Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Compositional Generalization using Iterated Learning and Simplicial Embeddings
Yi Ren, Samuel Lavoie, Mikhail Galkin, Danica J. Sutherland, Aaron Courville
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 iterated learning to SEM representations | GCN on ogbg-molhiv · ogbg-molhiv Test-full | AUROC: 78.40$\pm$0.67 -> 79.09$\pm$0.67 | Improved |
| Add iterated learning to SEM representations | GIN on ogbg-molhiv · ogbg-molhiv Test-full | AUROC: 78.21$\pm$0.78 -> 78.61$\pm$0.73 | Improved |
| Add iterated learning to SEM representations | GCN on ogbg-molpcba · ogbg-molpcba Test-full | Average precision: 25.89$\pm$0.71 -> 27.15$\pm$0.74 | Improved |
| Add iterated learning to SEM representations | GIN on ogbg-molpcba · ogbg-molpcba Test-full | Average precision: 25.66$\pm$0.47 -> 28.02$\pm$0.61 | Improved |
| Add iterated learning to SEM representations | GCN on PCQM4Mv2 · PCQM4Mv2 Valid | MAE: 0.106$\pm$0.002 -> 0.102$\pm$0.005 | Improved |
| Add iterated learning to SEM representations | GIN on PCQM4Mv2 · PCQM4Mv2 Valid | MAE: 0.106$\pm$0.004 -> 0.098$\pm$0.005 | Improved |
| Use SEM discretization without iterated learning (ablation) | 3dShapes compositional regression · 3dShapes compositional test | test error | No clear effect |
| Use iterated learning with continuous representations instead of SEM (ablation) | 3dShapes compositional regression · 3dShapes compositional test | test error | No clear effect |
| Use argmax pseudo-labels instead of sampling from the teacher distribution (ablation) | SEM-IL · 3dShapes | test performance | Hurt |
| Add iterated learning to SEM representations | Molecular graph probing on Molhiv-trained representations · Molecular graph probing (Molhiv-trained) | Average AUROC: 0.836 -> 0.878 | Improved |
| Add iterated learning to SEM representations | Molecular graph probing on Molpcba-trained representations · Molecular graph probing (Molpcba-trained) | Average AUROC: 0.926 -> 0.931 | Improved |
Closest papers by meaning
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.90)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.90)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.89)
- In-Context Compositional Learning vis Sparse Coding Transformer (NeurIPS 2025 · similarity 0.88)
- Scaling can lead to compositional generalization (NeurIPS 2025 · similarity 0.88)
- Provable Compositional Generalization for Object-Centric Learning (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.
