Research map / Learning theory and algorithms / Kernels and regression theory
Learning Kernels with Random Features
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| learn kernel distribution via chi2-divergence kernel alignment over random features (vs. unoptimized random features) | Gaussian kernel random features on adult · adult | test error (%) at best D: 15.54 vs 15.44 | No clear effect |
| learn kernel distribution via chi2-divergence kernel alignment over random features | linear kernel random features on reuters · reuters | test error (%) at best D: 9.27 vs 9.36 | Helped |
| learn kernel distribution via chi2-divergence kernel alignment over random features | arc-cosine kernel random features on buzz · buzz | test error (%) at best D: 4.92 vs 4.58 | Hurt |
| learn kernel distribution via alignment (two-phase) instead of unoptimized random features | Gaussian kernel on adult · adult | training time to best test error (s): 3.6 vs 43.1 s | Helped |
| learn kernel distribution via alignment instead of unoptimized random features | linear kernel random features on reuters · reuters | training time to best test error (s): 0.8 vs 295.9 s | Helped |
| learn kernel with joint optimization of kernel and SVM instead of two-phase alignment (ablation) | Gaussian, linear, arc-cosine kernels on subsampled data · adult, reuters, buzz (5000 train) | test error (%): joint slightly better on adult (16.31 vs 16.36), reuters (8.96 vs 9.66), buzz (7.08 vs 8.32) | Mixed |
| two-phase alignment-based kernel learning instead of joint kernel+SVM optimization (ablation) | subsampled benchmark datasets · adult, reuters, buzz (5000 train) | training time (s): 1.8 vs 198.1 s (adult); 0.4 vs 137.5 s (buzz) | Helped |
| learned kernel on synthetic data with poor base kernel (Gaussian kernel, d up to 15) | logistic regression on synthetic sign(||x||-sqrt(d)) · synthetic, d in 2..15 | misclassification error: degrades more slowly than Gaussian kernel as d grows | Helped |
| feature selection via learned sparse distribution q on aptamer k-gram features | linear kernel random features on aptamer data · aptamer (n=2900, d=105,476) | test error at small D: reduces to 379 important features; outperforms uniform random selection when D << d | Helped |
| learned kernel vs random features at very large D | all benchmarks · adult, reuters, buzz | test error (%): random features match or surpass learned kernel at very large D | Mixed |
Closest papers by meaning
- Quadrature-based features for kernel approximation (NeurIPS 2018 · similarity 0.90)
- An Empirical Study on The Properties of Random Bases for Kernel Methods (NeurIPS 2017 · similarity 0.90)
- Infinite Kernel Learning: Generalization Bounds and Algorithms (AAAI 2017 · similarity 0.89)
- Generalization in Kernel Regression Under Realistic Assumptions (ICML 2024 · similarity 0.87)
- Random features for Grassmannian kernel approximation with bounded rank-one projections (NeurIPS 2026 · similarity 0.87)
- The optimality of kernel classifiers in Sobolev space (ICLR 2024 · similarity 0.87)
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.