Pith. sign in

REVIEW 4 major objections 4 minor 36 references

Disentangling Granularity: An Implicit Inductive Bias in Factorized VAEs

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Factorized VAEs carry a fixed 'disentangling granularity' that biases them toward simple features; a tunable variant proposed here, β-STCVAE, broadens disentanglement and reaches higher ELBOs.

desk verdict The paper's central 'V-shaped optimal ELBO' trajectory and the claim that β-STCVAE beats β-TCVAE on the ELBO are artifacts of comparing a modified objective that drops nonnegative within-block total-correlation terms, not real representation gains. read the letter →

arxiv 2505.24684 v1 pith:VJVSWZHX submitted 2025-05-30 cs.LG cs.AI

Zihao Chen , Yu Xiang , Wenyong Wang This is my paper
classification cs.LGcs.AI
keywords variationalautoencoderdisentangledrepresentationlearningimplicitinductivebiastotalcorrelationfactorizationpriorβ-TCVAEβ-STCVAEdisentanglinggranularity
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 establish why factorized variational autoencoders can disentangle at all, given the known result that unsupervised disentanglement is impossible without some implicit inductive bias. Its answer is 'disentangling granularity': the scale at which the total-correlation penalty asks latent variables to be independent. The authors show that standard factorized VAEs such as β-TCVAE fix this scale at the finest possible level, biasing them toward low-complexity features, and they introduce β-STCVAE, which replaces the total-correlation term with a block-level version so the granularity can be tuned. Across more than 100K trained models they report a consistent V-shaped trajectory of the optimal ELBO in the plane of network capacity and granularity, with β-STCVAE reaching higher ELBOs and better disentanglement scores than β-TCVAE. If the claim is right, disentangling granularity is a controllable prior-side dial for representation learning, not an accidental property of the factorization.

What carries the argument

The load-bearing object is the iterative total-correlation decomposition, $$\mathcal{TC}(z) = \sum_{b} \mathcal{MU}_{\mathrm{joint},b}(z) + I(z_{f_1}; z_{f_2}),$$ which separates the independence penalty into within-block mutual information and a residual joint-distribution total correlation. From this the paper defines the disentangling granularity $g = \hat b / m$, where $\hat b$ is a grouping factor dividing the number of latent variables and $m$ is the largest such divisor; $g$ ranges from $1/m$ (the β-TCVAE limit, finest granularity) to $1$ (coarsest). The β-STCVAE loss substitutes $\mathcal{TC}_{\mathrm{joint},\hat b}(z)$ for the full $\mathcal{TC}(z)$, dropping the within-block mutual-information penalties. That substitution carries the argument: it converts a hidden bias into a tunable hyperparameter and, according to the paper, produces the V-shaped optimal-ELBO trajectory.

What would settle it

Estimate the marginal log-likelihood of held-out data for β-STCVAE and β-TCVAE under identical architecture and training budget using a granularity-independent estimator such as annealed importance sampling; if the empirical V-shaped trajectory and the β-STCVAE-over-β-TCVAE advantage disappear with this estimator, the claims describe the modified training loss rather than representation quality.

Watch

Extended reading notes

Core claim

The paper's central claim is that the factorization prior in β-TCVAE carries an implicit inductive bias—disentangling granularity—that fixes the coarseness of independence enforcement and thereby decides which feature complexities the model can disentangle. The total-correlation term $\mathcal{TC}(z)$ is iteratively decomposed into within-block mutual-information terms and a joint-distribution total correlation; this decomposition shows that setting the granularity to its finest value, as β-TCVAE does, targets only low-complexity features. The proposed β-STCVAE replaces $\mathcal{TC}(z)$ with $\mathcal{TC}_{\mathrm{joint},\hat b}(z)$, making granularity tunable, and the paper reports that the best ELBO over capacity and granularity forms a V-shaped curve on five datasets. The conclusion is that granularity should be coarsened when capacity is low, coarsened again when capacity is very high, and coarsened when the disentangling strength β is large.

Load-bearing premise

The load-bearing premise is that the β-STCVAE objective in Eq. (7) is still a valid ELBO and can be compared across different granularities; if dropping the within-block mutual-information terms simply inflates the objective by construction, the V-shaped optimal-ELBO trajectory and the reported gains would be artifacts of the loss definition rather than signs of better learned representations.

Editorial extensions

If this is right

  • Standard factorized VAEs are biased toward low-complexity features; increasing capacity alone does not remove that bias.
  • Tuning disentangling granularity can raise the ELBO without changing the network architecture, so model capacity and prior granularity should be treated as a joint axis.
  • At low capacity and at very high capacity, coarser granularity is optimal; at intermediate capacity the finest granularity is optimal, producing the V shape.
  • Stronger disentangling strength should be paired with coarser granularity to keep the model from over-constraining individual latent variables.
  • Sufficiently coarse granularity combined with high capacity can collapse a latent variable into an 'omniscient' near-zero-entropy unit, a failure mode located in the upper-right region of the parameter space.

Reading between the lines

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

  • Going beyond the paper, the grouping structure of the prior should become a first-class hyperparameter in disentanglement benchmarks, reported alongside latent dimension and β.
  • The V-shaped trajectory implies a testable law: the optimal grouping factor should track the intrinsic complexity of a dataset's generative factors, which would allow granularity to be scheduled during training rather than fixed.
  • The omniscient-latent-variable observation can serve as a practical diagnostic: a near-zero marginal entropy in one latent signals that granularity is too coarse for the current capacity, before any metric is computed.
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 proposes that factorized VAEs possess an implicit inductive bias called 'disentangling granularity,' arising from the total-correlation penalty in β-TCVAE. The authors derive an iterative decomposition of total correlation into pairwise mutual-information terms and a 'joint distribution total correlation,' and propose a new model, β-STCVAE, which replaces the full TC penalty with a block-wise joint TC penalty, thereby allowing the granularity of the independence constraint to be tuned. They report a V-shaped 'optimal ELBO trajectory' in the parameter space of network capacity versus granularity across over 100K experiments, and claim that β-STCVAE outperforms β-TCVAE both in ELBO and in disentanglement metrics. The paper argues that conventional factorized VAEs, with fixed granularity, tend to disentangle only low-complexity features, while tunable granularity enables disentanglement of high-complexity features.

Significance. If the central claims were correct, the paper would offer a practically useful mechanism for controlling the inductive bias in disentangled VAEs and a plausible explanation for the behavior of factorized priors. The authors are right to focus on the total-correlation term and to explore the effect of grouping structure in the latent space, and the paper contains a large amount of experimental effort. However, the main theoretical decomposition is not a valid identity for the true posterior, and the headline empirical claim—that β-STCVAE achieves higher ELBO than β-TCVAE—is substantially manufactured by the loss definition, since the training objective in Eq. (7) drops nonnegative terms without separate evaluation of the true ELBO. The disentanglement results provide some independent evidence of improved grouping, but the lack of rigorous derivation and the confounded ELBO comparison prevent the findings from being established.

major comments (4)
  1. [Section 3.1, Eq. (7)] The β-STCVAE objective is inflated relative to the β-TCVAE objective by construction. For any posterior q, the total correlation decomposes as TC(z) = TC_joint_bhat(z) + Σ_j TC(z_{B_j}), where the within-block terms are nonnegative. Substituting this into Eq. (7) gives L_STC = L_TCVAE + β Σ_j TC(z_{B_j}) for the same q and β. The paper never reports the omitted within-block terms or a separately evaluated true ELBO, so the 'higher ELBO' claim in Section 4.2 and the V-shaped optimal trajectory are not evidence of improved inference or representation quality; they may simply reflect the mechanical removal of penalty terms.
  2. [Section 2, Eq. (2) and Appendix A.1] The decomposition of TC(z) is not a valid identity under the true posterior q(z). The term I(z_i;z_j)|_{q(z)=q(z_i,z_j)q(z_{-i,j})} is defined as a mutual information computed under an independence assumption that does not hold for q(z). The derivation in Appendix A.1, Eqs. (7)-(8), shows that E_q(z)[log(q(z_i,z_j)/(q(z_i)q(z_j)))] equals ∫ q(z_i,z_j) q(z_{-i,j}|z_i,z_j) log(...) , which is not the mutual information I(z_i;z_j). Consequently, the 'always positive' claim for MU_joint and the inequality TC_joint_1(z) ≤ TC(z) are not established, and the iterative decomposition leading to Eq. (4) lacks a rigorous basis.
  3. [Section 4.2] The V-shaped optimal ELBO trajectory is determined by fitting binary functions to the granularity that maximizes the modified objective per network capacity, but the paper provides no error bars, confidence intervals, or statistical significance tests. Moreover, the red dotted reference line for β-TCVAE is the average of 1/m over different latent dimensions (reported as 0.0892), which mixes incomparable granularity scales because the set of possible granularity values depends on n. Without a separately evaluated held-out ELBO and with no measure of uncertainty, the V-shape is not demonstrated as a robust phenomenon across datasets.
  4. [Section 4.6 and Fig. 6] The comparison with HFVAE, ControlVAE, GuidedVAE, and DisCo lacks crucial experimental details: the hyperparameters, training budgets, model capacities, and selection procedures for these baselines are not fully specified, and the reported high variance of β-STCVAE makes the averaged MIG/Factor/SAP improvements unconvincing. In particular, the claim that β-STCVAE 'significantly improved' over factorized VAEs needs accompanying statistical tests or at least paired error bars.
minor comments (4)
  1. [Abstract] The word 'Keyords' in the manuscript should be corrected to 'Keywords'.
  2. [Section 2, Eq. (2)] The selection mechanism cp(i,j) is not formally defined before its first use; Table 1's explanation 'The final remaining variable r under the cp(i,j) of o∗' is unclear and needs a precise statement.
  3. [Appendix A.1, Eq. (5)] There is a typo in the denominator of the second term: 'q(z_4)q(z_4)' should be 'q(z_3)q(z_4)'.
  4. [Section 3.1, Eq. (6)] The constraint 1/m ≤ g ≤ 1 is redundant given that bhat is a divisor of n and m = max(S); consider simplifying the definition of g to avoid confusion.

Circularity Check

1 steps flagged · score 7.0 of 10

The V-shaped 'optimal ELBO' and β-STCVAE's higher ELBO than β-TCVAE are by construction: Eq. (7) deletes nonnegative within-block total-correlation penalties from the objective.

  1. self definitional [Section 3.1, Eq. (7); Section 4.2, 'Optimal ELBO Trajectory in Parameter Space' and Fig. 3]
    "Consequently, we propose modifying the 𝒯𝒞(𝑧) term in the β-TCVAE loss (1) by substituting it with 𝒯𝒞𝑗𝑜𝑖𝑛𝑡_𝑏̂(𝑧), thereby defining the loss function for the new model β-STCVAE: 𝐿𝛽−𝑆𝑇𝐶𝑉𝐴𝐸 ≔ 𝔼𝑞(𝑧|𝑥)𝑝(𝑥)[𝑙𝑜𝑔𝑝(𝑥|𝑧)] − 𝐼(𝑧 ; 𝑥) − 𝛽[𝒯𝒞𝑗𝑜𝑖𝑛𝑡_𝑏̂(𝑧)] − ∑ 𝒟𝐾𝐿(𝑞(𝑧𝑙)||𝑝(𝑧𝑙))𝑙 , (7). Fig. 3 shows that the new model with a reasonable disentangling granularity could always obtain better ELBO than β-TCVAE (red dotted line), regardless of disentangling strength anparameter capacities."

    For a partition of z into blocks B_j of size b̂, the total-correlation chain rule gives TC(z) = TC_joint_b̂(z) + Σ_j TC(z_{B_j}), where TC_joint_b̂(z) is exactly the D_KL term in Eq. (7). Substituting this into Eq. (1) yields L_STC = L_TCVAE + β Σ_j TC(z_{B_j}) for the same q and same β. The within-block terms are nonnegative and are never evaluated or reported. Therefore β-STCVAE's 'higher ELBO' relative to β-TCVAE is forced by deleting positive penalty terms, not by better inference. The paper itself notes that at g=1/m Eq. (7) is equivalent to Eq. (1), so every larger granularity inherits an objective inflated by β times the omitted within-block total correlations.

full rationale

The central empirical claim that β-STCVAE achieves a higher ELBO than β-TCVAE and exhibits a V-shaped optimal ELBO trajectory derives directly from the loss definition in Eq. (7), not from an independently evaluated ELBO. Because TC(z) = TC_joint_b̂(z) + Σ_j TC(z_{B_j}) with nonnegative within-block terms, Eq. (7) is the β-TCVAE objective plus β times the omitted penalties; hence any model trained with Eq. (7) has an inflated objective value by construction. The paper neither reports the omitted within-block TC values nor evaluates a separate held-out ELBO, so the V-shape and the superiority claim are not independent evidence about representation quality or inference quality. The MIG, Factor, and SAP evaluations in Sections 4.5-4.6 do use external metrics and provide independent content, which prevents the whole paper from being entirely circular; however the headline 'optimal ELBO trajectory' and the claim of better ELBO than β-TCVAE are artifacts of the modified loss. No self-citations are load-bearing here. Overall, one central 'prediction' reduces to a definitional change in the objective, warranting a score of 7.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central decomposition is a chain-rule identity; the main added assumption is the finite truncation of the improper integral in the appendix. The granularity normalization g=b̂/m and the loss in Eq (7) are ad hoc choices. There are no invented physical entities; 'disentangling granularity' and 'omniscient latent variable' are descriptive labels. The effective free parameter in the empirical analysis is the unspecified binary function curve used to produce the reported V-shaped trajectory.

free parameters (1)
  • binary function curve parameters = not reported
    The V-shaped optimal ELBO trajectory is obtained by fitting a binary function curve to the optimal granularity per capacity; the form and parameters of this curve are not specified in the paper.
assumptions (3)
  • ad hoc to paper The improper integral defining D_KL(q(z)||∏q(z_k)) can be replaced by a definite integral over finite unified bounds (v,u) with negligible error.
    Introduced in Appendix A.1, Eq (2), to justify the step-by-step decomposition. Latent variables are typically unbounded Gaussians, so the truncation is non-trivial and not justified.
  • ad hoc to paper The grouping factor b̂ is restricted to proper divisors of n and the normalization g=b̂/m with m=max(S) is a meaningful inductive bias scale.
    The definition of S in Eq (6) is ambiguous (whether b̂=n is included) and the normalization to 1/m≤g≤1 is chosen by hand to make β-TCVAE the g=1/m special case.
  • domain assumption The true data distribution factorizes into conditionally independent features v and conditionally dependent features w.
    Invoked in Section 3 as the conceptual motivation for matching granularity to feature complexity; this generative factorization is assumed rather than derived or tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangling Granularity: An Implicit Inductive Bias in Factorized VAEs." pith.science (2026). https://pith.science/paper/VJVSWZHX

@misc{pith2026250524684,
  author       = {Pith},
  title        = {Pith review of: Disentangling Granularity: An Implicit Inductive Bias in Factorized VAEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJVSWZHX}},
  note         = {Machine review of arXiv:2505.24684}
}
read the original abstract

Despite the success in learning semantically meaningful, unsupervised disentangled representations, variational autoencoders (VAEs) and their variants face a fundamental theoretical challenge: substantial evidence indicates that unsupervised disentanglement is unattainable without implicit inductive bias, yet such bias remains elusive. In this work, we focus on exploring the implicit inductive bias that drive disentanglement in VAEs with factorization priors. By analyzing the total correlation in \b{eta}-TCVAE, we uncover a crucial implicit inductive bias called disentangling granularity, which leads to the discovery of an interesting "V"-shaped optimal Evidence Lower Bound (ELBO) trajectory within the parameter space. This finding is validated through over 100K experiments using factorized VAEs and our newly proposed model, \b{eta}-STCVAE. Notably, experimental results reveal that conventional factorized VAEs, constrained by fixed disentangling granularity, inherently tend to disentangle low-complexity feature. Whereas, appropriately tuning disentangling granularity, as enabled by \b{eta}-STCVAE, broadens the range of disentangled representations, allowing for the disentanglement of high-complexity features. Our findings unveil that disentangling granularity as an implicit inductive bias in factorized VAEs influence both disentanglement performance and the inference of the ELBO, offering fresh insights into the interpretability and inherent biases of VAEs.

Figures

Figures reproduced from arXiv: 2505.24684 by the authors.

Figure 2
Figure 2. By tuning the disentangling granularity, we can control the disentangling expression of VAEs under different feature complexities. We use transparency to express the degree of disentangling attention that the model pays to the data features. The lower the transparency of the feature is, the more focused the disentangling attention from the model will be. As the disentangling granularity becomes smaller, the model fo… view at source ↗
Figure 3
Figure 3. The Fig. 3(a),(b) plot the optimal ELBO trajectory of the model in parameter space constructed [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. The left, right are the average MIG of the two models under different parameter capacities on [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The Fig. 5(a) [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: We trained five models on four experimental groups (3dc [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [1]

    X. WU, X. WEN, X. LIU and H. ZHAO, Masked scene contrast: A scalable framework for unsupervised 3d representation learning, Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2023, pp. 9415-9424

    work page 2023

  2. [2]

    ROMBACH, A

    R. ROMBACH, A. BLATTMANN, D. LORENZ, P. ESSER and B. OMMER, High-resolution image synthesis with latent diffusion models, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684-10695

    work page 2022

  3. [3]

    I. H. M. TAREK, M. M. RAHMAN and Z. M. ZAKI, Unsupervised Representation Learning using Wasserstein Generative Adversarial Network, International Journal of Computing

    work page

  4. [4]

    AI and Z.-H

    Y. AI and Z.-H. LING, Neural speech phase prediction based on parallel estimation architecture and anti-wrapping losses, ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1-5

    work page 2023

  5. [5]

    ZHENG, Y

    R.-C. ZHENG, Y. AI and Z.-H. LING, Speech reconstruction from silent tongue and lip articulation by pseudo target generation and domain adversarial training, ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , IEEE, 2023, pp. 1-5

    work page 2023

  6. [6]

    A. B. NASSIF, I. SHAHIN, I. ATTILI, M. AZZEH and K. SHAALAN, 18 Speech recognition using deep neural networks: A systematic review, IEEE access, 7 (2019), pp. 19143-19165

    work page 2019

  7. [7]

    K. XU, L. XU, G. HE, W. YU and Y. LI, Beyond Alignment: Blind Video Face Restoration via Parsing-Guided Temporal-Coherent Transformer, arXiv preprint arXiv:2404.13640 (2024)

    arXiv 2024

  8. [8]

    NGUYEN, C

    X.-B. NGUYEN, C. N. DUONG, X. LI, S. GAUCH, H.-S. SEO and K. LUU, Micron-bert: Bert-based facial micro-expression recognition, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 1482-1492

    work page 2023

Show all 36 references
  1. [9]

    M. T. H. FUAD, A. A. FIME, D. SIKDER, M. A. R. IFTEE, J. RABBI, M. S. AL-RAKHAMI, A. GUMAEI, O. SEN, M. FUAD and M. N. ISLAM, Recent advances in deep learning techniques for face recognition, IEEE Access, 9 (2021), pp. 99112-99142

    2021

  2. [10]

    L. LI, X. MU, S. LI and H. PENG, A review of face recognition technology, IEEE access, 8 (2020), pp. 139110-139120

    2020

  3. [11]

    J. LI, J. ZHANG, J. LI, G. LI, S. LIU, L. LIN and G. LI, Learning background prompts to discover implicit knowledge for open vocabulary object detection, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16678-16687

    2024

  4. [12]

    X. LIU, C. ZHENG, M. QIAN, N. XUE, C. CHEN, Z. ZHANG, C. LI and T. WU, Multi-View Attentive Contextualization for Multi-View 3D Object Detection, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16688-16698

    2024

  5. [13]

    S. ZHANG, Current status and development trend of target tracking methods based on deep learning, International Conference on Computational Modeling, Simulation, and Data Analysis (CMSDA 2021) , SPIE, 2022, pp. 489-493

    2021

  6. [14]

    R. TANG, L. LIU, A. PANDEY, Z. JIANG, G. YANG, K. KUMAR, P. STENETORP, J. LIN and F. TURE, What the daam: Interpreting stable diffusion using cross attention, arXiv preprint arXiv:2210.04885 (2022)

    2022 arXiv

  7. [15]

    XU, Research on deep learning in natural language processing, Advances in Computer and Communication, 4 (2023)

    Z. XU, Research on deep learning in natural language processing, Advances in Computer and Communication, 4 (2023)

    2023

  8. [16]

    ARKHANGELSKAYA and S

    E. ARKHANGELSKAYA and S. I. NIKOLENKO, Deep learning for natural language processing: a survey, Journal of Mathematical Sciences, 273 (2023), pp. 533-582

    2023

  9. [17]

    HIGGINS, L

    I. HIGGINS, L. MATTHEY, A. PAL, C. P. BURGESS, X. GLOROT, M. M. BOTVINICK, S. MOHAMED and A. LERCHNER, beta-vae: Learning basic visual concepts with a constrained variational framework, ICLR (Poster), 3 (2017)

    2017

  10. [18]

    KUMAR, P

    A. KUMAR, P. SATTIGERI and A. BALAKRISHNAN, Variational inference of disentangled latent concepts from unlabeled observations, arXiv preprint arXiv:1711.00848 (2017)

    2017 arXiv

  11. [19]

    KIM and A

    H. KIM and A. MNIH, Disentangling by factorising, International conference on machine learning, PMLR, 2018, pp. 2649-2658

    2018

  12. [20]

    R. T. CHEN, X. LI, R. B. GROSSE and D. K. DUVENAUD, Isolating sources of disentanglement in variational autoencoders, Advances in neural information processing systems, 31 (2018). 19

    2018

  13. [21]

    HAHN and H

    S. HAHN and H. CHOI, Disentangling latent factors with whitening, arXiv preprint arXiv:1811.03444, 2 (2018)

    2018 arXiv

  14. [22]

    X. WANG, H. CHEN, S. A. TANG, Z. WU and W. ZHU, Disentangled representation learning, IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  15. [23]

    LOCATELLO, S

    F. LOCATELLO, S. BAUER, M. LUCIC, G. RAETSCH, S. GELLY, B. SCHö LKOPF and O. BACHEM, Challenging common assumptions in the unsupervised learning of disentangled representations, international conference on machine learning, PMLR, 2019, pp. 4114-4124

    2019

  16. [24]

    LOCATELLO, S

    F. LOCATELLO, S. BAUER, M. LUCIC, G. Rä TSCH, S. GELLY, B. SCHö LKOPF and O. BACHEM, A sober look at the unsupervised learning of disentangled representations and their evaluation, Journal of Machine Learning Research, 21 (2020), pp. 1-62

    2020

  17. [25]

    X. YANG, W. BI, Y. SUN, Y. CHENG and J. YAN, Towards Better Understanding of Disentangled Representations via Mutual Information , arXiv preprint arXiv:1911.10922 (2019)

    2019 arXiv

  18. [26]

    CHEN and K

    J. CHEN and K. BATMANGHELICH, Robust ordinal VAE: employing noisy pairwise comparisons for disentanglement, arXiv preprint arXiv:1910.05898 (2019)

    2019 arXiv

  19. [27]

    ZIETLOW, M

    D. ZIETLOW, M. ROLINEK and G. MARTIUS, Demystifying inductive biases for (beta-) vae based architectures, International Conference on Machine Learning, PMLR, 2021, pp. 12945-12954

    2021

  20. [28]

    R. SHU, S. ZHAO and M. J. KOCHENDERFER, Rethinking style and content disentanglement in variational autoencoders, (2018)

    2018

  21. [29]

    WU and W.-S

    A. WU and W.-S. ZHENG, Factorized diffusion autoencoder for unsupervised disentangled representation learning, Proceedings of the AAAI Conference on Artificial Intelligence, 2024, pp. 5930-5939

    2024

  22. [30]

    VISHNUBHATLA, A

    S. VISHNUBHATLA, A. RAGLIN, R. MORAFFAH and H. LIU, Causal Inductive Biases for Cognitive Machine Learning, 2024 IEEE 6th International Conference on Cognitive Machine Intelligence (CogMI), IEEE, 2024, pp. 14-16

    2024

  23. [31]

    CAZENAVETTE and S

    G. CAZENAVETTE and S. LUCEY, On the bias against inductive biases, arXiv preprint arXiv:2105.14077 (2021)

    2021 arXiv

  24. [32]

    S. E. REED, Y. ZHANG, Y. ZHANG and H. LEE, Deep visual analogy- making, Advances in neural information processing systems, 28 (2015)

    2015

  25. [33]

    DENG, The mnist database of handwritten digit images for machine learning research [best of the web], IEEE signal processing magazine, 29 (2012), pp

    L. DENG, The mnist database of handwritten digit images for machine learning research [best of the web], IEEE signal processing magazine, 29 (2012), pp. 141-142

    2012

  26. [34]

    H. SHAO, S. YAO, D. SUN, A. ZHANG, S. LIU, D. LIU, J. WANG and T. ABDELZAHER, Controlvae: Controllable variational autoencoder, International conference on machine learning, PMLR, 2020, pp. 8655- 8664

    2020

  27. [35]

    X. REN, T. YANG, Y. WANG and W. ZENG, Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view, arXiv preprint arXiv:2102.10543 (2021)

    2021 arXiv

  28. [36]

    Layer” 25 represents the number of hidden layers in model; “ neuron_num

    B. ESMAEILI, H. WU, S. JAIN, A. BOZKURT, N. SIDDHARTH, B. PAIGE, D. H. BROOKS, J. DY and J.-W. MEENT, Structured 20 disentangled representations, The 22nd International Conference on Artificial Intelligence and Statistics, PMLR, 2019, pp. 2525-2534. Appendix Here we provide mo...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.