Pith. sign in

REVIEW 4 major objections 4 minor 56 references

Provable Affine Identifiability of Nonlinear CCA under Latent Distributional Priors

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Under a Gaussian latent prior and a spectral dominance condition, the optimal nonlinear CCA encoders are affine in the true latents, making them identifiable up to an orthogonal rotation.

desk verdict Plausible and potentially important identifiability result for nonlinear CCA, but the main proof is deferred and the key spectral assumption is never verified in the real-data experiment; still deserves refereeing. read the letter →

arxiv 2510.04758 v3 pith:F26VFAVG submitted 2025-10-06 cs.LG

Zhiwei Han , Stefan Matthes , Hao Shen This is my paper · ORCID
classification cs.LG
keywords nonlinearCCAaffineidentifiabilitydisentangledrepresentationlearningcanonicalcorrelationsHermiteexpansionwhiteninglatentdistributionalpriorsstatisticalconsistency
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

Nonlinear canonical correlation analysis (CCA) learns two maps that maximize correlation between two views of the same data. This paper proves that, under a latent generative model in which the two views share additive Gaussian (or near-Gaussian) factors, the optimal whitened CCA encoders must be affine in the true latent variables: they recover the latents up to an orthogonal rotation. The proof works by moving the analysis from observation space to the latent source space, then using classical Hermite/Mehler expansions of bivariate Gaussians to show that first-order correlations dominate higher-order ones when a spectral gap condition holds. It further proves that ridge-regularized empirical CCA converges to this population solution, and validates the theory on synthetic and rendered-image data, with ablations that confirm the predicted failure when the spectral condition is violated. If correct, this turns CCA from a mere alignment tool into a method with formal disentanglement guarantees, complementing contrastive learning.

What carries the argument

The load-bearing object is the Hermite (Mehler) expansion of the joint Gaussian latent pair. By representing the whitened encoder pair in the orthonormal basis of Hermite polynomials, the CCA objective decomposes into contributions from first-order components (linear in the latents) and higher-order components. The spectral separation condition ρ_dS > ρ_1^2 — that the smallest first-order canonical correlation exceeds the largest product ρ_i ρ_j of two first-order correlations — ensures the first-order contribution strictly dominates any higher-order contribution, forcing all maximizers to be linear in the sources. A second piece of machinery, Proposition 1, is a reparameterization invarianc

What would settle it

Take a Gaussian latent pair with d_S=2 and canonical correlations ρ1=0.9, ρ2=0.5, so ρ2=0.5 < 0.81 and Assumption 2 fails. Train the whitened nonlinear CCA objective with flexible encoders on many samples from the additive model. The theorem predicts that at least one canonical direction will be unrecoverable (the maximum principal angle will stay large); if all directions are recovered to within numerical error, the dominance condition is not necessary for affine identifiability. Alternatively, a direct population-level check: compute the objective J for a quadratic Hermite encoder pair and c

Watch

Extended reading notes

Core claim

Stated simply, the paper's central claim is that the population maximizer of the whitened nonlinear CCA objective — over a dense class of encoder pairs — is forced to be affine in the ground-truth latents. Specifically, for sources generated by the additive model s=a+c, s'=b+c with independent a,b,c, and with a non-degenerate Gaussian joint prior satisfying the first-order canonical dominance condition ρ_dS > ρ_1^2, every maximizer satisfies h* = f*∘g(s) = Q(s-μ_s) and h'* = f'*∘g'(s') = Q'(s'-μ_{s'}) for orthogonal matrices Q,Q'. The ambiguity is exactly an orthogonal transform, so the latent subspace — not the individual coordinates — is identifiable. The paper presents this as an extensio

Load-bearing premise

The load-bearing premise is the numerical spectral condition ρ_dS > ρ_1^2 — that the weakest first-order canonical correlation of the latent pair is stronger than the square of the strongest one; if real data do not satisfy this, the affine-identifiability guarantee collapses, as the paper's own ablations confirm.

Editorial extensions

If this is right

  • If the theorem holds, correlation-based self-supervised losses such as DeepCCA and W-MSE carry a formal disentanglement guarantee: they recover the true latent subspace up to rotation, not merely a set of correlated projections.
  • Whitening becomes a theoretically necessary component of these objectives; the proof predicts that non-whitened variants lose identifiability.
  • The finite-sample consistency theorem means practitioners can expect latent recovery to improve with dataset size at a rate governed by the ridge schedule, providing a practical recipe for tuning.
  • The additive latent model with distributional priors extends to several non-Gaussian families (e.g., Gamma, Poisson, negative binomial), suggesting that the affine-identifiability guarantee is not an artifact of Gaussianity alone.
  • The spectral condition ρ_dS > ρ_1^2 gives researchers a pre-training diagnostic: compute the canonical correlations of the latent prior; if the condition fails, at least one factor direction will be unidentifiable.

Reading between the lines

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

  • A natural extension the authors do not develop: use the same Hermite-dominance argument to predict the *number* of identifiable directions when ρ_dS ≤ ρ_1^2 — the ablation suggests a gradual loss of directions, which could be sharpened into a partial identifiability theorem.
  • The orthogonal (rather than permutation) ambiguity implies the learned coordinates are not the individual semantic factors but any rotation of them; an investor in causal representation learning would need an additional linear decoupling step to obtain factor-aligned axes.
  • One could test the necessity of the Gaussian prior: construct a non-Gaussian latent pair with identical first- and second-order statistics but different higher-order structure; the theorem's proof, which leans on Hermite expansions of Gaussians, predicts CCA may then prefer nonlinear maximizers.
  • The paper's experiments on latent-dimension mismatch suggest an over-complete setting (d_Z > d_S) may still identify the d_S-dimensional subspace; a formal statement bounding the identifiable dimension would be a direct follow-up.
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

4 major / 4 minor

Summary. The paper studies identifiability in nonlinear CCA for multi-view data generated by the additive latent model s=a+c, s'=b+c. Under joint Gaussian priors and a spectral separation condition on the canonical correlations (Assumption 2, ρ_dS > ρ_1^2), it claims (Theorem 1) that every population maximizer of the whitened CCA objective is affine in the ground-truth latents up to orthogonal transforms, thereby establishing affine identifiability. It further claims (Theorem 2) that ridge-regularized empirical CCA converges to the population solution and that the learned encoders recover the latents up to rotations. Experiments on synthetic data across five distributions and on 3DIdent report high R² values for DeepCCA and W-MSE, together with ablations varying the source dimension, the first-order canonical dominance ratio, and the encoder dimension.

Significance. If valid, the result would provide a population-level affine identifiability guarantee for nonlinear CCA, linking classical Lancaster/Eagleson theory to modern non-contrastive representation learning. The reparameterization strategy (Proposition 1) is principled and avoids constructing the proof specifically to match the target conclusion. The empirical work includes a deliberate ablation of the key spectral condition, and the reported R² values on synthetic data are strong. However, the central proof is not included in the submitted text, and several load-bearing assumptions are either not verified in the real-data experiment or appear inconsistent with the generative model as written. The contribution is potentially significant, but the current manuscript cannot be fully assessed without substantial revision.

major comments (4)
  1. [§3.3, Proposition 1 and Theorem 1] The complete proofs of Proposition 1 and Theorem 1 are deferred to a supplementary file that is not part of the submitted manuscript. Theorem 1 is the central contribution and depends on Hermite expansions, spectral truncation, and an approximation argument between dense encoder classes and whitened source-space functions. The provided sketch is only a few sentences and does not allow verification of the argument. Please include full proofs in the main text or as an appended supplement before the paper can be evaluated.
  2. [Eq. (1) and Assumption 1] The generative model states that a,b,c are i.i.d. If they are i.i.d. with common covariance V, then after standardizing s and s' to unit variance we have Var(s)=Var(s')=2V=I and Cov(s,s')=V=0.5I, so all canonical correlations equal 0.5. Thus Assumption 1's general full-rank Σ_ss' cannot be realized by the i.i.d. additive model, and Assumption 2 is automatically satisfied. If the authors intend independent but non-identically distributed components, Eq. (1) must be corrected; if they truly mean i.i.d., the setup is much more restrictive than stated and the ablation's varying ρ_i values do not correspond to the generative model.
  3. [§3.4, Theorem 2, Assumption A1] Assumption A1 explicitly assumes that the population maximizer exists and is unique up to post-orthogonal transforms. This uniqueness is precisely the content of Theorem 1 under Assumptions 1–2. If Theorem 1 is valid, A1 should be stated as a consequence, not as an additional assumption. If it is not a consequence, then Theorem 2's consistency statement is conditional on the identifiability result it is meant to transfer to finite samples. Please clarify the logical dependency and either prove A1 from Theorem 1 or state it as an independent lemma with its own proof.
  4. [§4.2, Table 2] The 3DIdent experiment is presented as validation of Theorem 1, but the paper never reports the canonical correlations of the ground-truth latent pair, nor the ratio ρ_dS / ρ_1^2 needed for Assumption 2. Since Theorem 1's identifiability conclusion is void when ρ_dS ≤ ρ_1^2, and the ablation in Fig. 5 shows that performance changes sharply with this ratio, the high R² on 3DIdent does not confirm that the theorem's assumptions hold in that setting. Please either report the canonical correlations of the latent factors or explicitly qualify 3DIdent as an out-of-assumption demonstration rather than a validation of the theorem's condition.
minor comments (4)
  1. [§3.2, Eq. (2)] The objective is written as the sum of σ_i(Cov(...)) without stating that the singular values are taken in decreasing order. Please make the ordering explicit.
  2. [§3.4, A4] In the approximate-maximization condition, the supremum is written over ~f'∈~f'_X; the first argument should also be a supremum over ~f∈~f_X. The notation in the displayed formula should be corrected.
  3. [§4.1, Metrics] The orbit-distance formula is confusingly written: if Z and ^Z are both learned latents, it is unclear which encoder outputs are being compared and why the denominator is n rather than the squared Frobenius norm without an extra factor. Please clarify the definition.
  4. [General] The paper contains several typos and incomplete phrases in the proof sketches, e.g., 'approximiated' and 'cofficients' in §3.3. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central identifiability theorem is grounded in external classical bivariate-distribution theory (Lancaster, Eagleson, Mehler) plus an explicit spectral-gap assumption; the only self-citation is used for experimental setup and framing, not in the proof.

full rationale

The claimed derivation chain is not circular. Theorem 1 is not obtained by fitting or defining the conclusion into the objective; it follows from a pushforward invariance argument (Proposition 1) and from classical external results on Gaussian canonical correlations and Hermite/Mehler expansions, with the additional spectral condition ρ_dS > ρ_1^2 (Assumption 2) used as a genuine sufficient condition. Proposition 1 is a change-of-variables statement, not a restatement of affine identifiability. Assumption 2 is explicit and empirically ablated; the ablation reports failure below the threshold, which is exactly what a non-circular sufficient condition should do. Theorem 2's A1 (realizability and orbit uniqueness) is a standard identifiability/regularity assumption for argmax consistency; because Theorem 1 is already assumed in Theorem 2, A1 is at most redundant rather than a circular substitute for the identifiability proof. The self-citation [44] appears in related-work framing and experimental setup, but it is not load-bearing in the proof of the main theorem, and the proof itself relies on Lancaster [36], Eagleson [17], and Mehler [45], which are external classical results rather than author-supplied uniqueness theorems. No prediction in the paper reduces by construction to a fitted parameter or to an author-imported uniqueness claim. The main proofs are deferred to the supplementary material; this is an omitted-details concern, not a circularity concern.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claim rests on classical probability results, a specific additive generative model, injective decoders, an idealized dense encoder class, and the spectral dominance assumption. The only hand-chosen numeric parameter in the empirical validation is the ridge schedule; no new entities are postulated.

free parameters (1)
  • ridge regularization schedule = ϵ = 0.01 · n^{-1/4}
    Hand-chosen schedule for empirical experiments (§4.2); not part of the theoretical derivation.
assumptions (5)
  • standard math Lancaster/Eagleson polynomial expansion results for bivariate Gaussian distributions [36, 17] hold as stated.
    Theorem 1's proof relies on these to argue first-order Hermite components dominate (§3.3 Sketch, and related work §2).
  • domain assumption The additive latent model s=a+c, s'=b+c with a,b,c i.i.d. (Equation 1) generates the data.
    Assumed throughout; the entire identifiability result is for this generative process (§3.1).
  • domain assumption The decoders g,g' are injective Borel-measurable maps, so the pushforward transformation is valid.
    Used in Proposition 1 to claim measurable inverses exist (§3.3).
  • ad hoc to paper The encoder classes are dense in L^2(P_x) and L^2(P_x') (universal approximators), enabling the transport to source space.
    Proposition 1.3; in practice neural networks are finite-dimensional, so exact maximizer characterizations rely on idealized density.
  • domain assumption Assumption 2 (first-order canonical dominance ρ_dS > ρ_1^2).
    This is the main new condition; the proof's dominance of first-order over second-order correlations depends on it (§3.3, Assumption 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Provable Affine Identifiability of Nonlinear CCA under Latent Distributional Priors." pith.science (2026). https://pith.science/paper/F26VFAVG

@misc{pith2026251004758,
  author       = {Pith},
  title        = {Pith review of: Provable Affine Identifiability of Nonlinear CCA under Latent Distributional Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F26VFAVG}},
  note         = {Machine review of arXiv:2510.04758}
}
read the original abstract

In this work, we establish the sufficient conditions under which nonlinear Canonical Correlation Analysis (CCA) recovers ground-truth latent factors up to an affine transformation. By transporting the analysis from the observation space to the source space, we extend classical statistical results on orthogonal polynomial expansions of bivariate distributions to representation learning, proving affine identifiability under specific distributional priors. We formally demonstrate that whitening is strictly necessary to ensure the boundedness and well-conditioning of the learned mappings. Furthermore, we bridge the gap between theory and practice by proving that ridge-regularized empirical CCA converges to its population counterpart in the finite-sample regime. Finally, our findings provide a rigorous theoretical foundation explaining the empirical success of recent correlation-based non-contrastive learning methods. Experiments on synthetic and rendered image datasets, alongside systematic ablations, validate the predicted recovery behavior and illustrate the failure modes that arise when the assumptions are violated.

Figures

Figures reproduced from arXiv: 2510.04758 by the authors.

Figure 2
Figure 2. The ℓ∞-gap in singular values and log orbit distance over training steps in Gaussian case. Shaded regions denote ±1 standard deviation across runs [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 1
Figure 1. Whitened latent learned by DeepCCA on synthetic data (dS = dZ = 2). Color gradients in different rows illustrate the variations along a single coordinate in S. tent learned by DeepCCA across four candidate dis￾tributions, omitting Hypergeometric due to collapse after whitening. Within each distribution, the two nearly perpendicular color gradients (one per source coordinate) are rotated relative to the x–y axes, and… view at source ↗
Figure 6
Figure 6. Log principal angles of encoder f in the under-complete setup (dS = 10, dZ = 7). Black dots denote principal angles and the shaded region indicates the log-standard deviation. over-complete setting confirm the recoverability of the source dimension, though the role of redundant coor￾dinates remains unclear. In contrast, under-complete CCA fails to recover the full dZ -dimensional subspace, though up to four canonica… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Ablation over the first order canonical ratio [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 6 linked inside Pith

  1. [1]

    Deep canonical correlation analy- sis

    Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analy- sis. InInternational conference on machine learn- ing, pages 1247–1255. PMLR, 2013

    2013

  2. [2]

    Kernel independent component analysis.Journal of ma- chine learning research, 3(Jul):1–48, 2002

    Francis R Bach and Michael I Jordan. Kernel independent component analysis.Journal of ma- chine learning research, 3(Jul):1–48, 2002

    2002

  3. [3]

    Randall Balestriero and Yann LeCun. Contrastive and non-contrastive self-supervised learning re- cover global and local spectral embedding meth- ods.Advances in Neural Information Processing Systems, 35:26671–26685, 2022

    2022

  4. [4]

    Vi- creg: Variance-invariance-covariance regulariza- tion for self-supervised learning.arXiv preprint arXiv:2105.04906, 2021

    Adrien Bardes, Jean Ponce, and Yann LeCun. Vi- creg: Variance-invariance-covariance regulariza- tion for self-supervised learning.arXiv preprint arXiv:2105.04906, 2021

    arXiv 2021

  5. [5]

    Finding minimum entropy codes.Neu- ral computation, 1(3):412–423, 1989

    Horace B Barlow, Tej P Kaushal, and Graeme J Mitchison. Finding minimum entropy codes.Neu- ral computation, 1(3):412–423, 1989

    1989

  6. [6]

    Representation learning: A review and new perspectives.IEEE transactions on pattern anal- ysis and machine intelligence, 35(8):1798–1828, 2013

    Yoshua Bengio, Aaron Courville, and Pascal Vin- cent. Representation learning: A review and new perspectives.IEEE transactions on pattern anal- ysis and machine intelligence, 35(8):1798–1828, 2013

    2013

  7. [7]

    Deep generalized canonical correla- tion analysis.arXiv preprint arXiv:1702.02519, 2017

    Adrian Benton, Huda Khayrallah, Biman Gu- jral, Dee Ann Reisinger, Sheng Zhang, and Ra- man Arora. Deep generalized canonical correla- tion analysis.arXiv preprint arXiv:1702.02519, 2017

    arXiv 2017

  8. [8]

    Provably learning object-centric representations

    Jack Brady, Roland S Zimmermann, Yash Sharma, Bernhard Sch¨ olkopf, Julius Von K¨ ugelgen, and Wieland Brendel. Provably learning object-centric representations. InInter- national Conference on Machine Learning, pages 3038–3062. PMLR, 2023

    2023

Show all 56 references
  1. [9]

    Weakly supervised causal representation learning.Advances in Neural In- formation Processing Systems, 35:38319–38331, 2022

    Johann Brehmer, Pim De Haan, Phillip Lippe, and Taco S Cohen. Weakly supervised causal representation learning.Advances in Neural In- formation Processing Systems, 35:38319–38331, 2022

    2022

  2. [10]

    Function classes for identifiable nonlinear independent component analysis.Ad- vances in Neural Information Processing Systems, 35:16946–16961, 2022

    Simon Buchholz, Michel Besserve, and Bern- hard Sch¨ olkopf. Function classes for identifiable nonlinear independent component analysis.Ad- vances in Neural Information Processing Systems, 35:16946–16961, 2022

    2022

  3. [11]

    Rate-optimal per- turbation bounds for singular subspaces with ap- plications to high-dimensional statistics

    T Tony Cai and Anru Zhang. Rate-optimal per- turbation bounds for singular subspaces with ap- plications to high-dimensional statistics. 2018

    2018

  4. [12]

    Blind signal separation: statistical principles.Proceedings of the IEEE, 86(10):2009– 2025, 1998

    J-F Cardoso. Blind signal separation: statistical principles.Proceedings of the IEEE, 86(10):2009– 2025, 1998

    2009

  5. [13]

    Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912– 9924, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912– 9924, 2020

    2020

  6. [14]

    Independent component analysis, a new concept?Signal processing, 36(3):287–314, 1994

    Pierre Comon. Independent component analysis, a new concept?Signal processing, 36(3):287–314, 1994

    1994

  7. [15]

    Identifiability results for multimodal contrastive learning.arXiv preprint arXiv:2303.09166, 2023

    Imant Daunhawer, Alice Bizeul, Emanuele Palumbo, Alexander Marx, and Julia E Vogt. Identifiability results for multimodal contrastive learning.arXiv preprint arXiv:2303.09166, 2023

    2023 arXiv

  8. [16]

    Compressed sensing.IEEE Transactions on information theory, 52(4):1289– 1306, 2006

    David L Donoho. Compressed sensing.IEEE Transactions on information theory, 52(4):1289– 1306, 2006

    2006

  9. [17]

    Polynomial expan- sions of bivariate distributions.The Annals of Mathematical Statistics, 35(3):1208–1215, 1964

    Geoffrey Kennedy Eagleson. Polynomial expan- sions of bivariate distributions.The Annals of Mathematical Statistics, 35(3):1208–1215, 1964

    1964

  10. [18]

    Whitening for self- supervised representation learning

    Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self- supervised representation learning. InInter- national conference on machine learning, pages 3015–3024. PMLR, 2021

    2021

  11. [19]

    Statistical consistency of kernel canon- ical correlation analysis.Journal of Machine Learning Research, 8(2), 2007

    Kenji Fukumizu, Francis R Bach, and Arthur Gretton. Statistical consistency of kernel canon- ical correlation analysis.Journal of Machine Learning Research, 8(2), 2007

    2007

  12. [20]

    Sparse cca: Adaptive estimation and computa- tional barriers

    Chao Gao, Zongming Ma, and Harrison H Zhou. Sparse cca: Adaptive estimation and computa- tional barriers. 2017

    2017

  13. [21]

    Independent mechanism analysis, a new concept?Advances in neural information process- ing systems, 34:28233–28248, 2021

    Luigi Gresele, Julius Von K¨ ugelgen, Vin- cent Stimper, Bernhard Sch¨ olkopf, and Michel Besserve. Independent mechanism analysis, a new concept?Advances in neural information process- ing systems, 34:28233–28248, 2021

    2021

  14. [22]

    Sparse canonical correlation analysis.Machine Learning, 83(3):331–353, 2011

    David R Hardoon and John Shawe-Taylor. Sparse canonical correlation analysis.Machine Learning, 83(3):331–353, 2011

    2011

  15. [23]

    beta- vae: Learning basic visual concepts with a con- strained variational framework

    Irina Higgins, Loic Matthey, Arka Pal, Christo- pher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta- vae: Learning basic visual concepts with a con- strained variational framework. InInternational conference on learning representations, 2017

    2017

  16. [24]

    Relations between two sets of variates.Biometrika, 28(3/4):321–377, 1936

    Harold Hotelling. Relations between two sets of variates.Biometrika, 28(3/4):321–377, 1936

    1936

  17. [25]

    Unsuper- vised feature extraction by time-contrastive learn- ing and nonlinear ica.Advances in neural infor- mation processing systems, 29, 2016

    Aapo Hyvarinen and Hiroshi Morioka. Unsuper- vised feature extraction by time-contrastive learn- ing and nonlinear ica.Advances in neural infor- mation processing systems, 29, 2016. Provable Affine Identifiability of Nonlinear CCA under Latent Distributional Priors

    2016

  18. [26]

    Nonlinear ica of temporally dependent stationary sources

    Aapo Hyvarinen and Hiroshi Morioka. Nonlinear ica of temporally dependent stationary sources. InArtificial intelligence and statistics, pages 460–

  19. [27]

    Independent component analysis: algorithms and applications

    Aapo Hyv¨ arinen and Erkki Oja. Independent component analysis: algorithms and applications. Neural networks, 13(4-5):411–430, 2000

    2000

  20. [28]

    Nonlinear independent component analysis: Existence and uniqueness results.Neural networks, 12(3):429– 439, 1999

    Aapo Hyv¨ arinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness results.Neural networks, 12(3):429– 439, 1999

    1999

  21. [29]

    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 conference on artificial intelligence and statistics, pages 859–868. PMLR, 2019

    2019

  22. [30]

    Revisiting deep generalized canonical correlation analysis.IEEE Transactions on Signal Process- ing, 71:4392–4406, 2023

    Paris A Karakasis and Nicholas D Sidiropoulos. Revisiting deep generalized canonical correlation analysis.IEEE Transactions on Signal Process- ing, 71:4392–4406, 2023

    2023

  23. [31]

    Variational au- toencoders and nonlinear ica: A unifying frame- work

    Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational au- toencoders and nonlinear ica: A unifying frame- work. InInternational conference on artificial in- telligence and statistics, pages 2207–2217. PMLR, 2020

    2020

  24. [32]

    Ice-beem: Identifi- able conditional energy-based deep models based on nonlinear ica.Advances in Neural Information Processing Systems, 33:12768–12778, 2020

    Ilyes Khemakhem, Ricardo Monti, Diederik Kingma, and Aapo Hyvarinen. Ice-beem: Identifi- able conditional energy-based deep models based on nonlinear ica.Advances in Neural Information Processing Systems, 33:12768–12778, 2020

    2020

  25. [33]

    Towards nonlinear disentanglement in natural data with temporal sparse coding.arXiv preprint arXiv:2007.10930, 2020

    David Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, and Dylan Paiton. Towards nonlinear disentanglement in natural data with temporal sparse coding.arXiv preprint arXiv:2007.10930, 2020

    2007 arXiv

  26. [34]

    Ad- ditive decoders for latent variables identifica- tion and cartesian-product extrapolation.Ad- vances in Neural Information Processing Systems, 36:25112–25150, 2023

    S´ ebastien Lachapelle, Divyat Mahajan, Ioan- nis Mitliagkas, and Simon Lacoste-Julien. Ad- ditive decoders for latent variables identifica- tion and cartesian-product extrapolation.Ad- vances in Neural Information Processing Systems, 36:25112–25150, 2023

    2023

  27. [35]

    Disentan- glement via mechanism sparsity regularization: A new principle for nonlinear ica

    S´ ebastien Lachapelle, Pau Rodriguez, Yash Sharma, Katie E Everett, R´ emi Le Priol, Alexan- dre Lacoste, and Simon Lacoste-Julien. Disentan- glement via mechanism sparsity regularization: A new principle for nonlinear ica. InConference on Causal Learning and Reasoning, pages...

    2022

  28. [36]

    The structure of bivari- ate distributions.The Annals of Mathematical Statistics, 29(3):719–736, 1958

    Henry Oliver Lancaster. The structure of bivari- ate distributions.The Annals of Mathematical Statistics, 29(3):719–736, 1958

    1958

  29. [37]

    Citris: Causal identifiability from temporal inter- vened sequences

    Phillip Lippe, Sara Magliacane, Sindy L¨ owe, Yuki M Asano, Taco Cohen, and Stratis Gavves. Citris: Causal identifiability from temporal inter- vened sequences. InInternational Conference on Machine Learning, pages 13557–13603. PMLR, 2022

    2022

  30. [38]

    Challenging com- mon assumptions in the unsupervised learning of disentangled representations

    Francesco Locatello, Stefan Bauer, Mario Lu- cic, Gunnar Raetsch, Sylvain Gelly, Bernhard Sch¨ olkopf, and Olivier Bachem. Challenging com- mon assumptions in the unsupervised learning of disentangled representations. Ininternational conference on machine learning, pages 4114–...

    2019

  31. [39]

    A sober look at the unsupervised learning of disentangled repre- sentations and their evaluation.Journal of Ma- chine Learning Research, 21(209):1–62, 2020

    Francesco Locatello, Stefan Bauer, Mario Lu- cic, Gunnar R¨ atsch, Sylvain Gelly, Bernhard Sch¨ olkopf, and Olivier Bachem. A sober look at the unsupervised learning of disentangled repre- sentations and their evaluation.Journal of Ma- chine Learning Research, 21(209):1–62, 2020

    2020

  32. [40]

    Weakly-supervised disentanglement without compromises

    Francesco Locatello, Ben Poole, Gunnar R¨ atsch, Bernhard Sch¨ olkopf, Olivier Bachem, and Michael Tschannen. Weakly-supervised disentanglement without compromises. InInternational conference on machine learning, pages 6348–6359. PMLR, 2020

    2020

  33. [41]

    Nonlinear multiview anal- ysis: Identifiability and neural network-assisted implementation.IEEE Transactions on Signal Processing, 68:2697–2712, 2020

    Qi Lyu and Xiao Fu. Nonlinear multiview anal- ysis: Identifiability and neural network-assisted implementation.IEEE Transactions on Signal Processing, 68:2697–2712, 2020

    2020

  34. [42]

    Understanding latent correlation-based multiview learning and self-supervision: An identifiability perspective.arXiv preprint arXiv:2106.07115, 2021

    Qi Lyu, Xiao Fu, Weiran Wang, and Songtao Lu. Understanding latent correlation-based multiview learning and self-supervision: An identifiability perspective.arXiv preprint arXiv:2106.07115, 2021

    2021 arXiv

  35. [43]

    Subspace perspec- tive on canonical correlation analysis: Dimension reduction and minimax rates

    Zhuang Ma and Xiaodong Li. Subspace perspec- tive on canonical correlation analysis: Dimension reduction and minimax rates. 2020

    2020

  36. [44]

    Towards a unified framework of contrastive learning for disentangled representations.Ad- vances in Neural Information Processing Systems, 36:67459–67470, 2023

    Stefan Matthes, Zhiwei Han, and Hao Shen. Towards a unified framework of contrastive learning for disentangled representations.Ad- vances in Neural Information Processing Systems, 36:67459–67470, 2023

    2023

  37. [45]

    Ueber die entwicklung einer function von beliebig vielen variablen nach laplaceschen functionen h¨ oherer ordnung

    F Gustav Mehler. Ueber die entwicklung einer function von beliebig vielen variablen nach laplaceschen functionen h¨ oherer ordnung. 1866

  38. [46]

    Nonparametric canonical correlation analysis

    Tomer Michaeli, Weiran Wang, and Karen Livescu. Nonparametric canonical correlation analysis. InInternational conference on machine learning, pages 1967–1976. PMLR, 2016

    1967

  39. [47]

    Inde- pendent component analysis with code informa- tion utilization in ds-cdma signal separation

    Tapani Ristaniemi and Jyrki Joutsensalo. Inde- pendent component analysis with code informa- tion utilization in ds-cdma signal separation. In Zhiwei Han, Steffan Matthes, Hao Shen Seamless Interconnection for Universal Services. Global Telecommunications Conference. GLOBE- CO...

    1999

  40. [48]

    Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    Bernhard Sch¨ olkopf, Francesco Locatello, Ste- fan Bauer, Nan Rosemary Ke, Nal Kalchbren- ner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    2021

  41. [49]

    Weakly super- vised disentangled generative causal representa- tion learning.Journal of Machine Learning Re- search, 23(241):1–55, 2022

    Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, and Tong Zhang. Weakly super- vised disentangled generative causal representa- tion learning.Journal of Machine Learning Re- search, 23(241):1–55, 2022

    2022

  42. [50]

    Weakly supervised disentanglement with guarantees.arXiv preprint arXiv:1910.09772, 2019

    Rui Shu, Yining Chen, Abhishek Kumar, Ste- fano Ermon, and Ben Poole. Weakly supervised disentanglement with guarantees.arXiv preprint arXiv:1910.09772, 2019

    1910 arXiv

  43. [51]

    The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells.Nature biotechnology, 32(4):381–386, 2014

    Cole Trapnell, Davide Cacchiarelli, Jonna Grimsby, Prapti Pokharel, Shuqiang Li, Michael Morse, Niall J Lennon, Kenneth J Livak, Tarjei S Mikkelsen, and John L Rinn. The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells.Natu...

    2014

  44. [52]

    Self- supervised learning with data augmentations provably isolates content from style.Advances in neural information processing systems, 34:16451– 16467, 2021

    Julius Von K¨ ugelgen, Yash Sharma, Luigi Gre- sele, Wieland Brendel, Bernhard Sch¨ olkopf, Michel Besserve, and Francesco Locatello. Self- supervised learning with data augmentations provably isolates content from style.Advances in neural information processing systems, 34:16...

    2021

  45. [53]

    Barlow twins: Self- supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann Le- Cun, and St´ ephane Deny. Barlow twins: Self- supervised learning via redundancy reduction. InInternational conference on machine learning, pages 12310–12320. PMLR, 2021

    2021

  46. [54]

    Identifiability guar- antees for causal disentanglement from soft inter- ventions.Advances in Neural Information Pro- cessing Systems, 36:50254–50292, 2023

    Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shan- mugam, and Caroline Uhler. Identifiability guar- antees for causal disentanglement from soft inter- ventions.Advances in Neural Information Pro- cessing Systems, 36:50254–50292, 2023

    2023

  47. [55]

    On the identifiability of nonlinear ica: Sparsity and beyond.Advances in neural information process- ing systems, 35:16411–16422, 2022

    Yujia Zheng, Ignavier Ng, and Kun Zhang. On the identifiability of nonlinear ica: Sparsity and beyond.Advances in neural information process- ing systems, 35:16411–16422, 2022

    2022

  48. [56]

    Contrastive learning inverts the data gener- ating process

    Roland S Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Bren- del. Contrastive learning inverts the data gener- ating process. InInternational conference on ma- chine learning, pages 12979–12990. PMLR, 2021

    2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.