Research map / Architectures and efficiency / RNNs and normalization
Initialization Schemes for Kolmogorov–Arnold Networks: An Empirical Study
Spyros Rigas, Dhruv Verma, Georgios Alexandridis, Yixuan Wang
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 |
|---|---|---|---|
| use power-law initialization instead of baseline | spline KAN · function fitting | runs improving loss and L2 vs baseline: 94.79%-100.00% of runs | Improved |
| use power-law initialization instead of baseline | spline KAN · forward PDEs | runs improving loss and L2 vs baseline: 73.61%-94.44% of runs | Improved |
| use Glorot-inspired initialization instead of baseline | spline KAN · function fitting | runs improving loss and L2 vs baseline: 60.41%-78.13% of runs | Improved |
| use Glorot-inspired initialization instead of baseline | spline KAN · forward PDEs | runs improving loss and L2 vs baseline: 36.11%-62.50% of runs | Mixed |
| use LeCun-numerical initialization instead of baseline | spline KAN · function fitting | runs improving loss and L2 vs baseline: 0.00%-8.33% of runs | Hurt |
| use LeCun-normalized initialization instead of baseline | spline KAN · forward PDEs | runs improving loss and L2 vs baseline: 0.00% of runs | Hurt |
| use normalized spline basis (LeCun-normalized) instead of numerical estimate | spline KAN · function fitting | runs improving loss and L2 vs baseline: e.g. 0.00 -> 5.21 (f2) | Improved |
| use normalized spline basis (LeCun-normalized) instead of numerical estimate | spline KAN · forward PDEs | runs improving loss and L2 vs baseline: 6.94 -> 0.00 (Burgers) | Hurt |
| fix power-law exponents at alpha=0.25, beta=1.0 rather than tuning per setting (ablation) | power-law init · forward PDEs | runs improving loss and L2 vs baseline: 94.44 -> 83.33 (Allen-Cahn); 73.61 -> 54.93 (Burgers) | Mixed |
| use power-law initialization (alpha=0.25, beta=1.75) instead of baseline | spline KAN, large architecture · Feynman subset | final training loss and L2 | Improved |
| use baseline instead of Glorot/power-law when scaling from small to large architecture | spline KAN · Feynman subset | final loss | Hurt |
Closest papers by meaning
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.91)
- Generalization Bounds and Model Complexity for Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.90)
- On the expressiveness and spectral bias of KANs (ICLR 2025 · similarity 0.90)
- PowerMLP: An Efficient Version of KAN (AAAI 2025 · similarity 0.90)
- Kolmogorov-Arnold Transformer (ICLR 2025 · similarity 0.89)
- Incorporating Arbitrary Matrix Group Equivariance into KANs (ICML 2025 · similarity 0.88)
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.
