-driven activity, adding more neurons would eventually offer diminishing returns and lead to an informational ceiling.
“A central question is why the brain invests so heavily in large numbers of neurons if shared fluctuations impose a strict information limit,” said S. Amin Moosavi, Ph.D., UCLA.
Mapping 20,000 Visual Cortex Neurons
An international research team from Kyoto University, Harvard University, and UCLA reanalyzed large-scale two-photon calcium imaging and electrophysiology recordings that captured between 18,000 and 21,000 neurons at once in mouse V1 as animals discriminated subtle visual differences.
They measured how much visual information could be linearly decoded while progressively sampling larger and larger subsets of neurons.
Using iterative subsampling and mathematical decomposition of population noise into eigenmodes, the investigators identified two consistent power-law relationships that remained invariant across different population sizes:
- Noise Amplitude Scaling: A power law governing the distribution of noise magnitudes across network modes.
- Signal–Noise Alignment: A power law describing how each noise mode’s orientation aligns with the stimulus-related signal axis.
With these scaling laws and careful subsampling analysis, the team extrapolated how information would scale beyond the physically recorded neural population.
Information Slows Down, But Never Stops Growing
Across all datasets, the empirically measured power-law exponents contradicted the traditional saturation view. Although the largest noise components did tend to align more with the stimulus, the sensory signal also occupied a broad subspace of lower-variability modes. In other words, quieter dimensions provided room for the signal to expand as more neurons were added.
As a result, correlated noise reduces the speed at which information accumulates with added neurons, but it does not halt information growth entirely.
“For thirty years many believed shared neural fluctuations would force information saturation,” said Hideaki Shimazaki, Ph.D., Kyoto University. “Our results show that saturation is not inevitable when the full high-dimensional noise structure is taken into account.”
Blueprints for Fault-Tolerant Computing
Beyond resolving a core question in sensory neurobiology, this mathematical framework offers a general theory of how linearly decodable information scales in complex stochastic networks. The scaling principles are relevant to neuromorphic hardware, distributed artificial intelligence systems, and other computing architectures that face correlated noise.
Engineers designing dense silicon circuits or noisy computing platforms can draw on these biological insights to build fault-tolerant systems where adding processing units continues to increase representational precision rather than hitting a fixed ceiling.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by staff editors.
About this Theoretical Neuroscience Research:
- Media Contact: Whitney Hubbell
- Source: Kyoto University
- Image Credit: Image credited to Neuroscience News
- Original Research (Open Access): Science Advances (Sept 25, 2026). Title: “Population coding under the scale invariance of high-dimensional noise.” Authors: S. Amin Moosavi, Sai Sumedh R. Hindupur, and Hideaki Shimazaki.
- DOI: 10.1126/sciadv.adz9632
Abstract
Population coding under the scale invariance of high-dimensional noise
Scale-invariant, high-dimensional neural activity appears across brain areas and species, but its consequences for information coding are not fully understood. We examined how stimulus information in mouse V1 scales with the number of neurons: does information saturate because of noise correlations, or can it grow without bound as subpopulations expand?
Contrary to earlier reports, we find that dominant noise components that grow linearly with population size are not aligned closely enough with the signal to impose a strict bound. This conclusion rests on two scale-invariant power-law properties of neuronal responses in mouse V1: the noise eigenspectrum and the alignment between noise components and the signal.
We show that population subsampling links observed power-law exponents to whether information is bounded, and that information scaling depends on the full eigenspectrum rather than only the leading modes. Finally, under subsampling, any information-limiting correlations, if present, must take the form of differential correlations.
These results clarify how sensory information scales in high-dimensional neuronal activity under scale-invariant noise and provide a unifying framework for understanding information growth in large networks.