pith. sign in

arxiv: 2605.16708 · v1 · pith:V7YLRDB4new · submitted 2026-05-15 · 💻 cs.LG · stat.ML

Isolating Nonlinear Independent Sources in fMRI with β-TCVAE Models

Qiang Li , Shujian Yu , Jesus Malo , Jingyu Liu , T\"ulay Adali , Vince D. Calhoun This is my paper

Pith reviewed 2026-05-20 18:39 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords nonlinear source separationfMRI analysisβ-TCVAEindependent componentsdefault mode networkfunctional connectivityvariational autoencoderbrain networks
0
0 comments X

The pith

Adapting β-TCVAE to fMRI recovers nonlinear independent sources that match known brain networks.

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

The paper adapts the β-TCVAE model, a variational autoencoder variant designed for disentangled representations, to real fMRI datasets in order to separate nonlinear mixtures of spatial and temporal brain signals. Standard independent component analysis assumes linear mixing, which restricts its ability to capture the complex dynamics of brain activity. By applying this framework the authors recover spatial components that align with established functional networks, including the default mode network, and confirm that the resulting latent structure exhibits coherent functional connectivity. This offers a route to modeling nonlinear organization in neuroimaging data where linear methods fall short.

Core claim

We adapt and modify the β-TCVAE framework to fMRI data for nonlinear source disentanglement, aiming to separate mixed spatial and temporal brain signals into interpretable components. The model recovers meaningful nonlinear spatial components with biological relevance, including well-established intrinsic connectivity networks such as the default mode network. Evaluation via functional network connectivity shows that the learned latent structure captures coherent and interpretable brain organization patterns.

What carries the argument

The β-TCVAE model, a refinement of β-VAE that disentangles latent factors by penalizing total correlation in the latent space without adding extra hyperparameters.

If this is right

  • Nonlinear mixing assumptions in fMRI can be relaxed to reveal independent sources that linear ICA cannot isolate.
  • Established networks such as the default mode network remain identifiable when the model accounts for nonlinear signal relationships.
  • The learned latent representations produce functional connectivity matrices that reflect coherent brain organization.
  • Deep representation learning approaches can be directly validated against real neuroimaging data rather than simulations alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adaptation could be tested on EEG or MEG recordings to check whether nonlinear disentanglement generalizes across modalities.
  • Hybrid pipelines that initialize linear ICA with β-TCVAE components might improve robustness on noisy clinical datasets.
  • Longitudinal fMRI studies could examine whether the nonlinear components track changes in network integrity more sensitively than linear ones.

Load-bearing premise

That components recovered from the adapted model correspond to real biological brain networks rather than artifacts produced by the training process itself.

What would settle it

If the spatial pattern of the recovered default mode network component shows low overlap with independently validated DMN maps obtained from the same fMRI scans using established linear methods, the claim of biological relevance would fail.

read the original abstract

Learning meaningful latent representations from nonlinear fMRI data remains a fundamental challenge in neuroimaging analysis. Traditional independent component analysis, widely used due to its ability to estimate interpretable functional brain networks, relies on a linear mixing assumption for latent sources, limiting its ability to capture the inherently nonlinear and complex organization of brain dynamics. More recently, deep representation learning methods have emerged as promising alternatives for modeling nonlinear latent structure. However, many of these approaches have been evaluated primarily on simulated datasets or natural image benchmarks, with comparatively limited validation on real-world neuroimaging data such as fMRI. In this work, we are motivated by the $\beta$-TCVAE (Total Correlation Variational Autoencoder), a refinement of the $\beta$-VAE framework for learning latent representations without introducing additional hyperparameters during training. We adapt and modify this model to fMRI data for nonlinear source disentanglement, aiming to separate mixed spatial and temporal brain signals into interpretable components. We show that the $\beta$-TCVAE framework can recover meaningful nonlinear spatial components with biological relevance, including well-established intrinsic connectivity networks such as the default mode network. Furthermore, we evaluate the learned representations using functional network connectivity, showing that the latent structure captures coherent and interpretable brain organization patterns. This study provides a pilot investigation that bridges nonlinear representation learning and fMRI analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper adapts the β-TCVAE framework to fMRI data for nonlinear source disentanglement, claiming that the model recovers meaningful nonlinear spatial components with biological relevance, including well-established networks such as the default mode network, and that the latent structure captures coherent brain organization as shown by functional network connectivity analysis.

Significance. If the central claims hold under rigorous validation, the work could offer a nonlinear alternative to linear ICA for modeling complex brain dynamics in neuroimaging. It bridges deep representation learning with fMRI analysis in a pilot study, but the current evidence base is primarily qualitative and does not yet demonstrate that the nonlinear capacity is actively exploited beyond what linear methods or regularization alone might achieve.

major comments (2)
  1. [Abstract and Results] The evaluation of recovered components relies on post-hoc visual inspection and functional network connectivity without quantitative metrics, error bars, statistical validation, or baseline comparisons (e.g., to standard ICA). This is load-bearing for the claim of recovering biologically relevant nonlinear sources.
  2. [Methods] No controlled simulation is reported in which known spatial sources are mixed by an explicit nonlinear function, recovered by the adapted β-TCVAE, and scored against ground truth using metrics such as the Amari index or component-wise correlation. Without this, it is not possible to confirm that the model isolates nonlinear independent sources due to the data-generating process rather than β-TC regularization or preprocessing.
minor comments (2)
  1. [Abstract and Methods] The abstract and methods lack details on the specific fMRI dataset(s) used, preprocessing pipeline, exact architectural modifications to β-TCVAE, hyperparameter choices, and training procedure.
  2. [Methods] Notation for the adapted model (e.g., how the total correlation term is implemented for spatial-temporal fMRI signals) should be clarified with explicit equations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the scope and validation needs for this pilot study adapting β-TCVAE to fMRI. We address each major comment below and describe the revisions planned to strengthen the evidence for nonlinear source disentanglement.

read point-by-point responses

  1. Referee: [Abstract and Results] The evaluation of recovered components relies on post-hoc visual inspection and functional network connectivity without quantitative metrics, error bars, statistical validation, or baseline comparisons (e.g., to standard ICA). This is load-bearing for the claim of recovering biologically relevant nonlinear sources.

    Authors: We agree that the present evaluation is primarily qualitative, centered on visual assessment of spatial maps and functional network connectivity to highlight recovery of known networks such as the default mode network. This is typical for initial real-data neuroimaging studies where ground truth is absent. To address the concern, we will add quantitative comparisons to standard linear ICA, including spatial correlation coefficients and mutual information measures between components, along with reproducibility metrics across subjects. Where feasible, we will report variability measures and basic statistical summaries to support the biological relevance claims. revision: yes

  2. Referee: [Methods] No controlled simulation is reported in which known spatial sources are mixed by an explicit nonlinear function, recovered by the adapted β-TCVAE, and scored against ground truth using metrics such as the Amari index or component-wise correlation. Without this, it is not possible to confirm that the model isolates nonlinear independent sources due to the data-generating process rather than β-TC regularization or preprocessing.

    Authors: We concur that a controlled simulation with explicit nonlinear mixing would provide direct evidence that the β-TCVAE isolates sources due to its nonlinear capacity rather than regularization or preprocessing alone. The current manuscript emphasizes real fMRI data to demonstrate practical utility and interpretability in a biologically relevant setting. We will incorporate a new simulation subsection that generates synthetic data via known nonlinear mixing functions applied to spatial sources, applies the adapted model, and evaluates recovery using the Amari index, component-wise correlations, and comparisons to linear ICA baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity: standard adaptation of β-TCVAE to fMRI with independent empirical evaluation

full rationale

The paper adapts the existing β-TCVAE framework (a refinement of β-VAE) to fMRI data for nonlinear source separation. The derivation consists of model training on real neuroimaging data followed by post-hoc interpretation of latent components via visual inspection and functional network connectivity metrics. No load-bearing step reduces by construction to its own inputs: there are no self-definitional equations, no fitted parameters renamed as predictions, no uniqueness theorems imported via self-citation, and no ansatz smuggled through prior work. The central claim rests on empirical outcomes against known brain networks rather than algebraic equivalence or statistical forcing from the training objective itself. This is a self-contained application of an established method to a new domain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central claim rests on the domain assumption that fMRI signals contain recoverable nonlinear independent sources and that biological relevance can be judged by matching known networks.

axioms (1)
  • domain assumption fMRI data consists of nonlinear mixtures of independent spatial and temporal sources that can be disentangled by a variational autoencoder
    This premise is required for the adaptation of β-TCVAE to be meaningful and is implicit in the motivation section of the abstract.

pith-pipeline@v0.9.0 · 5785 in / 1295 out tokens · 45327 ms · 2026-05-20T18:39:18.993485+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 3 internal anchors

  1. [1]

    INTRODUCTION The blind source separation (BSS) problem is a fundamental problem in signal processing, where the goal is to recover un- derlying source signals from observed mixtures without prior knowledge of the mixing process [1, 2, 3, 4]. Independent component analysis (ICA) [5, 6, 7] is a widely used com- putational method to address this problem, wit...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Model Overview We consider multi-subject fMRI data{x (s) t ∈R D}, where s∈ {1,

    METHODS 2.1. Model Overview We consider multi-subject fMRI data{x (s) t ∈R D}, where s∈ {1, . . . , S}indexes subjects andtindexes time. Our goal is to learn a low-dimensional latent representationz∈R K that captures shared latent structure across subjects while re- maining disentangled. We model the data using a subject- conditioned variational autoencod...

    work page

  3. [3]

    RESULTS 3.1. Disentangling Spatial and Temporal Components via β-TCV AE in fMRI Data We successfully disentangled spatial components and tem- poral dynamics in fMRI data using aβ-TCV AE framework. Fig.1A illustrates the training dynamics of theβ-TCV AE framework, including the total loss, MI, and TC. As training progresses, the total loss steadily decreas...

    work page

  4. [4]

    and orbitofrontal cortex (OFC; ICN 4), as illustrated in Fig.2A. The extracted DMN component exhibited clear spa- tial coverage across canonical DMN regions, including the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), and lateral parietal regions, demonstrating that the model effectively captured the core functional organization of th...

    work page

  5. [5]

    Compared with lin- ear ICA (InfoMax) [7], the learned representations are more spatially coherent and exhibit improved correspondence with established functional networks

    DISCUSSION Our results demonstrate thatβ-TCV AE can effectively learn biologically meaningful spatial components that are well aligned with canonical brain networks. Compared with lin- ear ICA (InfoMax) [7], the learned representations are more spatially coherent and exhibit improved correspondence with established functional networks. Importantly,β-TCV A...

    work page

  6. [6]

    The results show thatβ-TCV AE can effec- tively extract biologically meaningful spatial components

    CONCLUSION In this pilot investigation, we adapted and modified theβ- TCV AE framework to disentangle nonlinear representations in fMRI data. The results show thatβ-TCV AE can effec- tively extract biologically meaningful spatial components. In addition, we obtained corresponding time series associated with each spatial component, enabling further subject...

    work page

  7. [7]

    Blind source separation and indepen- dent component analysis: A review,

    Seungjin Choi, Andrzej Cichocki, Hyung-Min Park, and Soo-Young Lee, “Blind source separation and indepen- dent component analysis: A review,”Neural Informa- tion Processing-Letters and Reviews, vol. 6, no. 1, pp. 1–57, 2005

    work page 2005

  8. [8]

    Pierre Comon and Christian Jutten,Handbook of Blind Source Separation: Independent component analysis and applications, Academic press, 2010

    work page 2010

  9. [9]

    Blind source separation of real world signals,

    T-W Lee, Anthony J Bell, and Reinhold Orglmeister, “Blind source separation of real world signals,” inPro- ceedings of International Conference on Neural Net- works (ICNN’97). IEEE, 1997, vol. 4, pp. 2129–2134

    work page 1997

  10. [10]

    Blind source separation-semiparametric statistical approach,

    Shun-Ichi Amari and J-F Cardoso, “Blind source separation-semiparametric statistical approach,”IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. 2692–2700, 1997

    work page 1997

  11. [11]

    Independent compo- nent analysis: algorithms and applications,

    Aapo Hyv ¨arinen and Erkki Oja, “Independent compo- nent analysis: algorithms and applications,”Neural net- works, vol. 13, no. 4-5, pp. 411–430, 2000

    work page 2000

  12. [12]

    Fast and robust fixed-point algo- rithms for independent component analysis,

    Aapo Hyvarinen, “Fast and robust fixed-point algo- rithms for independent component analysis,”IEEE transactions on Neural Networks, vol. 10, no. 3, pp. 626–634, 1999

    work page 1999

  13. [13]

    An information-maximization approach to blind separation and blind deconvolution,

    Anthony J Bell and Terrence J Sejnowski, “An information-maximization approach to blind separation and blind deconvolution,”Neural computation, vol. 7, no. 6, pp. 1129–1159, 1995

    work page 1995

  14. [14]

    A re- view of group ica for fmri data and ica for joint inference of imaging, genetic, and erp data,

    Vince D Calhoun, Jingyu Liu, and T ¨ulay Adalı, “A re- view of group ica for fmri data and ica for joint inference of imaging, genetic, and erp data,”Neuroimage, vol. 45, no. 1, pp. S163–S172, 2009

    work page 2009

  15. [15]

    A method for making group inferences from functional mri data using independent component analysis,

    Vince D Calhoun, T ¨ulay Adalı, Godfrey D Pearlson, and James J Pekar, “A method for making group inferences from functional mri data using independent component analysis,”Human Brain Mapping, vol. 16, no. 2, pp. 131–131, 2002

    work page 2002

  16. [16]

    Neuromark: An auto- mated and adaptive ica based pipeline to identify repro- ducible fmri markers of brain disorders,

    Yuhui Du, Zening Fu, Jing Sui, Shuang Gao, Ying Xing, Dongdong Lin, Mustafa Salman, Anees Abrol, Md Ab- dur Rahaman, Jiayu Chen, et al., “Neuromark: An auto- mated and adaptive ica based pipeline to identify repro- ducible fmri markers of brain disorders,”NeuroImage: Clinical, vol. 28, pp. 102375, 2020

    work page 2020

  17. [17]

    The fas- tica algorithm with spatial constraints,

    Christian W Hesse and Christopher J James, “The fas- tica algorithm with spatial constraints,”IEEE signal processing letters, vol. 12, no. 11, pp. 792–795, 2005

    work page 2005

  18. [18]

    Spatiotemporal complexity in the psychotic brain,

    Qiang Li, Jingyu Liu, Godfrey D Pearlson, Jiayu Chen, Yu-Ping Wang, Jessica A Turner, and Vince D Cal- houn, “Spatiotemporal complexity in the psychotic brain,”Molecular Psychiatry, vol. 31, no. 4, pp. 2014– 2028, 2026

    work page 2014

  19. [19]

    Higher-order triadic interactions: Insights into the multiscale network organization in schizophrenia,

    Qiang Li, Shujian Yu, Jesus Malo, Godfrey D Pearlson, Yu-Ping Wang, and Vince D Calhoun, “Higher-order triadic interactions: Insights into the multiscale network organization in schizophrenia,”Human Brain Mapping, vol. 46, no. 16, pp. e70399, 2025

    work page 2025

  20. [20]

    Ap- plying fully tensorial ica to fmri data,

    Joni Virta, Sara Taskinen, and Klaus Nordhausen, “Ap- plying fully tensorial ica to fmri data,” in2016 IEEE Signal Processing in Medicine and Biology Symposium (SPMB). IEEE, 2016, pp. 1–6

    work page 2016

  21. [21]

    Variational autoencoders and nonlinear ica: A unifying framework,

    Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen, “Variational autoencoders and nonlinear ica: A unifying framework,” inInterna- tional conference on artificial intelligence and statistics. PMLR, 2020, pp. 2207–2217

    work page 2020

  22. [22]

    V1 non-linear proper- ties emerge from local-to-global non-linear ICA,

    Jes ´us Malo and Juan Guti ´errez, “V1 non-linear proper- ties emerge from local-to-global non-linear ICA,”Net- work: Computation in Neural Systems, vol. 17, no. 1, pp. 85–102, 2006

    work page 2006

  23. [23]

    Nonlinear ica using auxiliary variables and generalized contrastive learning,

    Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner, “Nonlinear ica using auxiliary variables and generalized contrastive learning,” inThe 22nd international con- ference on artificial intelligence and statistics. PMLR, 2019, pp. 859–868

    work page 2019

  24. [24]

    Unsupervised feature extraction by time-contrastive learning and non- linear ica,

    Aapo Hyvarinen and Hiroshi Morioka, “Unsupervised feature extraction by time-contrastive learning and non- linear ica,”Advances in neural information processing systems, vol. 29, 2016

    work page 2016

  25. [25]

    Misep–linear and nonlinear ica based on mutual information,

    Lu ´ıs B Almeida, “Misep–linear and nonlinear ica based on mutual information,”Journal of Machine Learning Research, vol. 4, no. Dec, pp. 1297–1318, 2003

    work page 2003

  26. [26]

    Deep deter- ministic nonlinear ica via total correlation minimization with matrix-based entropy functional,

    Qiang Li, Shujian Yu, Liang Ma, Chen Ma, Jingyu Liu, Tulay Adali, and Vince D Calhoun, “Deep deter- ministic nonlinear ica via total correlation minimization with matrix-based entropy functional,”arXiv preprint arXiv:2601.00904, 2025

    work page arXiv 2025

  27. [27]

    Isolating sources of disentangle- ment in variational autoencoders,

    Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud, “Isolating sources of disentangle- ment in variational autoencoders,”Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  28. [28]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  29. [29]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  30. [30]

    The wu-minn human connectome project: an overview,

    David C Van Essen, Stephen M Smith, Deanna M Barch, Timothy EJ Behrens, Essa Yacoub, Kamil Ugurbil, Wu- Minn HCP Consortium, et al., “The wu-minn human connectome project: an overview,”Neuroimage, vol. 80, pp. 62–79, 2013

    work page 2013

  31. [31]

    Mark Jenkinson, Christian F Beckmann, Timothy EJ Behrens, Mark W Woolrich, and Stephen M Smith, “Fsl,”Neuroimage, vol. 62, no. 2, pp. 782–790, 2012

    work page 2012

  32. [32]

    A short history of spm,

    Karl Friston, “A short history of spm,”Statistical para- metrical mapping: The analysis of functional brain im- ages, pp. 3–9, 2007

    work page 2007