{"id":"df102196-9270-43e8-ad5c-dce04258eb59","arxiv_id":"2509.04106","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Threshold adaptation keeps a recurrent neuron network near its critical point, where the variance (not the mean) of population firing encodes weak stimuli, a coding mode that survives across coupling strengths while rate coding is preserved for strong inputs.","lead":"This paper shows, in a mean-field neural network model, that neurons with slowly recovering firing thresholds can encode weak inputs through fluctuations of population activity while retaining rate coding for strong inputs. The result suggests a general mechanism by which ubiquitous spike-threshold adaptation, not fine-tuned connectivity, enables reliable coding of weak signals, with optimal timescales matching hippocampal recordings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Variance-encoding robustness rests on an unverified near-critical scaling: for finite tau the effective critical coupling shifts by O(1/tau), so Var(rho) may saturate as r->0 for off-critical J; the claimed J-robustness needs a direct test.","rationale":"The reader's weakest_assumption identifies precisely this: the near-critical scaling of Var(rho) as r -> 0 for a range of J is asserted rather than derived, and the effective distance to criticality as a function of J, tau, and r is not given. My linear-stability calculation supports the concern by showing that the adaptive feedback shifts the critical coupling by an amount proportional to 1/tau, so for finite tau the system is generally off-critical and the variance should saturate rather than diverge. The paper's own Methods section also acknowledges the transient nature of the rate-coding regime and the reliance on interpolated simulation curves for entropy and MI, which strengthens the need for a direct scaling test. The central claim may still hold if nonlinearities or finite-size effects produce the reported variance growth, but that is an empirical question the paper does not settle. Therefore the appropriate verdict remains CONDITIONAL: the idea is plausible and the simulations are suggestive, but the load-bearing scaling robustness is not established. My read does not change the reader's verdict, hence UNCHANGED.","tokens_in":1041,"tokens_out":1129,"duration_ms":118991,"concrete_test":"Run simulations for fixed J = 4, 5, 5.5 (and optionally 6) and tau = 100, 1000, 10000, N = 1e5, at input rates r = 1e-5, 1e-6, 1e-7, 1e-8 (with simulation time scaled as 1/r), and compute Var(rho) and the stationary distribution P(rho|r). If, at J = 5 and tau = 1000, Var(rho) stops growing below some r (i.e. Var(rho) -> const as r decreases), the 'near-critical' variance scaling is not supported. Equivalently, fit Var(rho) ~ r^{-a(J,tau)} over the smallest three r values; if a = 0 for any J != J_c, the robustness claim fails. Also check whether for J > J_c + 2/(Gamma u tau) a stable stationary distribution exists at all at r = 1e-6, since the linear analysis predicts an unstable fixed point there.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Results C is that threshold adaptation creates a weak-input stationary state rho ~ 1/tau that 'becomes critical' in the limit tau -> infinity and r -> 0, making Var(rho) grow as r vanishes, robustly for J around J_c = 1/Gamma = 5. This scaling is asserted, not derived. Linearizing Eq. (1) with the multiplicative adaptation rule, Eq. (5), around the fixed point rho* = 1/(u tau), theta* = 1 + (J - J_c) rho*, gives a Jacobian whose stability boundary is approximately Gamma(J - J_c) = 2/(u tau) + O(h), i.e. the effective critical coupling for finite tau is J_c + 2/(Gamma u tau), not J_c itself. Thus for any fixed finite tau and any J not exactly on this shifted line, the system is off-critical, and standard mean-field scaling predicts that Var(rho) saturates as r -> 0 rather than diverging. The entropy and MI curves in Figure 3 are interpolated from simulations at r = 1e-6 ms^-1 with no error bars, no finite-size scaling analysis, and no test against smaller r. The claimed 'robust as J is swept around the critical point' (Results C, Figure 3D) may therefore be an artifact of measuring above the saturation crossover. If Var(rho) saturates for off-critical J, the weak-input variance-coding mechanism and the tau ~ 1000 ms optimum are not robust across couplings, and the paper's main claim fails. This is the load-bearing condition, and the paper does not supply the derivation or data needed to secure it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":16180,"tokens_out":4729,"duration_ms":47646,"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":[{"comment":"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","section":"Results C / Discussion A"},{"comment":"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.","section":"Methods F / Eq. (13) / Figure 3"},{"comment":"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.","section":"Methods D / Eq. (14)"}],"minor_comments":[{"comment":"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.","section":"Figure 2/3 captions"},{"comment":"Equation (13) is garbled in the typeset text; the entropy functional should be displayed cleanly with the correct summation and normalization.","section":"Eq. (13)"},{"comment":"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.","section":"Results A"},{"comment":"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.","section":"Methods F"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised by the skeptic is substantive and should be the primary request for revision: the paper needs either a derivation of the r -> 0 variance scaling for finite tau and off-critical J, or direct simulations at smaller r with finite-size scaling. The paper is otherwise within scope for q-bio.NC and the proposed mechanism is interesting. No citation or novelty concerns beyond what is stated in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat'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.\n\nThat'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.","headline":"Novel variance-coding mechanism in adaptive networks, but the robustness-to-J claim needs a direct scaling test before it can be trusted.","tokens_in":16837,"tokens_out":5151,"would_cite":false,"duration_ms":50525,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive spike thresholds let weak inputs be encoded as firing-rate variance","keywords":["threshold adaptation","rate coding","pattern coding","variance coding","criticality","mutual information","dynamic range","recurrent excitatory network"],"falsifier":"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.","tokens_in":15670,"feed_emoji":"🧠","tokens_out":4100,"duration_ms":40482,"temperature":0.7,"pith_summary":"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.","feed_headline":"Adaptive thresholds code weak inputs as firing-rate variance","feed_subtitle":"Slow threshold recovery near criticality adds pattern coding to rate coding without synaptic fine-tuning.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the theoretical foundation that dynamic range is optimized at criticality and gives the Stevens-exponent relation m = 1/δ_h used to characterize rate coding.","marker":"Kinouchi & Copelli, 2006"},{"why":"Provides experimental evidence that cortical networks at criticality maximize dynamic range, the baseline against which nonadaptive and adaptive coding are compared.","marker":"Shew et al., 2009"},{"why":"Reports measured spike-threshold recovery times of 10²–10³ ms in hippocampal CA3 neurons and mossy cells, defining the biologically relevant τ range and supporting the adaptation model.","marker":"Trinh et al., 2023"},{"why":"Shows similar slow threshold adaptation in fish pallium, extending the hippocampal timescale result to another memory-related circuit.","marker":"Trinh et al., 2019"},{"why":"Introduces threshold fatigue as a mechanism for information transfer and supplies the mutual-information measure for systems with spontaneous activity used in the paper.","marker":"Chacron et al., 2007"},{"why":"Establishes the entropy–variance relationship for Gaussian signals that underlies the identification of firing-rate variance with pattern entropy.","marker":"Warland et al., 1996"},{"why":"Provides the unified theory that locates the mean-field directed-percolation critical point at J_c = 1/Γ for the constant-threshold model.","marker":"Girardi-Schappo et al., 2021"},{"why":"Defines self-organized quasicriticality and the ρ ∼ 1/τ stationary signature that the paper contrasts with its proposed adaptive near-critical coding.","marker":"Kinouchi et al., 2020"}],"fun_headline_variants":["Rate variance carries weak signals near criticality","Threshold adaptation adds pattern coding for weak inputs","Near-critical adaptive coding uses firing-rate variance","Weak inputs encoded by rate variance at criticality"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Rate variance carries weak signals near criticality","Threshold adaptation adds pattern coding for weak inputs","Near-critical adaptive coding uses firing-rate variance","Weak inputs encoded by rate variance at criticality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1312,"prompt_tokens":720,"completion_tokens":592,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":536}},"tokens_in":464,"tokens_out":592,"duration_ms":5544,"temperature":1.0,"reasoning_tokens":536,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:25:37.122234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces threshold fatigue as a mechanism for information transfer and supplies the mutual-information measure for systems with spontaneous activity used in the paper."}],"review_version":1}