{"id":"1d5d85ce-3c5f-47b4-a514-66f1d8dc7aa8","arxiv_id":"2507.02450","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A voltage-fluctuation-based change point detector, combined with time-delayed correlation reconstruction, detects abrupt rewiring in balanced spiking networks and lifts reconstruction AUC by 0.13 to 0.26 over static correlation.","lead":"This paper proposes a two-step method to detect when a balanced spiking network's wiring changes, using the statistics of voltage fluctuations, and then to reconstruct the new wiring from spike trains. It matters because real brain circuits are continually rewired by plasticity, so connectivity inference from recordings needs to know when the structure changed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The subspace-separation mechanism in Eq. 26 fails for full-rank sparse W, and the detector's reliance on a finite-sample minimum-variance direction creates a selection bias that no negative control currently rules out.","rationale":"The reader's weakest_assumption correctly identifies the full-rank column-space problem in Eq. 26, and this is the most load-bearing weakness in the paper's theoretical justification. I agree that the stated subspace-separation mechanism cannot hold for the random sparse connectivity matrices used here, and the situation is even clearer for the recorded subnetwork: with 200 observed neurons and N=4000, the observed W_sub is generically of rank 200, so col(W1_obs)=col(W2_obs)=R^200. The paper's own Fig. 2e contains a false peak at 500 ms, exactly at the boundary of the projection-estimation window, which is a signature of the finite-sample minimum-variance selection bias: \\hat v_n minimizes in-sample variance, so the F-test comparing the estimation window with the next window is biased away from 1 even under constant W. This is a concrete, testable concern rather than a mere absence of a comparison method. The empirical AUC improvements across LIF, Hodgkin-Huxley, and Morris-Lecar networks are encouraging and justify a conditional rather than reject verdict, but the detector's mechanism and false-positive behavior need to be established with a constant-W negative control and a baseline direction test. The proposed concrete test would settle whether the concern lands: if the constant-W control yields zero false positives and the fixed-direction baseline fails to detect true changes, the method's interpretation is supported; otherwise the reported detections may be artifacts of the projection procedure. Because the reader's verdict of CONDITIONAL already requires addressing these issues, my stress-test does not move the verdict.","tokens_in":21483,"tokens_out":6591,"duration_ms":90236,"concrete_test":"Simulate the LIF network of Fig. 3 with a fixed connectivity matrix W (no change) for 400 s, using the same CPD pipeline and parameters (10 s projection-estimation window, 400 ms F-test window, p<10^-20 threshold), over at least 20 random seeds; count detected change points. Additionally, in the change-point simulations, replace \\hat v_n by a fixed random unit vector or by the first principal component and compare F-statistic peaks; if the fixed-direction test detects the known change equally well, the claimed subspace/projection step is not the operative mechanism. A zero false-positive rate in the constant-W control and a failure of the fixed-direction baseline would support the paper's interpretation; any nonzero false-positive rate, especially concentrated at the end of the estimation window, would confirm the selection-bias artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 26 asserts that Δ²v_rec lies in col(W1) before the change and in col(W2) after, with these column spaces non-overlapping in high probability. For the simulated sparse random matrices (in-degree K=320, N=4000–32000), W is generically full rank, so col(W1)=col(W2)=R^N; for the recorded 200-neuron subnetwork, the observed W_sub is 200×N and generically has rank 200, so both column spaces are the entire observed space. The stated mechanism therefore cannot explain detection. The detector actually uses \\hat v_n, the right singular vector of the pre-change Δ²V^T with the smallest singular value, i.e. a finite-sample minimum-variance direction. Even with constant W, the in-sample variance along \\hat v_n is biased downward relative to future windows of the same process, so the F-test can show a spurious variance increase immediately after the estimation window; the paper's own Fig. 2e reports exactly such a false peak at 500 ms and dismisses it by 'adjusting the data length.' No constant-W negative control, no repeated-seed false-detection statistics, and no comparison against a trivial fixed-direction variance test are provided. Consequently, the central claim that detected change points are connectivity-induced is not yet supported: the empirical AUC gains may be real, but the detector's selectivity for connectivity changes versus generic nonstationarity or selection artifacts is unquantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-stage framework, CPD-TDCC, for reconstructing neuronal connectivity in balanced spiking networks that undergo abrupt structural changes. The first stage computes the second-order voltage difference Δ²v(t), projects it onto a direction \\v\\n estimated as the smallest right singular vector of pre-change data, and uses an F-test on consecutive windows to detect variance changes attributed to connectivity changes (Sec. 3.3, Eqs. 24–27). The second stage segments spike trains at detected change points and computes pairwise time-delayed correlation coefficients to reconstruct the adjacency matrix or coupling strengths (Secs. 2.2, 3.4–3.6). The manuscript reports AUC improvements from 0.809–0.862 to 0.985–0.990 for LIF networks, from 0.743 to 0.981 for Hodgkin-Huxley networks, and from 0.684 to 0.944 for Morris-Lecar networks, plus detection and reconstruction for heterogeneous coupling-weight changes.","tokens_in":21663,"tokens_out":8391,"duration_ms":98671,"significance":"If the method works as claimed, it addresses a real gap: most connectivity inference assumes stationarity, while synaptic plasticity can change structure on recording timescales. The paper has genuine strengths: the variance-dominance calculation (Eq. 22) is a parameter-free asymptotic prediction verified in Fig. 1; the TDCC-to-coupling-strength link (Eqs. 8–13) is a useful derivation; and the numerical demonstrations span three neuron models and two types of structural change. However, the central theoretical mechanism used to explain why the detector is sensitive to connectivity changes (Eq. 26) is not valid for the full-rank matrices simulated, and the paper lacks negative controls that would establish selectivity of the detector for connectivity changes as opposed to generic nonstationarity or in-sample selection effects. The empirical results are therefore promising but not yet supported by the stated mechanism.","major_comments":[{"comment":"The subspace-separation argument is not valid for the simulated connectivity matrices. For an N×N sparse random W with in-degree K=320 and N=4000–32000, W is generically of full rank, so col(W1)=col(W2)=R^N and the two column spaces overlap completely; for the recorded 200-neuron subnetwork, the 200×N matrix is generically of rank 200, so both column spaces coincide with R^200. Hence Δ²v_rec cannot be detected by an 'orthogonal to col(W1)' vector in the sense of Eq. (26). The detector actually uses \\v\\n, the smallest right singular vector of the pre-change Δ²V^T, which is a finite-sample minimum-variance direction; its sensitivity must be explained through the covariance of Δ²v, specifically the variance of \\v\\n^T W Δd, rather than through column-space non-overlap. This distinction matters because Eq. (26) is the paper's stated justification for interpreting F-statistic peaks as connectivity-induced. The authors should either derive the covariance-based statistic or state and verify a rank-deficient regime for W.","section":"§3.3, Eq. (26), Fig. 2a–b"},{"comment":"The paper does not establish that the reported F-statistic peaks are selective for connectivity changes. Because \\v\\n is estimated from the first segment of the same recording that is later tested, the in-sample variance along \\v\\n is minimized by construction; under constant W, a later window will tend to have larger variance even without any structural change. Figure 2e shows a spurious peak at 500 ms, which the text attributes to this estimation-window effect and says can be 'eliminated by adjusting the data length' without quantifying the adjustment or the false-alarm rate. In addition, the projection time series is temporally correlated, because consecutive Δ²v samples share voltage observations, so the claimed F-distribution with d1=d2=ΔT/Δt−1 should be calibrated under the null. The manuscript should report repeated-seed simulations with constant W and with the change scenarios, giving the distribution of F-statistic peaks and detection rates at a fixed threshold, and should include a baseline comparison against a fixed random projection direction or another standard detector. Without these controls, the central claim that detected change points are connectivity-induced is not fully supported.","section":"§3.3, Eq. (27), Fig. 2e"},{"comment":"The headline AUC numbers come from a single simulation per condition. For a stochastic method with tunable windows and thresholds, reporting one realization makes it impossible to judge whether the improvements over native TDCC are stable. I request repeated-seed means and standard deviations, or bootstrapped intervals, for the AUC values in Figs. 3l–o and 4e,h and for the correlation improvements in Fig. 5k. This is not a request for a new method, but for minimal statistical reporting of the existing experiments.","section":"§3.4–3.6, Figs. 3–5"}],"minor_comments":[{"comment":"In the recurrent term, [dj(t+Δt) − di(t)] should be [dj(t+Δt) − dj(t)]; the same typo appears in the third term of Eq. (21).","section":"Eqs. (19), (21)"},{"comment":"There are several typographical errors: 'causaly' should be 'causally', 'expend' should be 'expand' near Eq. (10), and Table 1's caption contains a formatting artifact ('T able 1').","section":"§2.2, Table 1"},{"comment":"The text describes 'a sharp transition of mean values' in the projection magnitude, but Eq. (27) tests a change in variance; because the plotted quantity is |⟨\\v\\n, Δ²v⟩|, a variance shift changes the mean of the absolute value, but the connection should be stated explicitly to avoid confusion.","section":"§3.4, Fig. 3b"},{"comment":"The Discussion appropriately lists limitations, but the claimed failure for external-input strength Fi=O(√K) is not tested; a single numerical experiment in that regime would make the stated scope concrete.","section":"§4, Discussion"},{"comment":"The data and code availability statement ('will be made available based on reasonable requests') is weak for a simulation paper; a repository link would improve reproducibility.","section":"Declarations"}],"recommendation":"major_revision","confidential_remarks":"I am not recommending rejection: the AUC gains are substantial and the variance-dominance check in Fig. 1 is solid. The load-bearing issue is that the theoretical justification and the empirical controls are out of sync: the manuscript claims a subspace mechanism that cannot operate for the full-rank simulated matrices, and it does not provide constant-W or repeated-seed baselines. If the authors supply those controls and revise the theory to a covariance-based statement, the paper could become acceptable. The journal should ask for these additions rather than a complete rewrite."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a solid empirical core and a broken theoretical justification for its central detector. The idea of segmenting spike-train recordings at detected connectivity change points before applying TDCC reconstruction is sensible, and the reported AUC gains are consistent and large across LIF, Hodgkin-Huxley, and Morris-Lecar networks (e.g., 0.809 to 0.985 for LIF). The variance-scaling analysis in Sec. 3.2 is a genuine parameter-free derivation, and Fig. 1 shows it matches simulations. That part deserves credit.\n\nThe problem is that the mechanism they give for why the detector works cannot hold for the systems they simulate. Eq. 26 asserts that the recurrent part of the second-order voltage difference lies in col(W1) before the change and col(W2) after, and that these column spaces are non-overlapping with high probability. But for sparse random connectivity with in-degree K=320 and N=4000–32000, W is generically full rank, so col(W1)=col(W2)=R^N. For the recorded 200-neuron subnetwork, the relevant 200×N matrix is generically full row rank, so both column spaces are all of R^200. The stated subspace-separation argument therefore cannot explain detection.\n\nWhat the detector actually uses is the right singular vector of the pre-change data with the smallest singular value—a finite-sample minimum-variance direction. That introduces an in-sample selection bias: even with constant W, the variance along that estimated direction is artificially low in the estimation window, so an F-test can show a spurious variance increase immediately afterward. Figure 2e displays exactly such a peak at 500 ms, which the paper dismisses by saying it can be eliminated by adjusting the data length. Without a constant-W negative control, or repeated-seed false-detection statistics, the claim that the detected change points are connectivity-induced is not established. The method may still work through a change in the covariance structure of voltage fluctuations, but the paper does not analyze that alternative.\n\nOther soft spots are addressable rather than fatal: the F-test window, estimation window, rejection threshold, and TDCC lag are chosen per simulation with knowledge of the ground truth; there are no error bars on any AUC or correlation value; no comparison with the cited change-point methods [26,49]; and no code or data are shipped. The limitations listed in Sec. 4—membrane-potential requirement, discrete changes, failure under strong external drive—are candid.\n\nThis paper deserves peer review because the empirical effect is interesting and the reconstruction improvement is real. But the current version should not pass without fixing the theoretical argument, adding negative controls, and reporting variability across seeds and parameter choices. The subspace story needs to be replaced or reframed as a covariance-change test.","headline":"Genuinely useful CPD-TDCC empirical result, but the load-bearing subspace theory is wrong for the simulated matrices and the missing negative controls leave detection selectivity unproven.","tokens_in":22328,"tokens_out":4106,"would_cite":false,"duration_ms":47547,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that connectivity changes in balanced spiking networks can be detected from voltage fluctuations, and that segmenting recordings at those change points sharply improves wiring reconstruction.","keywords":["network reconstruction","change point detection","balanced neuronal network","time-delayed correlation coefficient","voltage fluctuation","structural plasticity"],"falsifier":"Run the same F-statistic detector on a balanced leaky integrate-and-fire network whose connectivity matrix does not change, and compare the peak F values with the changing-network case; if the unchanged network produces peaks above the same threshold at a comparable rate, the detected variance shift does not single out structural changes.","tokens_in":21125,"feed_emoji":"🧠","tokens_out":11417,"duration_ms":114996,"temperature":0.7,"pith_summary":"Structural connectivity in cortex is not static: synaptic rewiring and strength changes alter the network while recordings are being taken, and standard reconstruction methods that assume a fixed wiring pattern degrade. The paper proposes detecting the moments of structural change from voltage fluctuations alone, then reconstructing the connectivity of each stable segment with pairwise time-delayed correlation (TDCC) applied to spike trains. It argues that in balanced excitatory-inhibitory networks the recurrent input dominates the second-order voltage difference, so the fluctuation statistics carry information about the current connectivity matrix. In simulations of leaky integrate-and-fire (LIF), Hodgkin-Huxley, and Morris-Lecar networks, segmenting at detected change points raises reconstruction AUC from roughly 0.81, 0.74, and 0.68 to roughly 0.99, 0.98, and 0.94, respectively. The practical significance is that if this holds, connectivity inference from sparse experimental recordings should be performed on segments separated by detected structural changes rather than on pooled data.","feed_headline":"Voltage fluctuations reveal when neuron wiring changes","feed_subtitle":"Segmenting spike data at detected shifts lifts reconstruction AUC from ~0.81 to ~0.99 in simulations.","key_machinery":"The load-bearing object is the second-order voltage difference $\\Delta^2 v(t) = v(t+2\\Delta t) - 2v(t+\\Delta t) + v(t)$, decomposed into ionic, external Poisson, and recurrent contributions. Under balanced-network scalings (coupling strengths $O(1/\\sqrt{K})$ and input rates $O(K)$), the recurrent term has variance of order $\\sqrt{K\\Delta t/N}$, which dominates the external term of order $\\sqrt{\\Delta t/N}$, so the population-averaged $\\Delta^2 v$ inherits the direction of the recurrent input and hence of the connectivity matrix $W$. The detector is the projection magnitude $|\\langle \\hat{v}_n, \\Delta^2 v\\rangle|$ onto the smallest-singular-value right singular vector estimated from pre-change data; an F-statistic on segment variances converts the projection shift into a change-point test. The reconstruction step uses the time-delayed correlation coefficient (TDCC) between binarized spike trains, which the paper shows is proportional to the underlying coupling strength in the weak-coupling limit.","core_discovery":"The central claim is that a connectivity change $W_1 \\to W_2$ at time $t_c$ produces a detectable change in the distribution of the second-order voltage difference $\\Delta^2 v(t) = v(t+2\\Delta t) - 2v(t+\\Delta t) + v(t)$. In the balanced regime the recurrent-synaptic term of $\\Delta^2 v$ dominates the external and ionic terms, so $\\Delta^2 v$ lies essentially in the column space of the active connectivity matrix: $\\Delta^2 v_{\\mathrm{rec}}(t) \\in \\operatorname{col}(W_1)$ before $t_c$ and $\\in \\operatorname{col}(W_2)$ after. The paper argues that a vector orthogonal to $\\operatorname{col}(W_1)$, estimated from pre-change data as the right singular vector of the $\\Delta^2 V$ matrix with the smallest singular value, will have near-zero projection of $\\Delta^2 v$ before the change and a significantly nonzero projection after, and an F-test on the variance of this projection locates the change point. Recursively segmenting the recording at such points and applying TDCC to spike trains within each segment reconstructs the successive connectivity matrices with AUC values around 0.94-0.99 in the paper's simulations, compared with 0.68-0.86 for unsegmented TDCC. The same pipeline is reported to work for changes in coupling weights with fixed topology when the weight heterogeneity exceeds a small threshold.","pith_inferences":["Beyond the paper: the stated subspace-separation story is not the only possible mechanism for the reported detections; for the sparse random connectivity matrices used, $W$ is generically full rank, so the empirical signal may actually come from a change in the covariance of the voltage fluctuations rather than from disjoint column spaces.","Beyond the paper: a direct test would be a constant-connectivity negative control; if the F-statistic peaks above threshold at a comparable rate with unchanged wiring, the variance shift is not specific to structural changes.","Beyond the paper: the projection-of-fluctuations idea could be extended to proxy voltage signals such as calcium imaging, which would address the paper's stated limitation that membrane-potential recordings are required.","Beyond the paper: since TDCC is linearly related to coupling strength in the weak-coupling limit, segmented reconstruction could in principle be used to estimate the magnitudes of synaptic weights, not just the adjacency pattern; the paper demonstrates correlation improvement but stops short of weight calibration."],"forward_implications":["Wiring inference from spike trains should be run on segments separated by detected structural change points rather than on the full pooled recording.","The detector is claimed to work for strength changes with fixed topology, with detectability increasing monotonically with coupling-weight heterogeneity and falling off below a heterogeneity of about 0.2 in the paper's LIF tests.","The pipeline operates on a 200-neuron subnetwork sampled from a 4000-neuron balanced network, so it is aimed at the sparse-recording regime typical of experiments.","Because the change-point statistic is an F-test, the method returns a p-value for each candidate change and can be applied recursively to find multiple successive changes."],"supporting_citations":[{"why":"supplies the balanced-network scaling used to argue that recurrent input dominates the voltage fluctuations.","marker":"[32]"},{"why":"establishes the asynchronous balanced cortical state that the fluctuation statistics assume.","marker":"[34]"},{"why":"provides the TDCC-to-connectivity relation that motivates the reconstruction step.","marker":"[17]"},{"why":"gives the balanced-network parameter settings used in the numerical simulations.","marker":"[39]"},{"why":"defines the Hodgkin-Huxley model used as a general neuron test case.","marker":"[41]"},{"why":"defines the Morris-Lecar model used as a second general neuron test case.","marker":"[44]"}],"fun_headline_variants":["Voltage shifts expose hidden rewiring of neuron networks","Spike segmentation reveals neural network rewiring moments","Voltage fluctuations pinpoint when neuron connections change","Detecting rewiring in balanced neural networks via voltage flickers","New method finds and reconstructs changing brain connectivity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The detection argument assumes that changing the wiring rotates the set of directions the voltage fluctuations can explore, so that a direction untouched before the change shows a clear response after it; if the old and new wiring produce overlapping direction sets, the stated test has nothing to detect.","fun_headline_variants_meta":{"raw":{"variants":["Voltage shifts expose hidden rewiring of neuron networks","Spike segmentation reveals neural network rewiring moments","Voltage fluctuations pinpoint when neuron connections change","Detecting rewiring in balanced neural networks via voltage flickers","New method finds and reconstructs changing brain connectivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1532,"prompt_tokens":1096,"completion_tokens":436,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":712,"tokens_out":436,"duration_ms":5321,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:32:36.649931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same F-statistic detector on a balanced leaky integrate-and-fire network whose connectivity matrix does not change, and compare the peak F values with the changing-network case; if the unchanged network produces peaks above the same threshold at a comparable rate, the detected variance shift does not single out structural changes.","supporting_citations":[{"cited_title":"Science 274(5293), 1724–1726 (1996) https://doi.org/10.1126/science.274.5293","cited_arxiv_id":null,"evidence_quote":"supplies the balanced-network scaling used to argue that recurrent input dominates the voltage fluctuations."},{"cited_title":"Proceedings of the National Academy of Sciences 121(14), 2305297121 (2024) https://doi.org/10.1073/pnas.2305297121","cited_arxiv_id":null,"evidence_quote":"provides the TDCC-to-connectivity relation that motivates the reconstruction step."},{"cited_title":"Frontiers in Computational Neuroscience 12, 47 (2018)","cited_arxiv_id":null,"evidence_quote":"gives the balanced-network parameter settings used in the numerical simulations."}],"review_version":1}