Research map / Optimization / Implicit bias and generalization
Separable Neural Networks: Approximation Theory, NTK Regime, and Preconditioned Gradient Descent
Yisi Luo, Deyu Meng
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 SepPGD preconditioning | SepNN · Plane image INR | PSNR: 26.48 -> 33.30 | Improved |
| use g1 modulation instead of g2 (ablation) | SepPGD · Plane image INR | PSNR: 31.58 -> 33.30 | Improved |
| increase modulated eigenvalues k from 30 to 80 (ablation) | SepPGD · Plane image INR | PSNR: 26.79 -> 34.28 | Improved |
| add SepPGD at rank R=100 (ablation) | SepNN · Plane image INR | PSNR: 26.36 -> 31.59 | Improved |
| update preconditioner less often (1100 vs 10 iterations) (ablation) | SepPGD · Plane image INR | PSNR: 33.30 -> 33.06 | No clear effect |
| add SepPGD with Fourier features + ReLU (ablation) | SepNN · Plane image INR | PSNR: 30.89 -> 40.49 | Improved |
| add SepPGD with ReLU activation (ablation) | SepNN · Plane image INR | PSNR: 18.20 -> 20.02 | Improved |
| add SepPGD at width 100 (ablation) | SepNN · Plane image INR | PSNR: 22.46 -> 22.79 | No clear effect |
| add SepPGD at width 700 (ablation) | SepNN · Plane image INR | PSNR: 27.20 -> 34.35 | Improved |
| add SepPGD under low noise std 0.01 (ablation) | SepNN · Plane image INR with noise | PSNR: 26.40 -> 33.26 | Improved |
| add SepPGD under high noise std 0.09 (ablation) | SepNN · Plane image INR with noise | PSNR: 25.29 -> 25.06 | Hurt |
| add SepPGD on high-rank image data (500) (ablation) | SepNN · Synthetic images | MSE: 0.066 -> 0.037 | Improved |
Closest papers by meaning
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.86)
- Deeper or Wider: A Perspective from Optimal Generalization Error with Sobolev Loss (ICML 2024 · similarity 0.85)
- Separation Power of Equivariant Neural Networks (ICLR 2025 · similarity 0.85)
- The Challenges of the Nonlinear Regime for Physics-Informed Neural Networks (NeurIPS 2024 · similarity 0.85)
- S$^{2}$-PINN: Stochastic Separable Physics-Informed Neural Networks (NeurIPS 2026 · similarity 0.85)
- Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks (NeurIPS 2025 · similarity 0.85)
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.
