A Scalable Exponential Random Graph Model: Amortised Hierarchical Sequential Neural Posterior Estimation with Applications in Neuroscience

Simon Richard White; Yefeng Fan

REVIEW 5 major objections 7 minor 34 references

A Scalable Exponential Random Graph Model: Amortised Hierarchical Sequential Neural Posterior Estimation with Applications in Neuroscience

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that Bayesian inference for multiple-network exponential random graph models can be made scalable by alternating neural posterior estimation for per-network parameters with closed-form variational updates for group-level…

desk verdict A credible scaling demo for hierarchical NPE on MN-ERGMs, but the convergence proof does not match the implemented point-estimate E-step, and the validation lacks code and non-visual baselines. read the letter →

arxiv 2506.04558 v1 pith:VWZIWBIY submitted 2025-06-05 stat.ME stat.AP

Yefeng Fan , Simon Richard White This is my paper

classification stat.MEstat.AP MSC 62F1562P10

keywords neuralposteriorestimationhierarchicalBayesianinferenceexponentialrandomgraphmodelsmultiple-networkERGMsamortisedsequentialvariationalbrainconnectivity

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

This paper proposes AHS-NPE (Amortised Hierarchical Sequential Neural Posterior Estimation), an algorithm that performs Bayesian inference for multiple-network exponential random graph models (MN-ERGMs) at a scale conventional Markov-chain fitting cannot reach. The method factorises the hierarchical posterior into per-network and group-level pieces: a neural conditional density estimator learns each network's local parameters, and closed-form Normal-inverse-Wishart updates refine the group mean and covariance in an expectation-maximisation loop. The authors show the fitted group-level and individual-level posteriors agree with a direct Bayesian fit on 100 resting-state fMRI brain networks, then apply the same procedure to 586 networks split into two age groups. If the central claim is right, MN-ERGM analyses that required millions of auxiliary network simulations become practical on large imaging cohorts, and the pattern generalises to any hierarchical simulator with a tractable upper level.

What carries the argument

The load-bearing mechanism is an expectation-maximisation loop under a mean-field variational approximation $q(\Theta_{\mathrm{all}})=q(\theta_g,\Sigma_g)\prod_i q(\theta_i)$. The E-step trains a conditional normalising-flow density estimator $q_\phi(\theta_i\mid x_i)$ with an SNPE-style ratio correction, allowing the proposal distribution to differ from the hierarchical prior; the M-step plugs the resulting posterior means and covariances into closed-form Normal-inverse-Wishart updates for $\theta_g$ and $\Sigma_g$, then fixes them at their maximum-a-posteriori values for the next E-step. ERGM-specific adjustments—a wide initial proposal, a burn-in period with a deliberately weak density estimator, then a refocused deeper estimator on data re-sampled near the target posterior—are what stabilise the estimator against boundary effects and multi-modality in ERGM summary statistics.

What would settle it

On data simulated from a known MN-ERGM with strong correlation between local parameters and the group mean, run AHS-NPE and a full Markov-chain Exchange-Algorithm fit with identical priors; if the AHS-NPE credible intervals for individual-level parameters and the estimated group covariance systematically fail to track the true generating values while the MCMC fit succeeds, the mean-field approximation is the cause.

Watch

Extended reading notes

Core claim

The central claim is that a hierarchical posterior over many ERGM networks can be approximated accurately by alternating a one-round sequential neural posterior estimator for the per-network parameters with analytical Normal-inverse-Wishart updates for the group-level mean and covariance, using a mixture-of-Normals proposal that reuses simulated data from previous rounds. Because the neural estimator is trained once and then evaluated on any new observed network, per-observation inference cost becomes negligible, so the method keeps the amortisation property that ordinary sequential NPE sacrifices. The paper demonstrates convergence of the group-level estimates to the corresponding Bayesian fit by iteration 9, posterior predictive alignment with observed network statistics, and a 586-network two-group application whose age-related connectivity differences agree with the earlier smaller-sample results.

Load-bearing premise

The approach depends on the assumption that each network's local parameters are independent of the group mean and covariance once everything else is known, and on the practical step of locking the group mean and covariance to point estimates while the neural network learns the local posteriors; if local and global parameters are strongly correlated, the local posterior approximations and the estimated group covariance will be biased.

Editorial extensions

If this is right

Group-level MN-ERGM inference becomes practical for cohorts of hundreds to thousands of networks: the paper's 586-network application used about 330,000 training simulations, versus an estimated at least 5,860,000 auxiliary simulations for conventional Bayesian fitting.
Because the conditional density estimator is amortised, adding a newly observed network to an analysis does not require retraining, removing the main barrier to analysing large imaging datasets.
AHS-NPE reproduces the qualitative conclusion of the earlier 100-network analysis: younger brains show more efficient, small-world-like connectivity, with the larger sample making the age-group differences clearer.
The algorithmic pattern transfers to any hierarchical simulation-based model whose upper-level prior admits conjugate variational updates, so it is not limited to network data.

Reading between the lines

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

A decisive test the paper does not run is to simulate from the generative model with a known group covariance and check the coverage of AHS-NPE's credible intervals; agreement with one Bayesian fit cannot by itself rule out systematic bias in $\Sigma_g$.
The wider individual-level spread and larger group covariance that AHS-NPE reports relative to the centred Bayesian fit are exactly what a mean-field, point-estimate E-step would predict when global parameters carry real posterior uncertainty; carrying that uncertainty into the E-step could remove the discrepancy.
The mixture-of-Normals proposal with a burn-in and refocused sampling can be read as a general recipe for simulation-based inference on weakly identified or multimodal summary statistics, not just ERGMs.
At biobank scale, the amortised property implies the marginal cost of one additional network approaches the cost of a single forward simulation, making group comparisons over thousands of scans computationally routine if estimation accuracy holds under greater heterogeneity.

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

Colleague—

Quick take: AHS-NPE is a credible, useful extension of hierarchical NPE to MN-ERGMs, and the 586-network Cam-CAN application is a genuine scaling demo. But the paper overstates its theoretical grounding: the implemented E-step uses point-estimate globals, so the ELBO monotonicity claim in the supplement does not actually cover what the code does. That is a fixable gap, not a fatal one.

What is actually new: the combination of one-round SNPE per EM iteration with an analytic NIW M-step and a mixture-of-normals proposal that keeps amortization across individuals. The parts are known (HNPE, HIM-LFVI), but this packaging preserves amortization better than HNPE and is simpler than HIM-LFVI. The M-step algebra (Section 7.3.3) is correct, and the method converged sensibly on 100 and 586 networks, with the young/old differences in the expected direction. The paper is honest that it is a variant of HNPE and an extension of HIM-LFVI—no overclaiming of novelty there.

Soft spots, in order of importance. First, the stress-test question: the mean-field ELBO requires q(theta_i) to average over q(theta_g, Sigma_g), but Algorithm 1 replaces q(theta_g, Sigma_g) with a MAP point mass. The Section 7.3.1 proof only covers the fixed-global case. The authors note this in their limitations paragraph but wave it away saying MAP is “adequate”. That is hand-waving. The inflated Sigma_g they see in Figure 4 could be exactly the artifact this produces. They need to either fix the algorithm to propagate global uncertainty (e.g., sample multiple global draws in the E-step) or provide calibration diagnostics (coverage, calibration plots) that show the point-mass approximation is not systematically overconfident. Second, the main comparison is against their own re-implemented “Bayesian fitting” rather than Lehmann et al.’s actual fit, and the Lehmann comparison is visual only. They should report proper numbers. Third, no code or data is shipped. For a methods paper whose value is amortized scalability, that is a gap; the 330,000-simulation claim is uncheckable without code.

Who should read this: anyone fitting ERGMs to network populations, and NPE users with hierarchical models. It deserves serious peer review, but only with a request for code, calibration checks, and a rewritten convergence proof that matches the algorithm actually run.

Referee Report

5 major / 7 minor

Summary. The paper proposes AHS-NPE, a hierarchical sequential neural posterior estimation procedure for multiple-network exponential random graph models (MN-ERGMs). The method alternates an NPE-based E-step for local network-level parameters with an analytically derived conjugate M-step for the global mean and covariance under a normal-inverse-Wishart prior. Several ERGM-specific heuristics are introduced, including wide initial proposal coverage, a burn-in period with a less expressive flow, covariance inflation, and a refined training data replacement. The method is validated by reproducing the young-group Cam-CAN analysis of Lehmann et al. (100 networks) and then scaled to 586 networks across two age groups, with reported age-related differences in brain connectivity. The paper claims both a scalable solution for MN-ERGMs and a general hierarchical NPE framework that preserves amortisation and sequential refinement.

Significance. If the result holds, AHS-NPE would be a valuable scalable alternative to MCMC for MN-ERGMs and a useful addition to the hierarchical simulation-based inference literature. The M-step algebra is standard and appears correct, the local posterior is amortised across individuals, and the real-data application demonstrates a large reduction in simulation cost relative to a conventional exchange-algorithm fit. The paper also gives explicit attention to ERGM-specific pathologies such as boundary effects, multi-modality, and leakage, and it states several limitations honestly in Section 6.1 and the Supplement. However, the manuscript's central statistical claim, that the reported posterior is a valid Bayesian posterior, is not yet established: the E-step conditions on point estimates of the global parameters, and no calibration study is provided. The inflated group-level covariance in Figure 4 is consistent with the artefact expected from that design decision.

major comments (5)

[Section 3, Algorithm 1, Supplement 7.3.1] The implemented E-step is not the mean-field variational Bayes algorithm whose ELBO is written in Section 3. In Algorithm 1, the local posterior q_phi(theta_i|x_i) is trained with the prior pi(theta_i|theta_g^(t),Sigma_g^(t)) using the MAP point estimates from the M-step, whereas the ELBO in Equation (3) requires an expectation over q(theta_g,Sigma_g). The Supplement itself concedes the relevant equality holds only “when keeping the global parameters theta_g and Sigma_g fixed”. Consequently, the ELBO-monotonicity argument does not apply to the implemented algorithm, and the reported q(theta_g,Sigma_g) is not the optimum of the stated objective. The paper should either describe the method as an empirical-Bayes/EM point-estimate procedure for the globals and validate it as such, or propagate global uncertainty in the E-step. This is load-bearing because the paper's central claim is accurate Bayesian posterior inference.
[Supplement 7.3.2] The proof that NPE minimises KL(q* || q_phi) is written only for the case where the proposal equals the prior: it sets p(x_i,theta_i) = p(x_i|theta_i) pi(theta_i|theta_g,Sigma_g). Algorithm 3 deliberately uses a mixture proposal different from the prior, and the atomic-loss correction is cited from [10,20] but not integrated into the proof. Please either extend the derivation to the actual proposal class or state the precise atomic-loss objective and prove that its optimum is the conditional posterior. Without this, the validity of the E-step under the proposed adjustment scheme is not established by the manuscript's own derivation.
[Section 5.1, Figure 4] The observation that AHS-NPE reports roughly double the group-level covariance of the full MCMC fit (for example, the diagonal component for Edges is 0.2056 versus 0.1101) is exactly the artefact expected from a point-estimate E-step: local posteriors trained against point-estimate globals do not propagate global uncertainty, and the M-step then converts their overdispersion into a larger Sigma_g. The two explanations offered in Section 5.1 are not disentangled, and the posterior-predictive comparisons in Figure 2 are checks of predictive means, not calibration. I request a simulation-calibration study: generate synthetic data from known global parameters, report frequentist coverage of 50% and 95% credible intervals for theta_i and Sigma_g, and compare against a version of the algorithm that averages over q(theta_g,Sigma_g) in the E-step.
[Section 4.1, Figure 1] The comparison with Lehmann et al. is explicitly based on “visual inspection of their Figure 2”, and the quantitative baseline “our Bayesian fitting” is the authors' own re-implementation under a different hyper-prior: theta_g is drawn conditionally on Sigma_g, whereas Lehmann et al. use a Normal hyper-prior for theta_g. This is not an independent benchmark. Please provide quantitative agreement measures against the MCMC baseline, such as maximum and root-mean-square differences of posterior means for all 100 networks, and interval coverage, rather than visual inspection of selected contours.
[Section 3.3.2, Algorithm 3] The ERGM-specific adjustment scheme has many free parameters, including N1=100,000, N_t=20,000, T_initial=4, the 5× covariance inflation, the flow architecture change, and the convergence threshold, but no sensitivity analysis is reported. Since the method's performance could plausibly depend on these choices, the claim that the paper provides “well-rounded guidelines” is supported by only a single application. At minimum, vary T_initial and the inflation factor and show that the Cam-CAN conclusions are unchanged, or state explicitly that these are application-specific heuristics rather than general recommendations.

minor comments (7)

[General] There are numerous typographical errors throughout, including “Kullbcak”, “armortisation”, “distrbution”, and “nising constants”; a careful proofread is needed.
[Supplement 7.3.3] The intermediate identification of the data factor as NIW with nu1 = n-d-2 is inconsistent with the stated inverse-Wishart normalisation; using nu1 = n-d-1 and the product rule nu_n = nu0+nu1+d+1 gives the same final update and avoids the off-by-one confusion.
[Section 4, Figure 5] Section 4 states that 100,000 posterior samples are drawn by default, but Figure 5 reports density plots based on 10,000 samples; please reconcile these numbers.
[Algorithm 3] After removing the initial wide component, the retained mixture weights are not explicitly re-normalised; please specify the exact weights after the replacement step, because the proposal density enters the atomic-loss correction.
[Section 4] The convergence criterion monitors only the relative change in theta_g; please also monitor Sigma_g, given that Figure 4 shows a large difference in Sigma_g from the MCMC baseline.
[Section 6] The stated 330,000 simulations are not derived in a transparent way from the algorithm parameters and iteration counts in Section 5.2; please clarify whether this is per age group, for both groups, or the effective count after discarding first-round data.
[General] A code and data availability statement is missing; providing code and processed data would substantially improve the reproducibility of the computational claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AHS-NPE is validated against an independent MCMC baseline and an external published fit; the mean-field/MAP gap is a limitation, not a self-referential derivation.

full rationale

The paper's central derivation is a variational EM scheme: the E-step trains an SNPE conditional density estimator toward the conditional posterior π(θ_i | θ_g, Σ_g, x_i) with the global parameters held at their current updates (Section 3.1, Algorithm 1), and the M-step computes the Normal-inverse-Wishart update from the NPE moments. The Supplement (Section 7.3.2) derives the NPE loss as equivalent to minimizing KL(q* || q_φ) using Bayes' rule and the conditional-posterior identity; no parameter is fitted and then reported as a prediction. The validation in Section 5.1 compares against 'our Bayesian fitting' (an independently run MCMC/exchange-algorithm fit with the same NIW hyper-prior) and against Lehmann et al.'s published fit, and Section 5.2 is an extrapolation to 586 networks whose neurological findings are checked against an external study. The only notable self-citation is [20], the authors' prior NPE-on-ERGM paper, used to motivate ERGM-specific adjustment choices (Section 3.3.2), but the paper's main accuracy claim does not reduce to that citation; the adjustments are presented as heuristics and the authors explicitly disclaim they are 'the definite truth'. The proof of validity has a known gap — the implemented E-step fixes θ_g, Σ_g at MAP point estimates rather than averaging over q(θ_g, Σ_g), and the Supplement itself states q*(θ_i)=π(θ_i|θ_g,Σ_g,x_i) holds only 'when keeping the global parameters fixed' and admits the MAP point estimate 'may also cause bias'. That is an approximation/coverage limitation, not a circular derivation: the target quantities are not defined in terms of the outputs, and the main comparisons are against external baselines.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

The central results depend on the standard ERGM simulation setup, the mean-field variational family, the SNPE ratio trick, and a set of hand-set tuning constants. No new entities are introduced.

free parameters (9)

Initial proposal N(theta|0,10I) = 0,10I
Chosen by hand as a wide initial proposal to cover possible posterior domains; affects coverage and boundary-effect behaviour.
First-round sample size N1 = 100,000
Large initial training dataset chosen to provide wide coverage; not derived from the model.
Per-round training sample size N_t = 20,000
Chosen as default training data size for each AHS-NPE round after round 1.
Refined sample size N_refined = 50,000
Training data size for the refined proposal component after burn-in; chosen by hand.
Burn-in rounds T_initial = 4
The number of early rounds using a less expressive MAF; chosen to avoid overfitting to ERGM boundary effects.
Refined covariance inflation factor = 5
Multiplier applied to the averaged covariance for Sigma_refined to broaden coverage; chosen by hand.
MAF architecture during and after burn-in = 32 units/5 transforms; 64 units/10 transforms
Small flow during burn-in, deeper flow afterwards; architecture choices made by the authors.
Convergence threshold = 0.01
Relative change in global parameters below 0.01 for two rounds defines convergence; an informal stopping rule.
Posterior sample count S = 100,000
Number of samples drawn from the trained flow to compute local moments; chosen as a convention.

assumptions (6)

domain assumption The ERGM likelihood is intractable but can be simulated exactly via MCMC (Algorithm 4).
Throughout; the training pipeline uses simulated networks and the exchange algorithm for the Bayesian baseline.
domain assumption The summary statistics h(y) (edges, GWESP, GWNSP) are Bayesian sufficient for the ERGM, so conditioning on x=h(y) is lossless for posterior inference.
Section 2.1.1; standard exponential-family fact, also relied on in the authors' prior work [20].
standard math The Normal-inverse-Wishart prior is conjugate and the M-step analytics are valid under the mean-field factorisation.
Section 7.3.3; uses standard NIW multiplication and completes the square.
domain assumption The joint posterior is approximated by the mean-field factorisation q(Theta_all)=q(theta_g,Sigma_g)prod q(theta_i).
Section 3; a deliberate variational approximation, acknowledged as approximate.
domain assumption The proposal distribution p_tilde(theta) is exactly the sampling distribution of the accumulated training dataset, making the SNPE ratio-trick correction (Eq. 2) valid.
Section 3.3.1 and Algorithm 2; if weights do not match data counts, the posterior estimate is biased. Ambiguous in Algorithm 3.
domain assumption The atomic loss can be applied so that the ratio-trick normalising constant Z_phi(x) is evaluable when the proposal differs from the prior.
Section 2.2, cited [10,20]; needed for the mixture-of-normals proposal.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Scalable Exponential Random Graph Model: Amortised Hierarchical Sequential Neural Posterior Estimation with Applications in Neuroscience." pith.science (2026). https://pith.science/paper/VWZIWBIY

@misc{pith2026250604558,
  author       = {Pith},
  title        = {Pith review of: A Scalable Exponential Random Graph Model: Amortised Hierarchical Sequential Neural Posterior Estimation with Applications in Neuroscience},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWZIWBIY}},
  note         = {Machine review of arXiv:2506.04558}
}

read the original abstract

Exponential Random Graph Models (ERGMs) are an inferential model for analysing statistical networks. Recent development in ERGMs uses hierarchical Bayesian setup to jointly model a group of networks, which is called a multiple-network Exponential Random Graph Model (MN-ERGMs). MN-ERGM has been successfully applied on real-world resting-state fMRI data from the Cam-CAN project to infer the brain connectivity on aging. However, conventional Bayesian ERGM estimation approach is computationally intensive and lacks implementation scalability due to intractable ERGM likelihood. We address this key limitation by using neural posterior estimation (NPE), which trains a neural network-based conditional density estimator to infer the posterior.\\ We proposed an Amortised Hierarchical Sequential Neural Posterior Estimation (AHS-NPE) and various ERGM-specific adjustment schemes to target the Bayesian hierarchical structure of MN-ERGMs. Our proposed method contributes to the ERGM literature as a very scalable solution, and we used AHS-NPE to re-show the fitting results on the Cam-CAN data application and further scaled it up to a larger implementation sample size. More importantly, our AHS-NPE contributes to the general NPE literature as a new hierarchical NPE approach that preserves the amortisation and sequential refinement, which can be applied to a variety of study fields.

Figures

Figures reproduced from arXiv: 2506.04558 by the authors.

**Figure 2.** Paired plots that compare AHS-NPE with our Bayesian fitting on the individual-level [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

**Figure 3.** Pairwise contour plots for posterior densities on individual-level parameters; the red [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Heat plot styled visualisation of MAP of group-level covariances [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]

**Figure 5.** Figure 5: Density plots of the posterior density estimates for group-level means; the young [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

[1]

Characterising group-level brain connectivity: a framework using Bayesian exponential random graph models

B. Lehmann, R. Henson, L. Geerligs, S. White, et al. “Characterising group-level brain connectivity: a framework using Bayesian exponential random graph models”. In:Neu- roImage225 (2021), p. 117480

work page 2021
[2]

Bayesian inference for exponential random graph models

A. Caimo and N. Friel. “Bayesian inference for exponential random graph models”. In: Social networks33.1 (2011), pp. 41–55

work page 2011
[3]

Bergm: Bayesian exponential random graph models in R

A. Caimo and N. Friel. “Bergm: Bayesian exponential random graph models in R”. In: arXiv preprint arXiv:1703.05144(2017)

work page Pith review arXiv 2017
[4]

Same but different: A comparison of estimation approaches for exponential random graph models for multiple networks

P. Tolochko and H. G. Boomgaarden. “Same but different: A comparison of estimation approaches for exponential random graph models for multiple networks”. In:Social Net- works76 (2024), pp. 1–11

work page 2024
[5]

Lusher, J

D. Lusher, J. Koskinen, and G. Robins.Exponential random graph models for social networks: Theory, methods, and applications. Cambridge University Press, 2013

work page 2013
[6]

Bayesian optimization for likelihood-free cosmological inference

F. Leclercq. “Bayesian optimization for likelihood-free cosmological inference”. In:Phys- ical Review D98.6 (2018), p. 063511

work page 2018
[7]

Cosmoabc: likelihood-free inference via population Monte Carlo ap- proximate Bayesian computation

E. E. Ishida et al. “Cosmoabc: likelihood-free inference via population Monte Carlo ap- proximate Bayesian computation”. In:Astronomy and Computing13 (2015), pp. 1–11

work page 2015
[8]

The frontier of simulation-based inference

K. Cranmer, J. Brehmer, and G. Louppe. “The frontier of simulation-based inference”. In:Proceedings of the National Academy of Sciences117.48 (May 2020), 30055–30062

work page 2020

Show all 34 references

[9]

Fastε-free inference of simulation models with bayesian conditional density estimation

G. Papamakarios and I. Murray. “Fastε-free inference of simulation models with bayesian conditional density estimation”. In:Advances in neural information processing systems 29 (2016). 26

2016
[10]

Automatic posterior transformation for likelihood-free inference

D. Greenberg, M. Nonnenmacher, and J. Macke. “Automatic posterior transformation for likelihood-free inference”. In:International Conference on Machine Learning. PMLR. 2019, pp. 2404–2414

2019
[11]

Real-time gravitational wave science with neural posterior estimation

M. Dax, S. R. Green, J. Gair, J. H. Macke, A. Buonanno, and B. Schölkopf. “Real-time gravitational wave science with neural posterior estimation”. In:Physical review letters 127.24 (2021), p. 241103

2021
[12]

Sequential neural likelihood: Fast likelihood- free inference with autoregressive flows

G. Papamakarios, D. Sterratt, and I. Murray. “Sequential neural likelihood: Fast likelihood- free inference with autoregressive flows”. In:The 22nd international conference on arti- ficial intelligence and statistics. PMLR. 2019, pp. 837–848

2019
[13]

ABC of the future

H. Pesonen et al. “ABC of the future”. In:International Statistical Review91.2 (2023), pp. 243–268

2023
[14]

Neural networks enable efficient and accurate simulation-based inference of evolutionary parameters from adaptation dynamics

G. Avecilla, J. N. Chuong, F. Li, G. Sherlock, D. Gresham, and Y . Ram. “Neural networks enable efficient and accurate simulation-based inference of evolutionary parameters from adaptation dynamics”. In:PLoS biology20.5 (2022), e3001633

2022
[15]

Likelihood approximation networks (LANs) for fast inference of simulation models in cognitive neuroscience

A. Fengler, L. N. Govindarajan, T. Chen, and M. J. Frank. “Likelihood approximation networks (LANs) for fast inference of simulation models in cognitive neuroscience”. In: Elife10 (2021), e65074

2021
[16]

Neural posterior estimation for exoplanetary atmospheric retrieval

M. Vasist, F. Rozet, O. Absil, P. Mollière, E. Nasedkin, and G. Louppe. “Neural posterior estimation for exoplanetary atmospheric retrieval”. In:Astronomy & Astrophysics672 (2023), A147

2023
[17]

Training deep neural density estimators to identify mechanistic models of neural dynamics

P. J. Gonçalves et al. “Training deep neural density estimators to identify mechanistic models of neural dynamics”. In:Elife9 (2020), e56261

2020
[18]

Robust neural posterior estimation and statistical model criticism

D. Ward, P. Cannon, M. Beaumont, M. Fasiolo, and S. Schmon. “Robust neural posterior estimation and statistical model criticism”. In:Advances in Neural Information Process- ing Systems35 (2022), pp. 33845–33859

2022
[19]

sbi: A toolkit for simulation-based inference

A. Tejero-Cantero et al. “sbi: A toolkit for simulation-based inference”. In:Journal of Open Source Software5.52 (2020), p. 2505

2020
[20]

Neural Posterior Estimation on Exponential Random Graph Mod- els: Evaluating Bias and Implementation Challenges

Y . Fan and S. White. “Neural Posterior Estimation on Exponential Random Graph Mod- els: Evaluating Bias and Implementation Challenges”. In:arXiv preprint arXiv:2504.09349 (2025)

2025 arXiv
[21]

An analytically tractable solution for hierarchical Bayesian model updating with variational inference scheme

X. Jia, W.-J. Yan, C. Papadimitriou, and K.-V . Yuen. “An analytically tractable solution for hierarchical Bayesian model updating with variational inference scheme”. In:Me- chanical Systems and Signal Processing189 (2023), p. 110060

2023
[22]

Amortized Bayesian Multilevel Models

D. Habermann, M. Schmitt, L. Kühmichel, A. Bulling, S. T. Radev, and P.-C. Bürkner. “Amortized Bayesian Multilevel Models”. In:arXiv preprint arXiv:2408.13230(2024)

2024 arXiv
[23]

Hierarchical Neural Simulation-Based Inference Over Event Ensembles

L. Heinrich, S. Mishra-Sharma, C. Pollard, and P. Windischhofer. “Hierarchical Neural Simulation-Based Inference Over Event Ensembles”. In:arXiv preprint arXiv:2306.12584 (2023)

2023 arXiv
[24]

HNPE: Leveraging global pa- rameters for neural posterior estimation

P. Rodrigues, T. Moreau, G. Louppe, and A. Gramfort. “HNPE: Leveraging global pa- rameters for neural posterior estimation”. In:Advances in Neural Information Processing Systems34 (2021), pp. 13432–13443

2021
[25]

Hierarchical implicit models and likelihood-free variational inference

D. Tran, R. Ranganath, and D. Blei. “Hierarchical implicit models and likelihood-free variational inference”. In:Advances in Neural Information Processing Systems30 (2017)

2017
[26]

The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) study protocol: a cross-sectional, lifespan, multidisciplinary examination of healthy cog- nitive ageing

M. A. Shafto et al. “The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) study protocol: a cross-sectional, lifespan, multidisciplinary examination of healthy cog- nitive ageing”. In:BMC Neurology14.1 (2014), p. 204

2014
[27]

J. K. Harris.An introduction to exponential random graph modeling. V ol. 173. Sage Pub- lications, 2013

2013
[28]

Masked autoregressive flow for density estimation

G. Papamakarios, T. Pavlakou, and I. Murray. “Masked autoregressive flow for density estimation”. In:Advances in neural information processing systems30 (2017). 27

2017
[29]

A theory of mean field approximation

T. Tanaka. “A theory of mean field approximation”. In:Advances in Neural Information Processing Systems11 (1998)

1998
[30]

A refined mean field approximation

N. Gast and B. Van Houdt. “A refined mean field approximation”. In:Proceedings of the ACM on Measurement and Analysis of Computing Systems1.2 (2017), pp. 1–28

2017
[31]

Truncated proposals for scalable and hassle-free simulation-based inference

M. Deistler, P. J. Goncalves, and J. H. Macke. “Truncated proposals for scalable and hassle-free simulation-based inference”. In:Advances in Neural Information Processing Systems35 (2022), pp. 23135–23149

2022
[32]

Mahalanobis distance

G. J. McLachlan. “Mahalanobis distance”. In:Resonance4.6 (1999), pp. 20–26

1999
[33]

The influence of ageing on complex brain networks: a graph theoretical analysis

S. Micheloyannis, M. V ourkas, V . Tsirka, E. Karakonstantaki, K. Kanatsouli, and C. J. Stam. “The influence of ageing on complex brain networks: a graph theoretical analysis”. In:Human brain mapping30.1 (2009), pp. 200–208

2009
[34]

Functional connectivity patterns of human magnetoencephalographic record- ings: a ‘small-world’network?

C. J. Stam. “Functional connectivity patterns of human magnetoencephalographic record- ings: a ‘small-world’network?” In:Neuroscience letters355.1-2 (2004), pp. 25–28. 28

2004

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Characterising group-level brain connectivity: a framework using Bayesian exponential random graph models

B. Lehmann, R. Henson, L. Geerligs, S. White, et al. “Characterising group-level brain connectivity: a framework using Bayesian exponential random graph models”. In:Neu- roImage225 (2021), p. 117480

work page 2021

[2] [2]

Bayesian inference for exponential random graph models

A. Caimo and N. Friel. “Bayesian inference for exponential random graph models”. In: Social networks33.1 (2011), pp. 41–55

work page 2011

[3] [3]

Bergm: Bayesian exponential random graph models in R

A. Caimo and N. Friel. “Bergm: Bayesian exponential random graph models in R”. In: arXiv preprint arXiv:1703.05144(2017)

work page Pith review arXiv 2017

[4] [4]

Same but different: A comparison of estimation approaches for exponential random graph models for multiple networks

P. Tolochko and H. G. Boomgaarden. “Same but different: A comparison of estimation approaches for exponential random graph models for multiple networks”. In:Social Net- works76 (2024), pp. 1–11

work page 2024

[5] [5]

Lusher, J

D. Lusher, J. Koskinen, and G. Robins.Exponential random graph models for social networks: Theory, methods, and applications. Cambridge University Press, 2013

work page 2013

[6] [6]

Bayesian optimization for likelihood-free cosmological inference

F. Leclercq. “Bayesian optimization for likelihood-free cosmological inference”. In:Phys- ical Review D98.6 (2018), p. 063511

work page 2018

[7] [7]

Cosmoabc: likelihood-free inference via population Monte Carlo ap- proximate Bayesian computation

E. E. Ishida et al. “Cosmoabc: likelihood-free inference via population Monte Carlo ap- proximate Bayesian computation”. In:Astronomy and Computing13 (2015), pp. 1–11

work page 2015

[8] [8]

The frontier of simulation-based inference

K. Cranmer, J. Brehmer, and G. Louppe. “The frontier of simulation-based inference”. In:Proceedings of the National Academy of Sciences117.48 (May 2020), 30055–30062

work page 2020