Research map / Learning theory and algorithms / Kernels and regression theory
Benign, Tempered, or Catastrophic: Toward a Refined Taxonomy of Overfitting
Neil Rohit Mallinar, James B Simon, Amirhesam Abedsoltan, Parthe Pandit, Misha Belkin, Preetum Nakkiran
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 |
|---|---|---|---|
| add a nonzero ridge (delta = 0.1) to Gaussian-kernel regression (ablation) | Gaussian kernel regression · Synthetic noise-only targets on sphere, d=5,10,15 | test MSE vs train samples | Improved |
| use Laplace kernel instead of Gaussian kernel, both ridgeless (ablation) | Ridgeless kernel regression · Synthetic noise-only targets on sphere | test MSE vs train samples | Improved |
| add a ridge to Laplace-kernel regression (ablation) | Laplace kernel regression · Synthetic noise-only targets on sphere | test MSE | Improved |
| increase input dimension d for ridgeless Laplace kernel regression | Laplace kernel regression · Synthetic Laplace-kernel experiments (Appendix B) | asymptotic excess MSE | Improved |
| increase spectral decay exponent alpha of a powerlaw kernel | Ridgeless KR with powerlaw spectrum · Synthetic exact powerlaw spectra | asymptotic test MSE | Hurt |
| increase label-flip noise in training set | Interpolating Wide ResNet 28x10 · Binary CIFAR-10 | clean test error | Hurt |
| increase train size n for interpolating ResNets | Wide ResNet 28x10 · Binary CIFAR-10, n=5k to 40k | clean test error under label noise | No clear effect |
| train MLP for one epoch instead of to interpolation (ablation) | MLP · Binary MNIST | clean test error under label noise | Improved |
| use k-NN with k ~ log n instead of 1-NN (ablation) | Nearest-neighbor classifier · Binary MNIST | clean test error under label noise | Improved |
| add early stopping to Wide ResNet training (ablation) | Wide ResNet · Binary SVHN | clean test error under label noise | Improved |
| train interpolating three-layer MLP on a constant target with slight label noise | Three-layer MLP · Synthetic S9 inputs, f*(x)=1 | test error | Hurt |
| continue training from early epochs to interpolation | MLP on synthetic noisy data · Appendix E synthetic | overfitting regime | Hurt |
Closest papers by meaning
- Noisy Interpolation Learning with Shallow Univariate ReLU Networks (ICLR 2024 · similarity 0.93)
- Rethinking Benign Overfitting in Two-Layer Neural Networks (ICML 2025 · similarity 0.91)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.90)
- Beyond Benign Overfitting in Nadaraya-Watson Interpolators (NeurIPS 2025 · similarity 0.90)
- Benign Overfitting in Out-of-Distribution Generalization of Linear Models (ICLR 2025 · similarity 0.89)
- An Agnostic View on the Cost of Overfitting in (Kernel) Ridge Regression (ICLR 2024 · 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.
