{"id":"9cd9bc5d-6aa5-4a5d-ad5d-827d6a104715","arxiv_id":"2509.02799","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A trained MLP reproduces QIF network mean-field dynamics, adds connection probability p as a whole-brain parameter, and improves parameter recovery in simulation-based inference.","lead":"This paper shows that a neural network can learn the averaged firing rate and voltage dynamics of a spiking neuron network, including a connection-sparsity parameter that no exact mathematical model currently handles. It then inserts this learned model into whole-brain simulations, where it improves parameter recovery in synthetic tests compared with the standard analytical model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear coupling used in the whole-brain/SBI pipeline is acknowledged to be invalid for the MLP; the resulting error is unquantified and may drive the reported results.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the whole-brain and SBI results depend on a linear-coupling identity that Appendix B shows the MLP violates. I agree this is the single most important threat to the paper's applied claims. The MLP's ability to reproduce MPR at p=1 and the recovered chaotic dynamics are credible internal validations, but they do not establish the validity of the coupled whole-brain model. The concrete test I propose directly compares the paper's linearized coupling with the true nonlinear input dependence, using the same figures and inference pipeline as the manuscript. If the test shows negligible differences, the concern is resolved and the current conditional verdict could move toward acceptance. If it shows material differences, the whole-brain and SBI sections would need to be reinterpreted or re-run. Since the concern is testable and the reader already conditioned the verdict on this issue, the appropriate recommendation is to keep the verdict unchanged rather than escalate or de-escalate without the test result.","tokens_in":15738,"tokens_out":5870,"duration_ms":73643,"concrete_test":"Run the TVB_QIF whole-brain simulations twice on the same connectome and parameter grid: once with the paper's linearized coupling, MLP(x,{k},0) + g sum_j W_ij psi_j, and once with the full nonlinear input coupling, MLP(x,{k}, g sum_j W_ij psi_j), with Iext fed inside the MLP. For both schemes, compute the Var(FCD_upper) and GC maps over the (g, eta) grid of Fig. 4 and repeat the SBI posterior estimation on the 500 synthetic observations. If the fluidity ridge shifts by more than one grid step or the posterior z-scores/shrinkage change materially (e.g., mean |Delta z| > 0.1), the linear-coupling assumption is load-bearing and the whole-brain/SBI conclusions require qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the linear-coupling approximation in the whole-brain/SBI pipeline. Equation (5) requires that the neural mass node be affine in the input current: N(psi,I)=N(psi,0)+I. This holds for MPR (Eq. 2: I enters linearly in the v equation), but Appendix B tests the MLP and finds that dr_dot/dIext and dv_dot/dIext are not constant over (r,v); hence MLP(r,v,eta,p,J,Iext) != MLP(r,v,eta,p,J,0)+Iext. The authors then build TVB_QIF using the linear form anyway and state only that the behavior 'still remains consistent,' with no quantitative error measure. All whole-brain results—the fluidity maps in Fig. 4, the p-sweep, and the SBI z-score/shrinkage comparisons in Fig. 5—are generated under this violated assumption. If the nonlinearity in Iext is significant in the operating regime of the coupled model, the recovered posteriors and the claimed improvement over TVB_MPR may be artifacts of the linearization rather than properties of the learned mean-field. The cusp bifurcation claim is also not independently verified against spiking simulations, but the linear-coupling issue is more immediately load-bearing for the whole-brain and SBI validation, and it is explicitly acknowledged in the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a data-driven framework in which a multi-layer perceptron (MLP) is trained to reproduce the macroscopic phase flow of a network of Quadratic-Integrate-and-Fire (QIF) neurons, using the population firing rate r, mean membrane potential v, the excitability parameter η, synaptic weight J, connection probability p, and an external input I(t) as inputs. The trained MLP is validated against the analytical Montbrió-Pazó-Roxin (MPR) mean-field model for p=1, including recovery of the phase diagram and a chaotic regime. Bifurcation analysis on the MLP is used to claim a cusp bifurcation in the (p,η) plane that is degenerate with J. The MLP is then embedded in a whole-brain model with linear structural coupling (TVB_QIF), compared to the standard TVB_MPR model, and used in simulation-based inference (SBI) to recover η, J, and p from synthetic and real fMRI data. The authors report that TVB_QIF recovers ground-truth parameters with lower z-scores than TVB_MPR, while TVB_MPR produces confidently wrong estimates.","tokens_in":16125,"tokens_out":6157,"duration_ms":68852,"significance":"If the claims hold, the paper offers a flexible and generic route from spiking-network simulations to mean-field models that are not constrained by analytical closure, and demonstrates a concrete use-case in whole-brain modeling and Bayesian parameter inference. The p=1 validation against the exact MPR solution and the reproduction of the chaotic regime are genuine strengths and give independent support for the MLP reconstruction method. The paper is also notably transparent in Appendix B about the failure of the linear-coupling assumption. However, the whole-brain and SBI results rest on an unquantified approximation, and the claim of a new cusp bifurcation in p has no independent ground truth. These issues are load-bearing for the paper's central claims and require additional validation before the results can be accepted.","major_comments":[{"comment":"The whole-brain model and all downstream SBI results rely on the linear-coupling form N(ψ,I) ≈ N(ψ,0)+I. Appendix B explicitly shows ∂ṙ/∂Iext and ∂v̇/∂Iext are not constant in (r,v), so this approximation is violated. The text only says that 'when enforcing it, the behavior of the model still remains consistent,' but no quantitative error measure is provided. Since the coupling term gΣW_ijS(ψ_j) is not small in the operating regime (Fig. 4 sweeps g around 0.55–0.66), the nonlinearity in Iext could materially change the whole-brain dynamics. The claimed SBI accuracy may therefore be an artifact of the linearized model rather than a property of the learned mean-field. Please provide a concrete error quantification over the operating range of (r,v) and coupling strengths, and ideally validate a coupled spiking-network simulation against TVB_QIF; alternatively, implement the coupling direct","section":"§III C and Appendix B, Eq. (5), Eq. (B1)"},{"comment":"The paper claims the trained MLP reveals a new cusp bifurcation in the (p,η) plane. However, for p<1 there is no analytical ground truth, and the MLP itself is the object under study, so the cusp is a property of the fitted function rather than of the spiking network. The text also acknowledges that p and J are 'highly degenerate' and that decreasing p has nearly the same effect as decreasing J. In the large-N limit the mean-field input should scale as pJ r, suggesting that the 'new' cusp may simply be the known (J,η) cusp reparameterized by the product pJ. Please test this directly: compare the cusp curves from Fig. 3(a) with those obtained by rescaling J by p, or train an MLP on the product pJ and show whether the two-parameter dependence is irreducible. Additionally, validate the cusp location at p<1 using direct spiking simulations at representative parameter points, and report varia","section":"§III B and Fig. 3"},{"comment":"The SBI ground-truth observations are generated with TVB_QIF, which is the same model family used for inference. This is a standard self-consistency test, but it does not validate the model against the actual spiking network. Since TVB_QIF is built on the linear-coupling approximation of Appendix B, the recovered posteriors for η, J, and p characterize the linearized model, not necessarily the QIF network. The comparison with TVB_MPR is informative as a model-misspecification example, but the conclusion that the data-driven model is more accurate for fMRI parameter recovery requires an end-to-end validation with data generated from a true coupled QIF network (or at least from a coupling scheme that avoids the linearization error). Without this, the reported improvement may be an artifact of the violated coupling assumption.","section":"§III D and Fig. 5"}],"minor_comments":[{"comment":"Grammar: 'existence of new cusp bifurcation' should be 'existence of a new cusp bifurcation'.","section":"Abstract"},{"comment":"The caption contains a typo: 'p = 0.92p = 0.88 simulated with simulated with and TVB QIF' appears garbled. Please correct.","section":"Figure 4 caption"},{"comment":"The caption says 'In (d),(h) and (e),(f),(h)'—the enumeration of panels is inconsistent and should be corrected.","section":"Figure 5 caption"},{"comment":"The text states 'a total of N = 220000 simulations.' Given the stated grid: η ∈ [−8,−1] with step 0.05, p ∈ [0.6,1] with step 0.1, J ∈ [5,15] with step 0.5, and 5 seeds, the number of simulations is approximately 74,000. The discrepancy should be clarified (e.g., number of training segments rather than simulations).","section":"§III A"},{"comment":"The cusp curves and fixed-point locations are shown without any measure of variability across MLP training runs or across noise seeds. A brief quantitative robustness statement would strengthen the presentation.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about the linear-coupling violation, which is a positive signal, but the whole-brain and SBI claims cannot be accepted as they stand without a quantitative validation of that approximation. The p<1 cusp claim should also be checked against independent spiking simulations, especially because p and J appear to be trivially degenerate in the mean-field limit. If code/data are not already public, the authors should be encouraged to share them to support reproducibility of the MLP training and SBI pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper has a good idea and some honest internal checks, but its main whole-brain validation leans on an approximation the authors admit is violated, and the flagship 'new cusp' is never checked against direct spiking simulations. I'd send it to review, but the revision needs to be substantial.\n\nWhat's new and valuable: using an MLP to learn the phase flow of a QIF network, with connection probability p as an input, is a genuinely useful move for mean-field approaches that are analytically intractable. For p=1 the trained MLP matches the MPR phase diagram and recovers the known chaotic regime even though that regime wasn't in the training data—real evidence the surrogate learned the dynamics, not just fitted a surface. The continuation analysis on the trained MLP is a sensible way to extract bifurcation structure. The paper is also honest enough to note that p and J are degenerate, and the point that inference with an incomplete model can be 'confidently wrong' is fair.\n\nWhere it gets shaky: the whole-brain model and the SBI comparison are built on Equation (5)'s linear coupling, and Appendix B plainly states the MLP does not satisfy the required affine-in-I condition. The authors say the behavior 'still remains consistent' but give no quantitative error check. That matters because the SBI results in Figure 5 are supposed to show the learned model recovers parameters better; if the coupling approximation distorts TVB_QIF and TVB_MPR differently, the comparison is compromised. The cusp in the (p, eta) plane is claimed as a new result, but it is derived from the fitted MLP itself, and with p and J effectively interchangeable, it may be little more than a re-expression of the known (J, eta) cusp. There is no independent verification against spiking simulations for p<1, no uncertainty quantification on the MLP, and no released code or data. The training benchmark in Appendix E is nice, but it doesn't address these issues.\n\nProportionally: none of this kills the core idea. The MLP surrogate for p=1 is well validated, and the framework is clearly transferable to other micro-scale parameters. The deficiencies are in validation of the novel parts, and they are fixable: quantify the linear-coupling error in the operating regime, test the predicted cusp against direct spiking simulations for p<1, and release the code.\n\nWho is this for? Computational neuroscientists working on whole-brain models and mean-field reductions. A serious referee should engage, but I would not cite it for the whole-brain claims until the linear-coupling issue is addressed.","headline":"Promising MLP-learned mean-field for QIF networks, but the whole-brain validation rests on a violated linear-coupling approximation and the new cusp is not independently verified.","tokens_in":16554,"tokens_out":2673,"would_cite":false,"duration_ms":29582,"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":"A multilayer perceptron trained directly on simulations of a quadratic integrate-and-fire network can learn its macroscopic mean-field dynamics, including sparse connectivity as a new parameter, and when embedded in a whole-brain model it r","keywords":["data-driven mean-field","QIF network","whole-brain model","multilayer perceptron","bifurcation analysis","simulation-based inference","connection probability","neural mass model"],"falsifier":"Measure the partial derivatives ∂ṙ/∂I_ext and ∂v̇/∂I_ext across the (r, v) phase space for the trained MLP; if their non-constant variation shifts the cusp location in the (p, η) plane or changes whole-brain functional connectivity by a detectable amount, the claimed phase diagram and inference results are not robust. Alternatively, simulate a two-node spiking network with true synaptic coupling and compare its joint activity to a two-node MLP whole-brain model under the same input; a clear mismatch in correlation structure would falsify the linear-coupling approximation.","tokens_in":15695,"feed_emoji":"🧠","tokens_out":5380,"duration_ms":58077,"temperature":0.7,"pith_summary":"The paper claims that a multilayer perceptron can learn the macroscopic dynamics of a network of quadratic integrate-and-fire (QIF) neurons directly from simulations, bypassing the restrictive all-to-all connectivity assumption of analytic mean-field theory. It introduces the connection probability p as a new parameter, performs bifurcation analysis on the trained network, and finds a cusp bifurcation in the (p, η) plane that reshapes the phase diagram in a way degenerate with synaptic coupling J. Embedding this learned mean-field in a whole-brain model, the authors use simulation-based inference on synthetic fMRI data and show accurate parameter recovery, while the analytic MPR model yields estimates that are confidently wrong. If correct, this offers a flexible data-driven bridge from micro-scale spiking rules to macro-scale brain dynamics, with the potential to build mean-field models for neuron models where analytic closure is intractable.","feed_headline":"MLP-learned brain model recovers fMRI parameters analytic model misses","feed_subtitle":"A trained MLP reproduces QIF mean-field, adds connection probability, and beats analytical MPR in inference.","key_machinery":"The central object is the trained MLP that represents the vector field ẋ = MLPΘ(x, η, J, p, I(t)) with state x = (r, v). It is trained by unrolling the MLP through a Heun integrator and minimizing the average Euclidean distance between predicted and observed population trajectories from QIF simulations. Once trained, the frozen weights define a deterministic dynamical system on which numerical continuation can be run to locate fixed points and bifurcation branches; the derivative with respect to the input current is computed via automatic differentiation to test coupling linearity. Adding p as an input parameter is the mechanism that bypasses the analytic all-to-all connectivity assumption,","core_discovery":"The central claim is that the phase flow of a network of QIF neurons—its mean firing rate r and mean membrane potential v—can be learned by an MLP from simulated time series without needing an analytic closure. The trained MLP reproduces the analytical MPR model almost exactly at p=1, and extends it to sparse connectivity p<1, which the analytic derivation cannot handle. Numerical continuation on the frozen MLP reveals a cusp bifurcation in the (η, p) plane with the same shape as the known cusp in (η, J), implying p and J are nearly degenerate: decreasing connection probability acts like decreasing synaptic coupling. The learned mean-field is then used as a neural mass inside a whole-brain m","pith_inferences":["The linear-coupling assumption (Appendix B) is violated for the MLP, yet the whole-brain simulations rely on it; a direct test comparing the MLP whole-brain model with the original spiking network under identical structural connectivity would show whether the parameter-recovery results survive without this approximation.","Because p and J are degenerate, the posterior in (p, J) is likely non-identifiable even in the MLP model; the improved inference may reflect a better match to finite-size effects rather than a true separation of p and J.","Extending the same pipeline to partially observed states (e.g., only firing rate) would require latent-state reconstruction or additional observables; the paper explicitly restricts to full observability, which limits direct application to real neural recordings.","The method naturally extends to detailed neuronal simulators (multi-compartment, conductance-based) where analytic mean-field is intractable; a concrete next step is to train on such a simulator and compare its macroscopic activity against the learned model."],"forward_implications":["A mean-field can be derived for any spiking network simulation with a fully observed macroscopic state, even when no analytic closure exists, such as with sparse connectivity or non-Lorentzian heterogeneity.","Connection probability becomes a first-class parameter in whole-brain models; because it is degenerate with synaptic coupling, inference must estimate both to avoid biased parameter recovery.","Simulation-based inference using the learned mean-field recovers ground-truth parameters from synthetic fMRI better than the analytical model, suggesting that data-driven mean-fields reduce model misspecification bias in whole-brain inversion.","The learned dynamics extrapolate beyond the training data: cusp branches and the chaotic regime under sinusoidal input emerge without explicit sampling, indicating the MLP captures the underlying vector field rather than memorizing examples."],"supporting_citations":[{"why":"Provides the exact analytical mean-field (MPR model) used as ground truth for p=1 and as the baseline that the MLP must match.","marker":"[23]"},{"why":"Supplies the spiking network simulator used to generate the training dataset of QIF population activity.","marker":"[31]"},{"why":"Delivers the numerical continuation toolkit used to compute fixed points and bifurcation branches of the trained MLP.","marker":"[36]"},{"why":"Defines the simulation-based inference framework used for posterior parameter estimation from fMRI features.","marker":"[48]"},{"why":"Supplies the whole-brain model settings and prior parameter ranges, notably the fixed global coupling and noise values used in the inference experiments.","marker":"[8]"},{"why":"Motivates the connection probability p as a parameter of interest by showing theoretically how it shapes population fluctuations.","marker":"[29]"}],"fun_headline_variants":["MLP-learned mean-field finds cusp bifurcation in whole-brain model","Data-driven brain model beats analytic at fMRI parameter recovery","Sparse brain dynamics learned by neural net, new bifurcation found","Whole-brain model gets MLP mean-field, better fMRI fits","Learned mean-field extends brain model to sparse connectivity"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole-brain results assume that coupling between brain regions can be treated as a linear input to the local neural mass, even though the learned MLP is not linear in that input; the paper forces the linear coupling anyway and asserts the behavior stays consistent.","fun_headline_variants_meta":{"raw":{"variants":["MLP-learned mean-field finds cusp bifurcation in whole-brain model","Data-driven brain model beats analytic at fMRI parameter recovery","Sparse brain dynamics learned by neural net, new bifurcation found","Whole-brain model gets MLP mean-field, better fMRI fits","Learned mean-field extends brain model to sparse connectivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1214,"prompt_tokens":739,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":483,"tokens_out":475,"duration_ms":5985,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:22:48.306580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the partial derivatives ∂ṙ/∂I_ext and ∂v̇/∂I_ext across the (r, v) phase space for the trained MLP; if their non-constant variation shifts the cusp location in the (p, η) plane or changes whole-brain functional connectivity by a detectable amount, the claimed phase diagram and inference results are not robust. Alternatively, simulate a two-node spiking network with true synaptic coupling and compare its joint activity to a two-node MLP whole-brain model under the same input; a clear mismatch in correlation structure would falsify the linear-coupling approximation.","supporting_citations":[{"cited_title":"Stimberg, R","cited_arxiv_id":null,"evidence_quote":"Supplies the spiking network simulator used to generate the training dataset of QIF population activity."},{"cited_title":"BifurcationKit.jl,","cited_arxiv_id":null,"evidence_quote":"Delivers the numerical continuation toolkit used to compute fixed points and bifurcation branches of the trained MLP."},{"cited_title":"Cranmer, J","cited_arxiv_id":null,"evidence_quote":"Defines the simulation-based inference framework used for posterior parameter estimation from fMRI features."},{"cited_title":"Breyton, J","cited_arxiv_id":null,"evidence_quote":"Supplies the whole-brain model settings and prior parameter ranges, notably the fixed global coupling and noise values used in the inference experiments."},{"cited_title":"How connec- tion probability shapes fluctuations of neural population dynamics,","cited_arxiv_id":null,"evidence_quote":"Motivates the connection probability p as a parameter of interest by showing theoretically how it shapes population fluctuations."}],"review_version":1}