Research map / Label-efficient and robust learning / Recognition and long-tail
Robust Disentanglement of a Few Factors at a Time using rPU-VAE
Benjamin Estermann, Markus Marks, Mehmet Fatih Yanik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use MIG instead of DCI Disentanglement as PBT eval function (ablation) | PBT-S-VAE · dsprites | MIG | Improved |
| Use MIG instead of DCI Disentanglement as PBT eval function (ablation) | PBT-S-VAE · dsprites | DCI Disentanglement | No clear effect |
| Use MIG instead of DCI as PBT eval function (ablation) | PBT-S-VAE · dsprites | score variance | Improved |
| Add PBT with 1000 labels versus Locatello et al. semi-supervised and model selection | 1000-PBT-S-VAE (MIG) · dsprites | MIG and DCI Disentanglement | Improved |
| Use all labels instead of 1000 labels in PBT-S-VAE (ablation) | PBT-S-VAE (MIG) · dsprites | MIG and DCI Disentanglement | Improved |
| Replace MIG eval with unsupervised UDR eval (PBT-U-VAE) | PBT-VAE · dsprites | MIG / DCI Disentanglement: MIG 0.058 to 0.343, DCI 0.165 to 0.451 | Hurt |
| Add recursive dataset reduction and surrogate labels (rPU-VAE) to PBT-U-VAE | PBT-U-VAE (UDR) · dsprites and shapes3d | MIG / DCI Disentanglement | Improved |
| rPU-VAE versus beta-VAE, FactorVAE, AnnealedVAE, DIP-VAE I/II, beta-TCVAE | rPU-VAE · dsprites and shapes3d | MIG / DCI variance | Improved |
| Increase number of leaf-runs from 0 to 1 (ablation) | rPU-VAE · dsprites and shapes3d | MIG / DCI Disentanglement | Improved |
| Increase number of leaf-runs from 1 to 3 (ablation) | rPU-VAE · dsprites and shapes3d | MIG / DCI Disentanglement | Improved |
| Increase leaf-runs (ablation) | rPU-VAE · dsprites and shapes3d | score variance | Improved |
| Unsupervised rPU-VAE versus supervised PBT-S-VAE upper bound | rPU-VAE · shapes3d | MIG | Hurt |
Closest papers by meaning
- A Sober Look at the Unsupervised Learning of Disentangled Representations and their Evaluation (JMLR 2020 · similarity 0.90)
- Disentangling by Factorising (ICML 2018 · similarity 0.90)
- Disentangling Factors of Variations Using Few Labels (ICLR 2020 · similarity 0.89)
- Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations (ICML 2019 · similarity 0.89)
- Demystifying Inductive Biases for (Beta-)VAE Based Architectures (ICML 2021 · 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.
