Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Parameters that produce the same dynamical behavior form structured viable manifolds, and conditional diffusion models can sample and measure that geometry.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 00:47 UTC pith:RHT3HS2Z

load-bearing objection Solid inverse-geometry framing plus usable diffusion sampling of viable sets; main caveats are local regularity and deferred code, not a broken claim. the 3 major comments →

arxiv 2607.03671 v1 pith:RHT3HS2Z submitted 2026-07-04 q-bio.QM cs.LGnlin.CDq-bio.NC

Diffusion learning reveals viable parameter manifolds and compensation geometry in biological dynamical systems

Ruilin Zhang , Louis Tao , Zhuo-Cheng Xiao This is my paper
classification q-bio.QM cs.LGnlin.CDq-bio.NC
keywords diffusion modelsviable parameter manifoldrobustnesscompensation geometrycomputational neurosciencedynamical systemsdegeneracyscore-based generative models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Complex dynamical models usually have many parameters but only a few measurable features, so many different parameter combinations can yield the same activity. This paper formalizes those compatible combinations as viable parameter manifolds: the inverse images of a target behavior under a map from parameters to dynamical statistics. What matters is not the raw number of reported features but the effective rank of that map, so co-varying observables leave more free directions than they appear to. Conditional score-based diffusion models trained on simulated parameter–feature pairs act as amortized samplers of these prior-weighted viable sets. In Lorenz, Izhikevich, and reduced spiking-network models, the learned geometry turns degeneracy, compensation, and hidden tradeoffs into objects that can be sampled, visualized, and measured.

Core claim

Compatible parameters that realize a fixed dynamical behavior form viable parameter manifolds whose local dimension equals the parameter count minus the effective rank of the parameter-to-feature map, and conditional score-based diffusion models trained on simulated pairs can sample those manifolds and expose their compensation geometry across Lorenz, Izhikevich, and dsODE network models.

What carries the argument

Viable parameter manifolds My = I^{-1}(y): the inverse images of target dynamical features under a parameter-to-feature map. Conditional score-based diffusion models serve as amortized samplers of prior-weighted, tolerance-thickened versions of these sets, so dimension, tangent compensation directions, and curvature become practical readouts.

Load-bearing premise

Within a chosen target neighborhood, long-time statistics depend smoothly on parameters inside a single dynamical regime, so a conditioned sample cloud can be treated as a smooth level-set manifold.

What would settle it

In a Lorenz or Izhikevich target known to sit in a single smooth regime, generate a diffusion cloud, re-simulate every sample, and check whether requested features are recovered and whether the participation-ratio dimension matches parameter count minus the effective rank of the feature map; systematic failure of either check refutes the central sampling-plus-geometry claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Degeneracy and compensation become measurable geometry: intrinsic dimension, tangent co-variation laws, and curvature rather than isolated best-fit points.
  • Feature design should add independent active constraints aligned with how function actually changes, not merely longer feature lists.
  • Irregular or multimodal generated clouds diagnose regime mixing or nearby bifurcations instead of being dismissed as sampler noise.
  • Higher-dimensional network models can be interrogated for hidden tradeoffs among synaptic strength, timescales, connection probability, and external drive.
  • One simulation library amortizes many inverse queries via conditional generation followed by direct re-simulation validation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Experimental feature choice can be treated as geometry design: pick observables that shrink or split the viable set along directions that matter for function.
  • Pairing the sampler with classical continuation could chart bifurcation surfaces precisely where the smooth-manifold assumption fails.
  • Measured ion-channel or synaptic co-variation in real neurons could be compared to learned tangent spaces to test whether homeostatic rules track viable-manifold directions.
  • The same pipeline could flag whether a model is over- or under-constrained by available data before expensive experimental parameter searches.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper formalizes compatible parameter sets of dynamical models as viable parameter manifolds My = I^{-1}(y), the inverse images of a parameter-to-feature map, and argues that the local dimension is k minus the effective rank of DI rather than the nominal feature count. Conditional score-based diffusion models trained on simulated (θ, I(θ)) pairs are used as amortized samplers of prior-weighted, tolerance-thickened viable sets. Across Lorenz, Izhikevich, and dsODE reductions of spiking networks, the generated clouds recover thin sheets, codimension-two compensation geometries, and E–I / S–τ tradeoffs, with validation by direct re-simulation and participation-ratio dimensions consistent with effective-rank expectations.

Significance. If the claims hold, the work turns biological degeneracy and robustness from qualitative observations into measurable inverse geometry (intrinsic dimension, tangent compensation laws, curvature, and branch structure). The three case studies are complementary and the re-simulation validation is a genuine strength: generated parameters are checked against requested features rather than accepted by density alone. The framing unifies sloppy-model, identifiability, and dynamical-compensation literatures as inverse geometry and supplies a practical generative tool for sampling high-dimensional viable sets that exhaustive sweeps cannot reach. Code and data availability are deferred, so full reproducibility is not yet demonstrated, but the conceptual and empirical contribution is substantial for computational neuroscience and dynamical systems.

major comments (3)
  1. §2.1 and Materials and Methods: the local-regularity approximation (smooth level sets inside a single regime, TθMy = ker DI) is load-bearing for the geometric claims, yet the manuscript provides no quantitative protocol for when a cloud is accepted as regular versus flagged as regime mixing. Participation-ratio dimensions and visual multimodality are reported, but there is no systematic check of local rank stability of DI, condition number of active singular values, or component-wise atlas construction. Without such criteria, the distinction between “smooth manifold” and “diagnostic irregularity” remains post hoc and weakens the strongest geometric interpretations (especially Izhikevich curvature and dsODE branch separation).
  2. §2.3–2.5 and §4.1: generated samples are explicitly prior-weighted and tolerance-thickened, not uniform on My, yet several geometric readouts (participation ratio, local PCA tangents, curvature labels in Fig. 3C) are treated as properties of the manifold itself. The paper correctly notes that density is not the scientific output, but it does not quantify how prior weighting or ε-thickening bias those estimators. A short sensitivity analysis (different priors or ε, or comparison to rejection sampling in low-d Lorenz/Izhikevich) would make the compensation-geometry claims more robust.
  3. Data/code availability and §4: the manuscript states that scripts and libraries will be released at submission, and the analysis is described as a “first-round demonstration” rather than exhaustively refined atlases. For a methods-plus-geometry paper whose central tool is a trained diffusion sampler, deferred code and incomplete iterative refinement are load-bearing for reproducibility and for claims about higher-dimensional (8–12D) manifolds. The scientific framing can stand, but the practical contribution cannot be fully assessed until the simulation libraries, training configs, and generation scripts are available.
minor comments (5)
  1. Fig. 2F caption: “Middel” should be “Middle”; several figure captions have minor typos and incomplete sentences.
  2. §2.1: “We fisrt construct” → “first”; a few other typos appear in the main text (e.g., “Gr”un” in the references).
  3. Feature scalings (rate/200, ISI/0.05, CV/5, Fano/5, rates/200 Hz) are listed in Methods but not motivated; a brief sentence on why these normalizations preserve relative feature importance would help.
  4. Fig. 7C is described as identifying empirical tangent space and an orthogonal direction, but the panel itself is only briefly discussed; a short caption expansion would improve readability.
  5. References: a few formatting inconsistencies (e.g., escaped umlauts, missing DOIs for some entries) should be cleaned for production.

Circularity Check

0 steps flagged

No significant circularity: viable manifolds are defined as inverse images and sampled by diffusion trained on independent forward simulations, with geometry validated by re-simulation rather than by construction.

full rationale

The paper defines My = {θ : I(θ) = y} (Eq. 1, 3) as the inverse image of a parameter-to-feature map and invokes the constant-rank theorem only under an explicit local-regularity approximation (§2.1). Training data are randomly sampled parameters with forward-simulated features; the conditional diffusion model is an amortized sampler of prior-weighted, tolerance-thickened sets, not a derivation that forces the geometry. Intrinsic dimension (participation ratio), tangent spaces (ker DI or local PCA), and curvature are computed post hoc from generated clouds and checked by direct re-simulation of held-out and generated parameters (Figs. 2E, 3D, 5, 6). Self-citations to Chang et al. (dsODE) and Xiao et al. (E–I landscape) supply the simulator and a known background strip that the paper recovers and then extends; they do not define or force the inverse-geometry claims. No fitted input is renamed a prediction, no uniqueness theorem is imported to forbid alternatives, and no ansatz is smuggled in as a first-principles result. The local-regularity caveat is stated openly and irregularity is treated as diagnostic, so the central claim remains empirical and non-circular under the paper’s own scope.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on a local regularity assumption that converts inverse images into smooth manifolds, on the constant-rank theorem for dimension, on the empirical success of score-based diffusion as a manifold-adapted sampler, and on the authors’ prior dsODE reduction as the network simulator. Free parameters are mainly numerical tolerances, feature scalings, and sampling ranges chosen for the demonstrations rather than fitted to external data. No new physical entities are postulated; the “viable manifold” is a definitional object whose utility is tested by re-simulation.

free parameters (5)
  • feature tolerance ε / target thickening
    Defines the thickened set My,ε used for practical conditioning; chosen by hand for each demonstration and affects apparent dimension and cloud width.
  • participation-ratio dimension estimator
    Empirical proxy dPR = (∑λi)²/∑λi² for intrinsic dimension; depends on sample size and normalization and is not a unique geometric invariant.
  • feature scalings (rate/200, ISI/0.05, CV/5, Fano/5, rates/200 Hz)
    Linear normalizations applied before diffusion training; alter relative weighting of constraints and therefore the learned geometry.
  • parameter box ranges (Lorenz log10 intervals, Izhikevich a,b,c,d bounds, dsODE S and τ intervals)
    Hand-chosen sampling domains that define the prior support of the training library and therefore the prior-weighted viable sets.
  • diffusion training and sampling hyperparameters (noise schedule, Langevin steps, network architecture)
    Not fully enumerated in the main text; control sample quality and coverage of the viable set.
axioms (5)
  • domain assumption Local regularity: within a chosen target neighborhood the long-time statistics I(θ) depend smoothly on parameters inside a single dynamical regime, so My is locally a level set.
    Stated explicitly in §2.1 as the working approximation; failures are treated as diagnostic of regime mixing but the main geometric claims assume it holds.
  • standard math Constant-rank theorem: if the effective rank ry of DI is locally constant and active singular values are bounded away from zero, then dim My = k − ry.
    Invoked in §2.1 to convert effective rank into local manifold dimension.
  • domain assumption Score-based diffusion can adapt to smooth low-dimensional manifold structure without explicit manifold estimation (citing Tang & Yang 2024).
    Used in §2.6 to justify why diffusion is a practical sampler of inverse geometry; the cited theory supplies the rate, not a proof for the present non-uniform, prior-weighted setting.
  • domain assumption dsODE is a faithful deterministic reduction of the underlying finite spiking network for the regimes studied.
    Relies on Chang et al. (2025) and Xiao et al. (2021); the network results inherit whatever fidelity limits those reductions possess.
  • ad hoc to paper Generated samples represent a prior-weighted, tolerance-thickened empirical atlas rather than a uniform measure on My.
    Acknowledged in §2.1 and Methods; all geometric claims are therefore about support and local structure, not density.
invented entities (2)
  • viable parameter manifold My no independent evidence
    purpose: Formal object of study: the inverse image of a target feature vector under the parameter-to-feature map, equipped with geometric readouts (dimension, tangent, curvature, components).
    Definitional rather than ontological; utility is demonstrated by re-simulation, but the entity itself has no independent existence outside the modeling framework.
  • compensation geometry (tangent directions + curvature of My) no independent evidence
    purpose: Interpret local co-variations that preserve target features as first- and second-order compensation laws.
    Extracted from the learned clouds via local PCA and curvature estimates; biological meaning is interpretive and validated only by re-simulation of generated points.

pith-pipeline@v1.1.0-grok45 · 20760 in / 3770 out tokens · 33572 ms · 2026-07-12T00:47:19.609894+00:00 · methodology

0 comments
read the original abstract

Models of complex systems often have many parameters, yet are constrained by far fewer experimentally accessible observables: similar activity can emerge from coordinated parameter changes. We formalize these compatible parameter sets as \emph{viable parameter manifolds}: the inverse images of a system's target dynamical behaviors under a parameter-to-feature map. The relevant codimension is not the number of reported features, but the effective rank of that map at the target scale. Co-varying features lower the codimension, while poor conditioning, high curvature, or regime mixing degrade learnability. We train conditional score-based diffusion models on simulated parameter--feature pairs and use them as amortized samplers of prior-weighted viable sets. In the Lorenz system, scalar trajectory statistics generate thin viable sheets, and two-feature conditioning localizes a transition-adjacent corridor. In the Izhikevich neuron model, four firing descriptors lie close to a nearly two-dimensional family of features, and the learned inverse images reveal distinct regular and irregular compensation geometries. In a recent ODE reduction of finite spiking networks, the same framework reveals excitatory--inhibitory compensation, timescale--coupling tradeoffs, and input-dependent viable manifolds across 4--12 parameter dimensions. In this view, robustness, compensation, and hidden parameter dependencies are organized as inverse geometry, with diffusion models providing practical tools for sampling, visualizing, and interrogating that geometry.

Figures

Figures reproduced from arXiv: 2607.03671 by Louis Tao, Ruilin Zhang, Zhuo-Cheng Xiao.

Figure 1
Figure 1. Figure 1: Diffusion learning of viable parameter manifolds. (A) Conceptual target-conditioned viable sets in parameter space. (B) Conditional score network and annealed Langevin sampler. (C) Reverse diffusion from a random Lorenz cloud to a target-conditioned manifold. (D) Geometric readouts used throughout: neighboring level sets, local tangent directions, participation-ratio dimension, and local curvature. 5 [PIT… view at source ↗
Figure 2
Figure 2. Figure 2: Viable manifolds in the Lorenz system. (A) Equations, parameter ranges, and representative trajec￾tories. (B and C) Scalar targets in log10 std(x) or log10 std(z) generate thin sheets in (log10 σ, log10 ρ, log10 β). (D) Simulated standard-deviation structure with subsets satisfying std(x) ≈ 10, std(z) ≈ 10, or both. (E) Re￾simulated validation for scalar-target generation. (F) Two-feature conditioning loca… view at source ↗
Figure 3
Figure 3. Figure 3: Viable manifolds in the Izhikevich neuron model. (A) Model, parameter ranges, and example responses under a step current. (B) Principal-component view of the four-feature response cloud. (C) Target-conditioned manifolds in projections of (a, b, c, d) with local curvature labeled by colored dots; stars mark held-out parameters and participation-ratio dimension is shown. (D) Re-simulated traces and feature d… view at source ↗
Figure 4
Figure 4. Figure 4: dsODE schematic and representative population activity. (A) Excitatory and inhibitory populations with recurrent coupling ranges. (B-D) Representative E/I rate traces (upper) and voltage-bin view of population activity at the time section indicated by black vertical lines (lower). 2.4 dsODE synaptic manifolds extend E–I landscape geometry into network inverse problems We next moved to dsODE, a deterministi… view at source ↗
Figure 5
Figure 5. Figure 5: Four-dimensional dsODE synaptic manifolds. (A) Mean-rate conditioned generations in the inhibition plane and direct validation. (B) Std-rate conditioned generations, overlap with mean-conditioned clouds near held-out parameters, and representative rate traces shows distinct 1-/2-beat oscillatory dynamics. (C) Generations conditioned on means and stds simultaneously; Validations of means and stds of E/I rat… view at source ↗
Figure 6
Figure 6. Figure 6: Higher-dimensional dsODE viable manifolds. (A) Eight-dimensional generations over synaptic strengths and timescales; pairwise densities show structured dependencies and validation confirms target recovery (red star indicating significant relations as p < 0.01). (B) Organization of the eight-dimensional manifolds in ratio and timescale-adjusted coordinates. (C) Ten-dimensional generations that additionally … view at source ↗
Figure 7
Figure 7. Figure 7: Geometry of target-conditioned viable manifolds across models. (A) In dsODE, successively richer target guides contract and separate four-dimensional synaptic clouds in ratio coordinates. (B) In Lorenz and Izhikevich, joint feature conditioning yields thinner clouds than scalar conditioning. (C) Local PCA identifies empirical tangent space and an orthogonal direction toward a nearby target-conditioned clou… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 14 canonical work pages

  1. [1]

    Prinz, Dirk Bucher, and Eve Marder

    Astrid A. Prinz, Dirk Bucher, and Eve Marder. Similar network activity from disparate circuit parameters.Nature Neuroscience, 7(12):1345–1352, 2004. doi: 10.1038/nn1352

    work page doi:10.1038/nn1352 2004

  2. [2]

    Variability, compensation and homeostasis in neuron and network function.Nature Reviews Neuroscience, 7(7):563–574, 2006

    Eve Marder and Jean-Marc Goaillard. Variability, compensation and homeostasis in neuron and network function.Nature Reviews Neuroscience, 7(7):563–574, 2006. doi: 10.1038/nrn1949

    work page doi:10.1038/nrn1949 2006

  3. [3]

    Eve Marder. Variability, compensation, and modulation in neurons and circuits.Proceedings of the National Academy of Sciences of the United States of America, 108(Supplement 3): 15542–15548, 2011. doi: 10.1073/pnas.1010674108

    doi:10.1073/pnas.1010674108 2011

  4. [4]

    Williams, Jonathan S

    Timothy O’Leary, Alex H. Williams, Jonathan S. Caplan, and Eve Marder. Correlations in ion channel expression emerge from homeostatic tuning rules.Proceedings of the National Academy of Sciences of the United States of America, 110(28):E2645–E2654, 2013. doi: 10.1073/pnas. 1309966110. 22

    doi:10.1073/pnas 2013

  5. [5]

    Ion channel degeneracy enables robust and tunable neuronal firing rates.Proceedings of the National Academy of Sciences of the United States of America, 112(38):E5361–E5370, 2015

    Guillaume Drion, Timothy O’Leary, and Eve Marder. Ion channel degeneracy enables robust and tunable neuronal firing rates.Proceedings of the National Academy of Sciences of the United States of America, 112(38):E5361–E5370, 2015. doi: 10.1073/pnas.1516400112

    doi:10.1073/pnas.1516400112 2015

  6. [6]

    Gutenkunst, Joshua J

    Ryan N. Gutenkunst, Joshua J. Waterfall, Fergal P. Casey, Kevin S. Brown, Christopher R. Myers, and James P. Sethna. Universally sloppy parameter sensitivities in systems biology models.PLoS Computational Biology, 3(10):e189, 2007. doi: 10.1371/journal.pcbi.0030189

    doi:10.1371/journal.pcbi.0030189 2007

  7. [7]

    Brown and James P

    Kevin S. Brown and James P. Sethna. Statistical mechanical approaches to models with many poorly known parameters.Physical Review E, 68(2):021904, 2003. doi: 10.1103/PhysRevE.68. 021904

    doi:10.1103/physreve.68 2003

  8. [8]

    Machta, Ritashree Chachra, Mark K

    Benjamin B. Machta, Ritashree Chachra, Mark K. Transtrum, and James P. Sethna. Parameter space compression underlies emergent theories and predictive models.Science, 342(6158): 604–607, 2013. doi: 10.1126/science.1238723

    doi:10.1126/science.1238723 2013

  9. [9]

    Transtrum, Benjamin B

    Mark K. Transtrum, Benjamin B. Machta, Kevin S. Brown, Bryan C. Daniels, Christopher R. Myers, and James P. Sethna. Perspective: Sloppiness and emergent theories in physics, biology, and beyond.The Journal of Chemical Physics, 143(1):010901, 2015. doi: 10.1063/1.4923066

    doi:10.1063/1.4923066 2015

  10. [10]

    On structural identifiability.Mathematical Bio- sciences, 7(3-4):329–339, 1970

    Richard Bellman and Karl Johan ˚Astr¨ om. On structural identifiability.Mathematical Bio- sciences, 7(3-4):329–339, 1970. doi: 10.1016/0025-5564(70)90132-X

    doi:10.1016/0025-5564(70)90132-x 1970

  11. [11]

    Eduardo D. Sontag. Dynamic compensation, parameter identifiability, and equivariances.PLoS Computational Biology, 13(4):e1005447, 2017. doi: 10.1371/journal.pcbi.1005447

    work page doi:10.1371/journal.pcbi.1005447 2017

  12. [12]

    Dynamical compensation in physiological circuits.Molecular Systems Biology, 12(11):886, 2016

    Omer Karin, Avi Swisa, Benjamin Glaser, Yuval Dor, and Uri Alon. Dynamical compensation in physiological circuits.Molecular Systems Biology, 12(11):886, 2016. doi: 10.15252/msb. 20167216

    doi:10.15252/msb 2016

  13. [13]

    Daniels, Yuhai Chen, James P

    Bryan C. Daniels, Yuhai Chen, James P. Sethna, Ryan N. Gutenkunst, and Christopher R. Myers. Sloppiness, robustness, and evolvability in systems biology.Current Opinion in Biotechnology, 19(4):389–395, 2008. doi: 10.1016/j.copbio.2008.06.008

    work page doi:10.1016/j.copbio.2008.06.008 2008

  14. [14]

    Robustness and evolvability: A paradox resolved.Proceedings of the Royal Society B: Biological Sciences, 275(1630):91–100, 2008

    Andreas Wagner. Robustness and evolvability: A paradox resolved.Proceedings of the Royal Society B: Biological Sciences, 275(1630):91–100, 2008. doi: 10.1098/rspb.2007.1137

    work page doi:10.1098/rspb.2007.1137 2008

  15. [15]

    Martin, and Andreas Wagner

    Stefano Ciliberti, Olivier C. Martin, and Andreas Wagner. Innovation and robustness in complex regulatory gene networks.Proceedings of the National Academy of Sciences of the United States of America, 104(34):13591–13596, 2007. doi: 10.1073/pnas.0705396104

    doi:10.1073/pnas.0705396104 2007

  16. [16]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840–6851, 2020

    2020

  17. [17]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  18. [18]

    Adaptivity of diffusion models to manifold structures

    Rong Tang and Yun Yang. Adaptivity of diffusion models to manifold structures. InProceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 238, pages 1648–1656. PMLR, 2024. 23

    2024

  19. [19]

    Edward N. Lorenz. Deterministic nonperiodic flow.Journal of the Atmospheric Sciences, 20 (2):130–141, 1963. doi: 10.1175/1520-0469(1963)020⟨0130:DNF⟩2.0.CO;2

    doi:10.1175/1520-0469(1963)020 1963

  20. [20]

    Izhikevich

    Eugene M. Izhikevich. Simple model of spiking neurons.IEEE Transactions on Neural Networks, 14(6):1569–1572, 2003. doi: 10.1109/TNN.2003.820440

    doi:10.1109/tnn.2003.820440 2003

  21. [21]

    Izhikevich

    Eugene M. Izhikevich. Which model to use for cortical spiking neurons?IEEE Transactions on Neural Networks, 15(5):1063–1070, 2004. doi: 10.1109/TNN.2004.832719

    doi:10.1109/tnn.2004.832719 2004

  22. [22]

    Minimizing informa- tion loss reduces spiking neuronal networks to differential equations.Journal of Computational Physics, 537:114117, 2025

    Jie Chang, Zhuoran Li, Zhongyi Wang, Louis Tao, and Zhuo-Cheng Xiao. Minimizing informa- tion loss reduces spiking neuronal networks to differential equations.Journal of Computational Physics, 537:114117, 2025. doi: 10.1016/j.jcp.2025.114117

    work page doi:10.1016/j.jcp.2025.114117 2025

  23. [23]

    Springer, New York, 1982

    Colin Sparrow.The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors, volume 41 ofApplied Mathematical Sciences. Springer, New York, 1982. doi: 10.1007/978-1-4612-5767-7

    work page doi:10.1007/978-1-4612-5767-7 1982

  24. [24]

    Softky and Christof Koch

    William R. Softky and Christof Koch. The highly irregular firing of cortical cells is inconsistent with temporal integration of random epsps.The Journal of Neuroscience, 13(1):334–350, 1993. doi: 10.1523/JNEUROSCI.13-01-00334.1993

    doi:10.1523/jneurosci.13-01-00334.1993 1993

  25. [25]

    Shadlen and William T

    Michael N. Shadlen and William T. Newsome. The variable discharge of cortical neurons: Im- plications for connectivity, computation, and information coding.The Journal of Neuroscience, 18(10):3870–3896, 1998. doi: 10.1523/JNEUROSCI.18-10-03870.1998

    work page doi:10.1523/jneurosci.18-10-03870.1998 1998

  26. [26]

    Martin P. Nawrot. Analysis and interpretation of interval and count variability in neural spike trains. In Sonja Gr”un and Stefan Rotter, editors,Analysis of Parallel Spike Trains, pages 37–58. Springer, New York, 2010. doi: 10.1007/978-1-4419-5675-0 3

    work page doi:10.1007/978-1-4419-5675-0 2010

  27. [27]

    Churchland, Margot Y

    Mark M. Churchland, Margot Y. Byron, John P. Cunningham, Leo P. Sugrue, Marlene R. Cohen, Greg S. Corrado, William T. Newsome, Anna M. Clark, Pooya Hosseini, Benjamin B. Scott, David C. Bradley, Matthew A. Smith, Adam Kohn, J. Anthony Movshon, K. Matthew Armstrong, Tirin Moore, Stephen W. C. Chang, Lawrence H. Snyder, Stephen G. Lisberger, Nicholas J. Pri...

    work page doi:10.1038/nn.2501 2010

  28. [28]

    Alfonso Renart and Christian K. Machens. Variability in neural activity and behavior.Current Opinion in Neurobiology, 25:211–220, 2014. doi: 10.1016/j.conb.2014.02.013

    work page doi:10.1016/j.conb.2014.02.013 2014

  29. [29]

    Dimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms.PNAS Nexus, 3(10): pgae415, 2024

    Arthur Fyon, Alessio Franci, Pierre Sacr´ e, and Guillaume Drion. Dimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms.PNAS Nexus, 3(10): pgae415, 2024. doi: 10.1093/pnasnexus/pgae415

    work page doi:10.1093/pnasnexus/pgae415 2024

  30. [30]

    Zhuo-Cheng Xiao, K. K. Lin, and Lai-Sang Young. A data-informed mean-field approach to mapping of cortical parameter landscapes.PLoS Computational Biology, 17(12):e1009718, 2021. doi: 10.1371/journal.pcbi.1009718

    work page doi:10.1371/journal.pcbi.1009718 2021

  31. [31]

    Multi-band oscillations emerge from a simple spiking network.Chaos: An Interdisciplinary Journal of Nonlinear Science, 33(4):043121, 2023

    Tianyi Wu, Yuhang Cai, Ruilin Zhang, Zhongyi Wang, Louis Tao, and Zhuo-Cheng Xiao. Multi-band oscillations emerge from a simple spiking network.Chaos: An Interdisciplinary Journal of Nonlinear Science, 33(4):043121, 2023. doi: 10.1063/5.0106884. 24

    work page doi:10.1063/5.0106884 2023

  32. [32]

    Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood.Bioinformatics, 25(15):1923–1929, 2009

    Andreas Raue, Clemens Kreutz, Thomas Maiwald, Ursula Klingm¨ uller, and Jens Timmer. Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood.Bioinformatics, 25(15):1923–1929, 2009. doi: 10.1093/ bioinformatics/btp358

    1923

  33. [33]

    Transtrum, Benjamin B

    Mark K. Transtrum, Benjamin B. Machta, and James P. Sethna. Geometry of nonlinear least squares with applications to sloppy models and optimization.Physical Review E, 83(3):036701,

  34. [34]

    doi: 10.1103/PhysRevE.83.036701

    doi:10.1103/physreve.83.036701

  35. [35]

    Springer, New York,

    John Guckenheimer and Philip Holmes.Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields, volume 42 ofApplied Mathematical Sciences. Springer, New York,

  36. [36]

    doi: 10.1007/978-1-4612-1140-2

    doi:10.1007/978-1-4612-1140-2

  37. [37]

    Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    Aapo Hyv”arinen. Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    2005

  38. [38]

    A connection between score matching and denoising autoencoders.Neural Computation, 23(7):1661–1674, 2011

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural Computation, 23(7):1661–1674, 2011. doi: 10.1162/NECO a 00142. 25

    doi:10.1162/neco 2011