Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

L-VAE: Variational Auto-Encoder with Learnable Beta for Disentangled Representation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A variational autoencoder that learns the weight of its own loss terms matches or beats hand-tuned beta-VAE.

desk verdict L-VAE is a simple Kendall-style weight learner for beta-VAE, but the 'consistently best or second best' claim contradicts the paper's own tables. read the letter →

arxiv 2507.02619 v1 pith:BIAZX77I submitted 2025-07-03 cs.LG cs.CV

classification cs.LGcs.CV
keywords variationalautoencoderdisentangledrepresentationlearninglearnablebetauncertaintyweightingbeta-VAEmulti-taskhyperparameteroptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that the $\beta$ hyperparameter of $\beta$-VAE, the weight balancing reconstruction quality against disentanglement in the loss, does not need to be tuned by hand. The proposed L-VAE learns two scale parameters sigma_0 and sigma_1 alongside the network weights, so that the ratio $sigma_0^{2}$/$sigma_1^{2}$ plays the role of $\beta$ automatically. Across dSprites, MPI3D-complex, Falcor3D, and Isaac3D, the paper reports that L-VAE gives the best or second-best scores on six disentanglement metrics, while also finding that the optimal effective $\beta$ is below 1 on some datasets. If correct, this removes the expensive $\beta$ search that practice requires, and it makes the trade-off adapt to the data rather than being fixed in advance.

What carries the argument

The mechanism is Kendall et al.'s uncertainty weighting for multi-task losses, transplanted from Gaussian log-likelihood tasks to the two terms of the VAE objective. Two learnable scalars sigma_0 and sigma_1 divide the reconstruction and KL terms, with a regularization term, here $sigma_0^{2}$ + $sigma_1^{2}$, that stops the weights from growing without bound; the ratio $sigma_0^{2}$/$sigma_1^{2}$ is the effective $\beta$. This turns hyperparameter search into an ordinary optimization problem, since the sigma_i sit in the optimizer's parameter set and are updated with the network weights. The paper's argument relies on the learned ratio being aligned with the empirically tuned $\beta$ of $\beta$-VAE.

What would settle it

Train L-VAE and $\beta$-VAE across a fine grid of $\beta$ on a fixed benchmark: if the learned ratio $sigma_0^{2}$/$sigma_1^{2}$ does not track the $\beta$ value that maximizes the chosen disentanglement metric, or a static $\beta$-VAE trained at the learned ratio matches L-VAE everywhere, the dynamic-benefit claim would fail. A direct check of the derivation is to train with the original log-$\sigma$ regularizer of Eq. 4 instead of the $\sigma$-squared term of Eq. 5 and compare the learned weights and scores.

Watch

Extended reading notes

Core claim

The paper's central claim is that the relative weight between the reconstruction term and the KL-divergence term of the ELBO can be learned by gradient descent instead of being set by search, and that doing so preserves or improves disentanglement while lowering reconstruction loss. L-VAE is $\beta$-VAE with the loss rewritten as (1/$sigma_0^{2}$) L_R + (1/$sigma_1^{2}$) D_KL + $sigma_0^{2}$ + $sigma_1^{2}$, where sigma_0 and sigma_1 are optimized together with the encoder and decoder. The learned ratio $\beta$-hat = $sigma_0^{2}$/$sigma_1^{2}$ lands where the optimal tuned $\beta$ lands in their experiments: 1.39 on dSprites, 1.05 on MPI3D, 2.34 on Falcor3D, and 0.95 on Isaac3D. An ablation that freezes this ratio into a plain $\beta$-VAE gives similar or worse disentanglement than the dynamic version, which the paper reads as evidence that continuously learning the weights, not just the final ratio, is doing the work.

Load-bearing premise

The load-bearing premise is that the uncertainty-weighting derivation, made for combining Gaussian log-likelihood task losses, still holds when one of the two terms is the KL divergence of the ELBO, and that replacing the derivation's log-sigma regularizer with sigma-squared leaves the learned trade-off intact.

Editorial extensions

If this is right

  • Users of beta-VAE can drop the beta grid search and initialize sigma_0 = sigma_1 = 1, letting the optimizer find the trade-off.
  • Datasets where low beta is better, such as MPI3D and Isaac3D, are handled automatically because the learned ratio can go below 1 without special-casing.
  • Because freezing the learned ratio into a static beta-VAE gives similar or slightly worse quality, continually learning the weights appears to be part of the gain, not just the final value.
  • The same parameterization transfers to other two-term losses, since the paper's derivation is not specific to VAEs.
  • Across the six reported metrics no single method wins everywhere, but L-VAE is best or second best on the aggregate, which is the paper's stated comparative claim.

Reading between the lines

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

  • The observation that beta below 1 can beat beta above 1 on real-image datasets is a testable challenge to the common assumption that stronger KL weighting always helps; applying L-VAE to new datasets would cheaply map where that assumption fails.
  • The paper swaps the log-sigma regularization of the source derivation for a sigma-squared term without justification, so an ablation of that regularizer choice would show whether the learned ratios are sensitive to it.
  • Model selection in the experiments uses the beta-VAE measure on the validation set, so rankings might shift under other selection criteria; reporting L-VAE's performance across selection metrics would test the strength of the comparison.
  • All sigma_i are initialized to 1 and learned with a OneCycle schedule, so L-VAE's dynamic behavior may interact with optimization schedules such as KL annealing; combining L-VAE with such schedules is a natural extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes L-VAE, a variational auto-encoder that learns the relative weight between the reconstruction term and the KL term by introducing two trainable parameters, σ0 and σ1, together with a quadratic regularizer. The effective β is defined as β̂ = σ0^2/σ1^2. The method is compared with VAE, β-VAE, ControlVAE, DynamicVAE, and σ-VAE on dSprites, MPI3D, Falcor3D, and Isaac3D using six disentanglement metrics and reconstruction loss, plus qualitative latent traversals on CelebA. The paper claims that L-VAE removes the need for empirical β search and consistently achieves the best or second-best performance among the compared methods, and that the learned β̂ aligns with the empirically tuned β.

Significance. If the claims were fully supported, the contribution would be practically useful: a simple, auto-differentiable way to replace the β search in β-VAE. The paper has genuine strengths: a systematic sensitivity analysis of β in Section 4.1 (Observations 1–4, Fig. 2), a broad evaluation across four datasets and six disentanglement metrics, and inclusion of several adaptive-β baselines. However, the central empirical claim is not established by the reported single-run tables, and the ablation in Table 3 does not support the importance of dynamic weighting. The significance of the paper is therefore conditional on substantial revision and additional evidence.

major comments (3)
  1. [Abstract and Section 1; Table 2] The abstract and the contributions list state that L-VAE 'consistently provides the best or the second best performances', but Table 2 contradicts this. On dSprites MIG, L-VAE scores 0.30, behind ControlVAE (0.34) and DynamicVAE (0.31), placing it third together with β-VAE. On Falcor3D MIG, L-VAE (0.05) trails β-VAE (0.07), ControlVAE (0.06), and DynamicVAE (0.06); on Falcor3D SAP it is tied for third with β-VAE (0.05) behind ControlVAE (0.07) and DynamicVAE (0.06). Section 6.3 reports hyperparameter searches but no multiple seeds, and Table 2 gives no error bars or standard deviations, so the small differences cannot be distinguished from noise. Section 7.2 uses the softer statement that L-VAE 'generally produces better or on par performance', which is consistent with the table but not with the abstract. The headline claim needs to be corrected or supported by repeated runs with statistical measures.
  2. [Section 5, Eqs. (4)-(5)] The derivation from Kendall et al. is not carried over correctly. Equation (4) is presented as the Kendall et al. multi-task loss with log(σ0) + log(σ1) regularizers, but Equation (5) replaces these with σ0^2 + σ1^2 without any derivation or justification. This is not a minor notational change: the two regularizers lead to different optimization landscapes and different interpretations of the learned σi. Moreover, β̂ = σ0^2/σ1^2 is by definition the effective weight on the KL term in Eq. (5), so the reported 'alignment' between β̂ and a searched β is a consistency check on fitted values, not an independent prediction. The fitted ratios (1.39, 1.05, 2.34, 0.95) differ from the searched β values (2 or 4) by up to about a factor of two, which further weakens the alignment claim in Section 7.2. The authors should either derive Eq. (5) from a specific noise model or explicitly state that the σi^2 penalty is a heuristic regularizer.
  3. [Section 7.3, Table 3] The ablation study does not support the claim that dynamic weighting is important. For Falcor3D, a standalone β-VAE trained with the learned β̂ = 2.34 outperforms L-VAE on β-VAE score (0.93 vs 0.88), FactorVAE (0.65 vs 0.47), IRS (0.35 vs 0.30), MIG (0.12 vs 0.05), and SAP (0.09 vs 0.05). For MPI3D, β-VAE with β̂ = 1.05 achieves equal or better scores on all six disentanglement metrics (e.g., β-VAE 0.74 vs 0.71, FactorVAE 0.46 tie, MIG 0.17 vs 0.16, SAP 0.21 vs 0.20) and a lower reconstruction loss (10.63 vs 10.79). The conclusion in Section 9 states that β-VAE with β̂ achieves better disentanglement 'with the cost of an increased reconstruction loss', but this is contradicted by the MPI3D row and by the dSprites row (reconstruction 18.13 vs 21.14). These results undercut the paper's central novelty, which is the benefit of learning the weights dynamically instead of fixing them.
minor comments (4)
  1. [Section 6.3] The text 'Kp to 0.0.01' appears to be a typo; presumably the intended value is 0.001 as in the ControlVAE baseline.
  2. [Title page and Figure 4] The affiliation contains the typo 'Universty' and the Figure 4 caption contains 'experimens'; these should be corrected.
  3. [Section 6.5] The text refers to 'Equation 6' for the overall loss, but the loss is given in Eq. (5) and the regularization term is not separately numbered. Cross-references to equations should be fixed.
  4. [Section 7.1] The statement that 'we select the σi values that maximize the β-VAE score' should clarify that this selection is performed on the validation set (as described in Section 6.5) and that the reported results are on a held-out test set, to avoid any appearance of selecting on the test metric.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; the learned-beta correspondence is a definitional consistency check, while the performance claims rest on independent disentanglement metrics.

  1. self definitional [Section 7.1 and Section 7.2, Eq. 5]
    "In all experiments, we select the σi values that maximize the β-VAE score. ... During the derivations of L-VAE, we mentioned that there is a correspondence between the empirically tuned β parameter of β-VAE model and the optimal ratio of the learned parameter of L-VAE model, β̂ = σ0^2/σ1^2."

    β̂ is not independently predicted: it is the effective KL coefficient obtained by rewriting Eq. 5 as reconstruction + (σ0^2/σ1^2)·DKL plus the regularizer. The paper's 'correspondence' with the beta-VAE hyperparameter is therefore a consistency check between two quantities that are each selected to maximize the same β-VAE score, not a derived result. This is a presentational circularity only; the central performance claims rest on external metrics in Table 2 and are not reduced by construction.

full rationale

Walking the derivation chain: Eq. 5 defines L-VAE's loss with learned weights 1/σi^2, so the effective beta is β̂ = σ0^2/σ1^2 by construction. The alignment reported in Sections 7.1 and 7.2 is thus a consistency check between a fitted ratio and a searched hyperparameter, not a prediction or first-principles result. It is minor and not load-bearing, because the paper's actual performance claims are assessed with six external disentanglement metrics and reconstruction loss in Table 2, plus qualitative CelebA traversals. No self-citations are used as load-bearing evidence; the only imported method, Kendall et al., is external, and the replacement of log σ regularization by σ^2 is an ansatz/design choice that raises correctness questions rather than circularity. The abstract's 'best or second best' phrasing is stronger than Table 2, but that contradiction is an empirical-support issue, not a circular derivation. Overall, the central derivation is self-contained and externally benchmarked, so the score reflects only the minor definitional beta-hat correspondence.

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

The paper introduces no new entities. Its free parameters are the two learned scale weights and the implicit regularizer coefficient, all of which are fitted to data. The central derivation hinges on the domain assumption that Kendall et al.'s weighting applies to a KL term and that the sigma^2 penalty behaves like the theoretically motivated log-sigma penalty without further justification.

free parameters (4)
  • sigma_0 (reconstruction loss weight) = learned per dataset; beta_hat ratio sigma_0^2/sigma_1^2 examples: 1.39 (dSprites), 1.05 (MPI3D), 2.34 (Falcor3D), 0.95…
    Learned via backpropagation alongside network weights; the effective beta is defined as the ratio of the two learned scales.
  • sigma_1 (KL divergence weight) = learned per dataset; see sigma_0 ratio examples
    Learned via backpropagation simultaneously with sigma_0.
  • Regularization coefficient of the sigma^2 penalty = 1 (implicit)
    The penalty in Eq. 5 is written as sum sigma_i^2 with an implicit coefficient of 1; this coefficient is not swept or justified, and changing it would alter the learned beta.
  • Initial value of sigma_i = 1 (for all experiments)
    Initialization is a hand-chosen value that affects the optimization path and final learned weights; no sensitivity analysis is provided.
assumptions (3)
  • domain assumption Kendall et al.'s uncertainty weighting formalism is valid for the ELBO loss with a KL-divergence term.
    Section 5 applies Eq. 4 (from Kendall et al., originally for Gaussian log-likelihood tasks) to the VAE loss without proof that it extends to a KL-divergence term.
  • ad hoc to paper The sigma^2 penalty is a suitable regularizer that preserves the intended beta trade-off.
    Eq. 4 (from Kendall et al.) uses log(sigma) terms, but Eq. 5 uses sigma^2; the replacement is announced in the text but never derived or ablated.
  • domain assumption The beta-VAE metric is a valid model-selection criterion for all compared methods.
    Section 6.5 selects hyperparameters based on the beta-VAE measure, admitting that metric choice biases results, but applies it uniformly without validating this choice against other criteria.

how reviews work

0 comments
Cite this review

Pith. "Pith review of L-VAE: Variational Auto-Encoder with Learnable Beta for Disentangled Representation." pith.science (2026). https://pith.science/paper/BIAZX77I

@misc{pith2026250702619,
  author       = {Pith},
  title        = {Pith review of: L-VAE: Variational Auto-Encoder with Learnable Beta for Disentangled Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIAZX77I}},
  note         = {Machine review of arXiv:2507.02619}
}
read the original abstract

In this paper, we propose a novel model called Learnable VAE (L-VAE), which learns a disentangled representation together with the hyperparameters of the cost function. L-VAE can be considered as an extension of \b{eta}-VAE, wherein the hyperparameter, \b{eta}, is empirically adjusted. L-VAE mitigates the limitations of \b{eta}-VAE by learning the relative weights of the terms in the loss function to control the dynamic trade-off between disentanglement and reconstruction losses. In the proposed model, the weight of the loss terms and the parameters of the model architecture are learned concurrently. An additional regularization term is added to the loss function to prevent bias towards either reconstruction or disentanglement losses. Experimental analyses show that the proposed L-VAE finds an effective balance between reconstruction fidelity and disentangling the latent dimensions. Comparisons of the proposed L-VAE against \b{eta}-VAE, VAE, ControlVAE, DynamicVAE, and {\sigma}-VAE on datasets, such as dSprites, MPI3D-complex, Falcor3D, and Isaac3D reveals that L-VAE consistently provides the best or the second best performances measured by a set of disentanglement metrics. Moreover, qualitative experiments on CelebA dataset, confirm the success of the L-VAE model for disentangling the facial attributes.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Variational Sparse Paired Autoencoders (vsPAIR) for Inverse Problems and Uncertainty Quantification

    cs.LG 2026-02 conditional novelty 5.0 of 10

    vsPAIR couples a Gaussian VAE over observations with a spike-and-slab sparse VAE over the quantity of interest via a learned latent mapping, yielding fast inverse reconstructions whose active latent dimensions can be ...

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    ISA-The Instrumentation, Sys- tems and Automation Society 14

    ˚Astr¨ om KJ, H¨ agglund T (2006) Advanced PID control. ISA-The Instrumentation, Sys- tems and Automation Society 14

  2. [2]

    IEEE transactions on pattern analysis and machine intelligence 35(8):1798– 1828

    Bengio Y, Courville A, Vincent P (2013) Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35(8):1798– 1828

  3. [3]

    arXiv preprint arXiv:180403599

    Burgess CP, Higgins I, Pal A, et al (2018) Understanding disentangling in β-vae. arXiv preprint arXiv:180403599

  4. [4]

    IEEE Transactions on Neural Networks and Learning Systems

    Carbonneau MA, Zaidi J, Boilard J, et al (2022) Measuring disentanglement: A review of metrics. IEEE Transactions on Neural Networks and Learning Systems

  5. [5]

    In: Advances in Neural Information Processing Systems, pp 2610– 2620

    Chen RT, Li X, Grosse RB, et al (2018) Isolating sources of disentanglement in varia- tional autoencoders. In: Advances in Neural Information Processing Systems, pp 2610– 2620

  6. [6]

    In: Advances in neural infor- mation processing systems, pp 2172–2180

    Chen X, Duan Y, Houthooft R, et al (2016) Infogan: Interpretable representation learn- ing by information maximizing generative adversarial nets. In: Advances in neural infor- mation processing systems, pp 2172–2180

  7. [7]

    In: Advances in neural information processing systems, pp 4414–4423

    Denton EL, et al (2017) Unsupervised learn- ing of disentangled representations from video. In: Advances in neural information processing systems, pp 4414–4423

  8. [8]

    arXiv preprint arXiv:190809961

    Do K, Tran T (2019) Theory and evaluation metrics for learning disentangled representa- tions. arXiv preprint arXiv:190809961

Show all 44 references
  1. [9]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, pp 10729–10736

    Dong Y, Liu Y, Zhang H, et al (2020) Fd- gan: Generative adversarial networks with fusion-discriminator for single image dehaz- ing. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp 10729–10736

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 7930–7938

    Duan B, Fu C, Li Y, et al (2020) Cross- spectral face hallucination via disentangling independent factors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 7930–7938

  3. [11]

    arXiv preprint arXiv:211214278

    Fil M, Mesinovic M, Morris M, et al (2021) beta-vae reproducibility: Challenges and extensions. arXiv preprint arXiv:211214278

  4. [12]

    Image and Vision Computing 135:104686

    Fu Z, Chen X, Liu D, et al (2023) Multi-level feature disentanglement network for cross- dataset face forgery detection. Image and Vision Computing 135:104686

  5. [13]

    Advances in Neural Information Processing Systems 32

    Gondal MW, Wuthrich M, Miladinovic D, et al (2019) On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Advances in Neural Information Processing Systems 32

  6. [14]

    In: International conference on learning repre- sentations

    Higgins I, Matthey L, Pal A, et al (2016) beta-vae: Learning basic visual concepts with a constrained variational framework. In: International conference on learning repre- sentations

  7. [15]

    arXiv preprint arXiv:181202230

    Higgins I, Amos D, Pfau D, et al (2018) Towards a definition of disentangled represen- tations. arXiv preprint arXiv:181202230

  8. [16]

    In: Advances in Neural Information Processing Systems, pp 517–526

    Hsieh JT, Liu B, Huang DA, et al (2018) Learning to decompose and disentangle representations for video prediction. In: Advances in Neural Information Processing Systems, pp 517–526

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 4401–4410

    Karras T, Laine S, Aila T (2019) A style- based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 4401–4410

  10. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 8110–8119

    Karras T, Laine S, Aittala M, et al (2020) Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 8110–8119

  11. [19]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 7482–7491

    Kendall A, Gal Y, Cipolla R (2018) Multi- task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 7482–7491

  12. [20]

    arXiv preprint arXiv:180205983

    Kim H, Mnih A (2018) Disentangling by factorising. arXiv preprint arXiv:180205983

  13. [21]

    arXiv preprint arXiv:14126980 15

    Kingma DP, Ba J (2014) Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980 15

  14. [22]

    arXiv preprint arXiv:13126114

    Kingma DP, Welling M (2013) Auto- encoding variational bayes. arXiv preprint arXiv:13126114

  15. [23]

    arXiv preprint arXiv:171100848

    Kumar A, Sattigeri P, Balakrishnan A (2017) Variational inference of disentangled latent concepts from unlabeled observations. arXiv preprint arXiv:171100848

  16. [24]

    International Journal of Computer Vision 128(10):2402–2417

    Lee HY, Tseng HY, Mao Q, et al (2020) Drit++: Diverse image-to-image trans- lation via disentangled representations. International Journal of Computer Vision 128(10):2402–2417

  17. [25]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 2080–2089

    Liu Y, Wei F, Shao J, et al (2018) Exploring disentangled feature representation beyond face identification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 2080–2089

  18. [26]

    In: Pro- ceedings of International Conference on Com- puter Vision (ICCV)

    Liu Z, Luo P, Wang X, et al (2015) Deep learning face attributes in the wild. In: Pro- ceedings of International Conference on Com- puter Vision (ICCV)

  19. [27]

    In: international conference on machine learning, pp 4114–4124

    Locatello F, Bauer S, Lucic M, et al (2019) Challenging common assumptions in the unsupervised learning of disentangled rep- resentations. In: international conference on machine learning, pp 4114–4124

  20. [28]

    In: International Conference on Machine Learning, PMLR, pp 6348–6359

    Locatello F, Poole B, R¨ atsch G, et al (2020) Weakly-supervised disentanglement without compromises. In: International Conference on Machine Learning, PMLR, pp 6348–6359

  21. [29]

    https://github.com/deepmind/dsprites- dataset/

    Matthey L, Higgins I, Hassabis D, et al (2017) dsprites: Disen- tanglement testing sprites dataset. https://github.com/deepmind/dsprites- dataset/

  22. [30]

    Nguyen-Phuoc T, Li C, Theis L, et al (2019) HoloGAN: Unsupervised learning of 3d rep- resentations from natural images. arXiv

  23. [31]

    In: Proceedings of the 37th Interna- tional Conference on Machine Learning, pp 7360–7369

    Nie W, Karras T, Garg A, et al (2020) Semi-supervised stylegan for disentanglement learning. In: Proceedings of the 37th Interna- tional Conference on Machine Learning, pp 7360–7369

  24. [32]

    ACM Transactions on Graphics (TOG) 39(6):1–14

    Nitzan Y, Bermano A, Li Y, et al (2020) Face identity disentanglement via latent space mapping. ACM Transactions on Graphics (TOG) 39(6):1–14

  25. [33]

    ACM Computing Surveys (CSUR) 51(5):1–36

    Pouyanfar S, Sadiq S, Yan Y, et al (2018) A survey on deep learning: Algorithms, tech- niques, and applications. ACM Computing Surveys (CSUR) 51(5):1–36

  26. [34]

    Advances in neural information process- ing systems 31

    Ridgeway K, Mozer MC (2018) Learning deep disentangled embeddings with the f-statistic loss. Advances in neural information process- ing systems 31

  27. [35]

    In: International Conference on Machine Learning, PMLR, pp 9179–9189

    Rybkin O, Daniilidis K, Levine S (2021) Simple and effective vae training with cali- brated decoders. In: International Conference on Machine Learning, PMLR, pp 9179–9189

  28. [36]

    Proceedings of the IEEE 109(5):612–634

    Sch¨ olkopf B, Locatello F, Bauer S, et al (2021) Toward causal representation learn- ing. Proceedings of the IEEE 109(5):612–634

  29. [37]

    In: International Conference on Machine Learning, PMLR, pp 8655–8664

    Shao H, Yao S, Sun D, et al (2020) Con- trolvae: Controllable variational autoencoder. In: International Conference on Machine Learning, PMLR, pp 8655–8664

  30. [38]

    IEEE transactions on pattern analysis and machine intelligence 44(12):9285–9297

    Shao H, Xiao Z, Yao S, et al (2021) Con- trolvae: Tuning, analytical properties, and performance analysis. IEEE transactions on pattern analysis and machine intelligence 44(12):9285–9297

  31. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 19250–19259

    Shao H, Yang Y, Lin H, et al (2022) Rethink- ing controllable variational autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 19250–19259

  32. [40]

    In: International Conference on Machine Learning, PMLR, pp 6056–6065

    Suter R, Miladinovic D, Sch¨ olkopf B, et al (2019) Robustly disentangled causal mech- anisms: Validating deep representations for interventional robustness. In: International Conference on Machine Learning, PMLR, pp 6056–6065

  33. [41]

    In: Proceedings of the IEEE conference on computer vision and pattern 16 recognition, pp 1415–1424

    Tran L, Yin X, Liu X (2017) Disentangled representation learning gan for pose-invariant face recognition. In: Proceedings of the IEEE conference on computer vision and pattern 16 recognition, pp 1415–1424

  34. [42]

    In: Proceed- ings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pp 9593–9602

    Yang M, Liu F, Chen Z, et al (2021) Causal- vae: disentangled representation learning via neural structural causal models. In: Proceed- ings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pp 9593–9602

  35. [43]

    Communications of the ACM 64(3):107–115

    Zhang C, Bengio S, Hardt M, et al (2021) Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3):107–115

  36. [44]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 6538–6547 17

    Zhu Y, Min MR, Kadav A, et al (2020) S3vae: Self-supervised sequential vae for representa- tion disentanglement and data generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 6538–6547 17

Pith tools

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