Illustration of information growth as neurons are added: continued growth (upper) or a ceiling (lower).
Illustration Credit: Robin Hoshino
Scientific Frontline: Extended "At a Glance" Summary: Neural Scaling Laws and Population Coding
The Core Concept: A mechanism where populations of neurons compensate for individual unreliability by collectively representing the same information, challenging previous theories that shared noise limits information capacity.
Key Distinction/Mechanism: Shared neuronal fluctuations (noise correlation) were thought to impose a ceiling on information representation. However, new research shows that while shared noise slows information growth as more neurons are added, it does not bring it to a halt; information continues to scale.
Major Frameworks/Components:
- Population coding: The strategy where neuron populations compensate for individual variability.
- Noise correlation: Shared variability when many neuronal signals rise and fall together.
- Neural scaling laws: Two power laws describing the distribution of noise strengths and how each noise pattern aligns with the signal.
Branch of Science: Neuroscience, Computational Biology, Information Theory.
Future Application: Informs the design of noisy computing systems by answering whether adding more components continues to improve accuracy or if shared noise imposes a ceiling.
Why It Matters: This fundamental discovery overturns three decades of assumptions about information saturation in the brain, proving that the addition of neurons continues to increase information capacity despite shared fluctuations.
Our brains internally represent the outside world through the coordinated activity of billions of neurons. A single neuron responds unreliably; even when the same stimulus is shown repeatedly, its activity varies from trial to trial. Neuron populations can compensate for this variability by representing the same information across many neurons—a strategy known as population coding.
However, neurons do not always fluctuate independently. When many neuronal signals rise and fall together, their shared variability—known as noise correlation—can overlap with the pattern of activity that carries information about a stimulus and may eventually cap the information the population can convey. This possibility puzzled an international team of researchers from Kyoto University, Harvard University, and the University of California, Los Angeles.
"We face a fundamental question: Why does the brain have so many neurons if shared fluctuations impose a ceiling on information?" asks S. Amin Moosavi of UCLA.
The researchers reanalyzed recordings of approximately 18,000 to 21,000 neurons from the primary visual cortex of each mouse viewing subtly different visual stimuli. They then examined how much information about these differences could be read from increasingly large groups of neurons.
By repeatedly drawing random subpopulations of different sizes and separating the noise in each into distinct activity patterns, the team found two power laws that retained the same form after adjusting for population size: one describing the distribution of noise strengths and the other describing how each noise pattern aligned with the signal. These scaling laws, combined with the constraints of subsampling, allowed the team to predict information growth beyond the observed population sizes.
In all five mice, the measured power-law exponents predicted that noise correlations would not place an upper limit on information. Although stronger noise components tended to align more closely with the signal, the signal also extended across less variable activity patterns. This indicates that shared noise slows the growth of information as more neurons are added but does not bring it to a halt.
"For three decades, shared neural fluctuations were widely expected to make information saturate," says Hideaki Shimazaki of Kyoto University. "Our results show that this is not inevitable."
This study also develops a general theory of how linearly readable information scales and may inform the design of noisy computing systems, which face a similar question: Will adding more components continue to improve accuracy, or will shared noise impose a ceiling?
Published in journal: Science Advances
Title: Population coding under the scale invariance of high-dimensional noise
Authors: S. Amin Moosavi, Sai Sumedh R. Hindupur, and Hideaki Shimazaki
Source/Credit: Kyoto University
Edited by: Scientific Frontline
Reference Number: ns092826_01