A multi-scale information geometry reveals the structure of mutual information in neural populations
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "A multi-scale information geometry reveals the structure of mutual information in neural populations".
Marcus: Here's the gist of the paper. A unique Riemannian geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining, resulting in a multi-scale extension of the Fisher information metric that captures encoding structure from fine stimulus details to coarse global distinctions <ref:2605.06304#pg2> Pruning redundant reasoning tokens The paper shows that a unique representational geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining <ref:2605.06304#pg3>.
Ines: This results in a multi-scale extension of the Fisher information metric, capturing encoding structure from fine stimulus details to coarse global distinctions <ref:2605.06304#pg2>.
Marcus: The resulting geometry is exactly related to the mutual information encoded by the population, where well encoded stimulus directions are expanded and poorly encoded ones are contracted <ref:2605.06304#pg2>.
Ines: The metric tensor can be estimated using diffusion models, making the framework practical for large neural populations and high-dimensional stimuli <ref:2605.06304#pg2>.
Paper summary: Ines: So wrapping up this discussion on the paper "A multi-scale information geometry reveals the structure of mutual information in neural populations", what it really does is move away from just thinking about local sensitivity when defining how neural codes work. They establish a unique geometry that depends on how stimulus resolution is lost.
Marcus: That unique geometry links directly to mutual information, showing it's not a single number but has an exact interpretation based on the expected rate of distance increase under stimulus changes. This means we have a way to quantify the structure of information encoding across multiple scales simultaneously.
Yuki: For us in population genetics, this suggests that the patterns we see in neural responses aren't just random noise; they follow a specific geometric structure dictated by how information is lost as detail is blurred. It supports the idea that important biological signals are those robust features that get stretched most strongly in this space.
Ines: The authors show it can be estimated using diffusion models, which gives us a concrete way to calculate these metrics for complex systems, even with large neural populations. They find that the leading eigenvectors of this geometry point to meaningful visual features across different areas of the brain.
Marcus: But they're honest about the limitations. Like any method based on learning a model, it still requires training a probabilistic encoding model describing how neurons respond to stimuli, and you need enough data to estimate that model reliably. Also, estimating this geometry is more computationally expensive than just calculating the standard Fisher information metric because it involves training a diffusion model and running multiple evaluations for those Monte Carlo estimates of the metric tensor.
Yuki: So while the theoretical framework seems powerful for understanding structure, the practical application still hinges on having good data and computational resources to estimate that underlying probabilistic model accurately.
Ines: Exactly. The paper provides a new mathematical lens—this multi-scale information geometry—to interpret how neural populations organize their responses across different levels of stimulus detail, giving us a way to look beyond just the local sensitivity we usually measure.
Conclusion: Ines: So we're wrapping up on this paper about "A multi-scale information geometry reveals the structure of mutual information in neural populations."
Marcus: It really boils down to them building a unique mathematical shape, a geometry, that describes how distances change when you lose detail from a stimulus.
Yuki: They connect this geometry directly to mutual information, which is basically how much info the population actually carries about the world.
Ines: It shows that stimulus directions that are most important—the ones contributing most to that total information—get stretched out more strongly in this geometry than less important ones.
Marcus: And they show it's not just a single number; it’s related to how fast squared distances increase when you slightly nudge the stimulus around.
Yuki: For population genetics, this means we can look at patterns in neural responses and see them as following a specific structure dictated by how information gets lost as detail is blurred.
Ines: The authors use diffusion models to estimate this geometry, which makes it practical for huge groups of neurons and complex stimuli.
Marcus: But they're clear about the catch, though; you still need a good model describing how those neurons respond to stimuli, and you need a lot of data to get that model right.
Yuki: And computationally, calculating this geometry is tougher than just using standard Fisher information because you have to train that diffusion model and run lots of tests for those measurements.
Ines: So they give us a tool to see the structure of information encoding across different scales, but it’s dependent on having a solid underlying model and enough processing power.
Marcus: It's a new way to look at how neural systems organize their signals, moving beyond just measuring local sensitivity.
Yuki: This sets up the next step for us—we need to figure out if this geometric structure holds up when we look at actual biological data from species or populations.
Institut de la Vision, INSERM, CNRS · Sorbonne University and École Normale Supérieure - PSL
q-bio.NC
Submitted: 2026-05-07
Updated: 2026-11-06
Code: https://github.com/pytorch/vision
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 89/100
The gist: A unique Riemannian geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining…
Key concepts
- Multi-scale Extension of Fisher Information Metric
- This is a geometric framework that extends the standard Fisher information metric to capture encoding structure at multiple levels of stimulus resolution. It models how representational distances contract as fine details are lost during coarse-graining, allowing analysis from fine features to broad global distinctions.
- Coarse-graining and Resolution Loss
- This process simulates losing stimulus resolution by averaging or mixing neighboring stimuli using a kernel like the Gaussian kernel. The geometry is uniquely determined by how discriminability changes under this loss, showing that finer details become harder to distinguish while distant stimuli remain relatively unaffected.
- Mutual Information Interpretation
- The derived geometry has an exact interpretation in terms of mutual information. Stimulus directions that contribute most significantly to the total mutual information are stretched more strongly by the geometry, meaning the metric quantifies how much stimulus variation contributes to encoding information.
Terminology
Summary
A unique Riemannian geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining, resulting in a multi-scale extension of the Fisher information metric that captures encoding structure from fine stimulus details to coarse global distinctions <ref:2605. p. 0 Here,
we show that a unique Riemannian representational geometry emerges from first
principles governing how distances contract as stimulus resolution is lost through
coarse-graining. This results in a multi-scale extension of the Fisher information
metric, capturing encoding structure from fine stimulus details to coarse global
distinctions.Where the paper says it
Pruning redundant reasoning tokens
The paper shows that a unique representational geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining <ref:2605.06304#pg3>. This results in a multi-scale extension of the Fisher information metric, capturing encoding structure from fine stimulus details to coarse global distinctions <ref:2605. p. 0 Here,
we show that a unique Riemannian representational geometry emerges from first
principles governing how distances contract as stimulus resolution is lost through
coarse-graining. This results in a multi-scale extension of the Fisher information
metric, capturing encoding structure from fine stimulus details to coarse global
distinctions. This results in a multi-scale extension of the Fisher information
metric, capturing encoding structure from fine stimulus details to coarse global
distinctions. The resulting geometry is exactly related to the mutual information
encoded by the population: well encoded stimulus directions — those contributing
more to mutual information — are expanded, whereas poorly encoded directions
are contracted. The resulting geometry is exactly related to the mutual information
encoded by the population: well encoded stimulus directions — those contributing
more to mutual information — are expanded, whereas poorly encoded directions
are contracted. The metric tensor can be estimated using diffusion models, making
the framework practical for large neural populations and high-dimensional stimuli.Where the paper says it
Where the paper says it
Where the paper says it
Deriving the geometry from coarse-graining
The geometry is uniquely determined by simple requirements on how representational distances change under stimulus coarse-graining, whereby fine stimulus details become progressively harder to discriminate <ref:2605.06304#pg3>. The Gaussian kernel mixes only an O(√t) neighbourhood around x, so in the limit t → 0 it reduces the ability of the neural code to discriminate infinitesimally close stimuli while leaving discrimination of distant stimuli unaffected <ref:2605. p. 2 (3)
This coarse-graining is the unique choice that is invariant with respect to translations and rotations of
stimulus space, and satisfies the semigroup property, whereby coarse-graining by t1 and then t2 is
equivalent to coarse-graining once by t1 + t2 [9]. The Gaussian kernel mixes only an O(
√
t) neighbourhood around x, so in the limit t → 0 it reduces
the ability of the neural code to discriminate infinitesimally close stimuli while leaving discrimination
of distant stimuli unaffected. Here we seek a geometry that reflects the ability of the neural code to discriminate stimuli, both near
and far. To uniquely define such a geometry, we consider how discriminability is affected by the loss
of stimulus resolution induced by coarse-graining with isotropic Gaussian noise:
xt = x +
√
t z, z ∼ N (0, I).Where the paper says it
Where the paper says it
Estimating the geometry using diffusion models
The squared infinitesimal line element dx⊤G(x)dx can be estimated by rewriting Eq. (10) as an expectation over z (with α = 1, following our convention) <ref:2605. p. 5 5 Estimating the multi-scale geometry from diffusion models
We can estimate the squared infinitesimal line element dx⊤G(x) dx by rewriting Eq. (10) as an
expectation over z (with α = 1, following our convention):
dx⊤G(x)dx =
Z ∞
0
Ez∼N(0,I)
h
dx⊤Jt(x +
√
tz) dxi
dt. (16)
Since xt does not depend on r, this yields
dx⊤Jt(xt)dx =
1
t
2
dx⊤Covpt(r|xt)
[ˆx(xt, r)]dx
=
1
2t
2
Er,r′∼pt(r|xt)
h(ˆx(xt, r) − xˆ(xt, r′
))⊤dx2
i
. (17)
Thus, the diffused Fisher along direction dx is the expected squared difference between two independently sampled posterior-mean estimates, projected onto dx. Equation (17) expresses dx⊤Jt(xt)dx in terms of the posterior mean xˆ(xt, r) = E[x | xt, r], which
we estimate using a conditional diffusion model [13, 14]. Because the estimator depends only on
differences between two samples from the same conditional model, systematic errors cancel; this
avoids the mismatch bias of approaches that subtract conditional from separately trained unconditional
estimators, which can become potentially unstable as t → 0 [11, 15–17].Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The multi-scale geometry admits an exact mutual information interpretation, where stimulus directions that contribute most to the total mutual information encoded by the population are stretched more strongly by the geometry <ref:2605.06304#pg5>. This identity shows that mutual information equals the expected rate at which squared distances increase under infinitesimal perturbations of the stimulus <ref:2605. p. 4 Formally, under mild regularity conditions on p(r, x) (Appendix A), the metric
tensor satisfies [11]
I(R; X) = 1
2
Ex [Tr (G(x))] . (11)
This identity has a natural geometric interpretation: mutual information equals the expected rate at
which squared distances increase under infinitesimal perturbations of the stimulus,
I(R; X) = 1
2
d
dεEx,z
d
2
x, x +
√
ε z
ε=0
, z ∼ N (0, I). Equivalently, for N stimuli sampled from p(x),
I(R; X) = nx
2
lim
N→∞
1
N
X
N
i=1
m2
i
, (13)
where nx is the number of stimulus dimensions, and mi = d
g
i
/di
is the local magnification at
xi
: the ratio of the geometric distance d
g
i
to the Euclidean distance di between xi and its nearest
neighbour (Appendix A). Mutual information is thus the mean squared magnification of stimulus
space induced by the neural code.Where the paper says it
Where the paper says it
A Relation between geometry and mutual information
The identity I(R; X) = 1/2 Ex [Tr (G(x))] is shown to be equivalent to Eqn (12) in the main text <ref:2605.06304#pg9>. This means that mutual information equals the expected rate at which squared distances increase under infinitesimal perturbations of the stimulus <ref:2605. p. 4 Formally, under mild regularity conditions on p(r, x) (Appendix A), the metric
tensor satisfies [11]
I(R; X) = 1
2
Ex [Tr (G(x))] . (11)
This identity has a natural geometric interpretation: mutual information equals the expected rate at
which squared distances increase under infinitesimal perturbations of the stimulus,
I(R; X) = 1
2
d
dεEx,z
d
2
x, x +
√
ε z
ε=0
, z ∼ N (0, I). Equivalently, for N stimuli sampled from p(x),
I(R; X) = nx
2
lim
N→∞
1
N
X
N
i=1
m2
i
, (13)
where nx is the number of stimulus dimensions, and mi = d
g
i
/di
is the local magnification at
xi
: the ratio of the geometric distance d
g
i
to the Euclidean distance di between xi and its nearest
neighbour (Appendix A). Mutual information is thus the mean squared magnification of stimulus
space induced by the neural code. The resulting geometry is exactly related to the mutual information
encoded by the population: well encoded stimulus directions — those contributing
more to mutual information — are expanded, whereas poorly encoded directions
are contracted. The metric tensor can be estimated using diffusion models, making
the framework practical for large neural populations and high-dimensional stimuli.Where the paper says it
Where the paper says it
Where the paper says it
Predictive neural models
The leading eigenvectors of the multi-scale metric tensor G(x) identify stimulus variations that contribute most strongly to encoded information, revealing which stimulus variations contribute most to mutual information. These directions yield interpretable image features that differ systematically across visual areas. p. 1 The eigenvectors of the metric
tensor identify the stimulus directions stretched by the geometry, revealing which stimulus variations
contribute most to the mutual information. These directions yield interpretable image features that
differ systematically across visual areas. Here we show that
for expressive deep models fit to V1 and V4 responses, the Fisher geometry can be poorly constrained,
yielding features that vary across architectures. In contrast, due to its link to mutual information, the
multi-scale geometry emphasizes stimulus directions that are both informative and well-constrained
by the data, producing features that are stable when different models are fitted to the same data.Where the paper says it
Where the paper says it
Comparing different geometries
The subspace alignment of A and B only constrains the angles between eigenvectors and is insensitive to the magnitude of the associated eigenvalues. The Bures–Wasserstein distance is argued to be the most appropriate choice for decodable information notions of representational similarity, since it has a direct optimal-transport interpretation between the two representational geometries. p. 7 For the Fisher information, the reverse was true: two models fit to the same area were less aligned
than identical architectures fit to different areas (Fig. 4C & E), implying that its leading eigenvectors
are driven more by the choice of encoding model than by the neural data (similar results were obtained
using Bures-Wasserstein distance [22] rather than subspace alignment, Appendix C. For the multi-scale metric,
alignment was higher for two models fit to the same area than for two models of the same architecture
fit to different areas (Fig. 4D), indicating that it captured structure intrinsic to the recorded population.Where the paper says it
Where the paper says it
A limitation of our approach
A limitation is that, like the Fisher information, it requires learning a probabilistic encoding model p(rx) describing how neurons respond to stimuli, and thus sufficient data to estimate this model reliably. A second limitation is computational: estimating the geometry is more expensive than the Fisher information, as it requires training a diffusion model and performing multiple evaluations to obtain Monte Carlo estimates of the metric tensor. p. 9 Increasingly, however, large-scale datasets of neural responses to natural stimuli
are becoming available [19, 34–37], making such approaches feasible in practice. A second limitation
is computational: estimating the geometry is more expensive than the Fisher information, as it requires
training a diffusion model and performing multiple evaluations to obtain Monte Carlo estimates of
the metric tensor.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Multi-scale geometric characterization of neural codes allows AI systems to identify stimulus variations that contribute most to information transmission, yielding interpretable features that are robust to modelling choices. This can be used to develop more principled feature extraction methods for visual recognition tasks, as the eigenvectors of the multi-scale metric tensor identify these directions which
contribute most to mutual information.
-
The framework enables the development of stimulus representation geometries that are invariant to sufficiency-preserving transformations of neural responses, ensuring that identified features
reflect robust properties of the neural code rather than details of a fitted model.
This robustness is crucial when training deep encoding models, as it suggests that features identified through this geometry are constrained by data rather than being artifacts of the specific model architecture used. -
The relationship between mutual information and stimulus space structure provides a method to predict perceptual judgments, as
distances in our geometry can only shrink as information passes through successive stages of processing, so the geometry defined by neural responses should, in principle, constrain perceptual geometry.
This suggests that models built on this geometric framework could be used to understand and potentially improve human-like visual processing. -
The estimation of the multi-scale Fisher information matrix using diffusion models allows for a practical approach to high-dimensional stimuli by estimating the
squared infinitesimal line element dx top G(x) dx
through Monte Carlo integration, making it feasible for large neural populations responding to complex natural stimuli. This provides a computationally tractable method for characterizing information transmission in complex visual data. -
The cross-architecture comparison of geometries, quantified by subspace alignment and Bures-Wasserstein distance, allows researchers to determine if the geometric features are intrinsic to the neural code or artifacts of the encoding model. For instance,
cross-model distances (same area, different backbone) lie below cross-area distances (same backbone, different area)
for the multi-scale metric G(x), indicating that these features are more stable when comparing models fitted to the same cortical area.
Abstract
Understanding how neural population responses represent sensory information is a central problem in systems neuroscience. One approach is to define a representational geometry on stimulus space in which distances reflect how reliably stimuli can be distinguished from neural activity. However, different constructions of these distances can lead to qualitatively different conclusions about the neural code. Here, we show that a unique Riemannian representational geometry emerges from first principles governing how distances contract as stimulus resolution is lost through coarse-graining. This results in a multi-scale extension of the Fisher information metric, capturing encoding structure from fine stimulus details to coarse global distinctions. The resulting geometry is exactly related to the mutual information encoded by the population: well-encoded stimulus directions (those contributing more to mutual information) are expanded, whereas poorly encoded directions are contracted. The metric tensor can be estimated using diffusion models, making the framework practical for large neural populations and high-dimensional stimuli. Applied to visual cortical responses to natural images, the eigenvectors of the metric tensor identify stimulus variations that contribute most to the encoded stimulus information, yielding interpretable features that are robust to modelling choices. Together, these results provide a principled, information-theoretic framework for characterising neural population codes.
Sources
- MMG: Mutual Information Estimation via the MMSE Gap in Diffusion
- The Sensorium competition on predicting large-scale mouse primary visual cortex activity
- OmniMouse: Scaling properties of multi-modal, multi-task Brain Models on 150B Neural Tokens
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks