Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Cells keep a stable phenotype under noisy signals by learning structured covariances between effectors and genes past a critical adaptation rate.

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 01:42 UTC pith:YG7T5YTO

load-bearing objection Solid online-learning import into a two-timescale morphogen pathway that yields a clean critical-η transition to specialized (T,Q,R) modes; the math checks out under its assumptions, but the global-mismatch GD rule and forced fixed-point GRNs remain untested against real pathways. the 3 major comments →

arxiv 2607.03545 v1 pith:YG7T5YTO submitted 2026-07-03 q-bio.MN cond-mat.dis-nnnlin.AO

Cellular Adaptation to Signal Fluctuations as Learning

Tuan Minh Pham , David Saad This is my paper
classification q-bio.MN cond-mat.dis-nnnlin.AO
keywords cellular adaptationmorphogen fluctuationsgene-regulatory networkscollective modesonline learningsymmetry breakingrobustnesstimescale separation
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.

Cells must interpret fluctuating chemical signals such as morphogens without losing their identity. This paper models that process as a closed feedback loop in which slow adaptive variables that map the signal onto effectors are updated by a learning rule that reduces mismatch with the fast gene-regulatory network. From the microscopic dynamics the authors derive closed equations for three collective modes: gene–gene covariances, effector–effector covariances, and the cross-covariances that link the two layers. Analysis of those equations shows that, above a critical adaptation rate, the system undergoes a transition from unstructured random correlations to specialized, time-invariant patterns in which each effector locks onto a particular target gene. The resulting structured state is what the paper identifies as robustness: not the suppression of variance, but the self-organized coordination that lets target genes determine their own attractors while the effectors remain consistent with those attractors.

Core claim

Robustness to non-instructive signal fluctuations is a dynamical phase transition: beyond a critical adaptation rate the collective modes (T, Q, R) change from random, vanishing correlations to structured, specialized covariances that remain stable under ongoing noise. Self-regulation, rather than simple noise filtering, produces this structured state.

What carries the argument

The macroscopic closed dynamics for the three collective covariance matrices T (genes), Q (effectors) and R (cross terms), obtained by self-averaging the fast–slow microscopic model in the infinite-signal limit and closed with Gaussian integrals.

Load-bearing premise

The cell’s adaptation of its signal-interpretation matrix is assumed to follow gradient descent on a single global mismatch between total effector output and total gene output, while the gene network always relaxes to a stable fixed point.

What would settle it

Measure gene–gene, effector–effector and cross-covariances in a morphogen-receiving tissue while systematically varying the temporal-integration timescale of the pathway; the predicted sharp onset of specialized diagonal-dominant R and a finite gap between diagonal and off-diagonal T and Q should appear only above a critical rate and should be absent when feedback is blocked.

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

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 models cellular robustness to non-instructive morphogen fluctuations as a multiscale feedback process. Fast GRN dynamics of K target genes (Eq. 3/8) are driven by M=K effectors whose interpretation matrix J is updated by gradient descent on the scalar mismatch ε = ½(Y-Z)² between summed effector and gene activations (Eqs. 4–6). In the N o∞ limit the authors derive closed macroscopic ODEs for the collective covariances T (gene–gene), Q (effector–effector) and R (cross) (Eqs. 14–15, Appendices A–C). Numerical integration and a symmetric-manifold analysis show that, beyond a critical adaptation rate η, the system undergoes a transition from vanishing/random correlations to structured, specialized modes (diagonal dominance of T, Q, R and emergent anti-correlations despite positive W), with an oscillatory regime at large η,γ. Robustness is thereby identified with the emergence of stable collective modes rather than mere noise suppression.

Significance. If the modeling premises hold, the work supplies a concrete, falsifiable link between online-learning theory and developmental robustness: a critical-η transition to specialized (T,Q,R) structure that can be compared with single-cell covariation data. Strengths include an explicit macroscopic closure (Gaussian integrals, Stein’s lemma), an analytically tractable σ o0 symmetric solution with a clear γ_c threshold, and phase diagrams that isolate the role of adaptation rate. The framing of phenotype maintenance as cooperative self-regulation rather than open-loop landscape deformation is a useful conceptual contribution to q-bio.MN.

major comments (3)
  1. [§II, Eqs. (4)–(6)] The central claim (Abstract; §IV; Figs. 2, 6) that robustness emerges as a critical-η transition to structured collective modes rests on the postulated learning rule (Eqs. 4–6): J is updated by gradient descent on the global scalar mismatch ε = ½(Y-Z)². No alternative objective (local matching, energy-based, or motif-based adaptation) is examined. Because asymptotic small ε is partly by construction of the rule, the nontrivial content of the transition must be shown to survive under other biologically plausible updates; otherwise the phase diagram of Fig. 6 remains an artifact of this particular objective.
  2. [§II; §IV; Eqs. (14)–(15)] The derivation and numerics are restricted a priori to the region where the fast GRN always converges to a stable fixed point under structural stability (§II last paragraph; §IV). The authors themselves note that structural stability “is not guaranteed after many iterations” and simply exclude limit-cycle/chaotic regimes. Because the macroscopic ODEs (14)–(15) and the reported transition presuppose this restriction, the claim that the transition accounts for phenotype maintenance is incomplete until the fate of (T,Q,R) is checked when the fixed-point assumption is relaxed (or when A is not Hurwitz).
  3. [§IV; Appendix B] All phase diagrams and the symmetric solution (Eqs. 16–19, Figs. 4–7) are obtained for random asymmetric W with i.i.d. positive entries drawn from U(0,1) and M=K. While this isolates emergent anti-correlations, it leaves open whether the critical-η transition and specialization of R persist for topologically realistic GRNs (e.g., the Shh neural-tube network of Balaskas et al.) or for M eq K. At least one such check is needed to support the claim of generic relevance.
minor comments (5)
  1. [Fig. 2] Figure 2 caption lists panels (a)–(i) but the body text refers to (a)–(c), (d)–(f), (g)–(i) inconsistently with the plotted η,γ pairs; the middle row is never labeled (d)–(f) in the figure itself.
  2. [§III] Notation for the slow time is τ_µ = µ/N then continuous τ; a single sentence clarifying that averages ⟨·⟩_h,s are with respect to the joint Gaussian at fixed τ would help readers unfamiliar with the online-learning literature.
  3. [§II] The claim that η ∝ 1/τ_0 (temporal integration timescale) is stated without a quantitative mapping to the Dessaud et al. experiments; a brief estimate would strengthen the experimental link.
  4. [Appendix C] Appendix C Jacobian elements are lengthy; a short statement of the leading eigenvalue that signals the transverse instability would make the onset of specialization more transparent.
  5. [throughout] Typos: “frow now on” (§IV), “predertemined” (§V), “asinglecell” (Abstract).

Circularity Check

1 steps flagged

Asymptotic mismatch reduction is partly by construction of the postulated GD rule; the critical-η transition to specialized (T,Q,R) modes is an independent dynamical result of the derived ODEs, not forced by definition or self-citation.

specific steps
  1. self definitional [§II, Eqs. (4)–(6) and surrounding text; Fig. 3(a)]
    "the update of J^µ to J^µ⁺¹ aims at reducing the following mismatch … ε(τ_µ) := ½(Y^µ - Z^µ)² … J_kα(τ_µ⁺¹) = J_kα(τ_µ) - (η/N) ∂ε/∂J_kα … One would expect an asymptotically diminishing value ε(τ_µ o ∞) o 0 … In both cases, ε relaxes towards zero"

    The continuous-time dynamics (7) are exactly gradient flow on ε. Under the paper’s maintained assumptions (small steps, structural stability of the fast fixed point), decrease of ε is therefore guaranteed by construction of the learning rule; reporting that ε o0 for moderate η merely reconfirms the definition rather than constituting an independent prediction.

full rationale

The microscopic model postulates that J evolves by gradient descent on the scalar global mismatch ε = ½(Y-Z)² (Eqs. 4–6). Consequently, under the continuous-time limit and the structural-stability restriction that keeps the fast GRN at fixed points, ε is expected to diminish; numerical observation that ε o0 (Fig. 3a) is therefore partly tautological for small η. However, the paper’s strongest claim—the existence of a critical adaptation rate beyond which permutation symmetry of the effectors breaks, producing specialized positive R_ii, emergent anti-correlations in T and Q despite non-negative W, and a transition from vanishing to structured collective modes (Figs. 2, 6; Eqs. 16–19)—is obtained by deriving and integrating the closed macroscopic ODEs (14)–(15) and by linear-stability analysis of the symmetric manifold. These steps do not reduce to the definition of ε, nor do they import a uniqueness theorem or fitted parameter from prior work. The citations to Saad & Solla (1995) supply only the standard Gaussian-integral technology for online learning; the biological model, the order parameters, and the phase diagram are new. No data are fitted and then re-predicted. The modeling assumptions (global-mismatch objective, forced fixed-point GRNs) are strong and untested against real pathways, but that is an external-validity issue, not circularity of the derivation chain. Hence only minor, non-load-bearing circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central claim rests on a postulated learning objective for pathway sensitivity, timescale separation with fixed-point GRN attractors, Gaussian collective statistics, and several free control parameters (η, γ, c, σ) scanned rather than measured. No new physical particle is invented, but the interpretation matrix as adaptive learning weights and the global mismatch ε are paper-specific modeling entities without independent experimental handles in this work.

free parameters (6)
  • η (adaptation rate)
    Bifurcation parameter controlling the random-to-structured transition; scanned by hand (e.g. 0.5–1.8), not fitted to measured morphogen integration times.
  • γ (effector control strength on genes)
    Sets impact of effectors on GRN and selects fixed vs oscillatory macroscopic regimes; free scan parameter.
  • c (relative cross-regulation strength)
    Fixed at 0.2 in main figures to keep A Hurwitz-stable; chosen, not measured.
  • σ (gene expression noise amplitude)
    Set to 0.01 to avoid attractor switching; small-noise regime is assumed.
  • W entry distribution / connectivity
    i.i.d. U(0,1) fully connected asymmetric matrices used as generic GRNs instead of data-inferred networks.
  • K (=M) system size
    Mainly K=4 to match four ventral neural-tube genes; also scanned to 8,16 for finite-size trends.
axioms (6)
  • domain assumption Fast GRN dynamics always relax to a stable fixed point (structurally stable under slow g changes); limit cycles/chaos excluded by parameter restriction.
    Stated in §II; required for nested integration and for T(τ) as phenotype covariance.
  • ad hoc to paper Effector fields h = Jx with uncorrelated white signal fluctuations; J is a coarse-grained interpretation matrix updated by gradient descent on ε.
    Eqs. (1)–(6); learning rule is postulated as the cell’s adaptation mechanism.
  • domain assumption Steady-state gene expression is approximately multivariate Gaussian; covariances close under linearization of the fast dynamics (A Hurwitz).
    §III, Eq. (11); standard OU/linear-noise assumption.
  • standard math N→∞ self-averaging yields closed ODEs for Q and R; averages expressible via Gaussian integrals of ϕ=erf.
    Online-learning style derivation §III and Appendix A.
  • ad hoc to paper M=K and additive linear control γg_i separate cis-drive from network interactions.
    §II explicitly notes additivity is artificial relative to microscopic cis-regulation models.
  • domain assumption No direct repressive W entries; anti-correlations, if any, must emerge from self-regulation.
    §IV setup contrasting with repressive-interaction explanations in prior Shh work.
invented entities (3)
  • Adaptive interpretation matrix J as online-learning weights for morphogen fluctuations no independent evidence
    purpose: Encode pathway sensitivity that cells can retune to maintain phenotype under non-instructive signal noise.
    Coarse-grained construct; not identified with specific molecular bindings in this paper.
  • Global mismatch objective ε = ½(Y−Z)² between total effector and gene activations no independent evidence
    purpose: Provide the scalar loss whose gradient defines adaptation of J.
    Postulated fluctuation-suppressing / matching principle; not derived from measured molecular cost functions.
  • Collective modes (T, Q, R) as macroscopic phenotype descriptors under fluctuating signal no independent evidence
    purpose: Closed description of ensemble covariations comparable in spirit to sequencing-based collective modes.
    Standard order parameters in learning theory, repurposed as cell-type signatures; experimental mapping left for future work.

pith-pipeline@v1.1.0-grok45 · 23585 in / 3842 out tokens · 35748 ms · 2026-07-12T01:42:54.628358+00:00 · methodology

0 comments
read the original abstract

Cells represent one of the most fundamental units of life. Underlying their robust performance against environmental variability, such as temporal fluctuations of chemical signals, between different cell types, is a dynamical interrelation between the two components of an intracellular pathway: a gene-regulatory network and its upstream signal transducers. To understand how a single cell utilizes this feedback to self-regulate its gene-expressions, we develop a multiscale model of the pathway's components, in which the adaptive variables responsible for signal interpretation follow a feedback-induced learning process. We then derive a macroscopic theory capturing the covariations between these components - so-called collective modes. Our theory shows how cells can achieve robust output against signal fluctuations via self-regulation rather than simple noise suppression. Such robustness corresponds to a transition from random- to structured collective modes beyond a critical adaptation rate.

Figures

Figures reproduced from arXiv: 2607.03545 by David Saad, Tuan Minh Pham.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. The target gene correlations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Symmetric solution as function of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Dependence of [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] 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

59 extracted references · 1 canonical work pages

  1. [1]

    snapshot

    =−1 + (2/π) 1/2 Z z −∞ duexp(−u 2/2). sj si ℓ τ x(η) x Microscopic model Jγ(η) Jk(η) hγ hk WijWji FastSignaling noise target genes Slow effectors Macroscopic theory Tij=⟨sisj⟩ Qkℓ=⟨hkhℓ⟩ Rik=⟨hisk⟩ Gene-gene covariances Effector covariances Intergroup covariances x Macroscopic theoryMicroscopic model sj si ℓ τ x(η) x Microscopic model Jγ(η) Jk(η) hγ hk WijW...

    2000

  2. [2]

    Kicheva and J

    A. Kicheva and J. Briscoe, Control of tissue development by morphogens, Annual Review of Cell and Developmental Biol- ogy39, 91 (2023)

    2023

  3. [3]

    M ¨uller, K

    P. M ¨uller, K. W. Rogers, S. R. Yu, M. Brand, and A. F. Schier, Morphogen transport, Development140, 1621 (2013)

    2013

  4. [4]

    Ebisuya and J

    M. Ebisuya and J. Briscoe, Timing and tempo in develop- ment, Current Opinion in Genetics & Development86, 102202 (2024)

    2024

  5. [5]

    L. Ham, T. E. Woodward, M. A. Coomer, and M. P. Stumpf, Mapping, modeling, and reprogramming cell-fate decision- making systems, Annual Review of Biomedical Data Science 8, 537–562 (2025)

    2025

  6. [6]

    F. Tang, K. Lao, and M. A. Surani, Development and applica- tions of single-cell transcriptome analysis, Nature Methods8, S6 (2021)

    2021

  7. [7]

    B. D. MacArthur and I. R. Lemischka, Statistical mechanics of pluripotency, Cell154, 484 (2013)

    2013

  8. [8]

    A. E. Teschendorff and A. P. Feinberg, Statistical mechanics meets single-cell biology, Nature Reviews Genetics22, 459 (2021)

    2021

  9. [9]

    Gorin, J

    G. Gorin, J. J. Vastola, and L. Pachter, Studying stochastic sys- tems biology of the cell with single-cell genomics data, Cell Systems14, 822 (2023)

    2023

  10. [10]

    C. H. Waddington,The strategy of the genes(George Allen & Unwin, 1957)

    1957

  11. [11]

    D. A. Rand, A. Raju, M. S ´aez, F. Corson, and E. D. Siggia, Geometry of gene regulatory dynamics, Proceedings of the Na- tional Academy of Sciences118, e2109729118 (2021)

    2021

  12. [12]

    L. Zhu, S. Yang, K. Zhang, H. Wang, X. Fang, and J. Wang, Uncovering underlying physical principles and driving forces of cell differentiation and reprogramming from single-cell tran- scriptomics, Proceedings of the National Academy of Sciences 121, e2401540121 (2024)

    2024

  13. [13]

    Zhou and T

    P. Zhou and T. Li, Construction of the landscape for multi-stable systems: Potential landscape, quasi-potential, a-type integral and beyond, The Journal of Chemical Physics144, 094109 (2016)

    2016

  14. [14]

    P. Zhou, S. Wang, T. Li, and Q. Nie, Dissecting transition cells from single-cell transcriptome data through multiscale stochas- tic dynamics, Nature Communications12, 5609 (2021)

    2021

  15. [15]

    Villarreal, P

    C. Villarreal, P. Padilla-Longoria, and E. R. Alvarez-Buylla, General theory of genotype to phenotype mapping: Derivation of epigenetic landscapes from n-node complex gene regulatory networks, Phys. Rev. Lett.109, 118102 (2012)

    2012

  16. [16]

    J. Wang, K. Zhang, L. Xu, and E. Wang, Quantifying the waddington landscape and biological paths for development and differentiation, Proceedings of the National Academy of Sciences108, 8257 (2011)

    2011

  17. [17]

    A. H. Lang, H. Li, J. J. Collins, and P. Mehta, Epigenetic landscapes explain partially reprogrammed cells and identify key reprogramming genes, PLOS Computational Biology10, 1 (2014)

    2014

  18. [18]

    Wolpert, Positional information and the spatial pattern of cellular differentiation, Journal of Theoretical Biology25, 1 (1969)

    L. Wolpert, Positional information and the spatial pattern of cellular differentiation, Journal of Theoretical Biology25, 1 (1969)

    1969

  19. [19]

    T. M. Jessell, Neuronal specification in the spinal cord: induc- tive signals and transcriptional codes, Nature Reviews Genetics 1, 20 (2000)

    2000

  20. [20]

    Cohen, A

    M. Cohen, A. Kicheva, A. Ribeiro, R. Blassberg, K. M. Page, C. P. Barnes, and J. Briscoe, Ptch1 and gli regulate shh sig- nalling dynamics via multiple mechanisms, Nature Communi- cations6, 6709 (2015)

    2015

  21. [21]

    Pezzotta and J

    A. Pezzotta and J. Briscoe, Optimal control of gene regulatory networks for morphogen-driven tissue patterning, Cell Systems 14, 940 (2023)

    2023

  22. [22]

    Balaskas, A

    N. Balaskas, A. Ribeiro, J. Panovska, E. Dessaud, N. Sasai, K. M. Page, J. Briscoe, and V . Ribes, Gene regulatory logic for reading the sonic hedgehog signaling gradient in the vertebrate neural tube, Cell148, 273–284 (2012)

    2012

  23. [23]

    Guillemin and M

    A. Guillemin and M. P. H. Stumpf, Non-equilibrium statistical physics, transitory epigenetic landscapes, and cell fate decision dynamics, Mathematical Biosciences and Engineering17, 7916 (2020)

    2020

  24. [24]

    Dessaud, L

    E. Dessaud, L. L. Yang, K. Hill, B. Cox, F. Ulloa, A. Ribeiro, 12 A. Mynett, B. G. Novitch, and J. Briscoe, Interpretation of the sonic hedgehog morphogen gradient by a temporal adaptation mechanism, Nature450, 717 (2007)

    2007

  25. [25]

    Engel and C

    A. Engel and C. Broeck,Statistical Mechanics of Learning (Cambridge University Press, 2001)

    2001

  26. [26]

    Gupta, J

    A. Gupta, J. D. Martin-Rufino, T. R. Jones, V . Subramanian, X. Qiu, E. I. Grody, A. Bloemendal, C. Weng, S.-Y . Niu, K. H. Min, A. Mehta, K. Zhang, L. Siraj, A. A. Khafaji, V . G. Sankaran, S. Raychaudhuri, B. Cleary, S. Grossman, and E. S. Lander, Inferring gene regulation from stochastic tran- scriptional variation across single cells at steady state, ...

    2022

  27. [27]

    D. J. Skinner, P. Lemaire, and M. Mani, Physical modeling of embryonic transcriptomes identifies collective modes of gene expression, PRX Life3, 033028 (2025)

    2025

  28. [28]

    Olmeda, Y

    F. Olmeda, Y . Dang, F. Rost, V . M. Schimmenti, I. Di Ter- lizzi, and S. Rulands, Coordinated gene expression vari- ability encodes the regulatory state of cells, bioRxiv 10.1101/2025.07.09.663203 (2026)

    work page doi:10.1101/2025.07.09.663203 2025

  29. [29]

    Goldt, M

    S. Goldt, M. S. Advani, A. M. Saxe, F. Krzakala, and L. Zde- borov´a, Dynamics of stochastic gradient descent for two-layer neural networks in the teacher–student setup, Journal of Statis- tical Mechanics: Theory and Experiment2020, 124010 (2020)

    2020

  30. [30]

    B. Yuan, C. Shen, A. Luna, A. Korkut, D. S. Marks, J. Ingra- ham, and C. Sander, Cellbox: Interpretable machine learning for perturbation biology with application to the design of can- cer combination therapy, Cell Systems12, 128 (2021)

    2021

  31. [31]

    Paczk ´o, D

    M. Paczk ´o, D. V ¨or¨os, P. Szab ´o, G. J ´ekely, E. Szathm ´ary, and A. Szil´agyi, A neural network-based model framework for cell- fate decisions and development, Communications Biology7, 323 (2024)

    2024

  32. [32]

    Reinitz and D

    J. Reinitz and D. H. Sharp, Mechanism of eve stripe formation, Mechanisms of Development49, 133 (1995)

    1995

  33. [33]

    Matsushita and K

    Y . Matsushita and K. Kaneko, Homeorhesis in waddington’s landscape by epigenetic feedback regulation, Phys. Rev. Res.2, 023083 (2020)

    2020

  34. [34]

    M. S. Sherman and B. A. Cohen, Thermodynamic state ensem- ble models of cis-regulation, PLOS Computational Biology8, 1 (2012)

    2012

  35. [35]

    Goutsias and S

    J. Goutsias and S. Kim, A nonlinear discrete dynamical model for transcriptional regulation: Construction and properties, Bio- physical Journal86, 1922 (2004)

    1922

  36. [36]

    M. A. Coomer, L. Ham, and M. P. Stumpf, Noise distorts the epigenetic landscape and shapes cell-fate decisions, Cell Sys- tems13, 83 (2022)

    2022

  37. [37]

    Sagner and J

    A. Sagner and J. Briscoe, Morphogen interpretation: concen- tration, time, competence, and signaling dynamics, WIREs De- velopmental Biology6, e271 (2017)

    2017

  38. [38]

    In biological terms, such multistability corresponds to the in- duction of a distinct set of target genes in different tissues in response to the same signal – a context-dependent interpreta- tion

  39. [39]

    S ´aez, R

    M. S ´aez, R. Blassberg, E. Camacho-Aguilar, E. D. Siggia, D. A. Rand, and J. Briscoe, Statistically derived geometrical land- scapes capture principles of decision-making dynamics during cell fate transitions, Cell Systems13, 12 (2022)

    2022

  40. [40]

    Godreche and J.-M

    C. Godreche and J.-M. Luck, Characterising the nonequilib- rium stationary states of ornstein–uhlenbeck processes, Journal of Physics A: Mathematical and Theoretical52, 035002 (2018)

    2018

  41. [41]

    C. W. H. Mace and A. C. C. Coolen, Statistical mechanical anal- ysis of the dynamics of learning in perceptrons, Statistics and Computing8, 55 (1998)

    1998

  42. [42]

    Saad and S

    D. Saad and S. A. Solla, Exact solution for on-line learning in multilayer neural networks, Phys. Rev. Lett.74, 4337 (1995)

    1995

  43. [43]

    Saad and S

    D. Saad and S. A. Solla, On-line learning in soft committee machines, Phys. Rev. E52, 4225 (1995)

    1995

  44. [44]

    Biehl, P

    M. Biehl, P. Riegler, and C. W ¨ohler, Transient dynamics of on- line learning in two-layered neural networks, Journal of Physics A: Mathematical and General29, 4769 (1996)

    1996

  45. [45]

    We emphasize that replacingWby its mean is a strong approx- imation and it is not exactly equivalent to the self-averaging as- sumption

  46. [46]

    Longitudinal perturbation does not push the solution leaving the symmetric manifold, so is not responsible for specialization

  47. [47]

    Fukumizu and S

    K. Fukumizu and S. Amari, Local minima and plateaus in hier- archical structures of multilayer perceptrons, Neural Networks 13, 317 (2000)

    2000

  48. [48]

    Inoue and K

    M. Inoue and K. Kaneko, Cooperative adaptive responses in gene regulatory networks with many degrees of freedom, PLOS Computational Biology9, 1 (2013)

    2013

  49. [49]

    C. J. Russo, K. Husain, R. Ranganathan, D. Pincus, and A. Mu- rugan, Simple biological controllers drive the evolution of soft modes, Proceedings of the National Academy of Sciences123, e2523032123 (2026)

    2026

  50. [50]

    Plugers and K

    D. Plugers and K. Kaneko, Evolution of robust cell differen- tiation under epigenetic feedback, Phys. Rev. Res.8, 013059 (2026)

    2026

  51. [51]

    A. H. L. West and D. Saad, Role of biases in on-line learning of two-layer networks, Phys. Rev. E57, 3265 (1998)

    1998

  52. [52]

    W. Ma, A. Trusina, H. El-Samad, W. A. Lim, and C. Tang, Defining network topologies that can achieve biochemical adaptation, Cell138, 760 (2009)

    2009

  53. [53]

    L. Qiao, W. Zhao, C. Tang, Q. Nie, and L. Zhang, Network topologies that can achieve dual function of adaptation and noise attenuation, Cell Systems9, 271 (2019)

    2019

  54. [54]

    T. M. Pham and K. Kaneko, Dynamical theory for adaptive sys- tems, Journal of Statistical Mechanics: Theory and Experiment 2024, 113501 (2024)

    2024

  55. [55]

    Pham and D

    T. Pham and D. Gupta, Entropy production rate in stochastically time-evolving asymmetric networks, arXiv2603.27658(2026)

    arXiv 2026

  56. [56]

    Saad and M

    D. Saad and M. Rattray, Globally optimal parameters for on- line learning in multilayer neural networks, Phys. Rev. Lett.79, 2578 (1997)

    1997

  57. [57]

    Rattray and D

    M. Rattray and D. Saad, Analysis of natural gradient descent for multilayer neural networks, Phys. Rev. E59, 4523 (1999)

    1999

  58. [58]

    Pratapa, A

    A. Pratapa, A. P. Jalihal, J. N. Law, A. Bharadwaj, and T. M. Murali, Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data, Nature Methods 17, 147 (2020)

    2020

  59. [59]

    Ramsden, Regulatory networks, inBioinformatics: An In- troduction(Springer International Publishing, Cham, 2023) pp

    J. Ramsden, Regulatory networks, inBioinformatics: An In- troduction(Springer International Publishing, Cham, 2023) pp. 311–338

    2023