Research map / Label-efficient and robust learning / Recognition and long-tail
Compositional Generalization from First Principles
Thaddäus Wiedemer, Prasanna Mayilvahanan, Matthias Bethge, Wieland Brendel
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 a compositional model instead of a monolithic model | Multi-sprite reconstruction with orthogonal training support · All latent space | MSE: 4.06e-02 -> 1.24e-03 | Improved |
| Use a compositional model instead of a monolithic model | Multi-sprite reconstruction with random training · All latent space | MSE: 1.73e-3 -> 1.07e-03 | Improved |
| Sample training from a normal distribution with orthogonal support (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.24e-03 -> 1.24e-03 | No clear effect |
| Use a broad diagonal training support (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.24e-03 -> 1.39e-03 | No clear effect |
| Broaden the orthogonal training support (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.24e-03 -> 1.16e-03 | No clear effect |
| Broaden the diagonal training support (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.39e-03 -> 1.13e-03 | No clear effect |
| Introduce gaps in the training support (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.24e-03 -> 2.64e-03 | Hurt |
| Narrow the diagonal training support so some background pixels remain occluded (ablation) | Compositional model · Multi-sprite, all latent space | MSE: 1.39e-03 -> 1.04e-02 | Hurt |
| Use RGBa component outputs and alpha compositing (ablation) | Compositional model · Multi-sprite, orthogonal support; all latent space | MSE: 1.24e-03 -> 5.24e-04 | Improved |
Closest papers by meaning
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.92)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.92)
- Provable Compositional Generalization for Object-Centric Learning (ICLR 2024 · similarity 0.91)
- Scalable Evaluation and Neural Models for Compositional Generalization (NeurIPS 2025 · similarity 0.90)
- Scaling can lead to compositional generalization (NeurIPS 2025 · similarity 0.89)
- Position: Compositional Generative Modeling: A Single Model is Not All You Need (ICML 2024 · similarity 0.89)
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.
