Research map / Label-efficient and robust learning / Recognition and long-tail
A Sober Look at the Unsupervised Learning of Disentangled Representations and their Evaluation
Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, Olivier Bachem
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 |
|---|---|---|---|
| add regularization strength to the objective function as predictor of disentanglement scores (ablation) | OLS variance analysis · all data sets | variance explained (%): 37% -> 59% | Improved |
| increase beta regularization strength | beta-VAE and beta-TCVAE · total correlation of mean representation | total correlation: increased regularization leads to substantially increased total correlation of mean representation | Hurt |
| increase beta regularization strength | beta-VAE and beta-TCVAE · total correlation of sampled representation | total correlation: decreases on all data sets as beta increased | Improved |
| add regularization strength in FactorVAE | FactorVAE · total correlation of sampled representation | total correlation: no decrease with increasing regularization | No clear effect |
| choose model by the same metric and data set (different seed) instead of random selection | hyperparameter selection · disentanglement metrics | probability of outperforming random: 52.7% -> 81.9% | Improved |
| transfer hyperparameters across data sets for the same metric | hyperparameter selection · disentanglement metrics | probability of outperforming random: 81.9% -> 59.6% | Mixed |
| use unsupervised scores (reconstruction, ELBO, total correlation) for model selection | model selection · disentanglement metrics | rank correlation | No clear effect |
| choose a different objective function (model) | unsupervised disentanglement models · disentanglement scores | disentanglement score | No clear effect |
| transfer hyperparameters from dSprites to Color-dSprites | hyperparameter transfer · Color-dSprites | rank correlation | Improved |
Closest papers by meaning
- Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations (ICML 2019 · similarity 0.99)
- On Disentangled Representations Learned from Correlated Data (ICML 2021 · similarity 0.93)
- Theory and Evaluation Metrics for Learning Disentangled Representations (ICLR 2020 · similarity 0.93)
- Disentangling Factors of Variations Using Few Labels (ICLR 2020 · similarity 0.93)
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.90)
- Transferring disentangled representations: bridging the gap between synthetic and real images (NeurIPS 2024 · similarity 0.90)
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.
