Research map / Detection and segmentation / Weakly supervised detection and segmentation
Refactoring Policy for Compositional Generalizability using Self-Supervised Object Proposals
Tongzhou Mu, Jiayuan Gu, Zhiwei Jia, Hao Tang, Hao Su
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 |
|---|---|---|---|
| Replace flattening with max pooling in the plain CNN (ablation) | Plain CNN · Multi-MNIST-CIFAR | Test Acc: 30.7(4.9) -> 20.2(3.6) | Hurt |
| Replace flattening with sum pooling in the plain CNN (ablation) | Plain CNN · Multi-MNIST-CIFAR | Test Acc: 30.7(4.9) -> 5.7(3.4) | Hurt |
| Replace flattening with max pooling in the plain CNN (ablation) | Plain CNN · Multi-MNIST-ImageNet | Test Acc: 12.0(2.1) -> 9.74(0.4) | Hurt |
| Replace flattening with sum pooling in the plain CNN (ablation) | Plain CNN · Multi-MNIST-ImageNet | Test Acc: 12.0(2.1) -> 10.2(1.8) | Hurt |
| Train the object-centric GNN by two-stage refactorization rather than end-to-end DQN with 25 false-positive proposals | Object-centric GNN policy · Pacman, 10 dots | Mean episode reward: 8.29 -> 4.43 | Improved |
| Train the object-centric GNN by end-to-end DQN with 50% of detected objects removed instead of refactorizing from demonstrations (ablation) | Object-centric GNN policy · Pacman, environments with more dots | Policy quality | Hurt |
Closest papers by meaning
- Not Just Object, But State: Compositional Incremental Learning without Forgetting (NeurIPS 2024 · similarity 0.86)
- Provable Compositional Generalization for Object-Centric Learning (ICLR 2024 · similarity 0.85)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.85)
- Scalable Evaluation and Neural Models for Compositional Generalization (NeurIPS 2025 · similarity 0.85)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.85)
- Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate (NeurIPS 2024 · 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.
