Research map / Label-efficient and robust learning / Recognition and long-tail
Weakly Supervised Disentanglement with Guarantees
Rui Shu, Yining Chen, Abhishek Kumar, Stefano Ermon, Ben Poole
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 restricted labeling on factor i (ablation) | GAN with weak supervision · Shapes3D | normalized consistency score | Improved |
| use share pairing on factor i (ablation) | GAN with weak supervision · Shapes3D | normalized consistency score | Improved |
| use rank pairing on factor i (ablation) | GAN with weak supervision · Shapes3D | normalized consistency score | Improved |
| use change pairing on factor i (ablation) | GAN with weak supervision · Shapes3D | normalized restrictiveness score | Improved |
| train on two change-pairing datasets (intersection of changed factors) (ablation) | GAN with weak supervision · Shapes3D | normalized restrictiveness score | Mixed |
| train with weak supervision targeting only consistency or only restrictiveness | GAN with weak supervision · Shapes3D (864 models) | consistency vs restrictiveness correlation | Mixed |
| use complete share pairing on all factors instead of unsupervised GAN | GAN with weak supervision · Shapes3D, dSprites, Scream-dSprites, SmallNORB, Cars3D | mutual information gap | Improved |
| use complete change pairing on all factors instead of unsupervised GAN | GAN with weak supervision · five disentanglement datasets | mutual information gap | Improved |
| use complete rank pairing on all factors instead of unsupervised GAN | GAN with weak supervision · five disentanglement datasets | mutual information gap | Improved |
| use weak supervision instead of fully-labeled GAN | GAN with weak supervision · five disentanglement datasets | disentanglement metrics | Mixed |
| restrict hyperparameter sweep to extra dense = False (ablation) | GAN with weak supervision · Shapes3D | normalized consistency/restrictiveness | Improved |
| select hyperparameters by average consistency and restrictiveness instead of supervised metrics | GAN with weak supervision · five disentanglement datasets | disentanglement metrics | Improved |
Closest papers by meaning
- Weakly Supervised Disentanglement by Pairwise Similarities (AAAI 2020 · similarity 0.92)
- Weakly-Supervised Disentanglement Without Compromises (ICML 2020 · similarity 0.91)
- 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 0.88)
- Disentangled Representation Learning with the Gromov-Monge Gap (ICLR 2025 · similarity 0.87)
- Disentangled Representation Learning via Flow Matching (NeurIPS 2026 · similarity 0.87)
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.
