Exploration of the generative capabilities of Boltzmann machines applied to social systems under the majority rule

Gonzalo A. Ruz; Mauricio A. Valle

REVIEW 3 major objections 5 minor 28 references

Exploration of the generative capabilities of Boltzmann machines applied to social systems under the majority rule

T0 review · 3 major / 5 minor · reviewed 2026-07-31 · deepseek-v4-flash

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

Pith's one-line read This paper argues that a deep belief network with Gaussian visible units can impute missing opinions in the three-state majority-vote model and produce reconstructions that remain in the critical phase, degrading gradually as more opinions

desk verdict DBN imputation for the MV3 model is a clean extension, but the criticality claim rests on a thermometer trained on the model's own outputs—worth publishing only if that circularity is confronted. read the letter →

arxiv 2607.23349 v1 pith:NFQPEXW5 submitted 2026-07-25 cs.LG cs.CY

Mauricio A. Valle , Gonzalo A. Ruz This is my paper

classification cs.LGcs.CY

keywords Boltzmannmachinesdeepbeliefnetworksmajority-rulemodelopiniondynamicscriticalityimputationcontrastivedivergencemean-fieldsampling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Generative models are usually tested on clean, structured data; this paper tries to show the same machinery works for a disordered, three-state opinion system near its phase transition. The authors train deep belief networks on Monte Carlo configurations of the three-state majority-vote model on a 28×28 lattice, then ask the network to 'dream' complete configurations from inputs where a fraction f of agent opinions has been erased. Their central result is that the dreamed samples remain in the critical phase — the regime between consensus and fragmentation — with the fraction of misclassified critical samples rising only gradually as f grows. If this holds, a trained Boltzmann machine could impute missing opinions in real-world social data, where most users are silent, and generate synthetic critical scenarios. The paper's own caveat is that the phase 'thermometer' used to certify criticality is weakest exactly at the critical point, so the criticality results are an upper bound.

What carries the argument

The load-bearing mechanism is a three-layer DBN (one Gaussian-Bernoulli RBM followed by two Bernoulli-Bernoulli RBMs) combined with mean-field imputation: a partially observed configuration is clamped at the visible layer, hidden units mix a bottom-up data signal and a top-down prior through parameters alpha_l, and the completed sample is generated from the top layer. Criticality of the completed sample is checked by a convolutional-network 'thermometer' trained to classify subcritical, critical, and supercritical states.

What would settle it

Run the same imputation experiments but score the dreamed samples with a thermodynamic observable instead of the learned thermometer: compute the mean absolute magnetization |m| = |Σ s_i|/N and fourth-order cumulant U_4 = 1 - <m^4>/(3<m^2>^2) for S'(p). If U_4 for dreamed samples at p≈0.8-0.85 does not match the critical value seen in the original MV3 samples (accounting for finite-size fluctuations), the claim of maintained criticality is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a DBN whose first visible layer is a Gaussian-Bernoulli RBM can approximate the equilibrium distribution of the three-state majority-vote model and, through a clamped mean-field imputation procedure, reconstruct a complete critical-phase sample S'(p) from a partially observed input S(p,f). The authors report low reconstruction errors (root-mean-square roughly 0.065-0.075 for expansive architectures) and, using a convolutional thermometer, find that reconstructions are still recognized as critical at missing fractions f up to 0.5, with gradual degradation of the misclassification rate. They also show that fully compressive architectures preserve criticality better in

Load-bearing premise

The load-bearing premise is that the CNN thermometer is trustworthy enough at the critical point to certify criticality — the paper reports its critical-state AUC as only 0.838 and says results should be read as upper bounds — together with the assumption that Gaussian visible units with a fixed empirical covariance faithfully encode the three-state opinion distribution.

Editorial extensions

If this is right

Expansive first-layer architectures (more hidden units than visible units) give the lowest reconstruction errors, while fully compressive architectures are more often recognized as critical by the thermometer, so architecture choice trades fidelity against phase preservation.
Imputation error grows with the missing fraction f and can be steered by the mixing parameters; data-driven weighting (higher alpha_1) works best for the expansive-contractive network.
Thermometer misclassification of critical reconstructions rises with f, reaching about 26% at f=0.3 for one expansive-contractive architecture, with compressive networks misclassified less often.
Degradation is gradual rather than abrupt as more opinions are erased, meaning the generative model fails softly.
Because the thermometer's critical-state recognition is imperfect, the reported criticality-maintenance numbers are upper bounds, not precise measurements.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Since the learned thermometer is the bottleneck at criticality (AUC 0.838), the true criticality preservation could be stronger than reported; a direct order-parameter check, such as comparing the fourth-order cumulant or largest-cluster distribution of S'(p) with the original critical samples, would test this without a learned classifier.
The fixed Gaussian encoding of the three discrete opinion states is a modeling choice; replacing it with a one-hot or softmax visible layer would clarify whether the imputation errors are intrinsic to the DBN or an artifact of the encoding.
The gradual-degradation curve suggests a practical imputation operating point at missing fractions below about 0.3-0.5, where completions still resemble critical configurations; this could be tested on social-media data with many silent users.
The authors' observation that more Gibbs steps do not improve reconstruction points to a memorization effect of contrastive divergence; a hybrid that refines DBN imputations with score-based diffusion might recover the conditional distribution better, as the paper itself suggests.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Punchline: the DBN imputation work on MV3 is a genuine extension of your Ising paper, and the experimental setup is transparent, but the claim that reconstructions stay critical rests on a CNN thermometer that was trained on GBRBM-generated samples, not on actual MV3 configurations. That is a circular validation, and it means Figure 10 cannot support the headline.

Credit: the architecture comparison (expansive vs contractive) and the mean-field imputation with mixing parameters are new, and the reconstruction-error numbers are a useful empirical baseline. The authors are also honest: they report the critical-class AUC of 0.838 and tell you to treat the results as upper bounds. That is more than many papers do.

Soft spots: First, the thermometer. Section II-C says it is trained on "synthetic samples S'(p) generated by a GBRBM fed by 24,000 simulated systems." Since a DBN is a stacked GBRBM, you are essentially using a model family to certify its own outputs. The labels come from the p of the source simulations, but the input distribution is the GBRBM's approximation, not the MV3 distribution. If the GBRBM is biased, the thermometer may be picking up reconstruction artifacts (e.g., rounding of Gaussian outputs) rather than physical criticality. They never evaluate the thermometer on held-out ground-truth MV3 critical samples, so the AUC of 0.838 is measured on GBRBM-generated test data, not on the real thing. That is load-bearing.

Second, the abstract says "several training sessions" but there are no error bars on imputation errors or misclassification rates. You cannot tell if differences between DBN1, DBN5, and DBN7 are real or noise.

Third, there is no baseline. A nearest-neighbor imputation on the lattice, or even per-agent majority vote, would show what the DBN adds. Without that, the practical value is unclear.

Fourth, the Gaussian visible encoding with fixed empirical covariance is never justified; a one-hot or Bernoulli encoding might behave differently. The rounding artifacts mentioned above are a real threat.

None of these flaws invalidate the reconstruction-error measurements or the imputation-error trends for DBN1 and DBN7. Those are legitimate empirical observations. But the central claim about maintaining criticality overreaches. The authors' own caveat is a symptom: "upper bounds" does not rescue the logic.

Who this is for: people working on generative models for opinion dynamics, and anyone who wants a clean example of circular validation in ML-for-physics. It is a good reading-group case.

Recommendation: send it to peer review, but the reviewers should require the authors to validate the thermometer on real MV3 critical samples and to report error bars plus a simple baseline. If they do that, the paper becomes a decent contribution; right now it is a solid first step that oversells the criticality claim.

Referee Report

3 major / 5 minor

Summary. The paper trains deep belief networks (DBNs) with a Gaussian-Bernoulli RBM first layer on configurations generated by the three-state majority-vote model (MV3), evaluates reconstruction and imputation errors for several architectures, and uses a convolutional 'thermometer' to claim that DBN-‘dreamed’ reconstructions of samples with missing opinions remain in the critical phase. The central claim, stated in the abstract and in Section II-E, is that DBN imputation can recover critical samples even at substantial missing fractions, with gradual degradation of physical observables.

Significance. If the criticality-preservation claim were established, this would be a useful contribution to generative modeling of opinion dynamics near phase transitions, with practical implications for imputation in social systems. The paper has several commendable features: systematic comparison of multiple DBN architectures, a normalized error metric that benchmarks imputation against worst-case errors, and explicit and honest caveats about the thermometer's limitations. However, the central criticality claim is not currently supported because the thermometer used to certify criticality is trained on outputs of the same Boltzmann-machine family that produces the reconstructions, and its critical-state discriminability is modest (AUC 0.838). The manuscript is a reasonable empirical study of reconstruction and imputation error, but the headline result about maintained criticality requires revalidation with a ground-truth-based thermometer or physical order parameters.

major comments (3)

[II-C and II-E] The thermometer is trained on synthetic samples S'(p) generated by a GBRBM fed from 24,000 simulated systems (Section II-C). A DBN is a stacked GBRBM/BBRBM, so the thermometer's 'critical' label is learned from the output distribution of the same generative-model family that is later used to produce reconstructions. Consequently, Figure 10 can at most show that DBN reconstructions resemble typical GBRBM outputs, not that they are in the physical MV3 critical phase. This is a circularity in the central claim. Concrete fix: retrain the thermometer on ground-truth MV3 configurations S(p) with known p labels, or evaluate S'(p) with physical order parameters (e.g., Binder cumulant, susceptibility, cluster-size distribution). The manuscript's own caveat at the end of II-E ('upper bounds on maintained criticality') is appropriate but does not resolve the circularity.
[II-E and Fig. 10] The critical-state AUC is 0.838, and the authors acknowledge the thermometer has 'reduced sensitivity at the critical point.' The misclassification rates in Fig. 10 are therefore uncalibrated: with AUC 0.838 and an operating point chosen to favor true-positive rate, a substantial misclassification fraction is expected even for perfectly reconstructed critical samples. No confusion matrices, confidence intervals, or multiple-seed variation are reported for the thermometer judgments. Without calibration against ground-truth critical samples, the specific misclassification numbers (e.g., 26% for DBN1 at f=0.3) cannot support the claim that some architectures 'preserve criticality' better than others. Please report thermometer performance on held-out ground-truth critical MV3 samples and use that to correct or bound the Fig. 10 results.
[II-B, Eq. (3), and II-D] The mapping from continuous Gaussian visible units to the discrete three-state space {0,0.5,1} is underspecified. The paper defines the GBRBM's visible distribution as Gaussian in Eq. (3), but imputation and error computations in Eqs. (5) assume discrete state values. How are continuous samples converted to discrete states (thresholding, rounding, or sampling)? This matters because the thermometer may classify based on artifacts of continuous-to-discrete rounding rather than genuine cluster statistics. The authors do not compare against a one-hot encoding of the three states, which could remove this ambiguity. Please specify the discretization procedure and, at minimum, report a comparison with one-hot visible units to show the results are not an artifact of the Gaussian encoding.

minor comments (5)

[II-B, Eq. (2)] Equation (2) contains a typo: the hidden-unit term is written as (h_j - c_j)^2 / s_j, but the conditional variance in Eq. (4) is denoted S^2. Define s_j consistently and state whether it is the variance or standard deviation.
[II-D, Eq. (6)] The mixing parameters are introduced with α1 + α2 = 1, but the notation is confusing because two layers are mixed while the equation writes a single sum over layers. Clarify how α_l is defined and normalized, particularly whether α1 and α2 are weights on bottom-up versus top-down contributions.
[II-A and Fig. 2] The paper states 'regular boundary conditions'; presumably periodic boundary conditions are intended. Also, the caption of Fig. 2 says 'from more to less noise' but the p values decrease from 0.97 to 0.65, which is actually from less noise to more noise. Please correct.
[II-C] Change 'RELU' to 'ReLU' in the architecture description. Also, Figure 4 labels the visible layer as 'nV=786', but the text says nV=784; one of these is a typo.
[Table I and Fig. 9] The reconstructed-error values in Table I are reported without error bars; since the paper says 'several training sessions' were used, report mean ± std over at least a few independent initializations. Similarly, Fig. 9 would benefit from error bars or at least a statement of how many imputation runs were averaged.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the evaluation is an empirical consistency check with acknowledged limitations, and the self-citation is not load-bearing.

full rationale

The derivation chain is: MV3 Monte Carlo simulations produce S(p); DBNs with a GBRBM first layer are trained on S(p); imputation of S(p,f) yields reconstructions S'(p); a CNN thermometer classifies these reconstructions as subcritical, critical, or supercritical. The only apparent circularity concern is that Section II-C states: 'The thermometer is trained on the basis of synthetic samples S′(p) generated by the GBRBM fed by a total of 24,000 simulated systems exclusively for this task and balanced for each of the three classes.' Since the DBN is also built on a GBRBM first layer, the thermometer may be recognizing features of GBRBM outputs rather than physical MV3 criticality. However, the thermometer's class labels are taken from the known p of the simulated systems that feed the GBRBM, not from the DBN being tested. The DBN outputs are not forced: Figure 10 reports misclassification rates that vary by architecture and missing fraction, so the test has empirical content. The paper itself flags the limitation in Section II-E: 'the thermometer’s reduced sensitivity at the critical point introduces uncertainty in the misclassification rates reported in Figure 10... These results should be interpreted as upper bounds on maintained criticality rather than precise measurements.' The mixing parameters α are tuned to minimize imputation error, but this is standard model selection and does not make the criticality evaluation definitionally equivalent to its inputs. The self-citation [13] appears only as background motivation ('what-if' scenario exploration) and is not load-bearing for the central claim. Overall, no step reduces by construction to its inputs; the concerns are measurement-validity limitations, not circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the MV3 simulation, the GBRBM/DBN model family, and the CNN thermometer. No new physical entities are introduced; the main freedom lies in α mixing, the hidden covariance scale s, the thermometer threshold, and the definition of the critical window.

free parameters (4)

Mixing parameters α1, α2 (α1+α2=1) = scanned values, e.g., α1=0.2, 0.3, 0.9
Chosen to minimize imputation error in Section II-D; not derived from theory.
Hidden covariance scale s = 0.5
'We found that setting s=0.5 yielded the most stable convergence' (Section II-B). Hand-tuned.
Thermometer decision threshold (operating point) = not specified numerically
'We adjusted the thermometer's operating point with the aim of maximizing the true positive rate and minimizing the false positive rate ... of samples at criticality' (Section II-E). Chosen on the same class of data used for evaluation.
Critical window definition (p boundaries) = p ∈ [0.75, 0.85]
Author-defined range for 'critical' used to label training data and select imputation test samples (Section II-C). Shifting the window changes the classification task.

assumptions (6)

domain assumption MV3 majority-rule dynamics (Section II-A) produce the ground-truth distribution of agent opinion states.
The simulated system is treated as the real system that the DBN must learn to reproduce.
domain assumption The GBRBM energy function (Eq. 2) and conditional distributions (Eqs. 3-4) can faithfully model the MV3 dataset.
The paper assumes Gaussian visible units with a fixed empirical covariance are an appropriate model for the tri-state data.
standard math Contrastive divergence training (k=1) produces a DBN that adequately approximates the data distribution.
Standard assumption in the RBM/DBN literature, invoked in Section II-B with refs [20], [21].
domain assumption The literature values of the critical point (q_c=0.106 for large lattices, finite-size shifted to q_c=0.151) are accurate for the simulated system.
Taken from ref [15] and used to define the critical region of interest.
domain assumption The CNN thermometer's classification is a valid measure of whether a reconstructed system is in a critical state.
The central evaluation relies on the thermometer, despite its acknowledged low sensitivity at criticality (Section II-C, II-E).
domain assumption The mean-field imputation formula (Eq. 6) with mixing parameters α_l is a valid approximation for layer-wise reconstruction in DBNs.
The imputation experiments in Section II-D depend on this heuristic; no formal justification is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploration of the generative capabilities of Boltzmann machines applied to social systems under the majority rule." pith.science (2026). https://pith.science/paper/NFQPEXW5

@misc{pith2026260723349,
  author       = {Pith},
  title        = {Pith review of: Exploration of the generative capabilities of Boltzmann machines applied to social systems under the majority rule},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NFQPEXW5}},
  note         = {Machine review of arXiv:2607.23349}
}

read the original abstract

We study the generative capabilities of Boltzmann machines to recover systems governed by the majority rule under critical conditions. To this end, we train deep belief networks (DBNs) with different configurations, where the first layer can use Gaussian visible units with more than two states (i.e., non-binary units). We then allow the DBN to "dream" samples conditioned on visible units that we keep fixed, and we measure the deviation of this dreamed system from the real one. We also corroborate, using a discrete thermometer based on a convolutional network, that the reconstructions remain in a critical state. Across several training sessions with different architectures, we show that, despite the complexity of the problem, the DBN can recover samples that remain critical even under input noise, with a gradual degradation of physical observables relative to the original sample.

Figures

Figures reproduced from arXiv: 2607.23349 by the authors.

**Figure 2.** Some examples of systems S(p) of L = 282 = 784 for different values of the parameter p, from more to less noise. where δ is the Kronecker delta, whose value is 1 if si = sj or 0 otherwise. To simulate the system, at each time, the states of the agents are updated simultaneously according to the following local rule: a) If the agent’s neighbors have a majority opinion, then the agent’s opinion changes to that of the … view at source ↗

**Figure 3.** Cluster behavior in the MV3 system of 28 × 28. Left: Average number of clusters found in the system for any opinion type. Right: Maximum cluster size found in the system independent of opinion. For each value of p, a simulation was performed 30 times to calculate a confidence band of ±1sd. TABLE I RECONSTRUCTION ERRORS MEASURED ON TEST DATASETS FOR DIFFERENT DBN ARCHITECTURES. Model Strategy nH1 nH2 nH3 epochs lr re… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Using the DBN to “dream” systems S ′ (p). Values of nH1, nH2 and nH3 represent the number of units in each stacked layer. In this case, the DBN has three layers, consisting of one GBRBM and then two BBRBMs. would be one in which the reconstruction fails for all agents …

**Figure 6.** Figure 6: ROC curves for the training and test sets in case B), corresponding [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: An example showing seven systems (first row [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 10.** Figure 10: Thermometer misclassification rate for reconstructed systems in [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

[1]

28: Opinion dynamics in social networks: from models to data,

A. F. Peralta, J. Kert ´esz, and G. I ˜niguez, “28: Opinion dynamics in social networks: from models to data,” inHandbook of Computational Social Science. Cheltenham, UK: Edward Elgar Publishing Limited, 2025, pp. 384 – 406

2025
[2]

Majority-vote model for financial markets,

A. L. Vilela, C. Wang, K. P. Nelson, and H. E. Stanley, “Majority-vote model for financial markets,”Physica A: Statistical Mechanics and its Applications, vol. 515, pp. 762–770, 2019

2019
[3]

The impact of external events on the emergence of social herding of economic sentiment,

M. Hohnisch, D. Stauffer, and S. Pittnauer, “The impact of external events on the emergence of social herding of economic sentiment,”
[4]

Herd behavior and financial crashes: An interacting particle system approach,

V . Crescimanna and L. Di Persio, “Herd behavior and financial crashes: An interacting particle system approach,”Journal of Mathematics, vol. 2016, no. 1, p. 7510567, 2016

2016
[5]

Statistical physics of social dynamics,

C. Castellano, S. Fortunato, and V . Loreto, “Statistical physics of social dynamics,”Rev. Mod. Phys., vol. 81, pp. 591–646, May 2009. [Online]. Available: https://link.aps.org/doi/10.1103/RevModPhys.81.591

doi:10.1103/revmodphys.81.591 2009
[6]

Phase transitions, collective emotions and decision-making problem in hetero- geneous social systems,

D. Tsarev, A. Trofimova, A. Alodjants, and A. Khrennikov, “Phase transitions, collective emotions and decision-making problem in hetero- geneous social systems,”Scientific Reports, vol. 9, no. 1, p. 18039, 2019

2019
[7]

Phase transition in the social impact model of opinion formation in scale-free networks: The social power effect,

A. Mansouri and F. Taghiyareh, “Phase transition in the social impact model of opinion formation in scale-free networks: The social power effect,”Journal of Artificial Societies and Social Simulation, vol. 23, no. 2, p. 3, 2020

2020
[8]

Feature extraction of machine learning and phase transition point of ising model,

S. S. Funai, “Feature extraction of machine learning and phase transition point of ising model,” 2021. [Online]. Available: https://arxiv.org/abs/2111.11166

arXiv 2021

Show all 28 references

[9]

Thermodynamics and feature extraction by machine learning,

S. S. Funai and D. Giataganas, “Thermodynamics and feature extraction by machine learning,”Phys. Rev. Res., vol. 2, p. 033415, Sep 2020. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevResearch.2.033415

2020 doi
[10]

Discovering phase transitions with unsupervised learning,

L. Wang, “Discovering phase transitions with unsupervised learning,” Phys. Rev. B, vol. 94, p. 195105, Nov 2016. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.94.195105

2016 doi
[11]

Discovering phases, phase transitions, and crossovers through unsupervised machine learning: A critical examination,

W. Hu, R. R. P. Singh, and R. T. Scalettar, “Discovering phases, phase transitions, and crossovers through unsupervised machine learning: A critical examination,”Phys. Rev. E, vol. 95, p. 062122, Jun 2017. [On- line]. Available: https://link.aps.org/doi/10.1103/PhysRevE.95.062122

2017 doi
[12]

Thermodynamics of the ising model encoded in restricted boltzmann machines,

J. Gu and K. Zhang, “Thermodynamics of the ising model encoded in restricted boltzmann machines,”Entropy, vol. 24, no. 12, 2022

2022
[13]

The capabilities of boltzmann machines to detect and reconstruct ising system’s configurations from a given temperature,

M. A. Valle, “The capabilities of boltzmann machines to detect and reconstruct ising system’s configurations from a given temperature,” Entropy, vol. 25, no. 12, 2023

2023
[14]

Ising percolation in a three-state majority vote model,

A. S. Balankin, M. Mart ´ınez-Cruz, F. Gayosso Mart ´ınez, B. Mena, A. Tobon, J. Pati ˜no-Ortiz, M. Pati ˜no-Ortiz, and D. Samayoa, “Ising percolation in a three-state majority vote model,”Physics Letters A, vol. 381, no. 5, pp. 440–445, 2017

2017
[15]

Three-state majority-vote model on square lattice,

F. Lima, “Three-state majority-vote model on square lattice,”Physica A: Statistical Mechanics and its Applications, vol. 391, no. 4, pp. 1753– 1758, 2012

2012
[16]

Three-state majority-vote model on small-world networks,

B. J. Zubillaga, A. M. Vilela, M. Wang, R. Du, G. Dong, and H. E. Stanley, “Three-state majority-vote model on small-world networks,” Scientific Reports, vol. 12, no. 1, p. 282, 2022

2022
[17]

Three-state majority-vote model on scale-free networks and the unitary relation for critical exponents,

A. M. Vilela, B. J. Zubillaga, C. Wang, M. Wang, R. Du, and H. E. Stanley, “Three-state majority-vote model on scale-free networks and the unitary relation for critical exponents,”Scientific Reports, vol. 10, no. 1, p. 8255, 2020

2020
[18]

Gaussian-bernoulli rbms without tears,

R. Liao, S. Kornblith, M. Ren, D. J. Fleet, and G. Hinton, “Gaussian-bernoulli rbms without tears,” 2022. [Online]. Available: https://arxiv.org/abs/2210.10318

2022 arXiv
[19]

Dynamical analysis of contrastive divergence learning: Restricted boltzmann machines with gaussian visible units,

R. Karakida, M. Okada, and S. ichi Amari, “Dynamical analysis of contrastive divergence learning: Restricted boltzmann machines with gaussian visible units,”Neural Networks, vol. 79, pp. 78–87, 2016

2016
[20]

Training restricted boltzmann machines: An introduction,

A. Fischer and C. Igel, “Training restricted boltzmann machines: An introduction,”Pattern Recognition, vol. 47, no. 1, pp. 25–39, 2014

2014
[21]

Training products of experts by minimizing contrastive divergence,

G. E. Hinton, “Training products of experts by minimizing contrastive divergence,”Neural Comput., vol. 14, no. 8, p. 1771–1800, Aug. 2002. [Online]. Available: https://doi.org/10.1162/089976602760128018

2002 doi
[22]

Machine learning phases of matter,

J. Carrasquilla and R. G. Melko, “Machine learning phases of matter,” Nature Physics, vol. 13, no. 5, pp. 431–434, 2017

2017
[23]

Equilibrium and non- equilibrium regimes in the learning of restricted boltzmann machines,

A. Decelle, C. Furtlehner, and B. Seoane, “Equilibrium and non- equilibrium regimes in the learning of restricted boltzmann machines,” Advances in neural information processing systems, vol. 34, pp. 5345– 5359, 2021

2021
[24]

Normalizing flows: An introduction and review of current methods,

I. Kobyzev, S. J. Prince, and M. A. Brubaker, “Normalizing flows: An introduction and review of current methods,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 11, pp. 3964– 3979, 2020

2020
[25]

Variational inference with normalizing flows,

D. Rezende and S. Mohamed, “Variational inference with normalizing flows,” inInternational conference on machine learning. PMLR, 2015, pp. 1530–1538

2015
[26]

Generative modeling by estimating gradients of the data distribution,

Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” inAdvances in Neural Information Processing Systems, vol. 32, 2019

2019
[27]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851

2020
[2008]

Available: https://arxiv.org/abs/physics/0606237

[Online]. Available: https://arxiv.org/abs/physics/0606237

arXiv

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [1]

28: Opinion dynamics in social networks: from models to data,

A. F. Peralta, J. Kert ´esz, and G. I ˜niguez, “28: Opinion dynamics in social networks: from models to data,” inHandbook of Computational Social Science. Cheltenham, UK: Edward Elgar Publishing Limited, 2025, pp. 384 – 406

2025

[2] [2]

Majority-vote model for financial markets,

A. L. Vilela, C. Wang, K. P. Nelson, and H. E. Stanley, “Majority-vote model for financial markets,”Physica A: Statistical Mechanics and its Applications, vol. 515, pp. 762–770, 2019

2019

[3] [3]

The impact of external events on the emergence of social herding of economic sentiment,

M. Hohnisch, D. Stauffer, and S. Pittnauer, “The impact of external events on the emergence of social herding of economic sentiment,”

[4] [4]

Herd behavior and financial crashes: An interacting particle system approach,

V . Crescimanna and L. Di Persio, “Herd behavior and financial crashes: An interacting particle system approach,”Journal of Mathematics, vol. 2016, no. 1, p. 7510567, 2016

2016

[5] [5]

Statistical physics of social dynamics,

C. Castellano, S. Fortunato, and V . Loreto, “Statistical physics of social dynamics,”Rev. Mod. Phys., vol. 81, pp. 591–646, May 2009. [Online]. Available: https://link.aps.org/doi/10.1103/RevModPhys.81.591

doi:10.1103/revmodphys.81.591 2009

[6] [6]

Phase transitions, collective emotions and decision-making problem in hetero- geneous social systems,

D. Tsarev, A. Trofimova, A. Alodjants, and A. Khrennikov, “Phase transitions, collective emotions and decision-making problem in hetero- geneous social systems,”Scientific Reports, vol. 9, no. 1, p. 18039, 2019

2019

[7] [7]

Phase transition in the social impact model of opinion formation in scale-free networks: The social power effect,

A. Mansouri and F. Taghiyareh, “Phase transition in the social impact model of opinion formation in scale-free networks: The social power effect,”Journal of Artificial Societies and Social Simulation, vol. 23, no. 2, p. 3, 2020

2020

[8] [8]

Feature extraction of machine learning and phase transition point of ising model,

S. S. Funai, “Feature extraction of machine learning and phase transition point of ising model,” 2021. [Online]. Available: https://arxiv.org/abs/2111.11166

arXiv 2021