mlsanityData to 14 Oct 2026

Operationalizing Quantized Disentanglement

arXiv 2511.20927 ↗First posted 25 Nov 2025cs.LG stat.MLcited by 0

Not on the map. Why?

Abstract (arXiv)

Recent theoretical work established the unsupervised identifiability of quantized factors under any diffeomorphism. The theory assumes that quantization thresholds correspond to axis-aligned discontinuities in the probability density of the latent factors. By constraining a learned map to have a density with axis-aligned discontinuities, we can recover the quantization of the factors. However, translating this high-level principle into an effective practical criterion remains challenging, especially under nonlinear maps. Here, we develop a criterion for unsupervised disentanglement by encouraging axis-aligned discontinuities. Discontinuities manifest as sharp changes in the estimated density of factors and form what we call cliffs. Following the definition of independent discontinuities from the theory, we encourage the location of the cliffs along a factor to be independent of the values of the other factors. We show that our method, Cliff, outperforms the baselines on all disentanglement benchmarks, demonstrating its effectiveness in unsupervised disentanglement.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 19 of its 26 references that are in the map. First posted on arXiv 25 Nov 2025.

Builds on

19 of 26 references.

References · 19

  1. 2017Neural Discrete Representation Learningcited by 4,712
  2. 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 941
  3. 2018Disentangling by Factorisingcited by 616
  4. 2023Finite Scalar Quantization: VQ-VAE Made Simplecited by 440
  5. 2021Disentanglement via Mechanism Sparsity Regularization: A New Principle for Nonlinear ICAcited by 160
  6. 2022Interventional Causal Representation Learningcited by 142
  7. 2020On Disentangled Representations Learned From Correlated Datacited by 94
  8. 2020The Hessian Penalty: A Weak Prior for Unsupervised Disentanglementcited by 92
  9. 2022Identifiability of deep generative models without auxiliary informationcited by 70
  10. 2022Weakly Supervised Representation Learning with Sparse Perturbationscited by 65
  11. 2023Additive Decoders for Latent Variables Identification and Cartesian-Product Extrapolationcited by 58
  12. 2022Function Classes for Identifiable Nonlinear Independent Component Analysiscited by 53
7 more
  1. 2023Disentanglement via Latent Quantizationcited by 40
  2. 2022Disentanglement of Correlated Factors via Hausdorff Factorized Supportcited by 37
  3. 2023Causal Representation Learning Made Identifiable by Grouping of Observational Variablescited by 32
  4. 2022DCI-ES: An Extended Disentanglement Framework with Connections to Identifiabilitycited by 23
  5. 2024Learning Discrete Concepts in Latent Hierarchical Modelscited by 17
  6. 2024Tripod: Three Complementary Inductive Biases for Disentangled Representation Learningcited by 13
  7. 2023On the Identifiability of Quantized Factorscited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2024Disentanglement with Factor Quantized Variational Autoencoders
  2. 2025Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states
  3. 2025Sculpting Latent Spaces With MMD: Disentanglement With Programmable Priors
  4. 2023Disentanglement of Latent Representations via Causal Interventions
  5. 2023DisDiff: Unsupervised Disentanglement of Diffusion Probabilistic Models
  6. 2026Saturation Makes Quantization Error Additive: A Coverage Model with a Certificate
  7. 2023Intriguing Properties of Quantization at Scale
  8. 2026Hidden not Deleted: How Networks Suppress Entangled Features