Amortizing personalization in virtual brain twins

Marmaduke M Woodman; Nina Baldy; Viktor K Jirsa

REVIEW 3 major objections 5 minor 20 references

Amortizing personalization in virtual brain twins

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Cohort-level simulation-based inference, with the connectome treated as a latent parameter under a cohort-averaged prior, matches per-subject inference (89% vs 95% true-value coverage) while cutting simulation cost from O(NM) to O(M).

desk verdict The amortized inference trick looks real; the 'anonymized' claim is where it falls down. read the letter →

arxiv 2506.21155 v1 pith:RPNHTYQH submitted 2025-06-26 q-bio.NC

Nina Baldy , Marmaduke M Woodman , Viktor K Jirsa This is my paper · ORCID

classification q-bio.NC

keywords virtualbraintwinsamortizedinferencesimulation-basedconnectomepersonalizedmedicineBayesianneuroimagingdataprivacy

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

The paper proposes "anonymized personalization" for virtual brain twins: instead of conditioning each individual model on that person's connectome, the connectome is treated as an inferred parameter drawn from a cohort-level prior over a low-dimensional latent code. A single amortized neural posterior estimator is trained once on simulations from that prior, and can then evaluate a personalized posterior over parameters such as coupling and noise for any new subject in near-real-time. On a test set of 230 subjects, true parameter values fell in the 90% credible interval for 89% of subjects with cohort-level inference versus 95% with subject-level inference, and the two inference methods showed highly correlated calibration scores. If this holds, the expensive, privacy-sensitive training step can be done once on shared infrastructure without touching personal data, while clinical inference becomes lightweight.

What carries the argument

The machinery has two parts: a cross-coder, a linear autoencoder trained across 20 parcellations to map each subject's connectome into a shared 16-dimensional latent code, whose mean and covariance over training subjects define a multivariate normal cohort-based prior; and amortized neural posterior estimation, a normalizing flow that maps data features to posterior distributions over the combined parameter space (latent connectome code, coupling scaling k, noise D). The cross-coder's role is to make the personal connectome a low-dimensional, samplable parameter; the amortized estimator's role is to train once from cohort-prior simulations and then evaluate a personalized posterior for any subject with a single forward pass.

What would settle it

Run a membership-inference attack: train the cross-coder and cohort prior as in the paper, then ask whether a given held-out subject's latent code reveals membership in the training cohort; if accuracy exceeds chance, anonymity is falsified. Separately, measure cohort-SBI true-value coverage among subjects whose latent codes are in the outer decile of the prior; if coverage falls well below 0.89, the approximation is not uniform across the population.

Watch

Extended reading notes

Core claim

Cohort-level simulation-based inference provides a useful approximation of subject-level SBI: absorbing the personal connectome into the parameter space, with a multivariate normal prior estimated from the latent codes of a training cohort, yields posteriors over coupling scaling k and noise D that contract toward true values nearly as well as per-subject inference. The amortized posterior is trained once from O(M) simulations and evaluated for any subject, reducing computational complexity from O(NM) to O(M), and the latent connectome parameters themselves are partially identifiable. The authors interpret the small coverage gap (89% vs 95%) as evidence that cohort-level inference is a strong compromise, not a replacement for costly MCMC workflows.

Load-bearing premise

The load-bearing premise is that a multivariate normal prior built by averaging latent connectome codes over training subjects is no longer personal data; if that average can be re-identified or reverse-engineered to individual connectomes, the anonymous-training claim fails even though the statistical approximation may hold.

Editorial extensions

If this is right

A trained cohort model can evaluate a personalized posterior for a new patient in near-real-time, enabling bedside or portable inference.
Training needs only O(M) simulations total rather than O(NM) for N subjects, so scaling to thousands of patients becomes feasible.
Clinical studies without diffusion imaging for a subject can still build a VBT by drawing the connectome from the cohort prior.
Amortized inference enables real-time model-based neurofeedback, where a subject's brain state must be inferred and fed back with minimal delay.
The latent connectome's weak identifiability can serve as a dynamical-sensitivity diagnostic: strong shrinkage of the connectome posterior flags parameter regimes where individual structure matters little.

Reading between the lines

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

If the cohort prior's anonymity is challenged, a membership-inference attack on the latent space could settle it: success beyond chance would invalidate the 'anonymized personalization' name without changing the statistical approximation result.
The 89%-vs-95% coverage gap may be larger for subjects whose latent codes lie far from the cohort mean; calibrating an atypicality score and reporting per-subject coverage would make the method safer for clinical outliers.
The same design -- a low-rank linear code of a high-dimensional personal input plus a population prior and one amortized estimator -- generalizes to other medical personalization settings, such as genomic or electronic-health-record features.
Because the mean-field model is not always sensitive to connectome details, the framework's practical value depends on the target dynamical regime; testing on resting-state and epilepsy models, as the authors propose, will reveal where personalization actually changes predictions.

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

The thing to know: the statistical core of this paper is plausible and worth engaging with, but the privacy half of the 'anonymized personalization' claim is not supported and, as stated, probably false. The reader's conditional verdict is about right.

What's genuinely new: treating the connectome as a latent variable with a cohort-derived Gaussian prior, rather than conditioning on a fixed connectome as in Hashemi et al. 2023, and the multi-parcellation cross-coder that maps across 20 parcellations into one latent space. The in silico comparison on 230 test subjects is clean: cohort-level SBI gets 89% coverage on the 90% CI versus 95% for subject-level SBI, with correlation 0.8 on calibration metrics. That is a reasonable demonstration that the amortized approximation works in this toy setting, and the authors are honest that cohort posteriors contract less. Code is available, which earns credit.

The soft spot is the privacy claim. Section 2.3 asserts without analysis that the MVN prior over latent codes 'is no longer personal data' because it averages over many subjects. But the cross-coder decoder is trained directly on personal connectomes and must be shared with the training infrastructure, so it is a model of personal data, not an anonymous aggregate. The reported 2.3% confusion rate cuts the wrong way: it means the latent codes preserve subject identity well enough that a membership-inference attack is plausible. No DP bound, no audit, no formal argument. This matters because the whole pitch is that you can train on shared HPC without personal data. That pitch does not survive contact with the method as written.

Two smaller issues: no population-average connectome baseline (which would test whether the cohort prior is actually needed), and the evaluation uses only one parcellation (79-node Shen) and one summary statistic (mean firing rate). Both are minor given the proof-of-concept nature, but they limit how far the headline claim generalizes.

Who is this for? Computational neuroscientists working on amortized inference for brain models. They will get a useful, concrete idea and a clear warning about privacy. It deserves peer review, but the referee should require a real privacy analysis (even a basic membership-inference test) and the missing baseline before acceptance. The statistical part can likely be salvaged; the privacy part needs to be redone or scoped down.

Referee Report

3 major / 5 minor

Summary. The paper introduces 'anonymized personalization' for virtual brain twins (VBTs). Instead of conditioning inference on a subject's measured structural connectome, the authors treat a low-dimensional latent representation of the connectome as an inferred parameter. A linear 'cross-coder' maps connectomes from 20 parcellations into a shared 16-dimensional latent space, and a multivariate normal prior over the latent codes is estimated from training subjects. A single amortized neural posterior estimator (NPE) is then trained on simulations that sample both the connectome latent codes and the biophysical parameters of interest (coupling scaling k and noise scaling D). At test time, the resulting amortized posterior can be applied to any new subject. The authors compare cohort-level SBI with subject-level SBI on 230 test subjects using the 79-node Shen 2013 parcellation, reporting 90% credible-interval coverage of 0.89 versus 0.95, similar shrinkage/z-score distributions, and a complexity reduction from O(NM) to O(M). The paper argues this enables training on shared infrastructure without personal data while keeping inference lightweight and personalized.

Significance. If the approximation holds, the amortized cohort-level approach is a useful step toward scalable and practical VBT inference, especially for settings where subject-specific connectomes are missing or where per-subject MCMC/SBI is too expensive. The use of a standard NPE pipeline, model-based calibration metrics, a held-out test set of 230 subjects, and publicly available data and code are strengths. The computational complexity reduction is real in the amortized setting, provided the cohort-level posterior requires no more simulations per subject to reach acceptable accuracy. However, the paper's second central claim, that training can be performed anonymously because the cohort prior averages away personal data, is not established and as stated is likely incorrect. The empirical evaluation is also narrow: one parcellation and one summary statistic, without uncertainty on the reported metrics. These issues materially affect the advertised contribution, though they do not undermine the core statistical experiment taken on its own.

major comments (3)

[Section 2.3] The statement that the cohort MVN prior 'averages over many subjects, it is no longer personal data subjected to privacy concerns' is an unsupported assertion and is likely not correct as stated. The MVN parameters are estimated from latent projections of personal connectomes, and the cross-coder decoder, which is trained directly on 1KB and HCP connectomes, must be released to the shared infrastructure in order to simulate connectomes from the prior. Sharing the decoder therefore means sharing a model fitted to personal data, and a linear autoencoder can memorize training points up to its latent dimensionality; no differential privacy bound, membership-inference audit, or formal anonymization argument is provided. Because the abstract and conclusion claim that training can be performed 'anonymously,' this is a load-bearing gap: either provide formal privacy guarantees (e.g., differentially private training of the cross-coder and differentially private estimation of the prior) or substantially weaken the privacy claims to 'no direct access to personal data during inference' while acknowledging that the shared artifacts are derived from personal data.
[Section 3.2 / Fig. 4] The central empirical claim that cohort-level SBI 'provides a useful approximation of subject-level SBI' is demonstrated for only one parcellation (79-node Shen 2013) and one summary statistic (mean steady-state firing rate), and the reported calibration metrics (coverage 0.89 vs 0.95, correlation 0.8) are point estimates without confidence intervals or significance tests. The coverage difference of 0.06 over N=230 has a standard error of roughly 0.025, so the result is suggestive but should be reported with intervals. More importantly, the abstract's 'demonstrate reliability' and the discussion's clinical generalizations go beyond what a single parcellation and feature can support. Please either add evaluations across additional parcellations and data features or explicitly restrict the reliability claim to the tested setting.
[Section 3.3 / Section 4.1] The paper acknowledges that cohort-level SBI does not strongly identify the latent connectome parameters and then interprets this as 'model degeneracy' rather than a limitation of the inference procedure. This interpretation is plausible but not established; the claim that the model dynamics are insensitive to fine-grained connectome details is testable, for example by comparing predictions made with the subject-specific connectome against predictions made with the cohort-mean connectome. Without such a test, the discussion in Section 4.1 goes beyond what the experiments demonstrate. I consider this a load-bearing point because the paper's proposed use case includes inference when a subject's connectome is unavailable, and the validity of that use case depends on whether the cohort prior adequately substitutes for the missing personal connectome.

minor comments (5)

[Section 2.2] The preprocessing list has a duplicated label: '(a) clipping... (a) normalizing... and (c)' should read (a), (b), (c).
[Section 2.3, Eq. (3)] Please clarify the vector orientations and dimensions in Eq. (3); as written, the transposes make it hard to verify that the loss is a scalar and that the decoder maps the latent vector back to the upper-triangle vector of each parcellation.
[Section 3.1] The 'confusion rate' should be defined more precisely: specify the dataset split, the number of subjects and parcellations included, and report the training-set confusion rate as well so that the reader can assess overfitting.
[Section 3.2 / Fig. 4] The sentence 'the metrics reveal a very similar (correlation 0.8) distribution' does not say which two quantities are correlated; state explicitly which metric is correlated across subjects and provide a confidence interval for that correlation.
[Section 4.2] The phrase 'without requiring access to sensitive personal data' is too strong given the training of the cross-coder and the prior on personal connectomes; please rephrase to indicate that personal data are not needed during inference and that the training artifacts themselves may still require data-sharing agreements.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: cohort-level SBI is benchmarked against held-out subject-level simulations, and no fitted input is relabeled as a prediction.

full rationale

The central claim that cohort-level amortized SBI provides a useful approximation of subject-level SBI is evaluated empirically on a held-out test set, not built into the prior. The cohort prior is a multivariate normal over latent connectome codes estimated from a training split, and the comparison with per-subject SBI uses independent calibration metrics (shrinkage, z-score, 90% credible-interval coverage: 0.89 vs 0.95). No target posterior is used to set the prior, and no estimated parameter is renamed as a prediction. Self-citations to the authors' earlier VBT/SBI work are background methodology or software, not load-bearing uniqueness claims. The Section 2.3 assertion that the cohort MVN prior 'is no longer personal data subjected to privacy concerns' and that the linear cross-coder has an 'inability to learn training data' are unsupported privacy claims, but those are external validity or security concerns, not circularity in the derivation. No equation-to-equation reduction or fitted-input-as-prediction step can be exhibited, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on standard SBI theory and the Montbrió mean-field model for whole-brain dynamics, both taken from cited prior work. The empirical MVN prior is fitted to the training half of the connectome data, and several hyperparameters, such as latent dimension and split ratio, are chosen without sensitivity analysis. The most fragile addition is the privacy axiom asserting that averaging removes personal data, which is stated without proof. No new physical entities are introduced.

free parameters (3)

Latent dimension of cross-coder = 16
Chosen without sensitivity analysis; it controls the capacity to preserve personalization and the scale of the cohort prior.
Train-test split ratio = 0.5
Chosen for cross-coder training; the MVN prior is computed on the training half, so the split influences the prior and the evaluation.
MVN prior mean and covariance = not reported
Estimated from training subjects' latent codes; these parameters directly shape cohort-level posteriors but their values are not disclosed.

assumptions (4)

domain assumption Montbrió et al. mean-field reduction (Eq. 1-2) provides an adequate model of whole-brain dynamics.
The simulator used for SBI is taken from reference [14]; the paper does not validate it against empirical EEG or fMRI data.
standard math Amortized NPE yields calibrated posteriors when the simulator is correct.
This is the standard theory of neural posterior estimation from references [15] and [18], on which the posterior evaluation rests.
ad hoc to paper Averaging latent connectome codes over subjects removes personal data.
The statement in Section 2.3 is asserted without privacy analysis; this is load-bearing for the anonymous training claim.
domain assumption The mean steady-state firing rate is sufficient to identify coupling and noise parameters.
The paper uses only this summary statistic in the SBI and does not examine posterior identifiability across other features.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Amortizing personalization in virtual brain twins." pith.science (2026). https://pith.science/paper/RPNHTYQH

@misc{pith2026250621155,
  author       = {Pith},
  title        = {Pith review of: Amortizing personalization in virtual brain twins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RPNHTYQH}},
  note         = {Machine review of arXiv:2506.21155}
}

read the original abstract

Virtual brain twins are personalized digital models of individual human subject or patient's brains, allowing for mechanistic interpretation of neuroimaging data features. Training and inference with these models however presents a pair of challenges: large shared infrastructure do not allow for use of personal data and inference in clinical applications should not require significant resources. We introduce "anonymized personalization" to address both by expanding model priors to include personalization which under amortized inference allows training to be performed anonymously, while inference is both personalized and lightweight. We illustrate the basic approach, demonstrate reliability in an example, and discuss the impact on both experimental and computational neuroscience. Code is available at https://github.com/ins-amu/apvbt.

Figures

Figures reproduced from arXiv: 2506.21155 by the authors.

**Figure 2.** Cross-coder architecture encodes connectomes. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Example of subject vs cohort-level inference. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Cohort-level SBI approximates subject-level SBI. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Cohort-level SBI partially identifies latent connectome parameters. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 15 canonical work pages

[1]

arXiv preprint arXiv:1803.08393 (2018)

Betancourt, M.: Calibrating model-based inferences and decisions. arXiv preprint arXiv:1803.08393 (2018)

arXiv 2018
[2]

Nature Re- views Neuroscience13, 336–349 (2012)

Bullmore, E., Sporns, o.: The economy of brain network organization. Nature Re- views Neuroscience13, 336–349 (2012). https://doi.org/10.1038/nrn3214

doi:10.1038/nrn3214 2012
[3]

Frontiers in aging neuroscience6, 149 (2014)

Caspers, S., Moebus, S., Lux, S., Pundt, N., Schütz, H., Mühleisen, T.W., Gras, V., Eickhoff, S.B., Romanzetti, S., Stöcker, T., et al.: Studying variability in hu- man brain aging in a population-based german cohort—rationale and design of 1000brains. Frontiers in aging neuroscience6, 149 (2014)

work page 2014
[4]

Proceedings of the National Academy of Sciences117(48), 30055–30062 (2020)

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

work page 2020
[5]

Scientific reports 14(1), 31970 (2024)

Fousek, J., Rabuffo, G., Gudibanda, K., Sheheitli, H., Petkoski, S., Jirsa, V.: Symmetry breaking organizes the brain’s resting state manifold. Scientific reports 14(1), 31970 (2024)

work page 2024
[6]

IEEE Reviews in Biomedical Engineering pp

Hashemi, M., Depannemaecker, D., Saggio, M., Triebkorn, P., Rabuffo, G., Fousek, J., Ziaeemehr, A., Sip, V., Athanasiadis, A., Breyton, M., Woodman, M., Wang, H., Petkoski, S., Sorrentino, P., Jirsa, V.: Principles and operation of virtual brain twins. IEEE Reviews in Biomedical Engineering pp. 1–25 (2025). https://doi.org/ 10.1109/RBME.2025.3562951

arXiv 2025
[7]

Neural Networks163, 178–194 (2023)

Hashemi, M., Vattikonda, A.N., Jha, J., Sip, V., Woodman, M.M., Bartolomei, F., Jirsa, V.K.: Amortized bayesian inference on generative dynamical network models of epilepsy using deep neural density estimators. Neural Networks163, 178–194 (2023)

work page 2023
[8]

Artificial Intelligence40(1), 185– 234 (1989)

Hinton, G.E.: Connectionist learning procedures. Artificial Intelligence40(1), 185– 234 (1989). https://doi.org/https://doi.org/10.1016/0004-3702(89)90049-0

work page doi:10.1016/0004-3702(89)90049-0 1989

Show all 20 references

[9]

JAX-authors: Jax: Composable transformations of python+numpy programs (2025), https://github.com/google/jax, accessed: 2025-02-06

2025
[10]

Journal of Physics: Complexity3(1), 015007 (2022)

Jirsa, V., Sheheitli, H.: Entropy, free energy, symmetry and dynamics in the brain. Journal of Physics: Complexity3(1), 015007 (2022)

2022
[11]

arXiv preprint arXiv:1412.6980 (2014)

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[12]

Baldy et al

Kyesam, J., Eickhoff, S.B., Popovych, O.V.: Parcellation-based structural and resting-state functional whole-brain connectomes of 1000brains cohort (v1.1) (2022) 12 N. Baldy et al

2022
[13]

NeuroImage283, 120403 (2023)

Lavanga, M., Stumme, J., Yalcinkaya, B.H., Fousek, J., Jockwitz, C., Sheheitli, H., Bittner, N., Hashemi, M., Petkoski, S., Caspers, S., et al.: The virtual ag- ing brain: Causal inference supports interhemispheric dedifferentiation in healthy aging. NeuroImage283, 120403 (2023)

2023
[14]

Physical Review X5(2), 021028 (2015)

Montbrió, E., Pazó, D., Roxin, A.: Macroscopic description for networks of spiking neurons. Physical Review X5(2), 021028 (2015)

2015
[15]

Advances in neural information processing systems 29 (2016)

Papamakarios, G., Murray, I.: Fast ε-free inference of simulation models with bayesian conditional density estimation. Advances in neural information processing systems 29 (2016)

2016
[16]

Frontiers in neuroinformatics7, 10 (2013)

Sanz Leon, P., Knock, S.A., Woodman, M.M., Domide, L., Mersmann, J., McIn- tosh, A.R., Jirsa, V.: The virtual brain: a simulator of primate brain network dynamics. Frontiers in neuroinformatics7, 10 (2013)

2013
[17]

NeuroImage251, 118973 (2022)

Schirner, M., Domide, L., Perdikis, D., Triebkorn, P., Stefanovski, L., Pai, R., Prodan, P., Valean, B., Palmer, J., Langford, C., et al.: Brain simulation as a cloud service: The virtual brain on ebrains. NeuroImage251, 118973 (2022)

2022
[18]

Journal of Open Source Software 5(52), 2505 (2020)

Tejero-Cantero, A., Boelts, J., Deistler, M., Lueckmann, J.M., Durkan, C., Gonçalves, P.J., Greenberg, D.S., Macke, J.H.: sbi: A toolkit for simulation- based inference. Journal of Open Source Software 5(52), 2505 (2020). https: //doi.org/10.21105/joss.02505, https://doi.org/1...

2020 doi
[19]

Neuroimage80, 62–79 (2013)

Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an overview. Neuroimage80, 62–79 (2013)

2013
[20]

Science Translational Medicine15(680), eabp8982 (2023)

Wang, H.E., Woodman, M., Triebkorn, P., Lemarechal, J.D., Jha, J., Dollomaja, B., Vattikonda, A.N., Sip, V., Medina Villalon, S., Hashemi, M., et al.: Delineating epileptogenic networks using brain imaging data and personalized modeling in drug-resistant epilepsy. Science Tran...

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

arXiv preprint arXiv:1803.08393 (2018)

Betancourt, M.: Calibrating model-based inferences and decisions. arXiv preprint arXiv:1803.08393 (2018)

arXiv 2018

[2] [2]

Nature Re- views Neuroscience13, 336–349 (2012)

Bullmore, E., Sporns, o.: The economy of brain network organization. Nature Re- views Neuroscience13, 336–349 (2012). https://doi.org/10.1038/nrn3214

doi:10.1038/nrn3214 2012

[3] [3]

Frontiers in aging neuroscience6, 149 (2014)

Caspers, S., Moebus, S., Lux, S., Pundt, N., Schütz, H., Mühleisen, T.W., Gras, V., Eickhoff, S.B., Romanzetti, S., Stöcker, T., et al.: Studying variability in hu- man brain aging in a population-based german cohort—rationale and design of 1000brains. Frontiers in aging neuroscience6, 149 (2014)

work page 2014

[4] [4]

Proceedings of the National Academy of Sciences117(48), 30055–30062 (2020)

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

work page 2020

[5] [5]

Scientific reports 14(1), 31970 (2024)

Fousek, J., Rabuffo, G., Gudibanda, K., Sheheitli, H., Petkoski, S., Jirsa, V.: Symmetry breaking organizes the brain’s resting state manifold. Scientific reports 14(1), 31970 (2024)

work page 2024

[6] [6]

IEEE Reviews in Biomedical Engineering pp

Hashemi, M., Depannemaecker, D., Saggio, M., Triebkorn, P., Rabuffo, G., Fousek, J., Ziaeemehr, A., Sip, V., Athanasiadis, A., Breyton, M., Woodman, M., Wang, H., Petkoski, S., Sorrentino, P., Jirsa, V.: Principles and operation of virtual brain twins. IEEE Reviews in Biomedical Engineering pp. 1–25 (2025). https://doi.org/ 10.1109/RBME.2025.3562951

arXiv 2025

[7] [7]

Neural Networks163, 178–194 (2023)

Hashemi, M., Vattikonda, A.N., Jha, J., Sip, V., Woodman, M.M., Bartolomei, F., Jirsa, V.K.: Amortized bayesian inference on generative dynamical network models of epilepsy using deep neural density estimators. Neural Networks163, 178–194 (2023)

work page 2023

[8] [8]

Artificial Intelligence40(1), 185– 234 (1989)

Hinton, G.E.: Connectionist learning procedures. Artificial Intelligence40(1), 185– 234 (1989). https://doi.org/https://doi.org/10.1016/0004-3702(89)90049-0

work page doi:10.1016/0004-3702(89)90049-0 1989