Research map / Learning theory and algorithms / Kernels and regression theory
Out-of-Distribution Generalization in Kernel Regression
Abdulkadir Canatar, Blake Bordelon, Cengiz Pehlevan
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 |
|---|---|---|---|
| Optimize test measure by gradient descent on theoretical E_g (beneficial shift) versus matched train/test | NTK regression, depth-3 ReLU NTK · MNIST 8s and 9s, M=1000 | generalization error | Improved |
| Optimize test measure by gradient ascent on theoretical E_g (detrimental shift) | NTK regression, depth-3 ReLU NTK · MNIST 8s and 9s, M=1000 | generalization error | Hurt |
| Optimize training distribution over MNIST digits at P=30 rather than uniform | NTK regression · MNIST 8s vs 9s, uniform test distribution | generalization error | Improved |
| Use a training distribution optimized for small P at larger P | optimized training measure · MNIST 8s vs 9s | generalization error | Hurt |
| Reduce training-measure rank M_r from 60 to 40 (target power concentrated in 40 directions) | linear-kernel regression · Synthetic linear target, N=60, D=120 | generalization error | Mixed |
| Reduce M_r below 40 (below the target's main directions) | linear-kernel regression · Synthetic linear target, N=60, D=120 | generalization error | Hurt |
| Increase effective regularization (via training variance and M_r) | linear regression with label noise · Synthetic, eps^2=0.1 | double-descent peak and learning speed | Mixed |
| Use optimal ridge lambda* = (M_r/D) eps^2 | linear regression · Synthetic, label noise | generalization error | Improved |
Closest papers by meaning
- Generalization in Kernel Regression Under Realistic Assumptions (ICML 2024 · similarity 0.89)
- Certifiable Out-of-Distribution Generalization (AAAI 2023 · similarity 0.89)
- Generalization error of spectral algorithms (ICLR 2024 · similarity 0.87)
- On Statistical Learning Theory for Distributional Inputs (ICML 2024 · similarity 0.87)
- Out-of-Variable Generalisation for Discriminative Models (ICLR 2024 · similarity 0.87)
- Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize (ICML 2024 · similarity 0.86)
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.
