Research map / Label-efficient and robust learning / Recognition and long-tail
On Disentangled Representations Learned from Correlated Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the strength of training-data correlations | Unsupervised VAE disentanglement methods · Shapes3D (A), object size–azimuth | pairwise entanglement score: 0.08 -> 0.38 | Hurt |
| Increase the strength of training-data correlations | Unsupervised VAE disentanglement methods · dSprites (B), orientation–position x | pairwise entanglement score: 0.11 -> 0.17 | Hurt |
| Increase the strength of training-data correlations | Unsupervised VAE disentanglement methods · MPI3D (C), first DOF–second DOF | pairwise entanglement score: 0.16 -> 0.2 | Hurt |
| Add post-hoc linear alignment using 100 labels for the correlated factors | Unsupervised VAE representations · Shapes3D (A), object size–azimuth, σ=0.2 | pairwise entanglement score: 0.38 -> 0.17 | Helped |
| Add post-hoc linear alignment using 100 labels for the correlated factors | Unsupervised VAE representations · Shapes3D (A), object size–azimuth, σ=0.4 | pairwise entanglement score: 0.26 -> 0.1 | Helped |
| Use weak supervision from pairs differing in one known-count factor | β-VAE · Shapes3D (A), correlated object size–azimuth | DCI score | Helped |
| Use weak supervision from pairs differing in one known-count factor | β-VAE · Shapes3D (A), correlated object size–azimuth | unfairness score | Helped |
| Use a nonlinear MLP substitution function instead of linear regression | Fast adaptation · Correlated factors without a unique natural ordering or isolated latent dimensions | pairwise entanglement | Helped |
| Train unsupervised on correlated observational data rather than uncorrelated data | Factorization-based VAE methods · Correlated factors of variation | pairwise latent entanglement | Hurt |
| Use common global disentanglement metrics to assess correlation-induced entanglement | Unsupervised disentanglement evaluation · Correlated datasets | sensitivity to correlation strength | No clear effect |
| Evaluate reconstructions on factor combinations outside the correlated training distribution | Unsupervised VAE representations · Shapes3D (A), OOD object size–azimuth combinations | OOD reconstruction | Helped |
Closest papers by meaning
- Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations (ICML 2019 · similarity 0.93)
- On the Fairness of Disentangled Representations (NeurIPS 2019 · similarity 0.92)
- Disentangling Factors of Variations Using Few Labels (ICLR 2020 · similarity 0.91)
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.90)
- Identifiability Guarantees for Causal Disentanglement from Purely Observational Data (NeurIPS 2024 · similarity 0.89)
- $\alpha$TC-VAE: On the relationship between Disentanglement and Diversity (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.