Research map / Learning theory and algorithms / Kernels and regression theory
Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models
Arnas Uselis, Andrea Dittadi, Seong Joon Oh
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 pretrained encoders instead of a random-init encoder (ablation) | CLIP/SigLIP/DINO embeddings · PUG-Animal, dSprites, MPI3D | linear factorization R2: 0.12-0.42 -> 0.4-0.6 | Improved |
| higher linear factorization (projected R2) across pretrained encoders | linear probes on frozen embeddings · PUG-Animal, dSprites, MPI3D (10% train combinations) | compositional accuracy on unseen combinations | Improved |
| pretraining versus random initialization for cross-concept orthogonality (ablation) | OpenCLIP ViT-L/14 · Pretrained vs random encoders | cross-concept cosine similarity: 0.32 -> 0.09-0.12 | Improved |
| use cross-entropy loss instead of BCE for minimum embedding dimension (ablation) | synthetic linearly compositional model · Synthetic data (Euclidean space) | embedding dimension needed: BCE approx 2k vs CE near bound k | Improved |
| use spherical (normalized) embedding geometry instead of Euclidean (ablation) | synthetic linearly compositional model · Synthetic data | embedding dimension needed: roughly one extra dimension | Hurt |
| increase number of concepts in synthetic training (ablation) | networks trained with classification losses · Synthetic data | emergence of linearity and orthogonality | Improved |
| zero-shot text-encoder probes instead of linear probes (ablation) | CLIP-style models · PUG-Animal, ImageNet-AO | R2 vs compositional accuracy trend | No clear effect |
Closest papers by meaning
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.93)
- Provable Compositional Generalization for Object-Centric Learning (ICLR 2024 · similarity 0.91)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · 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)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · 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.
