Research map / Label-efficient and robust learning / Test-time adaptation
Visual Representation Learning Does Not Generalize Strongly Within the Same Domain
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 |
|---|---|---|---|
| Evaluate representation-learning models on composition, interpolation, and extrapolation splits instead of random splits | 17 representation-learning approaches · dSprites, Shapes3D, MPI3D, and CelebGlow OOD splits | R-squared | Hurt |
| Move from artificial datasets toward more realistic datasets | Tested representation-learning models · dSprites, Shapes3D, MPI3D, and CelebGlow | R-squared | Hurt |
| Test only one factor out of distribution while keeping other factors in distribution | Tested models · MPI3D and dSprites single-factor interpolation and extrapolation | R-squared on OOD factors | Hurt |
| Keep other factors in distribution while one factor is OOD | Tested models · Single-factor OOD splits | R-squared on ID factors | No clear effect |
| Increase supervision signal, including transfer models pretrained on additional data | Models evaluated on MPI3D · MPI3D | R-squared | Helped |
| Use disentanglement models followed by a supervised readout rather than supervised models | Disentanglement models · OOD splits | R-squared | Hurt |
| Use a more disentangled representation | Disentanglement models with supervised readout · OOD downstream evaluation | R-squared: Pearson ρ=0.63; Spearman ρ=0.67 | Helped |
| Use ground-truth factors as the embedding for the readout MLP (ablation) | Readout MLP · dSprites, Shapes3D, and MPI3D-Real random, composition, interpolation, and extrapolation | R-squared: 1.000; 0.999 | Helped |
| Train the readout MLP on sign-flipped ground-truth factors (ablation) | Readout MLP · dSprites, Shapes3D, and MPI3D-Real | R-squared: 1.000; 0.999; 0.996; 0.997 | No clear effect |
| Vary CNN training iterations, regularization strength, and depth (ablation) | CNN · MPI3D extrapolation | R-squared | No clear effect |
| Replace real MPI3D images with synthetic MPI3D images (ablation) | Models evaluated on MPI3D · MPI3D | R-squared: 1-2sigma | No clear effect |
| Constrain dSprites rotations to [0, 90) to remove rotational ambiguity (ablation) | All tested models · dSprites random split | R-squared | Helped |
Closest papers by meaning
- Disentangling Factors of Variation by Mixing Them (CVPR 2018 · similarity 0.88)
- Lost in Latent Space: Examining failures of disentangled models at combinatorial generalisation (NeurIPS 2022 · similarity 0.87)
- Generative Interventions for Causal Learning (CVPR 2021 · similarity 0.87)
- Objective drives the consistency of representational similarity across datasets (ICML 2025 · similarity 0.86)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.86)
- Compositional Generalization via Forced Rendering of Disentangled Latents (ICML 2025 · similarity 0.85)
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.