Pith. sign in

REVIEW 3 major objections 5 minor 68 references

A Foundational Brain Dynamics Model via Stochastic Optimal Control

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

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

Pith's one-line read The paper claims that casting fMRI brain dynamics as a stochastic optimal control problem yields a foundation model that is state of the art on downstream brain-behavior prediction while using far fewer parameters than prior brain…

desk verdict Promising fMRI foundation model with a real gap between Theorem 4.2 and the implementation; the empirical claims are broader than the evidence shown. read the letter →

arxiv 2502.04892 v1 pith:2ULD47MB submitted 2025-02-07 cs.LG q-bio.NCstat.ML

Joonhyeong Park , Byoungwoo Park , Chang-Bae Bang , Jungwon Choi , Hyungjin Chung , Byung-Hoon Kim , Juho Lee This is my paper · ORCID
classification cs.LGq-bio.NCstat.ML
keywords braindynamicsfoundationmodelstochasticoptimalcontrolstatespacefMRIself-supervisedlearningamortizedinferencemaskedautoencoderlinearprobingtransferUKBiobankpretraining
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 sets out to show that resting-state fMRI brain dynamics are better modeled as a stochastic optimal control problem than by purely data-driven self-supervised learning. Its model, BDO, learns a control policy that steers a latent stochastic differential equation to reproduce masked brain signals, and the time-average of the control vectors becomes a transferable representation used for downstream prediction. Pre-trained on 41,072 UK Biobank scans, BDO reports state-of-the-art results on age and sex prediction, trait regression, and psychiatric diagnosis across five external datasets, while matching baselines at a fraction of the parameter count. The payoff, if true, is that a state-space backbone carrying a temporal inductive bias can do more with less data and compute than large masked-prediction transformers, and that the learned control signals carry clinically meaningful structure.

What carries the argument

Three mechanisms carry the argument. First, Proposition 4.1 ties stochastic optimal control to variational inference: minimizing $J(\alpha) = \mathbb{E}[\int_0^T \tfrac{1}{2}\|\alpha_t\|^2\,dt - \sum_t \log g(y_t|X_t^\alpha)]$ is maximizing an ELBO, so the optimal control $\alpha^\star$ embodies the posterior. Second, Theorem 4.2 makes the SDE simulation-free: with drift $-D_t X_t + \alpha_t$ and $D_t = V\Lambda_t V^\top$ for a fixed eigenbasis $V$, the marginal law of $X_t$ is Gaussian with mean and covariance given by the closed-form recursions (12)-(13), computable in $O(\log k)$ time via a parallel scan over the observation intervals. Third, the rescaled training objective (83) blends MAE-style reconstruction $\|y_t - D_\psi(z_t)\|^2$ with a JEPA-style regularization $\tau\|z_t - T_{\bar\theta}(t, Y_{\text{tar}})\|^2$ that pulls context-predicted latent states toward a slowly updated target encoder, and the downstream representation is the mean pooling $A = \frac{1}{|\mathcal{T}|}\sum_t \alpha_t$.

What would settle it

Run a Monte Carlo Euler-Maruyama simulation of the actual fitted SDE (11) and compare the sampled mean and covariance to the closed-form recursions (12)-(13): if they diverge, the learned drift matrices do not share the common eigenbasis assumed by Theorem 4.2 and the simulation-free inference is inexact. A second direct check is a full rerun of BrainLM and BrainJEPA under BDO's exact preprocessing and data splits, since the paper reports its BrainJEPA reproduction deviated on undocumented preprocessing steps.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a continuous-discrete state-space model whose posterior is approximated through stochastic optimal control, rather than through Bayesian filtering recursion, can serve as a foundation model for fMRI. The negation of the SOC cost equals the ELBO (Proposition 4.1), so minimizing the control cost is variational inference. Under a locally linear drift with a shared eigen-decomposition, the controlled SDE has closed-form Gaussian marginals (Theorem 4.2), making inference simulation-free and parallelizable. The model is trained with a masked-reconstruction objective that pairs a reconstruction term with a JEPA-style regularization term, and after training the mean control vector $A$ is extracted as the universal feature for downstream tasks. The paper reports that this single feature, especially under linear probing, matches or beats fine-tuned baselines across demographics, traits, and psychiatric diagnoses, with the smallest variant (5M parameters) reaching performance comparable to the 86M-parameter BrainLM and BrainJEPA models.

Load-bearing premise

The load-bearing premise is that all the drift matrices used across the time steps can be diagonalized on the same set of axes, so that the closed-form mean and covariance formulas exactly describe the trained model rather than an approximation to it.

Editorial extensions

If this is right

  • A frozen BDO encoder transfers to unseen datasets: linear probing on HCP-A, ABIDE, ADHD200, HCP-EP, and TCP reaches or beats training-from-scratch networks, so the control feature carries reusable clinical signal.
  • Larger models, more pre-training data, and longer training each monotonically improve downstream accuracy, supporting the claim that the architecture scales like a foundation model.
  • Psychiatric-diagnosis accuracy (autism, ADHD, early psychosis, transdiagnostic) is obtainable from a backbone pre-trained only on healthy UK Biobank participants, indicating transfer across diagnostic and demographic distributions.
  • The 5M-parameter variant reaches performance comparable to 86M-parameter baselines at a fraction of pre-training GPU time, suggesting the state-space inductive bias substitutes for raw model capacity in capturing temporal structure.
  • The learned feature space is visibly organized by age in PCA and UMAP projections, so deviations from the age axis could function as a brain-age-like biomarker.

Reading between the lines

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

  • Because the implementation parameterizes $D_t = \sum_l w_t^l D_l$ without constraining the $D_l$ to share eigenvectors, the condition in Theorem 4.2 may not hold for the fitted model; comparing the closed-form moments against Monte Carlo samples of the fitted SDE would reveal whether the simulation-free step is exact or an additional approximation.
  • The same SOC-plus-locally-linear machinery applies to any irregularly sampled multivariate time series, not just fMRI, so physiological signals with similar noise structure are a direct and cheap test of the method's generality.
  • Mean pooling discards temporal localization, so the representation cannot say when within a scan a clinical signal appears; attention-weighted or per-segment pooling of the control sequence is a natural test of whether diagnostic information concentrates in specific time windows.
  • If the per-ROI structure of the control signals is examined directly rather than through the pooled feature, the steering inputs may yield a dynamics-based connectivity map; the paper only visualizes the aggregate representation, leaving regional interpretability of the controls untested.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes BDO, a continuous-discrete state-space foundation model for fMRI that couples a stochastic-optimal-control (SOC) formulation with amortized inference and self-supervised learning. The model approximates the controlled SDE by a piecewise-linear dynamics, derives closed-form Gaussian moment recursions under a common-eigenbasis assumption on the drift matrices, and uses a parallel scan for efficient training. BDO is pretrained on 41,072 UK Biobank resting-state fMRI scans with a masked-reconstruction objective plus a JEPA-style regularization, and is evaluated on age/gender prediction, trait prediction, and psychiatric diagnosis across UKB held-out, HCP-A, ABIDE, ADHD200, HCP-EP, and TCP. The authors report strong performance and parameter/compute efficiency compared with BrainNetCNN, BrainGNN, BrainNetTF, BrainLM, and BrainJEPA.

Significance. If the theoretical and empirical issues identified below are resolved, the work would be a meaningful contribution to fMRI foundation models. Its main strengths are: (i) a principled SOC-based ELBO derivation (standard Girsanov argument) that is novel in this application area; (ii) a simulation-free latent-dynamics formulation that enables parallel scans and efficient training; (iii) a large-scale pretraining setup with 41,072 UKB scans and explicit data splits, seeds, and hyperparameters, which is valuable for reproducibility; and (iv) a clear efficiency argument against transformer-based baselines. However, the validity of the derived ELBO for the actually implemented model is compromised by the eigenbasis mismatch described below, and the SOTA claim is weakened by missing foundation-model baselines in the clinical evaluation and by mixed results on HCP-A gender classification.

major comments (3)
  1. [Section 4.1, Theorem 4.2 and Eq. (14)] Theorem 4.2 assumes that all drift matrices D_t share a common eigenbasis V, since the closed-form mean/covariance recursions (12)-(13) and the parallel-scan operators (Appendix B, Eqs. (84)-(88)) rely on composing diagonal matrices in a fixed coordinate frame. The implementation instead sets D_t = sum_l w_t^l D_l with learned matrices D_l and does not constrain the D_l to be simultaneously diagonalizable. For generic learned matrices, the transition matrices e^{-Delta_i D_{t_i}} at different times do not commute, so the product cannot be collapsed into the exponent sums used in Eqs. (12)-(13). Consequently, the moments produced by Algorithm 1 and sampled in Algorithm 3 are not the moments of the controlled SDE (11), and the objective in Eq. (83) is not a valid ELBO for the generative model actually trained. This is an internal inconsistency that is not acknowledged in Section 6 or in the limitations. Please either enforce a shared-eigenbasis parameterization (e.g., D_l = V diag(d_l) V^T with a fixed or learned V) or provide a formal characterization of the approximation error introduced by the unconstrained parameterization.
  2. [Table 3] The psychiatric diagnosis comparisons include only training-from-scratch baselines (BrainNetCNN, BrainGNN, BrainNetTF) and BDO linear probing; the two foundation models (BrainLM and BrainJEPA) are absent. Since the abstract and Section 5 claim state-of-the-art results 'across a variety of downstream tasks', the absence of foundation-model baselines for ABIDE, ADHD200, HCP-EP, and TCP leaves the diagnosis-related SOTA claim unsupported. The authors should add these baselines or, failing that, restrict the SOTA claim to the tasks for which such comparisons are provided.
  3. [Table 2, Gender row] The fine-tuned BDO reports 79.40% accuracy and 78.98% F1 on HCP-A gender classification, which is lower than BrainJEPA† (81.52% accuracy, 84.26% F1). In light of this result, the unqualified 'state-of-the-art' wording in the abstract and Section 5.1 overstates the empirical findings. Please qualify the claim by task and metric, and discuss the gender-specific gap.
minor comments (5)
  1. [Section 4.1, Theorem 4.2] The interval notation '[ti, ti−1)' appears to be a typo; the proof in Appendix A.2 uses [ti, ti+1). Please make the interval definition consistent throughout.
  2. [Section 4.1, Eq. (12)] The formula for mu_ti is missing the final V^T that appears in the analogous covariance formula (13) and in the derivation in Appendix A.2. Please correct the matrix multiplication.
  3. [Appendix A.1, Eq. (24)] In the definition of the exponential martingale M_t, the first stochastic integral is written as running from 0 to 1, while the second runs from 0 to t. The upper limit should presumably be t in both integrals.
  4. [Appendix C.1] The preprocessing description states that the first 10 volumes are discarded, but Table 5 does not list the effective number of retained time points per dataset. Reporting the final time-series length for each dataset would improve reproducibility.
  5. [Section 5.1, footnote 1] The disclosure that the reproduced BrainJEPA results may have deviated due to undocumented preprocessing is useful; consider moving this caveat from a table footnote into the experimental setup or a dedicated reproducibility statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core ELBO and moment recursions are derived in-paper, and downstream labels are used only for evaluation.

full rationale

We traced the main derivation chain. Proposition 4.1 derives the ELBO from the stochastic optimal control cost using Girsanov's theorem (Appendix A.1), and Theorem 4.2 derives the closed-form mean and covariance recursions from the linear SDE under the stated common-eigenbasis assumption (Appendix A.2); both arguments are self-contained and do not rest on the empirical results. The pretraining objective in Eqs. (21) and (83) is defined entirely on fMRI signals and the EMA target encoder, with no downstream labels; labels enter only in the linear-probe or fine-tuning evaluation of Algorithm 4, which is standard model selection and evaluation rather than a fitted input being renamed as a prediction. The citations to Park et al. (2024a,b) and Becker et al. are methodological antecedents, not load-bearing: the SOC-ELBO equivalence and the moment recursions are re-proved in the appendix rather than imported by citation. The implementation's parameterization D_t = sum_l w_t^l D_l in Eq. (14) may violate Theorem 4.2's common-eigenbasis condition, so the computed moments may not solve the stated SDE; this is an internal-consistency and correctness concern, not a circular reduction, because the theorem's assumptions are stated rather than fitted to the target result. No quantity derived from downstream performance is fed back into the pretraining objective or into the claimed representation A, so no step reduces to its own input by construction.

Assumptions & free parameters 5 free parameters · 7 assumptions · 1 invented entities

The central derivation leans on standard stochastic calculus. The main ad hoc elements are the linear-Gaussian approximating family, the shared-eigenbasis assumption that is not enforced, and hyperparameters chosen by validation. No new physical entities are posited.

free parameters (5)
  • Latent dimension d = 192 / 384 / 768 for 5M / 21M / 85M variants
    Model capacity hyperparameter, chosen to create three sizes; performance improves with size (Figure 3).
  • Number of drift basis matrices L = 100
    Fixed number of learned SPD basis matrices in D_t = sum_l w_t^l D_l; no scaling study is reported for L.
  • Observation time scale factor = 0.1
    Multiplied to observation times for numerical stability; chosen by hand, not motivated by data.
  • Mask ratio gamma = 0.75
    Tuned on validation through ablation (Figure 5); optimal masking ratio.
  • Regularization balance tau = Not reported exactly; ablation suggests tau around 0.03
    Controls weight of JEPA-style regularizer relative to reconstruction; selected by validation from {0, 0.03, 0.1}.
assumptions (7)
  • standard math Girsanov theorem and Novikov condition
    Used to prove Proposition 4.1, the equivalence between SOC cost and ELBO.
  • standard math Ito isometry and martingale property of Wiener process
    Used in deriving the covariance recursion in Theorem 4.2.
  • domain assumption Prior and controlled SDE measures are mutually absolutely continuous
    Required for the Radon-Nikodym change of measure in the ELBO proof; assumed to hold for the controlled linear SDE.
  • ad hoc to paper Common eigenbasis V for all D_t
    Theorem 4.2 assumes all D_t are simultaneously diagonalizable, but Eq (14) does not enforce this in the learned parameterization.
  • domain assumption Linear-Gaussian controlled dynamics sufficiently approximate the posterior
    The locally linear approximation with piecewise-constant drift and unit diffusion replaces the true posterior; the paper acknowledges the resulting variational gap in the limitations.
  • domain assumption Gaussian factorized observation likelihood
    The decoder models y_t given z_t as Gaussian with diagonal covariance, and the likelihood factorizes over time points; this is a modeling choice for noisy BOLD signals.
  • domain assumption ROI parcellation and preprocessing preserve brain dynamics of interest
    The model operates on mean ROI time series after standard preprocessing; if the parcellation discards relevant signal, the foundation model claim is weakened.
invented entities (1)
  • Universal feature A (mean-pooled control sequence)
    purpose: Aggregates the learned optimal control signals into a single transferable representation for downstream tasks.
    A is a deterministic function of learned controls; its usefulness is demonstrated only through the paper's downstream evaluations, with no external falsifiable prediction attached.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Foundational Brain Dynamics Model via Stochastic Optimal Control." pith.science (2026). https://pith.science/paper/2ULD47MB

@misc{pith2026250204892,
  author       = {Pith},
  title        = {Pith review of: A Foundational Brain Dynamics Model via Stochastic Optimal Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ULD47MB}},
  note         = {Machine review of arXiv:2502.04892}
}
read the original abstract

We introduce a foundational model for brain dynamics that utilizes stochastic optimal control (SOC) and amortized inference. Our method features a continuous-discrete state space model (SSM) that can robustly handle the intricate and noisy nature of fMRI signals. To address computational limitations, we implement an approximation strategy grounded in the SOC framework. Additionally, we present a simulation-free latent dynamics approach that employs locally linear approximations, facilitating efficient and scalable inference. For effective representation learning, we derive an Evidence Lower Bound (ELBO) from the SOC formulation, which integrates smoothly with recent advancements in self-supervised learning (SSL), thereby promoting robust and transferable representations. Pre-trained on extensive datasets such as the UKB, our model attains state-of-the-art results across a variety of downstream tasks, including demographic prediction, trait analysis, disease diagnosis, and prognosis. Moreover, evaluating on external datasets such as HCP-A, ABIDE, and ADHD200 further validates its superior abilities and resilience across different demographic and clinical distributions. Our foundational model provides a scalable and efficient approach for deciphering brain dynamics, opening up numerous applications in neuroscience.

Figures

Figures reproduced from arXiv: 2502.04892 by the authors.

Figure 1
Figure 1. Conceptual illustration of our proposed Brain Dynamics with Optimal control (BDO). The ROI signals observed at discrete time points are encoded into an opti￾mal control policy, which steers the continuous latent state dynamics. The pre-trained optimal control policy is then utilized for various downstream tasks. et al., 2021; Taghia et al., 2018). A central goal of fMRI analysis is to extract, interpret, and underst… view at source ↗
Figure 2
Figure 2. Our BDO surpasses other foundation models, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Scalability results of HCP-A age regression in LP. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: BDO captures a latent space that encodes clinically relevant information from fMRI recordings. For each fMRI [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: (Left) Training curve (Right) Pearson correlation [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Age distribution across training, validation, and test splits for the UKB held-out age regression task under three [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Age distribution across training, validation, and test splits for the HCP-A age regression task under three different [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Label distributions across six classification tasks (UKB held-out gender, HCP-A gender, ABIDE autism, ADHD200 [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: Reconstruction quality of BDO in the UKB held-out subset (internal dataset). Five samples are randomly drawn [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Reconstruction quality of BDO in HCP-A (external dataset). Five samples are randomly drawn for visualization, [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 52 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey, D. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020

    arXiv 2010

  3. [3]

    K., Andersson, J

    Alfaro-Almagro, F., Jenkinson, M., Bangerter, N. K., Andersson, J. L., Griffanti, L., Douaud, G., Sotiropoulos, S. N., Jbabdi, S., Hernandez-Fernandez, M., Vallee, E., et al. Image processing and quality control for the first 10,000 brain imaging datasets from uk biobank. Neuroimage, 166: 0 400--424, 2018

    work page 2018

  4. [4]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., and Ballas, N. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15619--15629, 2023

    2023

  5. [5]

    Stochastic Calculus: An Introduction Through Theory and Exercises

    Baldi, P. Stochastic Calculus: An Introduction Through Theory and Exercises. Universitext. Springer International Publishing, 2017. ISBN 9783319622262. URL https://books.google.co.kr/books?id=fO09DwAAQBAJ

    work page 2017

  6. [6]

    J., and Neumann, G

    Becker, P., Pandya, H., Gebhardt, G., Zhao, C., Taylor, C. J., and Neumann, G. Recurrent kalman networks: Factorized inference in high-dimensional deep feature spaces. In International conference on machine learning, pp.\ 544--552. PMLR, 2019

    work page 2019

  7. [7]

    and Hashemi, F

    Behrouz, A. and Hashemi, F. Brain-mamba: Encoding brain activity via selective state space models. In Conference on Health, Inference, and Learning, pp.\ 233--250. PMLR, 2024

    work page 2024

  8. [8]

    Behzadi, Y., Restom, K., Liau, J., and Liu, T. T. A component based noise correction method (compcor) for bold and perfusion based fmri. Neuroimage, 37 0 (1): 0 90--101, 2007

    work page 2007

Show all 68 references
  1. [9]

    S., and Craddock, R

    Bellec, P., Chu, C., Chouinard-Decorte, F., Benhajali, Y., Margulies, D. S., and Craddock, R. C. The neuro bureau adhd-200 preprocessed repository. Neuroimage, 144: 0 275--286, 2017

    2017

  2. [10]

    Blelloch, G. E. Prefix sums and their applications. School of Computer Science, Carnegie Mellon University, 1990

    1990

  3. [11]

    Y., Salat, D

    Bookheimer, S. Y., Salat, D. H., Terpstra, M., Ances, B. M., Barch, D. M., Buckner, R. L., Burgess, G. C., Curtiss, S. W., Diaz-Santos, M., Elam, J. S., et al. The lifespan human connectome project in aging: an overview. Neuroimage, 185: 0 335--348, 2019

    2019

  4. [12]

    R., Sidhu, G

    Brown, M. R., Sidhu, G. S., Greiner, R., Asgarian, N., Bastani, M., Silverstone, P. H., Greenshaw, A. J., and Dursun, S. M. Adhd-200 global competition: diagnosing adhd using personal characteristic data can outperform resting state fmri measurements. Frontiers in systems neur...

    2012

  5. [13]

    L., Duberg, K., Pennington, B., Hinshaw, S

    Cai, W., Warren, S. L., Duberg, K., Pennington, B., Hinshaw, S. P., and Menon, V. Latent brain state dynamics distinguish behavioral variability, impaired decision-making, and inattention. Molecular Psychiatry, 26 0 (9): 0 4944--4957, September 2021. ISSN 1476-5578. doi:10.103...

    2021 doi

  6. [14]

    Lectures on BSDEs, stochastic control, and stochastic differential games with financial applications

    Carmona, R. Lectures on BSDEs, stochastic control, and stochastic differential games with financial applications. SIAM, 2016

    2016

  7. [15]

    O., de Oliveira Fonseca, A

    Caro, J. O., de Oliveira Fonseca, A. H., Rizvi, S. A., Rosati, M., Averill, C., Cross, J. L., Mittal, P., Zappala, E., Dhodapkar, R. M., Abdallah, C., et al. Brainlm: A foundation model for brain activity recordings. In The Twelfth International Conference on Learning Represen...

    2024

  8. [16]

    Emerging properties in self-supervised vision transformers

    Caron, M., Touvron, H., Misra, I., J \'e gou, H., Mairal, J., Bojanowski, P., and Joulin, A. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 9650--9660, 2021

    2021

  9. [17]

    D., Bodien, Y

    Chakravarty, S., Threlkeld, Z. D., Bodien, Y. G., Edlow, B. L., and Brown, E. N. A state-space model for dynamic functional connectivity. In 2019 53rd Asilomar Conference on Signals, Systems, and Computers, pp.\ 240--244. IEEE, 2019

    2019

  10. [18]

    An empirical study of training self-supervised vision transformers

    Chen, X., Xie, S., and He, K. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 9640--9649, 2021

    2021

  11. [19]

    Chopin, N., Fulop, A., Heng, J., and Thiery, A. H. Computational doob's h-transforms for online filtering of discretely observed diffusions, 2023

    2023

  12. [20]

    transdiagnostic connectome project

    Chopra, S., Cocuzza, C. V., Lawhead, C., Ricard, J. A., Labache, L., Patrick, L., Kumar, P., Rubenstein, A., Moses, J., Chen, L., Blankenbaker, C., Gillis, B., Germine, L. T., Harpaz-Rote, I., Yeo, B. T., Baker, J. T., and Holmes, A. J. "transdiagnostic connectome project", 2024 a

    2024

  13. [21]

    V., Lawhead, C., Ricard, J

    Chopra, S., Cocuzza, C. V., Lawhead, C., Ricard, J. A., Labache, L., Patrick, L. M., Kumar, P., Rubenstein, A., Moses, J., Chen, L., et al. The transdiagnostic connectome project: a richly phenotyped open dataset for advancing the study of brain-behavior relationships in psych...

    2024

  14. [22]

    S., Lewis, J

    Craddock, C., Benhajali, Y., Chu, C., Chouinard, F., Evans, A., Jakab, A., Khundrakpam, B. S., Lewis, J. D., Li, Q., Milham, M., et al. The neuro bureau preprocessing initiative: open sharing of preprocessed neuroimaging data and derivatives. Frontiers in Neuroinformatics, 7 0...

    2013

  15. [23]

    S., Yan, C., Li, Q., Lurie, D., Vogelstein, J., Burns, R., et al

    Craddock, C., Sikka, S., Cheung, B., Khanuja, R., Ghosh, S. S., Yan, C., Li, Q., Lurie, D., Vogelstein, J., Burns, R., et al. Towards automated analysis of connectomes: The configurable pipeline for the analysis of connectomes (c-pac). Front Neuroinform, 42 0 (10.3389), 2013 b

  16. [24]

    E., and Friston, K

    Daunizeau, J., Stephan, K. E., and Friston, K. J. Stochastic dynamic causal modelling of fmri data: should we care about neural noise? Neuroimage, 62 0 (1): 0 464--481, 2012

    2012

  17. [25]

    Davatzikos, C., Xu, F., An, Y., Fan, Y., and Resnick, S. M. Longitudinal progression of alzheimer's-like patterns of atrophy in normal older adults: the spare-ad index. Brain, 132 0 (8): 0 2026--2035, 2009

    2026

  18. [26]

    X., Alaerts, K., Anderson, J

    Di Martino, A., Yan, C.-G., Li, Q., Denio, E., Castellanos, F. X., Alaerts, K., Anderson, J. S., Assaf, M., Bookheimer, S. Y., Dapretto, M., et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Molecular...

    2014

  19. [27]

    Diederik, P. K. Adam: A method for stochastic optimization. (No Title), 2014

    2014

  20. [28]

    T., Chong, J

    Dong, Z., Ruilin, L., Wu, Y., Nguyen, T. T., Chong, J. S. X., Ji, F., Tong, N. R. J., Chen, C. L. H., and Zhou, J. H. Brain-jepa: Brain dynamics foundation model with gradient positioning and spatiotemporal masking. In The Thirty-eighth Annual Conference on Neural Information ...

    2024

  21. [29]

    L., Belsky, D

    Elliott, M. L., Belsky, D. W., Knodt, A. R., Ireland, D., Melzer, T. R., Poulton, R., Ramrakha, S., Caspi, A., Moffitt, T. E., and Hariri, A. R. Brain-age in midlife is associated with accelerated biological aging and cognitive decline in a longitudinal birth cohort. Molecular...

    2021

  22. [30]

    J., Blair, R

    Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik, A. I., Erramuzpe, A., Kent, J. D., Goncalves, M., DuPre, E., Snyder, M., et al. fmriprep: a robust preprocessing pipeline for functional mri. Nature methods, 16 0 (1): 0 111--116, 2019

    2019

  23. [31]

    Fleming, W. H. and Soner, H. M. Controlled Markov processes and viscosity solutions, volume 25. Springer Science & Business Media, 2006

    2006

  24. [32]

    Dynamic causal modelling

    Friston, K., Harrison, L., and Penny, W. Dynamic causal modelling. NeuroImage, 19 0 (4): 0 1273--1302, 2003. ISSN 1053-8119. doi:https://doi.org/10.1016/S1053-8119(03)00202-7. URL https://www.sciencedirect.com/science/article/pii/S1053811903002027

    2003 doi

  25. [33]

    K., Dinga, R., Hahn, T., Ching, C

    Han, L. K., Dinga, R., Hahn, T., Ching, C. R., Eyler, L. T., Aftanas, L., Aghajani, M., Aleman, A., Baune, B. T., Berger, K., et al. Brain aging in major depressive disorder: results from the enigma major depressive disorder working group. Molecular psychiatry, 26 0 (9): 0 512...

    2021

  26. [34]

    Masked autoencoders are scalable vision learners

    He, K., Chen, X., Xie, S., Li, Y., Doll \'a r, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

    2022

  27. [35]

    N., Deligiannidis, G., and Doucet, A

    Heng, J., Bishop, A. N., Deligiannidis, G., and Doucet, A. Controlled sequential monte carlo. The Annals of Statistics, 48 0 (5): 0 2904--2929, 2020

    2020

  28. [36]

    R., Coleman, M

    Jacobs, G. R., Coleman, M. J., Lewandowski, K. E., Pasternak, O., Cetin-Karayumak, S., Mesholam-Gately, R. I., Wojcik, J., Kennedy, L., Knyazhanskaya, E., Reid, B., et al. An introduction to the human connectome project for early psychosis. Schizophrenia Bulletin, pp.\ sbae123, 2024

    2024

  29. [37]

    F., Behrens, T

    Jenkinson, M., Beckmann, C. F., Behrens, T. E., Woolrich, M. W., and Smith, S. M. Fsl. Neuroimage, 62 0 (2): 0 782--790, 2012

    2012

  30. [38]

    Brain network transformer

    Kan, X., Dai, W., Cui, H., Zhang, Z., Guo, Y., and Yang, C. Brain network transformer. Advances in Neural Information Processing Systems, 35: 0 25586--25599, 2022

    2022

  31. [39]

    Kappen, H. J. and Ruiz, H. C. Adaptive importance sampling for control and inference. Journal of Statistical Physics, 162: 0 1244--1266, 2016

    2016

  32. [40]

    J., Miller, S

    Kawahara, J., Brown, C. J., Miller, S. P., Booth, B. G., Chau, V., Grunau, R. E., Zwicker, J. G., and Hamarneh, G. Brainnetcnn: Convolutional neural networks for brain networks; towards predicting neurodevelopment. NeuroImage, 146: 0 1038--1049, 2017

    2017

  33. [41]

    and Platen, E

    Kloeden, P. and Platen, E. Numerical Solution of Stochastic Differential Equations. Stochastic Modelling and Applied Probability. Springer Berlin Heidelberg, 2013. ISBN 9783662126165. URL https://books.google.co.kr/books?id=r9r6CAAAQBAJ

    2013

  34. [42]

    A path towards autonomous machine intelligence version 0.9

    LeCun, Y. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62 0 (1): 0 1--62, 2022

    2022

  35. [43]

    B., Yuan, R., Ryman, S., Kim, J., Santini, V., Henderson, V

    Lee, B., Cai, W., Young, C. B., Yuan, R., Ryman, S., Kim, J., Santini, V., Henderson, V. W., Poston, K. L., and Menon, V. Latent brain state dynamics and cognitive flexibility in older adults. Progress in neurobiology, 208: 0 102180, January 2022. ISSN 1873-5118 0301-0082. doi...

    2022

  36. [44]

    L., Chen, R

    Li, X., Wong, T.-K. L., Chen, R. T. Q., and Duvenaud, D. Scalable gradients for stochastic differential equations. International Conference on Artificial Intelligence and Statistics, 2020

    2020

  37. [45]

    H., Ventola, P., and Duncan, J

    Li, X., Zhou, Y., Dvornek, N., Zhang, M., Gao, S., Zhuang, J., Scheinost, D., Staib, L. H., Ventola, P., and Duncan, J. S. Braingnn: Interpretable brain graph neural network for fmri analysis. Medical Image Analysis, 74: 0 102233, 2021

    2021

  38. [46]

    Liu, G.-H., Lipman, Y., Nickel, M., Karrer, B., Theodorou, E., and Chen, R. T. Q. Generalized schr\"odinger bridge matching. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=SoismgeX7z

    2024

  39. [47]

    Decoupled weight decay regularization

    Loshchilov, I. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  40. [48]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016

    2016 arXiv

  41. [49]

    and Wang, Y

    Lu, J. and Wang, Y. Guidance for twisted particle filter: a continuous-time perspective. arXiv preprint arXiv:2409.02399, 2024

    2024 arXiv

  42. [50]

    nilearn , 2025

    Nilearn contributors . nilearn , 2025. URL https://github.com/nilearn/nilearn

    2025

  43. [51]

    Spectral dynamic causal modeling: A didactic introduction and its relationship with functional connectivity

    Novelli, L., Friston, K., and Razi, A. Spectral dynamic causal modeling: A didactic introduction and its relationship with functional connectivity. Network Neuroscience, 8 0 (1): 0 178--202, 2024

    2024

  44. [52]

    M., Kay, A

    Ogawa, S., Lee, T. M., Kay, A. R., and Tank, D. W. Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proceedings of the National Academy of Sciences, 87 0 (24): 0 9868--9872, 1990. doi:10.1073/pnas.87.24.9868. URL https://www.pnas.org/doi/abs/10.10...

    1990 doi

  45. [53]

    Stochastic optimal control for diffusion bridges in function spaces

    Park, B., Choi, J., Lim, S., and Lee, J. Stochastic optimal control for diffusion bridges in function spaces. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a . URL https://openreview.net/forum?id=WyQW4G57Zd

    2024

  46. [54]

    Amortized control of continuous state space Feynman-Kac model for irregular time series

    Park, B., Lee, H., and Lee, J. Amortized control of continuous state space Feynman-Kac model for irregular time series. arXiv preprint arXiv:2410.05602, 2024 b

    2024 arXiv

  47. [55]

    and Shenton Martha; Breier, A

    Prunier, N. and Shenton Martha; Breier, A. Human connectome project for early psychosis -- release 1.1, 2021

    2021

  48. [56]

    Bayesian Filtering and Smoothing

    S \"a rkk \"a , S. Bayesian Filtering and Smoothing. Bayesian Filtering and Smoothing. Cambridge University Press, 2013. ISBN 9781107030657. URL https://books.google.co.kr/books?id=5VlsAAAAQBAJ

    2013

  49. [57]

    and Garc \' a-Fern \'a ndez, \'A

    S \"a rkk \"a , S. and Garc \' a-Fern \'a ndez, \'A . F. Temporal parallelization of bayesian smoothers. IEEE Transactions on Automatic Control, 66 0 (1): 0 299--306, 2020

    2020

  50. [58]

    M., Laumann, T

    Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X.-N., Holmes, A. J., Eickhoff, S. B., and Yeo, B. T. T. Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity mri. Cerebral Cortex, 28 0 (9): 0 3095--3114, 07 2017. ISSN 1047-3...

    2017 doi

  51. [59]

    Modeling irregular time series with continuous recurrent units

    Schirmer, M., Eltayeb, M., Lessmann, S., and Rudolph, M. Modeling irregular time series with continuous recurrent units. In International conference on machine learning, pp.\ 19388--19405. PMLR, 2022

    2022

  52. [60]

    Uk biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age

    Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P., Danesh, J., Downey, P., Elliott, P., Green, J., Landray, M., et al. Uk biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS medicine, 12 0 (3): 0...

    2015

  53. [61]

    Uncovering hidden brain state dynamics that regulate performance and decision-making during cognition

    Taghia, J., Cai, W., Ryali, S., Kochalka, J., Nicholas, J., Chen, T., and Menon, V. Uncovering hidden brain state dynamics that regulate performance and decision-making during cognition. Nature Communications, 9 0 (1): 0 2505, June 2018. ISSN 2041-1723. doi:10.1038/s41467-018-04723-6

    2018 doi

  54. [62]

    Theodorou, E. A. Nonlinear stochastic control and information theoretic dualities: Connections, interdependencies and thermodynamic interpretations. Entropy, 17 0 (5): 0 3352--3375, 2015

    2015

  55. [63]

    S., Breakspear, M., and Zalesky, A

    Tian, Y., Margulies, D. S., Breakspear, M., and Zalesky, A. Topographic organization of the human subcortex unveiled with functional connectivity gradients. Nature Neuroscience, 23 0 (11): 0 1421--1432, Nov 2020. ISSN 1546-1726. doi:10.1038/s41593-020-00711-6. URL https://doi....

    2020 doi

  56. [64]

    Linearly-solvable markov decision problems

    Todorov, E. Linearly-solvable markov decision problems. Advances in neural information processing systems, 19, 2006

    2006

  57. [65]

    Triantafyllopoulos, K. et al. Bayesian inference of state space models. Springer, 2021

    2021

  58. [66]

    E., Triebkorn, P., Breyton, M., Dollomaja, B., Lemarechal, J.-D., Petkoski, S., Sorrentino, P., Depannemaecker, D., Hashemi, M., and Jirsa, V

    Wang, H. E., Triebkorn, P., Breyton, M., Dollomaja, B., Lemarechal, J.-D., Petkoski, S., Sorrentino, P., Depannemaecker, D., Hashemi, M., and Jirsa, V. K. Virtual brain twins: from basic neuroscience to clinical use. National Science Review, 11 0 (5): 0 nwae079, 02 2024. ISSN ...

    2024 doi

  59. [67]

    Hierarchical spatio-temporal state-space modeling for fmri analysis

    Wei, Y., Abrol, A., Hassanzadeh, R., and Calhoun, V. Hierarchical spatio-temporal state-space modeling for fmri analysis. arXiv preprint arXiv:2408.13074, 2024

    2024 arXiv

  60. [68]

    S., C ukur, T., and Gallant, J

    Zhang, T., Gao, J. S., C ukur, T., and Gallant, J. L. Voxel-based state space modeling recovers task-related cognitive states in naturalistic fmri experiments. Frontiers in neuroscience, 14: 0 565976, 2021

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.