pith. sign in

arxiv: 2605.23449 · v1 · pith:XGV63IKGnew · submitted 2026-05-22 · 💻 cs.LG · cs.CV· math.AG

Commutator-Induced Uncertainty in VAEs

Tahereh Dehdarirad , Michael Felsberg , Gabriel Eilertsen , Ziliang Xiong This is my paper

Pith reviewed 2026-05-25 04:40 UTC · model grok-4.3

classification 💻 cs.LG cs.CVmath.AG
keywords variational autoencoderLie groupnon-commutativityBaker-Campbell-Hausdorfflatent spacedeformation stabilityrepresentation learningorder sensitivity
0
0 comments X

The pith

A Lie Group VAE diagnoses latent non-commutativity through Baker-Campbell-Hausdorff deviations and order-swap tests, then aligns decoder sensitivity with a calibrated deformation-stability constraint.

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

Standard VAEs and symmetry-aware variants often force commutativity in latent spaces, which can erase meaningful non-commutative structure when it is intrinsic to the data. The paper instead trains first without constraints to measure algebraic non-commutativity via finite Baker-Campbell-Hausdorff deviations and decoder order sensitivity via reconstruction order-swap tests. A second phase adds a deformation-stability constraint whose single data-driven calibration constant brings the two into alignment. On dSprites, 3DShapes, 3DCars and CelebA the resulting model produces higher-quality reconstructions and decoder behavior that respects order-dependent latent compositions.

Core claim

The central claim is that non-commutativity should be explicitly diagnosed rather than suppressed: measure it algebraically with Baker-Campbell-Hausdorff deviations and behaviorally with order-swap tests, then enforce consistency through a deformation-stability constraint whose calibration constant is chosen from the data so that decoder reconstructions reflect the measured non-commutativity.

What carries the argument

The two-phase Lie Group VAE training procedure that first quantifies algebraic non-commutativity and decoder order sensitivity, then applies a data-driven deformation-stability constraint to align them.

If this is right

  • Reconstruction quality improves over beta-VAE, CLG-VAE and CFASL on synthetic benchmarks.
  • Decoder reconstructions exhibit clearer order-dependent latent compositions.
  • Reconstructions become more stable under small perturbations of the latent code.
  • On CelebA the model produces more faithful images and factor-specific traversals than CFASL while displaying meaningful order-dependent interactions between latent directions.

Where Pith is reading between the lines

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

  • The same diagnostic-plus-calibration pattern could be applied to other generative models that currently regularize toward commutativity.
  • Extending the order-swap test to sequential data would test whether the framework captures non-commutative temporal structure.
  • If the calibration constant proves stable across datasets, the method offers a route to parameter-light incorporation of Lie-group structure without hand-crafted symmetries.

Load-bearing premise

A single data-driven calibration constant can be chosen to align decoder order sensitivity with algebraic non-commutativity without introducing post-hoc bias or requiring dataset-specific retuning.

What would settle it

Train the same architecture on dSprites with and without the deformation-stability constraint; the version with the constraint should show a statistically significant reduction in the mismatch between measured Baker-Campbell-Hausdorff deviations and order-swap reconstruction error.

Figures

Figures reproduced from arXiv: 2605.23449 by Gabriel Eilertsen, Michael Felsberg, Tahereh Dehdarirad, Ziliang Xiong.

Figure 1
Figure 1. Figure 1: Commutator-aware VAE with discrete–continuous factorization. Discrete latents provide a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Phase-aware diagnostics on Cars3D and 3DShapes. Top: [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative reconstruction comparisons. (a) On 3DShapes, our model reconstructs more [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Single-latent traversals and order-dependent composition on dSprites. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Latent traversal comparison between our model and CFASL. In each row, the first image is [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional phase-aware diagnostic for dSprites. Top: [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Order-dependent interaction between two learned latent directions. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
read the original abstract

Variational autoencoders (VAEs) often struggle to represent non-commutative structure in learned latent spaces. Symmetry-aware VAEs commonly address this issue by enforcing commutativity through algebraic regularization, which is appropriate for commutative transformation groups but can suppress meaningful non-commutative structure when it is intrinsic to the data. We argue that non-commutativity should instead be explicitly diagnosed and reflected in reconstruction behavior. We introduce a Lie Group VAE framework that combines geometric and algebraic perspectives on uncertainty while separating discrete generative factors from continuous geometric transformations. In a first phase, the model is trained without structural constraints while algebraic non-commutativity is measured through finite Baker-Campbell-Hausdorff deviations and decoder order sensitivity is measured through reconstruction order-swap tests. These diagnostics reveal a scale mismatch between latent non-commutativity and reconstruction behavior under unconstrained training. In a second phase, we introduce a deformation-stability constraint with a data-driven calibration constant that aligns decoder sensitivity with algebraic non-commutativity. We evaluate the framework on dSprites, 3DShapes, 3DCars, and CelebA against generic and symmetry-aware baselines, including beta-VAE, CLG-VAE, and CFASL. Across synthetic benchmarks, the method improves reconstruction quality and yields decoder-level behavior more consistent with latent non-commutative structure. Qualitative analyses show clearer order-dependent latent compositions and more stable reconstructions. On CelebA, the model yields more faithful reconstructions and factor-specific latent traversals than CFASL, while also exhibiting meaningful order-dependent interactions between learned latent directions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper introduces a Lie Group VAE framework with two phases: an unconstrained first phase that measures algebraic non-commutativity via finite Baker-Campbell-Hausdorff deviations and decoder order sensitivity via reconstruction order-swap tests, revealing a scale mismatch; and a second phase that adds a deformation-stability constraint whose single data-driven calibration constant is chosen to align decoder sensitivity with the measured non-commutativity. The method is evaluated on dSprites, 3DShapes, 3DCars, and CelebA against baselines including beta-VAE, CLG-VAE, and CFASL, with claims of improved reconstruction quality, more consistent decoder behavior with latent non-commutative structure, clearer order-dependent compositions, and better factor-specific traversals on CelebA.

Significance. If the calibration constant can be shown to be a fixed, non-tuned function of phase-1 diagnostics that transfers across datasets without using evaluation reconstruction metrics as feedback, the framework would offer a principled alternative to commutativity-enforcing regularization in symmetry-aware VAEs, allowing intrinsic non-commutative structure to be preserved and reflected in decoder behavior.

major comments (2)
  1. [Phase 2 (deformation-stability constraint)] The description of the second phase states that the deformation-stability constraint uses a data-driven calibration constant chosen to align decoder order sensitivity with algebraic non-commutativity, but provides no explicit procedure, equation, or table demonstrating that the constant is computed solely from phase-1 diagnostics without reference to the reconstruction metrics later used for evaluation; this directly undermines the claim that the resulting decoder behavior is genuinely consistent with latent non-commutativity rather than an artifact of per-dataset fitting.
  2. [Evaluation on synthetic benchmarks] No quantitative tables, error bars, ablation results on the calibration constant, or cross-dataset transfer experiments are referenced in support of the alignment claim or the reported improvements in reconstruction quality and order-dependent behavior; without these, the central consistency claim rests on unshown evidence.
minor comments (1)
  1. The abstract would be strengthened by a brief mention of the specific quantitative metrics (e.g., reconstruction error or consistency scores) that support the improvement claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments below and will revise the manuscript accordingly to strengthen the presentation of the calibration procedure and supporting quantitative evidence.

read point-by-point responses

  1. Referee: [Phase 2 (deformation-stability constraint)] The description of the second phase states that the deformation-stability constraint uses a data-driven calibration constant chosen to align decoder order sensitivity with algebraic non-commutativity, but provides no explicit procedure, equation, or table demonstrating that the constant is computed solely from phase-1 diagnostics without reference to the reconstruction metrics later used for evaluation; this directly undermines the claim that the resulting decoder behavior is genuinely consistent with latent non-commutativity rather than an artifact of per-dataset fitting.

    Authors: We agree that the current manuscript lacks an explicit equation, derivation, or table for the calibration constant. The description states that the constant is chosen from phase-1 diagnostics (finite BCH deviations and order-sensitivity metrics) but does not demonstrate the computation or confirm independence from evaluation metrics. In the revised manuscript we will add a dedicated subsection with the precise formula for the constant, a table of per-dataset values, and a short proof that only phase-1 quantities are used. This revision will directly address the concern that the alignment may be an artifact of per-dataset fitting. revision: yes

  2. Referee: [Evaluation on synthetic benchmarks] No quantitative tables, error bars, ablation results on the calibration constant, or cross-dataset transfer experiments are referenced in support of the alignment claim or the reported improvements in reconstruction quality and order-dependent behavior; without these, the central consistency claim rests on unshown evidence.

    Authors: We acknowledge that the present version relies primarily on qualitative results and does not reference the requested quantitative tables, error bars, ablations, or transfer experiments. In the revision we will insert new tables reporting reconstruction metrics with standard-error bars over multiple seeds, an ablation study on the calibration constant, and cross-dataset transfer results in which the constant calibrated on one dataset is applied unchanged to the others. These additions will supply the missing quantitative support for the alignment and consistency claims. revision: yes

Circularity Check

1 steps flagged

Data-driven calibration constant aligns sensitivities by construction in phase 2

specific steps
  1. fitted input called prediction [Abstract (second phase description)]
    "In a second phase, we introduce a deformation-stability constraint with a data-driven calibration constant that aligns decoder sensitivity with algebraic non-commutativity."

    The constant is chosen data-driven specifically to align the two quantities whose mismatch was diagnosed in phase 1; the claimed consistency of decoder-level behavior with non-commutative structure is therefore achieved by construction via the fit rather than emerging from unconstrained training or external validation.

full rationale

The paper's central claim of decoder behavior consistent with latent non-commutativity rests on a second-phase deformation-stability constraint whose single calibration constant is explicitly data-driven to force alignment between measured algebraic non-commutativity (BCH deviations) and decoder order sensitivity. This reduces the reported consistency to a fitted parameter chosen on the evaluation data rather than an independent derivation. The first-phase diagnostics are used to set the constant, making the phase-2 outcome statistically forced. No other circular steps are present; the initial measurement phase and baseline comparisons remain independent.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the framework rests on the domain assumption that finite BCH deviations and order-swap tests are valid proxies for latent non-commutativity, plus one fitted calibration constant whose value is chosen to enforce alignment.

free parameters (1)
  • deformation-stability calibration constant
    Data-driven constant introduced in the second phase to align decoder sensitivity with measured algebraic non-commutativity.
axioms (1)
  • domain assumption Finite Baker-Campbell-Hausdorff deviations and decoder order-swap tests reliably diagnose the scale of latent non-commutativity
    Invoked to justify the first-phase diagnostics and the subsequent need for the calibration constant.

pith-pipeline@v0.9.0 · 5824 in / 1573 out tokens · 73306 ms · 2026-05-25T04:40:25.759307+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · 1 internal anchor

  1. [1]

    ArXiv , year=

    Disentangling images with Lie group transformations and sparse coding , author=. ArXiv , year=

    work page

  2. [2]

    and Clements, William R

    Quessard, Robin and Barrett, Thomas D. and Clements, William R. , title =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =. 2020 , isbn =

    work page 2020

  3. [3]

    2025 , eprint=

    Non(anti)Commutative Superspace, Baker-Campbell-Hausdorff Closed Forms, and Dirac-K\"ahler Twisted Supersymmetry , author=. 2025 , eprint=

    work page 2025

  4. [4]

    Audenaert , keywords =

    David Wenzel and Koenraad M.R. Audenaert , keywords =. Impressions of convexity: An illustration for commutator bounds , journal =. 2010 , issn =. doi:https://doi.org/10.1016/j.laa.2010.06.039 , url =

    work page doi:10.1016/j.laa.2010.06.039 2010

  5. [5]

    Symmetry-Based Disentangled Representation Learning requires Interaction with Environments , url =

    Caselles-Dupr\'. Symmetry-Based Disentangled Representation Learning requires Interaction with Environments , url =. Advances in Neural Information Processing Systems , editor =

    work page

  6. [6]

    Linear disentangled representations and unsupervised action estimation , year =

    Painter, Matthew and Hare, Jonathon and Pr\". Linear disentangled representations and unsupervised action estimation , year =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    work page

  7. [7]

    arXiv preprint arXiv:2011.06070 , year =

    Quantifying and Learning Linear Symmetry-Based Disentanglement , author =. arXiv preprint arXiv:2011.06070 , year =

    work page arXiv 2011

  8. [8]

    International Conference on Machine Learning , year=

    Commutative Lie Group VAE for Disentanglement Learning , author=. International Conference on Machine Learning , year=

    work page

  9. [9]

    and Bauer, S

    Locatello, F. and Bauer, S. and Lucic, M. and Raetsch, G. and Gelly, S. and Sch. Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations , booktitle =. 2019 , slug =

    work page 2019

  10. [10]

    Theory of Trotter Error with Commutator Scaling , author =. Phys. Rev. X , volume =. 2021 , month =. doi:10.1103/PhysRevX.11.011020 , url =

    work page doi:10.1103/physrevx.11.011020 2021

  11. [11]

    The Frobenius norm and the commutator , journal =

    Albrecht Böttcher and David Wenzel , keywords =. The Frobenius norm and the commutator , journal =. 2008 , issn =. doi:https://doi.org/10.1016/j.laa.2008.05.020 , url =

    work page doi:10.1016/j.laa.2008.05.020 2008

  12. [12]

    Journal of Computational and Applied Mathematics , year =

    Zhang Wen and Zhang Chaolu , title =. Journal of Computational and Applied Mathematics , year =. doi:10.1016/j.cam.2021.113921 , url =

    work page doi:10.1016/j.cam.2021.113921 2021

  13. [13]

    2025 , eprint=

    Directional Non-Commutative Monoidal Structures for Compositional Embeddings in Machine Learning , author=. 2025 , eprint=

    work page 2025

  14. [14]

    International Conference on Learning Representations (ICLR) , year =

    Wenxuan Zhao and Xudong Han and Tengyu Ma and Jiaming Song and Denny Zhou and Honglak Lee , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  15. [15]

    Higham , title =

    Nicholas J. Higham , title =. Functions of Matrices: Theory and Computation , chapter =. 2008 , address =

    work page 2008

  16. [16]

    Scientific Reports , volume =

    Yunlong Guo , title =. Scientific Reports , volume =. 2016 , doi =

    work page 2016

  17. [17]

    K. M. R. Audenaert , title =. Linear Algebra and its Applications , volume =. 2010 , doi =

    work page 2010

  18. [18]

    arXiv preprint arXiv:2405.17823 , year =

    Davide Pastorello and Simone Ghosh , title =. arXiv preprint arXiv:2405.17823 , year =

    work page arXiv

  19. [19]

    International Conference on Learning Representations , year=

    beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework , author=. International Conference on Learning Representations , year=

    work page

  20. [20]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    Disentangling by Factorising , author =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , editor =

    work page 2018

  21. [21]

    2025 , eprint=

    CAD-VAE: Leveraging Correlation-Aware Latents for Comprehensive Fair Disentanglement , author=. 2025 , eprint=

    work page 2025

  22. [22]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Disentanglement Learning via Topology , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    work page 2024

  23. [23]

    Menkovski and Portegies, Jacobus W

    Mahefa Ravelonanosy and V. Menkovski and Portegies, Jacobus W. Topological degree as a discrete diagnostic for disentanglement, with applications to the ΔVAE. 2024

    work page 2024

  24. [24]

    ArXiv , year=

    Explorations in Homeomorphic Variational Auto-Encoding , author=. ArXiv , year=

    work page

  25. [25]

    Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS) , series=

    Adversarial robustness of VAEs through the lens of local geometry , author=. Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS) , series=

    work page

  26. [26]

    Entropy , year=

    Probabilistic Autoencoder Using Fisher Information , author=. Entropy , year=

    work page

  27. [27]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Riemannian Hamiltonian Variational Auto-Encoders , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  28. [28]

    2020 , journal =

    Geometry-Aware Hamiltonian Variational Auto-Encoder , author =. 2020 , journal =

    work page 2020

  29. [29]

    Advances in Neural Information Processing Systems , volume =

    Learning Manifold Dimensions with Conditional Variational Autoencoders , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    work page 2022

  30. [31]

    Journal of Machine Learning Research , year =

    Mikhail Belkin and Partha Niyogi and Vikas Sindhwani , title =. Journal of Machine Learning Research , year =

    work page

  31. [32]

    2022 , howpublished =

    Hamza Fawzi , title =. 2022 , howpublished =

    work page 2022

  32. [33]

    ArXiv ,year=

    Spectral Normalization for Lipschitz-Constrained Policies on Learning Humanoid Locomotion ,author=. ArXiv ,year=

    work page

  33. [34]

    CFASL: Composite Factor-Aligned Symmetry Learning for Disentanglement in Variational AutoEncoder , author=. Trans. Mach. Learn. Res. , year=

    work page

  34. [35]

    ArXiv , year=

    Optimal transport for causal discovery , author=. ArXiv , year=

    work page

  35. [36]

    International Conference on Learning Representations , year =

    Latent Space Oddity: on the Curvature of Deep Generative Models , author =. International Conference on Learning Representations , year =

    work page

  36. [38]

    Advances in Neural Information Processing Systems , editor=

    A Geometric Perspective on Variational Autoencoders , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    work page 2022

  37. [39]

    Milone and Rodrigo Echeveste , title=

    Josefina Catoni and Enzo Ferrante and Diego H. Milone and Rodrigo Echeveste , title=. CoRR , volume=. 2024 , cdate=

    work page 2024

  38. [40]

    Gaussian Process Encoders: VAEs with Reliable Latent-Space Uncertainty

    B \"u tepage, Judith and Maystre, Lucas and Lalmas, Mounia. Gaussian Process Encoders: VAEs with Reliable Latent-Space Uncertainty. Machine Learning and Knowledge Discovery in Databases. Research Track. 2021

    work page 2021

  39. [41]

    NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations , year=

    Optimal Latent Transport , author=. NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations , year=

    work page 2022

  40. [42]

    Predicting and Enhancing the Fairness of DNNs With the Curvature of Perceptual Manifolds , year=

    Ma, Yanbiao and Jiao, Licheng and Liu, Fang and Wen, Maoji and Li, Lingling and Ma, Wenping and Yang, Shuyuan and Liu, Xu and Chen, Puhua , journal=. Predicting and Enhancing the Fairness of DNNs With the Curvature of Perceptual Manifolds , year=

    work page

  41. [43]

    Proceedings of the 32nd International Conference on Neural Information Processing Systems , pages =

    Dupont, Emilien , title =. Proceedings of the 32nd International Conference on Neural Information Processing Systems , pages =. 2018 , publisher =

    work page 2018

  42. [44]

    3D Shapes Dataset , author=

    work page

  43. [45]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Higgins, Irina and Matthey, Loic and Pal, Arka and Burgess, Christopher and Glorot, Xavier and Botvinick, Matthew and Mohamed, Shakir and Lerchner, Alexander , title =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

    work page

  44. [46]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou , title =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    work page

  45. [47]

    and Zhang, Yi and Zhang, Yuting and Lee, Honglak , title =

    Reed, Scott E. and Zhang, Yi and Zhang, Yuting and Lee, Honglak , title =. Advances in Neural Information Processing Systems , volume =

    work page

  46. [48]

    Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , year =

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , author =. Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , year =

    work page 2017

  47. [49]

    2026 , eprint =

    A Unified Latent Space Disentanglement VAE Framework with Robust Disentanglement Effectiveness Evaluation , author =. 2026 , eprint =

    work page 2026

  48. [50]

    Latent space oddity: on the curvature of deep generative models

    Georgios Arvanitidis, Lars Kai Hansen, and S ren Hauberg. Latent space oddity: on the curvature of deep generative models. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=SJzRZ-WCZ

    work page 2018

  49. [51]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, May 2021. Section 3: “Symmetries, Representations, and Invariance”

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  50. [52]

    3d shapes dataset

    Chris Burgess and Hyunjik Kim. 3d shapes dataset. https://github.com/deepmind/3dshapes-dataset/, 2018

    work page 2018

  51. [53]

    Gaussian process encoders: Vaes with reliable latent-space uncertainty

    Judith B \"u tepage, Lucas Maystre, and Mounia Lalmas. Gaussian process encoders: Vaes with reliable latent-space uncertainty. In Nuria Oliver, Fernando P \'e rez-Cruz, Stefan Kramer, Jesse Read, and Jose A. Lozano, editors, Machine Learning and Knowledge Discovery in Databases. Research Track, pages 84--99, Cham, 2021. Springer International Publishing. ...

    work page 2021

  52. [54]

    Milone, and Rodrigo Echeveste

    Josefina Catoni, Enzo Ferrante, Diego H. Milone, and Rodrigo Echeveste. Uncertainty in latent representations of variational autoencoders optimized for visual tasks. CoRR, abs/2404.15390, 2024. URL https://doi.org/10.48550/arXiv.2404.15390

    work page doi:10.48550/arxiv.2404.15390 2024

  53. [55]

    Geometry-aware hamiltonian variational auto-encoder

    Clément Chadebec, Clément Mantoux, and Stéphanie Allassonnière. Geometry-aware hamiltonian variational auto-encoder. HAL Open Science, 2020. https://hal.science/hal-02974244

    work page 2020

  54. [56]

    Olshausen

    Ho Yin Chau, Frank Qiu, Yubei Chen, and Bruno A. Olshausen. Disentangling images with lie group transformations and sparse coding. ArXiv, abs/2012.12071, 2020. URL https://api.semanticscholar.org/CorpusID:229348879

    work page arXiv 2012

  55. [57]

    Learning disentangled joint continuous and discrete representations

    Emilien Dupont. Learning disentangled joint continuous and discrete representations. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS'18, page 708–718, Red Hook, NY, USA, 2018. Curran Associates Inc

    work page 2018

  56. [58]

    Directional non-commutative monoidal structures for compositional embeddings in machine learning, 2025

    Mahesh Godavarti. Directional non-commutative monoidal structures for compositional embeddings in machine learning, 2025. URL https://arxiv.org/abs/2505.15507

    work page arXiv 2025

  57. [59]

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

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In Proceedings of the International Conference on Learning Representations (ICLR), 2017

    work page 2017

  58. [60]

    Nicholas J. Higham. The matrix exponential. In Functions of Matrices: Theory and Computation, chapter 10, pages 239--262. Society for Industrial and Applied Mathematics, Philadelphia, PA, 2008. ISBN 978-0-898716-46-7. URL https://epubs.siam.org/doi/book/10.1137/1.9780898717778

    work page doi:10.1137/1.9780898717778 2008

  59. [61]

    Cfasl: Composite factor-aligned symmetry learning for disentanglement in variational autoencoder

    Heeseung Jung, Jaehyoung Jeong, and Kangil Kim. Cfasl: Composite factor-aligned symmetry learning for disentanglement in variational autoencoder. Trans. Mach. Learn. Res., 2024. URL https://api.semanticscholar.org/CorpusID:267028511

    work page 2024

  60. [62]

    Adversarial robustness of vaes through the lens of local geometry

    Asif Khan and Amos Storkey. Adversarial robustness of vaes through the lens of local geometry. In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 206 of Proceedings of Machine Learning Research, pages 8954--8967, 2023

    work page 2023

  61. [63]

    Disentangling by factorising

    Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2649--2658. PMLR, 10--15 Jul 2018. URL https://proceedings.mlr.press/v80/kim18b.html

    work page 2018

  62. [64]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3730--3738, 2015

    work page 2015

  63. [65]

    Munthe-Kaas, Ari Stern, and Olivier Verdier

    Hans Z. Munthe-Kaas, Ari Stern, and Olivier Verdier. Invariant connections, lie algebra actions, and foundations of numerical integration on manifolds. SIAM Journal on Applied Algebra and Geometry, 4 0 (1): 0 49--68, 2020. doi:10.1137/19M1252879. URL https://doi.org/10.1137/19M1252879

    work page doi:10.1137/19m1252879 2020

  64. [66]

    Linear disentangled representations and unsupervised action estimation

    Matthew Painter, Jonathon Hare, and Adam Pr\" u gel-Bennett. Linear disentangled representations and unsupervised action estimation. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc. ISBN 9781713829546

    work page 2020

  65. [67]

    Barrett, and William R

    Robin Quessard, Thomas D. Barrett, and William R. Clements. Learning disentangled representations and group structure of dynamical environments. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc. ISBN 9781713829546

    work page 2020

  66. [68]

    Reed, Yi Zhang, Yuting Zhang, and Honglak Lee

    Scott E. Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In Corinna Cortes, Neil Lawrence, D. D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015

    work page 2015

  67. [69]

    Johannes Zacherl, Philipp Frank, and Torsten A. Ensslin. Probabilistic autoencoder using fisher information. Entropy, 23, 2021. URL https://api.semanticscholar.org/CorpusID:240070326

    work page 2021

  68. [70]

    Yijia Zheng, Tong He, Yixuan Qiu, and David P. Wipf. Learning manifold dimensions with conditional variational autoencoders. In Advances in Neural Information Processing Systems, volume 35, 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/e04101138a3c94544760c1dbdf2c7a2d-Abstract-Conference.html

    work page 2022

  69. [71]

    Commutative lie group vae for disentanglement learning

    Xinqi Zhu, Chang Xu, and Dacheng Tao. Commutative lie group vae for disentanglement learning. In International Conference on Machine Learning, 2021. URL https://api.semanticscholar.org/CorpusID:235359173

    work page 2021