{"id":"0cc696c4-508a-4b59-8dc4-b7cb819b8b40","arxiv_id":"1908.04103","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In a plastic network of Hodgkin-Huxley neurons, frequency-synchronized clusters of different sizes self-organize and produce slow modulations of the population mean field.","lead":"This paper shows that a network of Hodgkin-Huxley neurons with spike timing-dependent plasticity can split into a few groups that fire at different frequencies. The authors build a simple two-cluster model to explain the stability of this split and to show how it could create slow brain rhythms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Common-period approximation in Eq. (8)-(9) underlies the Fig. 13(A) stability boundary; the paper does not report the frequency mismatch or STDP step needed to verify it.","rationale":"The paper's central claim has two parts: an empirical observation (frequency clusters, hierarchical in size, in HH networks with STDP) and a mechanistic explanation (the two-dimensional reduced model with sigma_dot = epsilon*G(phi)). The observation is supported by several simulations (Figs. 3, 5, 6, 9) and is likely robust. The load-bearing part for the claimed mechanism is the derivation of G(phi) in Eqs. (8)-(9), because it determines the sigma-nullcline and hence the predicted stability regions in Fig. 13(A) and the claimed criteria for plasticity functions. The derivation assumes a common period T for both clusters; for the HH two-cluster states the frequencies are not exactly equal, and the paper does not report the frequency mismatch used in the comparison nor the STDP step delta. Without these, the regime of validity of the common-T approximation is unverified, and the correspondence in Fig. 13 could be fortuitous. The reader's weakest assumption identifies the same point. I agree, and the recommended verdict stays CONDITIONAL: the numerical phenomenon is credible, but the mechanistic claim should be conditional on a quantitative check of the common-T approximation and on reporting of delta and omega.","tokens_in":14830,"tokens_out":10306,"duration_ms":112463,"concrete_test":"Recompute the stability boundary in Fig. 13(A) using the exact average STDP update for two periodic spike trains with periods T1=2*pi/omega1 and T2=2*pi/omega2, i.e., simulate the circle map Delta_{n+1}=(Delta_n+T2-T1) mod T2 and average W over one beat cycle, using the frequency mismatch implied by Fig. 7(A) for Ns=7, N=50. If this boundary shifts by more than the width of the white region in Fig. 13(A), the common-T approximation is load-bearing and the HH comparison in Fig. 13(B) does not validate the reduced model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reduced model's sigma-nullcline, and hence the predicted stability boundary in Fig. 13(A), rests on Eqs. (8)-(9), where the STDP update rate is computed as (delta/T)[W(T-DeltaT)+W(DeltaT)] under the assumption that both clusters share a common period T approximately 2*pi/omega_bar, with DeltaT approximately (phi mod 2*pi)/omega_bar. This is exact only when the two cluster frequencies are equal. For a frequency mismatch omega, the true sequence of inter-cluster spike intervals follows a circle map with periods T1 and T2, and the average update rate over a beat cycle is not generally equal to (delta*omega_bar/(2*pi))*G(phi). The error is of order (omega/omega_bar) times the slope of W, and can be sizeable near phi=0, where W is steepest and the potentiation that determines cluster fusion or separation is strongest. The paper never reports the frequency mismatch used in the HH simulations underlying Figs. 12-13, nor the STDP step delta, so it is unverified that the simulations lie in the regime where the common-T approximation is accurate. If this approximation is inaccurate, the sigma-nullcline G(phi)=0 and the criterion sigma(phi*)=sigma* that determines the white/black regions in Fig. 13(A) could shift, and the claimed correspondence with Fig. 13(B) would not constitute evidence for the mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports numerical observations of frequency clustering in a Hodgkin-Huxley network with symmetric spike timing-dependent plasticity (STDP). Depending on initial conditions, the network either fully synchronizes or splits into a few clusters that are synchronized internally at different frequencies and are hierarchical in size; the mean synaptic activity then exhibits slow amplitude modulations on a timescale much longer than individual spiking. The authors propose a two-dimensional phenomenological model for two clusters, with phase difference phi and mean inter-cluster coupling sigma, in which the coupling evolves as sigma_dot = epsilon G(phi), where G is derived from the symmetric plasticity function W. They analyze the phase portraits of this reduced model, compare them with the HH simulations, and use the model to predict regions in the (cp, cd) plane of the plasticity function where stable clusters, full synchronization, or full decoupling occur.","tokens_in":15161,"tokens_out":9411,"duration_ms":103692,"significance":"If correct, the paper offers a concrete and physiologically plausible mechanism by which a population of tonically firing neurons with adaptive synapses can generate slow modulations of the population mean field, relevant to LFP/EEG/BOLD observations. The two-cluster reduction is attractive because it reduces a high-dimensional spiking network to a planar system with an explicit design criterion for plasticity rules that produce clusters. The paper gives credit to the approximate nature of the reduction and supports the main phenomenon with several numerical realizations and a robustness test under random inputs. However, the quantitative correspondence claimed in Fig. 13 is not yet fully established because the averaging step used to derive G is not validated for the reported detuning, and because key parameter values (delta, omega, and sometimes epsilon) are not reported.","major_comments":[{"comment":"The derivation of sigma_dot = epsilon G(phi) assumes that, for a fixed phase difference phi, the two inter-cluster spike intervals per cycle are DeltaT and T-DeltaT with DeltaT approximately (phi mod 2pi)/omega_bar. For clusters with different frequencies, the phase difference at successive spike encounters advances by about 2pi omega/omega_bar, so the actual sequence of intervals follows a circle map and the average STDP update over a beat cycle is not generally equal to (delta omega_bar / 2pi) G(phi). The error is of order (omega/omega_bar) times the slope of W, which is largest near phi=0 where W is steep and where the decision between cluster fusion and separation is made. Because the paper does not report the effective frequency mismatch omega of the HH two-cluster states underlying Figs. 12-13, the claimed correspondence between the reduced-model boundary in Fig. 13(A) and the HH results in Fig. 13(B) is not yet quantitatively verified.","section":"Model derivation (Eqs. (8)-(9))"},{"comment":"The STDP update size delta in Eq. (2) is never specified, and the frequency mismatch omega used for the reduced model is not given in the captions of Figs. 12 and 13; the Fig. 11 captions state omega = 0.037 kHz, omega = 0.06 kHz, and epsilon = 0.08, but Fig. 13 does not state whether the same values were used. The reduced-model fixed point and the stability criterion sigma(phi*) = sigma* depend on omega through alpha = asin(omega/sigma_max), and the value of sigma(phi*) obtained by integrating system (10)-(11) from (-phi*,0) depends on epsilon. Without these parameter values, the simulations cannot be reproduced and the comparison in Fig. 13 cannot be checked.","section":"Simulation parameters (Eq. (2); Figs. 12-13)"},{"comment":"The HH classification labels a parameter set as stable-two-clusters if the clusters 'stay apart after the simulation time 3000 ms.' Earlier in the paper, the smaller cluster is reported to reach its final state only at t approximately 17000 for another realization, so a 3000 ms cutoff may misclassify slowly fusing clusters as stable. The authors should justify this threshold or report that the classification is unchanged for longer simulation times.","section":"Criteria for the emergence of clusters (Fig. 13(B))"}],"minor_comments":[{"comment":"The nullcline labels appear to be swapped: from the equations, G(phi)=0 gives the sigma-nullcline as the vertical lines phi = phi* and phi = -phi*, while sigma = omega/sin(phi+alpha) is the phi-nullcline; the text states the opposite.","section":"Properties of the model (paragraph after Eq. (11))"},{"comment":"Please state the values of omega, epsilon, delta, and sigma_max used for panel (A), and the precise numerical protocol used for panel (B), including the integration time and how the final state was classified.","section":"Fig. 13 caption"},{"comment":"The entry 'omega approximately 0.455 Hz' appears to be in different units from the other entries in the same caption, which are given in kHz; please correct the units.","section":"Fig. 11 caption"},{"comment":"The term 'hierarchical in size' is used descriptively; a short sentence defining the sense in which the cluster structure is hierarchical (e.g., each newly formed cluster is markedly smaller than the preceding one) would improve precision.","section":"Results, 'Numerical observation'"}],"recommendation":"major_revision","confidential_remarks":"The paper reports an interesting and potentially impactful phenomenon, and the reduced model is a useful conceptual tool. However, the central quantitative comparison in Fig. 13 currently depends on unreported parameters and on an averaging approximation that is not validated for the detuned case; both issues are fixable with additional reporting and analysis. I would encourage the editor to require a complete parameter table and a robustness check of the common-period approximation before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a solid, competent computational study of STDP-induced frequency clusters in Hodgkin-Huxley neurons, with a useful reduced two-cluster model. It is not a breakthrough—frequency clustering with STDP has been reported before for phase oscillators and for Morris-Lecar neurons. The new bits are the HH setting, the symmetric plasticity rule, and the two-dimensional phenomenological model that explains why clusters of sufficiently different sizes are stable.\n\nWhat the paper does well: the numerics are convincing. The cluster states are shown with multiple realizations, the time course of cluster formation is interesting (the small cluster takes much longer to form), and the robustness to independent random input is a nice addition. The reduced model is a sensible phase-oscillator approximation, and the qualitative match in Fig. 12 is credible. The authors also honestly note that the cluster state coexists with complete synchrony, which is the right thing to flag.\n\nThe main soft spot is in the derivation of the sigma-nullcline. Equation (8) computes the STDP update rate assuming both clusters share a common period T ≈ 2π/ωbar. With a frequency mismatch ω, the actual sequence of inter-spike intervals follows a circle map, and the average update is not exactly (δωbar/2π)G(φ); the error is of order (ω/ωbar) times the slope of W, which is steepest near φ=0. The paper never reports the ω used in the HH simulations behind Figs. 12–13, nor the STDP step δ, so the quantitative boundary in Fig. 13(A) is not fully pinned down. Given the frequencies quoted elsewhere (e.g., ~0.065 kHz with a difference ~0.0004 kHz), the common-period approximation is probably good to a few percent, so this is not likely fatal. But the missing parameters should be given so the comparison can be checked.\n\nTwo smaller issues. The abstract says the paper determines the set of plasticity functions that lead to \"multi-clustering,\" but the actual analysis is limited to two clusters of fixed sizes; the multi-cluster case is only observed numerically. That is an overstatement. Also, no code or data are provided, which is a gap for a computational paper.\n\nBottom line: the central phenomenon is likely real, and the reduced model is a helpful explanatory tool. I would send this to a serious referee. A revision should specify δ and ω, discuss the common-period approximation, and soften the abstract. Worth a look if you work on adaptive networks or slow oscillations.","headline":"A plausible, well-illustrated computational study of STDP-driven frequency clusters in Hodgkin-Huxley neurons, with a useful reduced model; the missing simulation parameters make the quantitative comparison hard to verify but not a fatal flaw.","tokens_in":15667,"tokens_out":3551,"would_cite":true,"duration_ms":37699,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34C15","34D06","37N25","92C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper reports that a Hodgkin-Huxley network with symmetric spike-timing-dependent plasticity can self-organize into hierarchical frequency clusters, and that a two-variable phase-and-coupling model captures when those clusters…","keywords":["spike timing-dependent plasticity","Hodgkin-Huxley neurons","frequency clustering","hierarchical clusters","slow oscillations","mean field modulation","adaptive network","synchronization"],"falsifier":"Track the projected phase difference $\\varphi_{\\mathrm{HH}}$ and mean inter-cluster coupling $\\sigma_{\\mathrm{HH}}$ in the full network near the predicted fusion boundary and check the reduced model's condition $\\sigma(\\varphi^*) < \\sigma^*$ for stable clusters; a surviving cluster with $\\sigma(\\varphi^*) > \\sigma^*$, or a fusing pair with $\\sigma(\\varphi^*) < \\sigma^*$, would refute the mechanism.","tokens_in":14640,"feed_emoji":"🧠","tokens_out":10598,"duration_ms":101227,"temperature":0.7,"pith_summary":"This paper reports that a population of Hodgkin-Huxley neurons with symmetric spike-timing-dependent plasticity can self-organize into frequency clusters of different sizes: the largest cluster fires slowest, each smaller cluster fires faster, and connections between clusters are depressed while connections within a cluster are potentiated. The authors' central move is to compress two interacting clusters into a two-dimensional system for their phase difference and mean coupling, with the coupling adapting on a much slower timescale through a plasticity function G(φ). That model explains why clusters of sufficiently different sizes are stable and similar-sized clusters merge, and it predicts the plasticity windows that allow clustering. The wider consequence is that cluster dynamics can amplitude-modulate the population's mean activity at sub-Hertz timescales even though every neuron fires tonically, offering a candidate mechanism for slow fluctuations in LFP, EEG, and BOLD signals.","feed_headline":"Spike-timing plasticity splits neurons into frequency clusters","feed_subtitle":"Clusters beating at different rates slow-modulate the population signal, a candidate source of slow EEG and BOLD rhythms.","key_machinery":"The central object is the two-cluster phenomenological model: $\\dot\\varphi = \\omega - \\sigma \\sin(\\varphi+\\alpha)$ and $\\dot\\sigma = \\varepsilon G(\\varphi)$, where $G(\\varphi) = W\\big((2\\pi-(\\varphi \\bmod 2\\pi))/\\bar\\omega\\big) + W\\big((\\varphi \\bmod 2\\pi)/\\bar\\omega\\big)$ sums the two STDP updates per common period $T\\approx 2\\pi/\\bar\\omega$, and $W$ is the double-exponential symmetric plasticity window. The small parameter $\\varepsilon$ separates the fast phase dynamics from the slow coupling adaptation. The phase portrait is organized by a saddle fixed point $(\\varphi^*,\\sigma^*)$; the condition $\\sigma(\\varphi^*) = \\sigma^*$ marks the boundary between fusion and stable coexistence, and the paper uses it to map which plasticity windows $W$ produce stable clusters.","core_discovery":"The paper claims that a network of identical Hodgkin-Huxley neurons with symmetric spike-timing-dependent plasticity does not settle into a single fate: depending on initial conditions and plasticity parameters, it can reach full synchrony or a hierarchical cluster state in which each group fires at its own frequency and the largest group fires slowest. The stabilization of these clusters is controlled by the slow adaptation of inter-cluster synaptic weights. When two clusters come into phase, their mutual coupling is potentiated, but if their frequency difference is large enough the in-phase episode is brief and the subsequent out-of-phase interval depresses the inter-cluster weights back toward zero, keeping the clusters apart. The paper reduces this to the two-dimensional system $\\dot\\varphi = \\omega-\\sigma\\sin(\\varphi+\\alpha)$, $\\dot\\sigma = \\varepsilon G(\\varphi)$ and shows that the same phase portraits—fusion, coexistence of synchrony with stable clusters, and full decoupling—appear in the full network when its cluster phases and mean inter-cluster coupling are projected onto the $(\\varphi,\\sigma)$ plane. It also shows that the mean synaptic activity of the two-cluster state is modulated on a timescale of seconds, about two orders of magnitude slower than the tens-of-milliseconds firing periods, which the authors connect to slow LFP, EEG, and BOLD oscillations.","pith_inferences":["A natural extension is to test the same slow-fast phase-and-coupling reduction in other conductance-based spiking models; the mechanism should survive as long as within-cluster synchrony and weak inter-cluster coupling hold.","The hierarchical size distribution implies a discrete set of reachable cluster partitions; mapping which initial conditions lead to which partition would treat the network as a multi-stable encoder of its own history.","The model predicts a sharp experimental signature: blocking the potentiation of inter-cluster synapses during in-phase episodes should remove the slow mean-field modulation while leaving tonic firing intact, distinguishing this mechanism from slow-bursting generators."],"forward_implications":["Stable frequency clusters require sufficiently different cluster sizes; similar-sized clusters fuse into global synchrony.","The population mean field of a two-cluster state is amplitude-modulated on a timescale set by the cluster frequency difference (about 2.5 seconds in the example), two orders of magnitude slower than the roughly 15-millisecond firing period.","There is a coexistence regime in which complete synchrony and stable clusters are both attractors, so initial conditions or perturbations determine which state is reached.","For plasticity windows with $c_d \\geq c_p$ and $\\tau_d \\geq \\tau_p$, the update function $G(\\varphi)$ is non-positive everywhere and the network decouples instead of clustering.","Weak random synaptic input preserves clusters, while stronger input destroys the smaller clusters first, meaning cluster stability is size-dependent."],"supporting_citations":[{"why":"Supplies the experimentally observed symmetric STDP window used in the simulations.","marker":"[1]"},{"why":"Shows how an effectively symmetric plasticity function can arise from asymmetric STDP, motivating the choice of W.","marker":"[2]"},{"why":"Provides the Hodgkin-Huxley network with STDP whose self-organized dynamics this paper extends to frequency clustering.","marker":"[28]"},{"why":"Supplies the two-oscillator long-term plasticity setup behind the slow coupling update in the reduced model.","marker":"[29]"},{"why":"Establishes multi-cluster states in adaptively coupled phase oscillators, the phenomenon the full network is compared with.","marker":"[37]"},{"why":"Contains the stability analysis of hierarchical frequency clusters in phase-oscillator networks that motivates the size-difference condition.","marker":"[38]"},{"why":"Reports cluster formation and slow mean-field modulation in bursting neurons with STDP, the prior neural result this paper builds on.","marker":"[39]"},{"why":"Defines the Hodgkin-Huxley neuron model used for the full network.","marker":"[44]"},{"why":"Provides the weak-coupling phase dynamics for Hodgkin-Huxley neurons that justifies reducing clusters to phase oscillators.","marker":"[45]"},{"why":"Supplies the standard phase and synchronization formalism used to extract cluster phases from spike times.","marker":"[47]"}],"fun_headline_variants":["Plasticity splits neural networks into frequency-tuned clusters","Neural plasticity creates hierarchical frequency clustering","Spike-timing plasticity yields multi-frequency neural clusters","Slow brain rhythms from plastic neural frequency clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that during the brief moments when two clusters align, their spiking periods are close enough to a single common value that the plasticity update can be written as a fixed function of their phase difference alone; if the periods diverge during those episodes, the predicted boundary between stable and fusing clusters may not carry over from the reduced model to the full network.","fun_headline_variants_meta":{"raw":{"variants":["Plasticity splits neural networks into frequency-tuned clusters","Neural plasticity creates hierarchical frequency clustering","Spike-timing plasticity yields multi-frequency neural clusters","Slow brain rhythms from plastic neural frequency clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000667,"raw_usage":{"total_tokens":3039,"prompt_tokens":939,"completion_tokens":2100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":555,"tokens_out":2100,"duration_ms":15195,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:51:19.808249+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track the projected phase difference $\\varphi_{\\mathrm{HH}}$ and mean inter-cluster coupling $\\sigma_{\\mathrm{HH}}$ in the full network near the predicted fusion boundary and check the reduced model's condition $\\sigma(\\varphi^*) < \\sigma^*$ for stable clusters; a surviving cluster with $\\sigma(\\varphi^*) > \\sigma^*$, or a fusing pair with $\\sigma(\\varphi^*) < \\sigma^*$, would refute the mechanism.","supporting_citations":[{"cited_title":"Malleability of Spike-Timing-Dependent Plasticity at the CA3-CA1 Synapse","cited_arxiv_id":null,"evidence_quote":"Supplies the experimentally observed symmetric STDP window used in the simulations."},{"cited_title":"Spike-timing-dependent plasticity: common themes and divergent vistas","cited_arxiv_id":null,"evidence_quote":"Shows how an effectively symmetric plasticity function can arise from asymmetric STDP, motivating the choice of W."},{"cited_title":"Self-organized noise resistance of oscillatory neural networks with spike timing-dependent plasticity","cited_arxiv_id":null,"evidence_quote":"Provides the Hodgkin-Huxley network with STDP whose self-organized dynamics this paper extends to frequency clustering."},{"cited_title":"Noise-enhanced cou- pling between two oscillators with long-term plasticity","cited_arxiv_id":null,"evidence_quote":"Supplies the two-oscillator long-term plasticity setup behind the slow coupling update in the reduced model."},{"cited_title":"Multi-clusters in networks of adaptively coupled phase oscillators networks","cited_arxiv_id":null,"evidence_quote":"Establishes multi-cluster states in adaptively coupled phase oscillators, the phenomenon the full network is compared with."},{"cited_title":"Hierarchical frequency clusters in adaptive networks of phase oscillators","cited_arxiv_id":null,"evidence_quote":"Contains the stability analysis of hierarchical frequency clusters in phase-oscillator networks that motivates the size-difference condition."},{"cited_title":"The Spacing Principle for Un- learning Abnormal Neuronal Synchrony","cited_arxiv_id":null,"evidence_quote":"Reports cluster formation and slow mean-field modulation in bursting neurons with STDP, the prior neural result this paper builds on."},{"cited_title":"A quantitative description of membrane current and its application to conduction and excitation in nerve","cited_arxiv_id":null,"evidence_quote":"Defines the Hodgkin-Huxley neuron model used for the full network."},{"cited_title":"Phase Dynamics for Weakly Cou- pled Hodgkin-Huxley Neurons","cited_arxiv_id":null,"evidence_quote":"Provides the weak-coupling phase dynamics for Hodgkin-Huxley neurons that justifies reducing clusters to phase oscillators."},{"cited_title":"Synchronization.{A} Universal Con- cept in Nonlinear Sciences","cited_arxiv_id":null,"evidence_quote":"Supplies the standard phase and synchronization formalism used to extract cluster phases from spike times."}],"review_version":1}