Research map / Label-efficient and robust learning / OOD and anomaly detection
Diagnosing Failures in Generalization from Task-Relevant Representational Geometry
Chi-Ning Chou, Artem Kirsanov, Yao-Yuan Yang, SueYeon Chung
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 ID effective manifold dimension and utility instead of ID test accuracy to rank pretrained weights | model selection · ImageNet v1/v2 transfer to 9 datasets | OOD prediction accuracy: 37.22% -> 73.02% | Improved |
| prefer v1 weights with higher Deff despite v2 having higher ImageNet accuracy | ImageNet-pretrained models · OOD linear-probe transfer | accuracy: 14 of 20 architectures | Improved |
| use ID test accuracy as predictor of OOD performance on class-disjoint shift | CIFAR-10 trained CNNs · CIFAR-100 linear probe | correlation with OOD | No clear effect |
| use participation ratio, effective dimension and utility as OOD predictors | CIFAR-10 trained CNNs · CIFAR-100 linear probe | Pearson correlation | Improved |
| use ID test accuracy to predict OOD on corrupted CIFAR-10C | CIFAR-10 trained CNNs · CIFAR-10C | accuracy | Improved |
| use geometric manifold compression to predict OOD on corrupted data | CIFAR-10 trained CNNs · CIFAR-10C | correlation | No clear effect |
| use numerical rank (Tunnel Effect) as OOD predictor | CIFAR-10 trained CNNs · CIFAR-100 and ImageNet linear probe | correlation | Mixed |
| use Neural Collapse and logit-based scores as OOD predictors | CIFAR-10 trained CNNs · OOD linear probe | correlation | Hurt |
Closest papers by meaning
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.86)
- Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking (NeurIPS 2025 · similarity 0.86)
- Feature Learning beyond the Lazy-Rich Dichotomy: Insights from Representational Geometry (ICML 2025 · similarity 0.86)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.86)
- Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers (NeurIPS 2026 · similarity 0.85)
- Fisher information and the geometry of memorization in neural networks (NeurIPS 2026 · 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.
