Research map / Learning theory and algorithms / Kernels and regression theory
Why Linear Probing Works: Non-Vacuous Generalization Bounds via Effective Dimension
Dongxin Guo, Jikun Wu, SM Yiu
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 spectral (covariance-aligned) prior instead of isotropic prior (ablation) | PAC-Bayes bound, DINOv2-B · ImageNet-1K | bound: 2.34 -> 0.114 | Improved |
| use spectral prior instead of PCA to d_eff dims plus standard bound (ablation) | PAC-Bayes bound, DINOv2-B · ImageNet-1K | bound: 0.172 -> 0.114 | Improved |
| add l2 regularisation (lambda>0) (ablation) | linear probe bound, DINOv2-B · ImageNet-1K | bound: 0.205 -> 0.114 | Improved |
| use full spectrum rather than top-50 eigenvectors only (ablation) | spectral bound, DINOv2-B · ImageNet-1K | bound: 0.138 -> 0.114 | Improved |
| use discriminative (DINOv2-B) rather than generative (MAE-B) pretraining | spectral bound · ImageNet-1K | bound: vacuous (>1) -> 0.114 | Improved |
| scale model from DINOv2-S to DINOv2-g | spectral bound · ImageNet-1K | bound: 0.175 -> 0.112 | Improved |
| use S = gamma^2/d_eff instead of PACTran-N for ranking | model ranking · 12 models, ImageNet-1K | Spearman rho: 0.82 -> 0.91 | Improved |
| use d_eff alone (zero labels) instead of label-based metrics | model ranking · 12 models, ImageNet-1K | Spearman rho: 0.85 vs 0.82 (PACTran-N), 0.91 (S) | Mixed |
| increase distribution shift from ImageNet source (Delta_shift 0 -> 234.1) | DINOv2-B transfer bound · 7 datasets | bound: 0.114 -> 0.501 | Hurt |
| increase shift on target dataset | DINOv2-B linear probe · 7 datasets | accuracy: rho = -0.29, n.s. | No clear effect |
Closest papers by meaning
- Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models (NeurIPS 2026 · similarity 0.87)
- Understanding prompt engineering may not require rethinking generalization (ICLR 2024 · similarity 0.86)
- Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors (NeurIPS 2026 · similarity 0.85)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
- LR0.FM: Low-Res Benchmark and Improving robustness for Zero-Shot Classification in Foundation Models (ICLR 2025 · similarity 0.84)
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.
