Optimal rate-variance coding due to firing threshold adaptation near criticality

Andr\'e Longtin; Leonard Maler; Mauricio Girardi-Schappo

REVIEW 3 major objections 4 minor 2 references

Optimal rate-variance coding due to firing threshold adaptation near criticality

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Adaptive spike thresholds let weak inputs be encoded as firing-rate variance

desk verdict Novel variance-coding mechanism in adaptive networks, but the robustness-to-J claim needs a direct scaling test before it can be trusted. read the letter →

arxiv 2509.04106 v1 pith:BWQ6KCHB submitted 2025-09-04 q-bio.NC cond-mat.dis-nncond-mat.stat-mechnlin.AOphysics.bio-ph

Mauricio Girardi-Schappo , Leonard Maler , André Longtin This is my paper · ORCID

classification q-bio.NCcond-mat.dis-nncond-mat.stat-mechnlin.AOphysics.bio-ph

keywords thresholdadaptationratecodingpatternvariancecriticalitymutualinformationdynamicrangerecurrentexcitatorynetwork

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that adding spike-triggered threshold adaptation to a recurrent excitatory network gives the network a dual coding scheme: weak stimuli, too faint to be distinguished by mean firing rate, are encoded in the variance of the population firing rate, while stronger stimuli are encoded by the rate itself. This matters because rate coding alone breaks down for weak inputs, and the proposed mechanism does not require fine-tuning of synaptic coupling, unlike nonadaptive networks that only work at a specific critical point. The optimal adaptation timescale falls at roughly 1,000–1,100 ms, matching measured threshold recovery times in hippocampal neurons, so the authors identify a concrete biological implementation.

What carries the argument

The mechanism is the adaptive-threshold mean-field dynamics ρ(t+1) = [1−ρ(t)][I(r) + Jρ(t) − θ(t)]Γ, θ(t+1) = θ(t) − θ(t)/τ + uF(ρ(t),θ(t)), where θ is a spike-adapted threshold and τ its slow recovery timescale. The threshold acts as negative feedback that suppresses high firing rates, leaving a low-rate state ρ ∼ 1/τ whose fluctuations become near-critical as τ → ∞ and r → 0. Because the network is all-to-all, a spatial pattern is fully determined by ρ, so the variance of ρ quantifies pattern diversity; this links entropy, variance and pattern coding.

What would settle it

Simulate the adaptive network at a clearly subcritical coupling (e.g., J = 3) with τ ≈ 1,000 ms and measure Var(ρ) as the input rate r is reduced from 10⁻³ to 10⁻⁶ spikes/ms; if Var(ρ) does not keep growing as r → 0, the central claim of robust weak-input variance coding fails. A complementary check would measure the entropy and mutual information at several J values to see whether their maxima persist in the same τ range.

Watch

Extended reading notes

Core claim

The central claim is that firing-threshold adaptation creates a stationary weak-input regime in which the population firing rate sits near ρ ∼ 1/τ, and the fluctuations of that rate — not its mean — encode the input intensity. As τ grows and the input rate r goes to zero, these fluctuations approach the critical fluctuations of the underlying mean-field directed-percolation point, making the variance of ρ grow as r vanishes. In simulations, the conditional entropy (pattern diversity) peaks at τ ≈ 1,000 ms and the mutual information at τ ≈ 1,100 ms, and these maxima persist as the synaptic coupling J is swept across the critical value. Nonadaptive networks show peak performance only at the tu

Load-bearing premise

The mechanism rests on the claim that the weak-input state maintained by adaptation is genuinely near critical, so that the firing-rate variance grows as the input vanishes for a wide range of synaptic couplings; this scaling is not derived from the dynamics in the main text.

Editorial extensions

If this is right

Weak stimuli below the rate-coding range can be transmitted by pattern/variance coding in adaptive networks, extending the usable input range.
Hub-like networks with adaptive thresholds do not need precise synaptic tuning to code weak inputs, since the coding maxima survive variation in coupling strength.
The predicted optimal adaptation timescale of roughly 10²–10³ ms matches the measured threshold recovery times of hippocampal CA3 and mossy cells, tying the theory to memory circuits.
The same dual-rate/variance coding scheme could be exploited in reservoir computing and artificial sensors by adjusting threshold adaptation timescales.
Nonadaptive networks, by contrast, achieve comparable variance coding only at the critical point, which explains why pure criticality is a fragile coding strategy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: If the variance-encoding claim transfers to sparsely connected networks, the spatial pattern distribution could be even richer than the all-to-all reduction suggests, but the entropy–variance identification would need to be replaced by a full spatial entropy.
Editorial inference: The predicted coding optimum could be tested experimentally by recording from a recurrent population with known threshold recovery times and checking whether weak inputs produce increased firing-rate variance with a peak near τ ≈ 1 s.
Editorial inference: The robustness result likely depends on the population-level mean-field closure; finite-size effects or structured connectivity may narrow the J range over which variance coding remains effective.
Editorial inference: The paper's mechanism is distinct from self-organized quasicriticality because J remains a free parameter; this suggests threshold adaptation might generally regularize excitable networks toward a near-critical operating regime without explicit synaptic tuning.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my read. The new piece is the claim that firing-threshold adaptation creates a weak-input regime where the variance of the population rate carries information, with optima at recovery timescales of 100–1000 ms. That's a genuinely novel mechanism, and it's presented in a clean model with a transparent mean-field framework. The dynamic-range derivation is solid, and the authors are honest about the limitations of the all-to-all reduction and the transient nature of rate coding under the multiplicative rule.

What's not solid is the robustness claim. The paper asserts that weak-input variance grows as r→0 because the adaptive network 'becomes critical' as τ→∞ and r→0, but it never derives or tests the finite-τ scaling. A linearization around the adaptive fixed point gives an effective critical coupling shifted by O(1/τ), meaning that for fixed finite τ and off-critical J, the variance should saturate at sufficiently small r. The entropy and MI curves are interpolated from simulations at a single r = 10^-6, with no error bars and no check against even weaker inputs. So the J-robustness and the τ optimum are not established. The 'spatial spike patterns' language in the abstract also overpromises, since the all-to-all model reduces patterns to population-rate fluctuations.

That's a real problem, but it's fixable. The core mechanism—adaptation can create a variance-based weak-input channel—probably survives. A serious referee should push for a direct test of the variance scaling across J and τ, error bars on the information measures, and a careful statement of what 'pattern coding' means in a mean-field model. This paper deserves peer review because the idea is new and the framework is usable, but I wouldn't cite it in its current form.

Referee Report

3 major / 4 minor

Summary. This paper studies a dense all-to-all network of stochastic excitatory integrate-and-fire neurons, with and without spike-threshold adaptation, using a mean-field description and N = 10^5 simulations. The central claim is that threshold adaptation creates a weak-input stationary state rho ~ 1/tau whose fluctuations grow as the input rate r decreases, allowing weak inputs to be encoded by the variance of the population firing rate. The authors report that entropy and mutual information are maximized at adaptation timescales tau ~ 10^2-10^3 ms, that these optima persist when the coupling J is swept around the constant-threshold critical point J_c = 1/Gamma = 5, and that nonadaptive networks exhibit variance coding only at criticality. They connect the optimal timescales to experimentally measured hippocampal threshold recovery times.

Significance. If the central scaling claim is established, this is a valuable conceptual contribution: it offers a concrete, analytically tractable mechanism for dual rate/variance coding and links optimal adaptation timescales to experimentally observed hippocampal values. The model is clearly specified, the comparison between adaptive and nonadaptive networks is instructive, and the analytical dynamic-range calculation for adaptive networks is a strength. The paper is also explicit about the all-to-all reduction and about the definitional character of the entropy-as-pattern-capacity identification. However, the main robustness claim currently rests on an asserted near-critical scaling rather than a derived one, and the entropy/MI curves are interpolation-based with limited error characterization. The significance of the paper is therefore real but presently not fully secured.

major comments (3)

[Results C / Discussion A] The central robustness claim is that threshold adaptation makes Var(rho) grow as r -> 0 for a range of couplings J around J_c (Results C, Figure 3D), because the weak-input state 'becomes critical in the limit tau -> infinity and r -> 0'. No derivation of this scaling is given in the main text. In fact, linearizing Eqs. (1) and (5) around the stationary state rho* = 1/(u tau) indicates that the stability boundary is shifted from J_c by an amount of order 1/(u tau); Discussion A explicitly states that the dependence on J - J_c is weakened only 'to be of order 1/tau'. For any fixed finite tau and any J not exactly on this shifted line, mean-field scaling predicts a finite susceptibility, so Var(rho) should saturate as r -> 0 rather than diverge. The paper does not provide the r-scaling of Var(rho), finite-size scaling, or simulations at r smaller than 10^-6 ms^-1 that would distinguish div
[Methods F / Eq. (13) / Figure 3] The entropy in Eq. (13) is computed from the histogram of the population rate rho, not from spatial patterns; the combinatorial multiplicity Omega is dismissed because all-to-all connectivity makes all patterns with the same rho equivalent. Thus 'pattern coding' is a definitional identification of entropy with population-rate variance, not a measured property of spatial patterns. This identification needs defense: a downstream reader that cannot resolve individual cells may lack access to the full pattern multiplicity, and the entropy is sensitive to histogram binning. The entropy and MI curves in Figure 3B/D are interpolated from simulations at a single r = 10^-6 ms^-1 with no error bars, binning analysis, or finite-size scaling. Please quantify the estimation error and test robustness to bin count and simulation time.
[Methods D / Eq. (14)] For the multiplicative adaptation rule, strong inputs cause the threshold to grow without bound and the network activity to eventually shut down; RC, entropy, and MI are therefore measured over a transient metastable state of finite duration D ~ 1/u (Methods D). The mutual information definition in Eq. (14) is written for stationary distributions, but the adaptive entropies are measured over this nonstationary window. If the duration D or the chosen measurement window varies with tau, the reported MI maximum near tau = 1100 ms and the entropy maximum near tau = 1000 ms could be artifacts of the measurement procedure. Please show that the reported values are converged over the measurement window and robust to the choice of D and the transient cutoff.

minor comments (4)

[Figure 2/3 captions] The figures report error bars for simulation symbols but the entropy and MI curves are described as 'inter/extrapolation'; please state the number of independent simulation runs, the binning parameters, and how the solid curves are computed.
[Eq. (13)] Equation (13) is garbled in the typeset text; the entropy functional should be displayed cleanly with the correct summation and normalization.
[Results A] The statement 'r = 10^-6 spikes/ms/neuron = 1 spike every 10 ms in the population' is only true for N = 10^5; make the population size explicit to avoid confusion.
[Methods F] The phrase 'the fluctuations of rho(t) for a given r dictate the capacity of the network to generate patterns' is a definitional assumption, not a consequence; consider clarifying this in the main text so that readers do not mistake it for a measured spatial-pattern property.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; minor self-citations are not load-bearing.

full rationale

The paper's derivation chain is largely self-contained. The dynamic range is computed analytically from the mean-field equation (1), while the entropy and mutual information are measured from simulations and then interpolated; they are not derived from their own definitions. The identification of pattern-coding capacity with conditional entropy (Eq. 13) and with the variance of the population firing rate is an openly stated modeling assumption, not a hidden circular step: the paper does not fit a parameter and then rename it as a prediction. The explanation that threshold adaptation creates near-critical fluctuations is an interpretation of the observed variance growth, not a quantity derived from the input parameters; whether the finite-tau system is truly near critical for all J is a scientific validity concern, but not a circularity. Self-citations (Girardi-Schappo et al. 2021 for the constant-threshold critical point, Trinh et al. 2023 for hippocampal timescales) provide background and experimental context and are not the sole justification for the central claim. The central finding—that adaptive thresholds yield robust weak-input variance coding across coupling strengths—is an empirical simulation result with no fitted parameters, so it does not reduce to its inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the standard mean-field directed-percolation framework (the critical point J_c = 1/Gamma is cited from the authors' prior work but is a standard, re-derivable result), on the specific multiplicative threshold dynamics (F = rho*theta) used in the main text, on the Gaussian entropy-variance relation, and on the hand-chosen fatigue strength u = 0.1. The identification of the histogram entropy with pattern-coding capacity is a definitional assumption (Methods F). No new physical entities (particles, forces, fields) are introduced; 'Adaptive Near-Critical Coding' is a label for the observed mechanism.

free parameters (4)

fatigue strength u = 0.1
Fixed homogeneous value; sets the weak-input stationary rate rho ~ 1/(u*tau) and thereby influences the position of the entropy/MI peaks in tau. The biological alignment of the peak with hippocampal timescales is conditional on this hand-chosen value, since the peak shifts to tau = 10^2 ms for alternative adaptation rules.
gain Gamma = 0.2
Sets the critical coupling J_c = 1/Gamma = 5; a standard model parameter, not fitted to data, but it rescales the dynamic range results and the position of the MF-DP critical point.
weak-input reference rate r = 10^-6 ms^-1
Chosen as the practical minimum to measure entropy and r_min 'to avoid running simulations for computationally prohibitive durations' (Methods E); the claim that weak-input entropy peaks at tau ~ 10^3 ms is evaluated at this specific r.
histogram binning for entropy = not stated in main text
H[P(rho|r)] in Eq. (13) requires binning the simulated rho(t) series; the bin width sets the entropy scale and is not reported in the main text (deferred to supplementary), so the entropy and MI magnitudes are not fully specified.

assumptions (5)

domain assumption Mean-field/all-to-all network reduction: a spike pattern is fully determined by the population firing rate rho
Methods F: 'in a MF network, a pattern is completely determined by rho'; this turns the abstract's spatial-pattern coding into population-rate variance coding and is load-bearing for the PC claims.
standard math Gaussian relation between entropy and variance of rho
Results: 'We know from Gaussian signals that both the entropy and the variance of rho are related'; used to identify entropy (PC capacity) with log Var(rho).
standard math Underlying MF-DP critical point at J_c = 1/Gamma
Used to anchor the 'near-critical' explanation; the critical point is cited from Girardi-Schappo et al. 2021, is a standard mean-field directed-percolation result, and is re-derivable from the constant-threshold limit of Eq. (1).
domain assumption Threshold dynamics F = rho*theta (multiplicative) as the main adaptation rule
Eqs. (1) and (5); the stationary solution rho ~ 1/tau follows from this choice, and the transient shutdown of RC for strong inputs is a consequence of this rule; alternative rules are relegated to supplementary material.
domain assumption Rates and parameters self-average over the population (single-peaked distributions with small variance)
Methods D: all parameters independently obey sharp Gaussian distributions so population means capture the dynamics; required for the mean-field equations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimal rate-variance coding due to firing threshold adaptation near criticality." pith.science (2026). https://pith.science/paper/BWQ6KCHB

@misc{pith2026250904106,
  author       = {Pith},
  title        = {Pith review of: Optimal rate-variance coding due to firing threshold adaptation near criticality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BWQ6KCHB}},
  note         = {Machine review of arXiv:2509.04106}
}

abstract

Recurrently connected neuron populations play key roles in sensory perception and memory storage across various brain regions. While these populations are often assumed to encode information through firing rates, this method becomes unreliable with weak stimuli. We propose that in such cases, information can be transmitted via spatial spike patterns, employing a sparse or combinatorial coding based on firing rate variance. Around the critical point of a stochastic recurrent excitable network, we uncover a synergistic dual-coding scheme, enabled by single-cell threshold adaptation. This scheme optimizes variance coding for weak signals without compromising rate coding for stronger inputs, thus maximizing input/output mutual information. These optimizations are robust across adaptation rules and coupling strengths through self-suppression of internal noise, particularly around the network's phase transition, and are linked to threshold recovery times observed in hippocampal memory circuits (~$10^2$-$10^3$ms). In contrast, nonadaptive networks perform similarly only at criticality, suggesting that threshold adaptation is essential for reliable encoding of weak signals into diverse spatial patterns. Our results imply a fundamental role for near-critical latent adaptive dynamics enabled by dual coding in biological and artificial neural networks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Benda J, Maler L & Longtin A. (2010). Linear Versus Nonlinear Signal Transmission in Neuron Models With Adaptation Currents or Dynamic Thresholds. J Neurophysiol 104, 2806–2820-2806–2820. Buendía V, di Santo S, Bonachela JA & Muñoz MA. (2020). Feedback Mechanisms for Self-Organization to the Edge of a Phase Transition. Frontiers in Physics 8,

work page 2010
[333]

𝐻[𝒫(𝜌|𝑟)] = − ∑ 𝒫(𝜌|𝑟) log2 𝒫(𝜌|𝑟){𝜌} , (13) Chacron MJ, Lindner B & Longtin A. (2007). Threshold fatigue and information transfer. Journal of Computational Neuroscience 23, 301-311. Girardi-Schappo M, Galera EF, Carvalho TTA, Brochini L, Kamiji NL, Roque AC & Kinouchi O. (2021). A unified theory of E/I synaptic balance, quasicritical neuronal avalanches ...

work page arXiv 2007

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Benda J, Maler L & Longtin A. (2010). Linear Versus Nonlinear Signal Transmission in Neuron Models With Adaptation Currents or Dynamic Thresholds. J Neurophysiol 104, 2806–2820-2806–2820. Buendía V, di Santo S, Bonachela JA & Muñoz MA. (2020). Feedback Mechanisms for Self-Organization to the Edge of a Phase Transition. Frontiers in Physics 8,

work page 2010

[2] [333]

𝐻[𝒫(𝜌|𝑟)] = − ∑ 𝒫(𝜌|𝑟) log2 𝒫(𝜌|𝑟){𝜌} , (13) Chacron MJ, Lindner B & Longtin A. (2007). Threshold fatigue and information transfer. Journal of Computational Neuroscience 23, 301-311. Girardi-Schappo M, Galera EF, Carvalho TTA, Brochini L, Kamiji NL, Roque AC & Kinouchi O. (2021). A unified theory of E/I synaptic balance, quasicritical neuronal avalanches ...

work page arXiv 2007