Research map / Label-efficient and robust learning / Recognition and long-tail
Global Geometry Is Not Enough for Vision Representations
Jiwan Chung, Seon Joo Kim
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 |
|---|---|---|---|
| measure Jacobian effective rank instead of geometric metrics (G.Iso) | encoder diagnostics · Attribute Binding | Pearson r with binding: +0.11 -> +0.69 | Improved |
| compute local isotropy with k=8 instead of k=32 (ablation) | local isotropy · Attribute Binding | Pearson r with binding: +0.045 vs +0.017 | No clear effect |
| use effective rank instead of variance formulation for local geometry (ablation) | local isotropy · Attribute Binding | Pearson r with binding: +0.020 -> -0.031 | No clear effect |
| evaluate JER on Gaussian noise instead of natural images (ablation) | JER · Attribute Binding | Pearson r with binding: +0.69 -> +0.62 | Hurt |
| evaluate JER on random-init weights (ablation) | JER · Attribute Binding | Pearson r with binding: +0.69 -> -0.16 | Hurt |
| add Same/Different accuracy to JER in regression | JER · Attribute Binding | R^2: 0.47 -> 0.78 | Improved |
| use kNN or local PCA readout instead of cosine similarity (ablation) | frozen embeddings · Attribute Binding | binding accuracy | Hurt |
| control for log param count in JER-binding correlation (ablation) | JER · Attribute Binding | Pearson r: +0.69 -> +0.54 | No clear effect |
| control for ImageNet accuracy and architecture in JER-SugarCrepe (ablation) | JER · SugarCrepe | Pearson r: +0.69 -> +0.70 | No clear effect |
| control for JER and architecture in ImageNet-SugarCrepe (ablation) | ImageNet accuracy · SugarCrepe | Pearson r: +0.67 -> +0.34 | Hurt |
| compare pairwise CKA and Procrustes distances with binding differences (ablation) | inter-model shape metrics · Attribute Binding | Mantel r: +0.014 and +0.144 | No clear effect |
Closest papers by meaning
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.87)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.86)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.86)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.85)
- Unlocking Compositional Generalization in Continual Few-Shot Learning (NeurIPS 2026 · similarity 0.85)
- Neural Isometries: Taming Transformations for Equivariant ML (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.
