Research map / Learning theory and algorithms / Kernels and regression theory
Efficient Second-Order Online Kernel Learning with Adaptive Embedding
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 |
|---|---|---|---|
| Use budget-capped PROS-N-KONS (B-KONS, max 100 dictionary points) instead of unbounded dictionary (ablation) | PROS-N-KONS on ijcnn1 with alpha=gamma=0.01 · ijcnn1 binary classification | accuracy error rate: 4.85 vs 10.73 for PROS-N-KONS | Helped |
| Use CON-KONS (no solution reset at epoch change) instead of resetting at each new dictionary element (ablation) | PROS-N-KONS on ijcnn1 with alpha=gamma=0.01 · ijcnn1 binary classification | accuracy error rate: 6.23 vs 10.73 for PROS-N-KONS | Helped |
| Use CON-KONS (no reset) instead of PROS-N-KONS reset (ablation) | PROS-N-KONS on cadata regression · cadata | avg. squared loss: 0.02850 vs 0.03095 | Helped |
| Use CON-KONS (no reset) instead of PROS-N-KONS reset on cod-rna with alpha=gamma=1 (accuracy worse, high variance) (ablation) | PROS-N-KONS on cod-rna · cod-rna (alpha=1, gamma=1) | accuracy error rate: 18.99 vs 13.95 | Hurt |
| Use PROS-N-KONS (adaptive Nystrom embedding + second-order updates) instead of first-order embedding methods FOGD/NOGD on small datasets | PROS-N-KONS vs FOGD/NOGD · parkinson and cpusmall | avg. squared loss: 0.05798 vs 0.04896 (NOGD) on parkinson; 0.02494 vs 0.02559 on cpusmall is better | Hurt |
| Use PROS-N-KONS second-order adaptive embedding instead of first-order baselines on larger datasets | PROS-N-KONS vs NOGD/FOGD · cadata and casp regression | avg. squared loss: 0.03095 vs 0.03983 (cadata); 0.06773 vs 0.07844 (casp) | Helped |
| Use PROS-N-KONS on the very large regression dataset year (adaptive dictionary, no budget) | PROS-N-KONS · year | avg. squared loss: 0.01450 vs 0.01427 (NOGD) | Hurt |
| Use budget-capped B-KONS instead of unbounded PROS-N-KONS on year (ablation) | PROS-N-KONS on year · year | avg. squared loss: 0.01302 vs 0.01450 | Helped |
| Increase regularization from alpha=gamma=1 to alpha=gamma=0.01 on binary classification (ablation) | PROS-N-KONS · cod-rna | accuracy error rate: 4.91 vs 13.95 | Helped |
| Decrease alpha and gamma from 1 to 0.01 (more SVs and runtime) on ijcnn1 (ablation) | PROS-N-KONS · ijcnn1 | accuracy error rate: 9.70 at alpha=gamma=1 vs 10.73 at 0.01 | Mixed |
Closest papers by meaning
- Second-Order Kernel Online Convex Optimization with Adaptive Sketching (ICML 2017 · similarity 0.88)
- On Statistical Learning Theory for Distributional Inputs (ICML 2024 · similarity 0.85)
- Limited Memory Online Gradient Descent for Kernelized Pairwise Learning with Dynamic Averaging (AAAI 2024 · similarity 0.85)
- Ahpatron: A New Budgeted Online Kernel Learning Machine with Tighter Mistake Bound (AAAI 2024 · similarity 0.84)
- Scalable and adaptive prediction bands with kernel sum-of-squares (NeurIPS 2025 · similarity 0.84)
- On the Consistency of Kernel Methods with Dependent Observations (ICML 2024 · 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.