A new mathematical framework reveals how populations of neurons encode sensory information, providing a principled way to identify which stimulus features are most reliably represented. The work, led by Simone Azeglio and colleagues, extends the classical Fisher information metric to multiple scales, directly linking geometric distances in stimulus space to the mutual information carried by neural activity.
The research
Published on arXiv (May 7, 2026), the study proposes a Riemannian geometry that emerges from a coarse-graining process: as stimulus resolution is lost, distances between stimuli contract in a way that reflects how reliably they can be discriminated. This multi-scale metric tensor exactly relates to mutual information—well-encoded directions are expanded, poorly encoded directions contracted. The team tested their model on visual cortical responses to natural images, where eigenvectors of the metric identified interpretable stimulus variations that contribute most to information transmission. Importantly, the metric can be estimated using diffusion models, making it practical for large neural populations and high-dimensional stimuli. The approach unifies representational geometries that previously gave conflicting views of neural coding.
Why it matters
For anyone interested in how their brain processes the world, this research provides a deeper understanding of the principles underlying perception. By revealing which features—like edges or textures—are most reliably encoded, it hints at why certain visual patterns are more memorable or easier to recognize. This framework could eventually inform brain-computer interfaces or training programs that optimize sensory learning.
What you can do
While this is foundational research, you can explore your own cognitive abilities by taking a free adaptive IQ test or engaging in brain training exercises that challenge sensory discrimination—a key component of fluid intelligence.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
