Random neural networks match observed dimensionality of neural population recordings and motivate stronger experimental tests

Ilya M Nemenman; Michael J Pasek; Zehui Zhao

arxiv: 2605.26551 · v1 · pith:I2YXE5JVnew · submitted 2026-05-26 · 🧬 q-bio.NC · cond-mat.dis-nn· physics.bio-ph

Random neural networks match observed dimensionality of neural population recordings and motivate stronger experimental tests

Zehui Zhao , Michael J Pasek , Ilya M Nemenman This is my paper

Pith reviewed 2026-06-29 15:08 UTC · model grok-4.3

classification 🧬 q-bio.NC cond-mat.dis-nnphysics.bio-ph

keywords neuraldimensionalityrandomnetworkspopulationrecordingsstructureacross

0 comments

The pith

Random neural networks with finite recording time and context variability match observed neural population dimensionality, but current data lengths prevent dimensionality from discriminating connectivity structures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Brain recordings often show that the collective activity of hundreds of neurons can be captured by just a handful of dimensions, as if the population lives on a low-dimensional surface. Earlier theory suggested that randomly wired networks could produce this, but it was unclear if the numbers matched real data. The authors extend dynamical mean-field theory to include the fact that recordings last only a finite time and that the animal switches between different behaviors. With these additions, the predicted dimensionality falls in the same range as measured values. However, the match holds for a range of network types, so dimensionality by itself does not reveal whether the real brain wiring is random or structured. A different measure, how similarly the activity patterns point in different behaviors, turns out to be more sensitive to the wiring details.

Core claim

when these factors are included, the dimensionality measured from large-scale recordings is consistent with the values predicted by random models. However, current recording durations make it difficult to use dimensionality to discriminate among connectivity structures.

Load-bearing premise

The model assumes that variability across behavioral contexts can be represented as additional fluctuating external input to the random network and that the dynamical mean-field theory approximations remain accurate for the finite-time statistics of interest.

read the original abstract

Randomly connected neural networks have long served as a theoretical tool for studying collective dynamics in neural populations, yet quantitative comparisons to experiments remain limited. Recent technological advances have made it possible to resolve population-wide correlations across neurons, and minimal models such as random neural networks predict their generic structure. Whether the two agree quantitatively remains untested. In this work, we examine whether a minimally structured random neural network can account for the low dimensionality of activity in neural population recordings by building on recent developments in Dynamical Mean-Field Theory and incorporating two additional experimentally relevant features into the model: finite measurement time and variability across behavioral contexts. We show that, when these factors are included, the dimensionality measured from large-scale recordings is consistent with the values predicted by random models. However, current recording durations make it difficult to use dimensionality to discriminate among connectivity structures. We further show that analytically predicted dimensionality varies non-monotonically with external input strength, and that the orientation similarity between neural manifolds recorded under different behavioral contexts can be more sensitive to network structure than dimensionality is. Together, these results provide quantitative guidance for experimental design to infer the connectivity structure underlying population activity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Random networks match recorded dimensionality once finite time and context variability are folded into DMFT, but current data lengths still cannot distinguish connectivity structures.

read the letter

The main point is that random networks, after adding finite measurement time and context-to-context input fluctuations, give dimensionality values that line up with large-scale recordings. At the same time the work shows that existing recording durations are too short for dimensionality to tell random connectivity apart from other structures.

The new pieces are the non-monotonic dependence of dimensionality on external input strength and the claim that orientation similarity across contexts is more sensitive to network details than dimensionality itself. Both come from the DMFT treatment and supply concrete numbers for how long recordings need to be and which metric to prioritize.

The match to data still requires tuning the strength of the fluctuating external input, so it is not an out-of-sample prediction. The abstract gives no sign that the finite-time DMFT expressions were cross-checked against direct network simulations, which leaves open whether the mean-field closure holds for the covariance statistics at the relevant timescales. That is the main soft spot, and it is moderate rather than minor.

The paper is aimed at computational neuroscientists who want quantitative rules for choosing recording length and analysis metrics when trying to infer connectivity from population activity. It is worth sending to referees because the guidance on experimental design is specific enough to be tested and the analytical extensions are clearly stated even if the validation steps need tightening.

Referee Report

2 major / 0 minor

Summary. The manuscript claims that Dynamical Mean-Field Theory (DMFT) analysis of minimally structured random neural networks, after incorporating finite measurement time and behavioral context variability modeled as fluctuating external input, yields dimensionality values consistent with those measured in large-scale neural population recordings. It further concludes that current recording durations limit the use of dimensionality to discriminate connectivity structures, while analytically predicted dimensionality varies non-monotonically with external input strength and manifold orientation similarity across contexts may be a more sensitive probe of network structure.

Significance. If the central results hold, the work supplies a quantitative baseline showing that random connectivity plus realistic experimental factors can account for observed low dimensionality without invoking additional structure, while identifying concrete limitations of dimensionality as a discriminator and proposing manifold orientation similarity as an alternative experimental test. The use of analytical DMFT derivations and emphasis on falsifiable predictions for experimental design are positive features.

major comments (2)

[DMFT finite-time analysis] The claim that random-network dimensionality matches experimental values once finite time and context variability are included rests on DMFT remaining accurate for finite-time covariance statistics under fluctuating external input; however, the manuscript provides no direct numerical simulations validating the mean-field closure against full network dynamics for the relevant timescales and input strengths (see skeptic concern on finite-time predictions).
[Comparison to experimental data] The reported consistency with data involves varying external input strength (listed among free parameters), which functions as a fitted rather than independently predicted quantity; this introduces moderate circularity in the comparison to recordings and weakens the assertion that random models quantitatively account for the observations without additional tuning.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard DMFT approximations for large random networks plus two modeling extensions whose parameters are not independently constrained by the abstract.

free parameters (1)

external input strength
Varied analytically to demonstrate non-monotonic effect on dimensionality; value chosen to match observed range.

axioms (1)

domain assumption Dynamical mean-field theory provides accurate statistics for finite-time correlations in large random networks
Invoked to derive the dimensionality formula used for comparison with recordings.

pith-pipeline@v0.9.1-grok · 5745 in / 1151 out tokens · 28936 ms · 2026-06-29T15:08:12.556625+00:00 · methodology

Random neural networks match observed dimensionality of neural population recordings and motivate stronger experimental tests

Core claim

Load-bearing premise

discussion (0)