Pith. sign in

REVIEW 3 major objections 6 minor 40 references

ST-VTD, a tensor decomposition with low-rank spatial and LSTM temporal priors, recovers latent fMRI-like factors with spatial and temporal correlations near 0.98, outperforming classical and probabilistic decomposition baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:18 UTC pith:YFBBQ4YU

load-bearing objection A genuinely new synthesis of LL1-style spatial priors, LSTM temporal priors, and LPALM amortized variational inference, but the central empirical claim is undercut by the GICA warm-start confound and the synthetic-only evaluation. the 3 major comments →

arxiv 2607.22262 v1 pith:YFBBQ4YU submitted 2026-07-24 stat.ML cs.LGeess.SP

Variational Low-rank Tensor Decomposition for Multisubject Spatiotemporal Data Analysis

Laura M. Montaldo , Ricardo A. Borsoi , Sebastian Miron , Tulay Adali This is my paper
classification stat.ML cs.LGeess.SP
keywords tensor decompositionvariational inferencemultisubject fMRIspatiotemporal datalow-rank priorLSTM prioralgorithm unrollinggroup ICA warm start
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that combining a tensor-factorization generative model with structured spatial and temporal priors lets a variational method recover subject-specific brain networks and their time courses more faithfully than standard decomposition tools. The proposed ST-VTD represents each subject's data as a sum of outer products of spatial maps and temporal courses, regularizes the spatial maps with a low-rank LL1-style prior and the time courses with an LSTM-based prior, and learns the posterior with an amortized variational network built by unrolling an optimization algorithm. On a realistic synthetic fMRI dataset with 100 subjects, it reports the highest spatial and temporal correlations (0.983 and 0.985) and the lowest factor errors among five methods, while matching the best reconstruction error. The authors argue this shows the structured variational design is effective not only for reconstruction but also for capturing more faithful spatiotemporal representations.

Core claim

The central claim is that subject-specific spatial maps and temporal dynamics in multisubject spatiotemporal data can be jointly modeled as latent factors with structured priors — low-rank spatial maps inspired by LL1 and LSTM-driven temporal courses — and inferred with an amortized variational posterior built by unrolling LPALM iterations. If this is right, ST-VTD recovers the underlying components more accurately than GICA, LL1, Bayesian CP factorization, and PARAFAC2 on the synthetic benchmark: corr_z=0.9829, corr_c=0.9845, re_z=0.1578, re_c=0.1605, against best-baseline correlations around 0.93 and errors above 0.25. The ablation supports the design: dropping the LSTM prior or the low-ra

What carries the argument

The central mechanism is a bilinear generative model X^{(n)} = sum_k Z_k^{(n)} ∘ c_k^{(n)} + Gaussian noise, combined with two structured priors: each spatial map is centered at a low-rank product U_k V_k^T with a shared per-component variance (LL1-inspired), and each temporal course is centered at the output of a learned LSTM with per-component variance. The posterior is made tractable by a conditional mean-field variational approximation whose means are computed by unrolling I=50 iterations of LPALM, the linearized proximal alternating least-squares update, with a hard low-rank SVD projection on the spatial columns; small MLPs output the log-variances. The GICA warm start provides the init

Load-bearing premise

The load-bearing premise is that the synthetic evaluation is a valid substitute for real multisubject fMRI: the benchmark data are generated from the same bilinear low-rank plus Gaussian-noise model ST-VTD assumes, so violations such as motion artifacts, scanner effects, heavy-tailed noise, or missing data could erase the reported gains.

What would settle it

Generate a multisubject spatiotemporal benchmark from realistic fMRI-like spatial templates with heavy-tailed noise, motion-related outliers, and a percentage of missing time points, then fit ST-VTD and the four baselines; if ST-VTD no longer achieves the highest correlations and lowest factor errors, the paper's central claim about latent recovery on its own benchmark would not generalize.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, ST-VTD recovers latent fMRI-like components with correlations around 0.98 on synthetic multisubject data, substantially above the tested baselines.
  • The framework supports swapping priors, update rules, and projection constraints within the same variational formulation, so it can be adapted without redesigning inference.
  • The ablations indicate the LSTM temporal prior, the low-rank spatial prior, and the hard low-rank projection each contribute to the reported gains.
  • The GICA-based warm start is important for the non-convex optimization; random initialization leads to a large degradation in component recovery.
  • Because the posterior means are used as MMSE estimates, the model naturally produces uncertainty intervals alongside point estimates for each spatial and temporal factor.

Where Pith is reading between the lines

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

  • Editorial inference: the modularity of the design suggests a direct stress test the paper does not run — keep the LPALM-unrolled posterior exactly as is, replace the LSTM temporal prior with a simpler linear Gaussian prior, and measure how much temporal recovery drops; this would isolate how much of the gain comes from the temporal prior versus the unrolled inference.
  • Editorial inference: since GICA initialization is nearly load-bearing, a spectral or multi-start initialization selected by the ELBO could make the method applicable where group ICA models are invalid (e.g., data with nonstationary mixing), a route the paper leaves open.
  • Editorial inference: if the reported correlations transfer to real fMRI, the per-component variational variances could be used as statistical confidence maps for spatial-network discovery, a clinical use case the paper does not evaluate.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces ST-VTD, a variational framework for multisubject spatiotemporal tensor data that combines a bilinear outer-product generative model (Eq. (1)) with a low-rank spatial prior inspired by LL1 and an LSTM-based temporal prior. Posterior inference is amortized and implemented by unrolling LPALM iterations, with a GICA-based warm start. Experiments on a synthetic fMRI-like dataset (N=100, T=150, 30×30 spatial resolution, K=10 components) compare ST-VTD with GICA, LL1, BCPF, and PARAFAC2. The central claim is that ST-VTD achieves the best latent-factor recovery, with corr_z=0.9829, corr_c=0.9845, re_z=0.1578, re_c=0.1605, while also matching the best reconstruction error. The paper includes an ablation study of modeling choices and a random-initialization check, and the code is publicly available.

Significance. If the central claim holds, ST-VTD offers a useful contribution to multisubject fMRI-style tensor decomposition: it combines structured spatial and temporal priors with a parameter-efficient amortized variational inference scheme and demonstrates strong latent recovery on a synthetic benchmark. The paper's strengths include a reproducible experimental setup, public code, a clear ablation of the proposed priors and projection scheme, and a transparent discussion of the GICA-warm-start dependence. However, the empirical claim currently rests on a controlled synthetic dataset generated from the same bilinear outer-product model family assumed by ST-VTD, and the comparison is complicated by an asymmetric initialization: ST-VTD is initialized from GICA solutions while baselines are run standalone. The missing warm-start control and the absence of statistical significance testing are load-bearing for the claim that the proposed variational framework, rather than fine-tuning of a privileged initialization, is responsible for the reported gains.

major comments (3)
  1. [§III-C, §III-D, Table I] The warm-start asymmetry is not controlled. ST-VTD is initialized from GICA outputs (Z(0)=Z_GICA+Δz, C(0)=C_GICA+Δc) and then trained for 150 Adam epochs, while GICA and the other baselines are run standalone. The ablation in §III-C shows that random initialization degrades re_z and re_c by ≈335% and correlations by ≈30%, implying that without the GICA warm start ST-VTD would be worse than GICA (re_z≈0.68 vs 0.525 in Table I). To support the claim that the variational framework, the LSTM prior, and the low-rank spatial prior provide the benefit, the authors should add a control that refines the GICA solution with the same unrolled LPALM updates and Adam training but without the variational KL terms and LSTM prior. Without this control, the Table I ordering may be attributable to additional fine-tuning of a strong initialization rather than to the proposed model.
  2. [§III-A, §IV] The evaluation is conducted under model match. Section IV states that the benchmark is 'a controlled synthetic dataset generated according to the proposed generative model,' i.e., the same bilinear outer-product structure as Eq. (1) with subject-specific factors, Gaussian noise, and smooth spatial/temporal components. Since the ground truth is generated from exactly the model family that ST-VTD assumes, the reported advantage over baselines may not transfer to real fMRI data containing non-Gaussian artifacts, missing observations, scanner effects, or structures outside the low-rank + LSTM prior family. The authors should add model-mismatch experiments (e.g., non-Gaussian noise, missing entries, different generating mechanisms) or an application to real data before claiming that ST-VTD 'significantly improves latent factor recovery' in a general sense. The acknowledgment in the conclusion
  3. [§III-D, Table I] No statistical significance testing is reported. Table I lists means and standard deviations, but the central claim is based on point estimates. For example, the reconstruction error difference between ST-VTD and LL1 is 0.2524 vs 0.2527, well within the reported standard deviations, and several baseline correlation metrics have large variability (e.g., LL1 corr_z = 87.4±24.9). The authors should provide paired tests across repeated synthetic instantiations or across subjects, and report effect sizes or confidence intervals. Without this, the statement that ST-VTD 'outperforms all baselines' is not statistically supported, especially where the differences are small relative to variance.
minor comments (6)
  1. [§II-A, Eq. (3)] The notation New(·,·) is defined, but it would help to state explicitly that p(Z_k) in Eq. (3) is a matrix-variate Gaussian written entrywise; the low-rank mean Uk V_k^T plus λ_k^2 I is a valid prior, but the identifiability of U_k and V_k under the likelihood is not discussed. A brief note on this would improve clarity.
  2. [§II-B] The unrolled LPALM updates use a matrix representation X∈R^{T×V}, Z∈R^{V×K}, C∈R^{T×K}; however, the generative model in Eq. (1) defines X^{(n)} as a third-order tensor. The reshaping from tensor to matrix is implicit and should be stated explicitly, especially because the low-rank projection P_LR reshapes columns of Z back into Vx×Vy matrices.
  3. [§III-C] The description of the learnable offset Δz, Δc says it is 'shared across subjects,' but the later text refers to 'subject-shared learnable offset.' This is consistent, but the phrase 'small subject-shared' could confuse; suggest rewording to 'shared across all subjects.'
  4. [§III-D, Fig. 2] The caption says the subject is 'n=1 across models' and that ST-VTD achieved the median reconstruction error, but it is unclear whether n=1 is the same subject for all methods or whether each method's median subject is displayed. Please clarify.
  5. [Table I] The table caption states 'All values are multiplied by 100.' For readability, consider using percentages or fractions consistently, and note whether the standard deviations are over subjects or over independent synthetic trials.
  6. [§II-A] The LSTM prior in Eq. (4) sets p(c_k)=New(w(t,θ_k), γ_k). The input t is a vector of time coordinates, but the mechanism by which the LSTM produces a deterministic mean vector of length T is not specified; a sentence on the LSTM architecture (input, hidden state, output layer) would help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the paper's empirical benchmark is model-matched but not definitionally equivalent, and the GICA warm-start is a confound, not a circular reduction.

full rationale

I found no step in which a claimed prediction reduces, by the paper's own equations or by a self-citation chain, to its input. The generative model in Eq. (1) is used as the likelihood for ST-VTD and also as the template for the synthetic benchmark, but the recovery of latent factors from noisy data is still a nontrivial inference problem, and all baselines are evaluated on the same data. Section IV's statement that the evaluation used 'a controlled synthetic dataset generated according to the proposed generative model' is an explicitly acknowledged limitation on external validity, not a circular construction: the ground-truth factors are not fed into the estimator, and the reported correlations/errors are measured against known targets. The GICA warm-start described in Section III-C is a legitimate concern for attribution — the ablation shows random initialization degrades recovery by roughly 335% and correlations drop by about 30% — but this is an experimental confound (missing a control where GICA is refined with the same unrolled updates without the variational terms), not a definitional equivalence. No equation in the paper makes the final ST-VTD estimate equal to the GICA input, nor is any benchmark metric forced by construction. Self-citations in the references are background citations for existing tensor-decomposition and fMRI methods; they do not carry the load of the paper's central empirical claim. Therefore the paper is not circular.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests on the generative assumptions in Eqs. (1)-(4): bilinear factorization, rank-L spatial structure, and LSTM temporal structure. The benchmark data are generated under the same family, which is both an axiom and a source of partial circularity. Free parameters are learned or hand-chosen; no new physical entities are introduced.

free parameters (7)
  • Spatial low-rank rank L = floor(min(Vx,Vy)/K) = 3 (for Vx=Vy=30, K=10)
    Chosen by hand via heuristic formula; controls low-rank prior and projection. If L is wrong, the spatial model is misspecified.
  • Number of components K = 10
    Set equal to ground-truth component count in the synthetic data; no model-order selection.
  • Spatial prior factors U_k, V_k and variances lambda_k^2 = learned during training (not reported)
    Parameters of the low-rank spatial prior p(Z_k); optimized through the ELBO.
  • LSTM temporal prior weights theta_k and variances gamma_k = learned during training (not reported)
    Parameters of the temporal prior p(c_k); shape shared time-course dynamics.
  • Likelihood noise variance sigma^2 = not reported; presumably optimized as part of theta
    Appears in the likelihood (Eq. 2); estimation procedure not described.
  • KL annealing coefficients beta_max_z, beta_max_c and warm-up steps W_z, W_c = 5, 5, 50, 50
    Hand-chosen training hyperparameters that weigh the KL terms and affect final performance.
  • Variational encoder parameters (MLP variance heads, unrolled LPALM offsets) = learned (~19,300 total trainable parameters)
    Amortized inference network weights that determine posterior means/variances.
axioms (6)
  • domain assumption Each subject tensor equals a sum of K outer products of spatial and temporal factors plus i.i.d. Gaussian noise (Eq. 1).
    Core generative model; real fMRI contains artifacts, non-Gaussian noise, and missing data that violate this.
  • domain assumption Spatial maps live near a rank-L manifold with L = floor(min(Vx,Vy)/K) (Eq. 3 and projection P_LR).
    Low-rank prior and hard SVD truncation assume the spatial structure is captured by this specific rank.
  • domain assumption Temporal courses can be represented by a shared-parameter LSTM with Gaussian residual (Eq. 4).
    Temporal prior assumes nonlinear dynamics in the LSTM family; other dynamics would be misspecified.
  • domain assumption The Neuromark-based synthetic data with perturbations and Gaussian noise is a realistic proxy for real fMRI.
    The only evaluation is on this synthetic dataset, admitted in Section IV to be generated according to the proposed generative model.
  • domain assumption Mean-field conditional Gaussian variational family is sufficiently expressive for the posterior.
    q(Z|X)q(C|X) assumes posterior independence across factors; posterior correlations are ignored.
  • domain assumption Truncated SVD column-wise projection is the correct low-rank structure for spatial maps.
    P_LR connects the LL1-inspired prior to the unrolled optimization; alternative low-rank structures would change results.

pith-pipeline@v1.3.0-alltime-deepseek · 10146 in / 18680 out tokens · 147911 ms · 2026-08-01T05:18:06.889883+00:00 · methodology

0 comments
read the original abstract

Modeling shared and subject-specific structure in multisubject spatiotemporal data remains challenging, particularly in neuroimaging, where both spatial and temporal patterns exhibit rich variability across subjects. Existing matrix and tensor decompositions provide interpretable factorizations, but rely on fixed multilinear structures or coupling schemes that may limit their flexibility in capturing complex variability. In this work, we introduce a spatiotemporal variational tensor decomposition (ST-VTD) framework that combines a tensor factorization generative model with structured priors to jointly represent spatial maps and temporal dynamics. Spatial factors are regularized to promote a low-rank structure inspired by the LL1 decomposition, while temporal factors are modeled using a learned Long short-term memory (LSTM)-based prior, enabling flexible and adaptive dynamics. Posterior inference is performed using an amortized variational formulation by unrolling iterations of an optimization algorithm, leading to an interpretable and parameter-efficient architecture. The proposed inference framework employs a warm-start strategy based on group independent component analysis, which we found to improve optimization performance. Experiments on a realistic synthetic functional MRI (fMRI) dataset demonstrate that the proposed approach significantly improves latent factor recovery compared with representative classical and probabilistic decomposition benchmarks.

Figures

Figures reproduced from arXiv: 2607.22262 by Laura M. Montaldo, Ricardo A. Borsoi, Sebastian Miron, Tulay Adali.

Figure 1
Figure 1. Figure 1: Ground-truth spatial map for component k = 7 (a), its corresponding time course (b), the observed data slice at time t = 150 for subject n = 1 (c), and the 10 selected spatial maps overlaid on the data slice (d). SVD, FastICA [39] is applied to U⊤Xg, and the resulting spatial maps Z are back-projected to pixel space, column￾normalized to fix scale ambiguity, and used to estimate subject￾specific time cours… view at source ↗
Figure 2
Figure 2. Figure 2: Ground truth (GT) and estimated component ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 1 linked inside Pith

  1. [1]

    Empirical orthogonal functions and related techniques in atmospheric science: A review,

    A. Hannachi, I. T. Jolliffe, and D. B. Stephenson, “Empirical orthogonal functions and related techniques in atmospheric science: A review,” International Journal of Climatology, vol. 27, no. 9, pp. 1119–1152, 2007

    2007

  2. [2]

    Blind fMRI source unmixing via higher-order tensor decompositions,

    C. Chatzichristos, E. Kofidis, M. Morante, and S. Theodoridis, “Blind fMRI source unmixing via higher-order tensor decompositions,”Journal of Neuroscience Methods, vol. 315, pp. 17–47, 2019

    2019

  3. [3]

    A review of change detection in multitemporal hyperspectral images: Current techniques, applications, and challenges,

    S. Liu, D. Marinelli, L. Bruzzone, and F. Bovolo, “A review of change detection in multitemporal hyperspectral images: Current techniques, applications, and challenges,”IEEE Geoscience and Remote Sensing Magazine, vol. 7, no. 2, pp. 140–158, 2019

    2019

  4. [4]

    Reproducibility in Matrix and Tensor Decompositions: Focus on model match, interpretability, and uniqueness,

    T. Adali, F. Kantar, M. A. B. S. Akhonda, S. Strother, V . D. Calhoun, and E. Acar, “Reproducibility in Matrix and Tensor Decompositions: Focus on model match, interpretability, and uniqueness,”IEEE Signal Processing Magazine, vol. 39, no. 4, pp. 8–24, 2022

    2022

  5. [5]

    Personalized coupled tensor decomposition for multimodal data fusion: Uniqueness and algo- rithms,

    R. A. Borsoi, K. Usevich, D. Brie, and T. Adali, “Personalized coupled tensor decomposition for multimodal data fusion: Uniqueness and algo- rithms,”IEEE Transactions on Signal Processing, vol. 73, pp. 113–129, 2024

    2024

  6. [6]

    Multimodal Data Fusion: An Overview of Methods, Challenges, and Prospects,

    D. Lahat, T. Adali, and C. Jutten, “Multimodal Data Fusion: An Overview of Methods, Challenges, and Prospects,”Proceedings of the IEEE, vol. 103, no. 9, pp. 1449–1477, 2015

    2015

  7. [7]

    Data fusion in metabolomics using coupled matrix and tensor factorizations,

    E. Acar, R. Bro, and A. K. Smilde, “Data fusion in metabolomics using coupled matrix and tensor factorizations,”Proceedings of the IEEE, vol. 103, no. 9, pp. 1602–1620, 2015

    2015

  8. [8]

    Coupled tensor decomposition for hyperspectral and multispectral image fusion with inter-image variability,

    R. A. Borsoi, C. Pr ´evost, K. Usevich, D. Brie, J. C. Bermudez, and C. Richard, “Coupled tensor decomposition for hyperspectral and multispectral image fusion with inter-image variability,”IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 3, pp. 702–717, 2021

    2021

  9. [9]

    Coupled tensor decompositions for data fusion,

    C. Chatzichristos, S. Van Eyndhoven, E. Kofidis, and S. Van Huffel, “Coupled tensor decompositions for data fusion,” inTensors for data processing. Elsevier, 2022, pp. 341–370

    2022

  10. [10]

    A review of group ICA for fMRI data and ICA for joint inference of imaging, genetic, and erp data,

    V . D. Calhoun, J. Liu, and T. Adalı, “A review 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

    2009

  11. [11]

    Tensor decompositions and applications,

    T. G. Kolda and B. W. Bader, “Tensor decompositions and applications,” SIAM Review, vol. 51, no. 3, pp. 455–500, 2009

    2009

  12. [12]

    Tensor decomposition for signal processing and machine learning,

    N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex- akis, and C. Faloutsos, “Tensor decomposition for signal processing and machine learning,”IEEE Transactions on signal processing, vol. 65, no. 13, pp. 3551–3582, 2017

    2017

  13. [13]

    Coupled canonical polyadic decompositions and (coupled) decompositions in multilinear rank- (Lr,n, Lr,n,1)terms–part I: Uniqueness,

    M. Sørensen and L. D. De Lathauwer, “Coupled canonical polyadic decompositions and (coupled) decompositions in multilinear rank- (Lr,n, Lr,n,1)terms–part I: Uniqueness,”SIAM Journal on Matrix Analysis and Applications, vol. 36, no. 2, pp. 496–522, 2015

    2015

  14. [14]

    Learning spatiotemporal brain dynamics in adolescents via multimodal MEG and fMRI data fusion using joint tensor/matrix decomposition,

    I. Belyaeva, B. Gabrielson, Y .-P. Wang, T. W. Wilson, V . D. Calhoun, J. M. Stephen, and T. Adali, “Learning spatiotemporal brain dynamics in adolescents via multimodal MEG and fMRI data fusion using joint tensor/matrix decomposition,”IEEE Transactions on Biomedical Engineering, vol. 71, no. 7, pp. 2189–2200, 2024

    2024

  15. [15]

    An AO-ADMM approach to constraining PARAFAC2 on all modes,

    M. Roald, C. Schenker, V . D. Calhoun, T. Adali, R. Bro, J. E. Cohen, and E. Acar, “An AO-ADMM approach to constraining PARAFAC2 on all modes,”SIAM Journal on Mathematics of Data Science, vol. 4, no. 3, pp. 1191–1222, 2022

    2022

  16. [16]

    Coupled CP tensor decomposition with shared and distinct components for multi-task fMRI data fusion,

    R. A. Borsoi, I. Lehmann, M. A. B. S. Akhonda, V . D. Calhoun, K. Usevich, D. Brie, and T. Adali, “Coupled CP tensor decomposition with shared and distinct components for multi-task fMRI data fusion,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 51, 2023, pp. 1–5

    2023

  17. [17]

    Exploring multimodal data fusion through joint decompositions with flexible couplings,

    R. C. Farias, J. E. Cohen, and P. Comon, “Exploring multimodal data fusion through joint decompositions with flexible couplings,”IEEE Transactions on Signal Processing, vol. 64, no. 18, pp. 4830–4844, 2016

    2016

  18. [18]

    Bayesian CP factorization of incomplete tensors with automatic rank determination,

    Q. Zhao, L. Zhang, and A. Cichocki, “Bayesian CP factorization of incomplete tensors with automatic rank determination,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 37, no. 9, pp. 1751–1763, 2015

    2015

  19. [19]

    Diversity in Independent Com- ponent and Vector Analyses: Identifiability, algorithms, and applications in medical imaging,

    T. Adali, M. Anderson, and G.-S. Fu, “Diversity in Independent Com- ponent and Vector Analyses: Identifiability, algorithms, and applications in medical imaging,”IEEE Signal Processing Magazine, vol. 31, no. 3, pp. 18–33, 2014

    2014

  20. [20]

    Neural topographic factor analysis for fMRI data,

    E. Sennesh, Z. Khan, Y . Wang, J. B. Hutchinson, A. Satpute, J. Dy, and J.-W. van de Meent, “Neural topographic factor analysis for fMRI data,”Advances in Neural Information Processing Systems, vol. 33, pp. 12 046–12 056, 2020

    2020

  21. [21]

    Deep markov factor analysis: Towards concurrent temporal and spatial analysis of fMRI data,

    A. Farnoosh and S. Ostadabbas, “Deep markov factor analysis: Towards concurrent temporal and spatial analysis of fMRI data,”Advances in Neural Information Processing Systems, vol. 34, pp. 17 876–17 888, 2021

    2021

  22. [22]

    Undirected probabilistic model for tensor decomposition,

    Z. Tao, T. Tanaka, and Q. Zhao, “Undirected probabilistic model for tensor decomposition,”Advances in Neural Information Processing Systems, vol. 36, pp. 25 837–25 853, 2023

    2023

  23. [23]

    Unrolling PALM for sparse semi-blind source separation,

    M. Fahes, C. Kervazo, J. Bobin, and F. Tupin, “Unrolling PALM for sparse semi-blind source separation,” inInternational Conference on Learning Representations, 2022

    2022

  24. [24]

    Tensors: a brief introduction,

    P. Comon, “Tensors: a brief introduction,”IEEE Signal Processing Magazine, vol. 31, no. 3, pp. 44–53, 2014

    2014

  25. [25]

    Decompositions of a higher-order tensor in block terms—part II: Definitions and uniqueness,

    L. De Lathauwer, “Decompositions of a higher-order tensor in block terms—part II: Definitions and uniqueness,”SIAM Journal on Matrix Analysis and Applications, vol. 30, no. 3, pp. 1033–1066, 2008

    2008

  26. [26]

    On uniqueness and computation of the decomposition of a tensor into multilinear rank-(1, L r, Lr)terms,

    I. Domanov and L. D. Lathauwer, “On uniqueness and computation of the decomposition of a tensor into multilinear rank-(1, L r, Lr)terms,” SIAM Journal on Matrix Analysis and Applications, vol. 41, no. 2, pp. 747–803, 2020

    2020

  27. [27]

    Hyperspectral super-resolution accounting for spectral variability: Coupled tensor LL1-based recovery and blind unmixing of the unknown super-resolution image,

    C. Pr ´evost, R. A. Borsoi, K. Usevich, D. Brie, J. C. Bermudez, and C. Richard, “Hyperspectral super-resolution accounting for spectral variability: Coupled tensor LL1-based recovery and blind unmixing of the unknown super-resolution image,”SIAM Journal on Imaging Sciences, vol. 15, no. 1, pp. 110–138, 2022

    2022

  28. [28]

    Fast and structured block-term tensor decomposition for hyperspectral unmixing,

    M. Ding, X. Fu, and X.-L. Zhao, “Fast and structured block-term tensor decomposition for hyperspectral unmixing,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 1691–1709, 2023

    2023

  29. [29]

    Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition,

    H. Sak, A. Senior, and F. Beaufays, “Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition,” 2014. [Online]. Available: https://arxiv.org/abs/1402.1128

    Pith/arXiv arXiv 2014

  30. [30]

    C. M. Bishop,Pattern Recognition and Machine Learning. Springer, 2006

    2006

  31. [31]

    Deep unrolling of the multiplicative updates algorithm for blind source separation, with appli- cation to hyperspectral unmixing,

    C. Kervazo, A. Chetoui, and J. E. Cohen, “Deep unrolling of the multiplicative updates algorithm for blind source separation, with appli- cation to hyperspectral unmixing,” in32nd European Signal Processing Conference (EUSIPCO). EURASIP, 2024, pp. 656–660

    2024

  32. [32]

    Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,

    V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,”IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021

    2021

  33. [33]

    K. P. Murphy,Machine learning: A probabilistic perspective. MIT press, 2012

    2012

  34. [34]

    An introduction to variational autoen- coders,

    P. K. Diederik and W. Max, “An introduction to variational autoen- coders,”F oundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307–392, 2019

    2019

  35. [35]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in International Conference on Learning Representations, 2014

    2014

  36. [36]

    beta-V AE: Learning basic visual concepts with a constrained variational framework,

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner, “beta-V AE: Learning basic visual concepts with a constrained variational framework,” inInterna- tional Conference on Learning Representations, 2016

    2016

  37. [37]

    Cyclical annealing schedule: A simple approach to mitigating KL vanishing,

    H. Fu, C. Li, X. Liu, J. Gao, A. Celikyilmaz, and L. Carin, “Cyclical annealing schedule: A simple approach to mitigating KL vanishing,” in Proc. NAACL-HLT 2019, V olume 1 (Long and Short Papers), 2019, pp. 240–250

    2019

  38. [38]

    Neuromark: An automated and adaptive ICA based pipeline to identify reproducible fMRI markers of brain disorders,

    Y . Du, Z. Fu, J. Sui, S. Gao, Y . Xing, D. Lin, M. Salman, A. Abrol, M. A. Rahaman, J. Chenet al., “Neuromark: An automated and adaptive ICA based pipeline to identify reproducible fMRI markers of brain disorders,” NeuroImage: Clinical, vol. 28, p. 102375, 2020

    2020

  39. [39]

    Independent component analysis: algorithms and applications,

    A. Hyv ¨arinen and E. Oja, “Independent component analysis: algorithms and applications,”Neural networks, vol. 13, no. 4-5, pp. 411–430, 2000

    2000

  40. [40]

    Vervliet, O

    N. Vervliet, O. Debals, L. Sorber, M. Van Barel, and L. De Lathauwer. (2016) Tensorlab 3.0. [Online]. Available: https://www.tensorlab.net

    2016