Reproduce “Reduced Effectiveness of Kolmogorov-Arnold Networks on Functions with Noise” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoran Shen, Chen Zeng, Jiahui Wang, Qiao Wang
Research quantifies the performance degradation of Kolmogorov-Arnold Networks (KANs) when trained on noisy data. It demonstrates that oversampling training data significantly improves KAN robustness, with test loss decreasing asymptotically by O(r^(-1/2)) as data size increases, while kernel filtering shows limited and context-dependent benefits.
Key insights
- Kolmogorov-Arnold Networks (KANs) exhibit substantial performance degradation even with small amounts of noise in training data.
- Oversampling training data is an effective strategy for mitigating noise sensitivity in KANs, improving performance across various noise levels.
- Kernel filtering can reduce noise impact in high noise scenarios, but its effectiveness depends critically on hard-to-determine optimal parameters and can be counterproductive at lower noise levels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.