文章背景与核心概要
本文介绍了 ER-KANs(Efficient and Robust Kolmogorov-Arnold Networks,高效且鲁棒的柯尔莫哥洛夫-阿诺德网络),这是一种专为数据稀缺和带噪科学机器学习(SciML)任务设计的新型神经网络架构。尽管现有的大多数高效 KAN 变体在干净的基准测试中表现良好,但在噪声环境下性能会发生灾难性退化。ER-KAN 通过结合共享高斯 RBF 基、课程噪声注入以及熵加权自适应正则化这三项核心设计,显著提升了模型在噪声和数据稀缺场景下的鲁棒性。
该模型在多项解析函数基准、阻尼谐振子物理信息神经网络(PINN)以及伯格斯方程(Burgers' equation)中进行了严格评估。结果表明,ER-KAN 相比于标准多层感知机(MLP)和传统的高效 KAN 变体具有更低的测试均方误差(MSE)以及卓越的可靠性。同时,作者还引入了“噪声退化率”(noise degradation ratio)作为评估高效 KAN 研究的标准化指标。
ER-KANs:Efficient and Robust Kolmogorov-Arnold Networks for Data-Scarce Scientific Machine Learning

📋 Executive Summary
The paper introduces ER-KANs (Efficient and Robust Kolmogorov-Arnold Networks), a novel neural network architecture designed specifically for data-scarce and noisy scientific machine learning (SciML) tasks.
本文介绍了 ER-KANs(Efficient and Robust Kolmogorov-Arnold Networks),这是一种专为数据稀缺且带噪的科学机器学习(SciML)任务设计的新型神经网络架构。
While existing efficient Kolmogorov-Arnold Network (KAN) variants (such as Chebyshev, wavelet, and radial-basis-function models) perform well on clean benchmarks, they suffer massive performance degradation under noisy conditions. For instance, ChebyKAN's test mean squared error (MSE) spikes by 10.6× under a \(\sigma = 0.1\) noise level, compared to 1.4× for ER-KAN.
尽管现有的高效柯尔莫哥洛夫-阿诺德网络(KAN)变体(如切比雪夫、小波和径向基函数模型)在干净的基准测试中表现良好,但它们在噪声条件下会遭受巨大的性能退化。例如,在 \(\sigma = 0.1\) 的噪声水平下,ChebyKAN 的测试均方误差(MSE)飙升了 10.6倍,而 ER-KAN 仅增大了 1.4倍。
ER-KAN achieves this resilience by combining three core design elements: 1. Shared Gaussian RBF bases across layer edges for locality and efficient parameterization. 2. Curriculum noise injection during training to explicitly train robustness. 3. Entropy-weighted adaptive regularization to prevent overfitting in data-scarce regimes (\(N\)).
ER-KAN 通过结合三个核心设计元素实现了这种极强的韧性: 1. 跨层边缘的共享高斯 RBF 基,用于实现局部性并进行高效的参数化。 2. 训练期间的课程噪声注入,以显式训练鲁棒性。 3. 熵加权自适应正则化,用于在数据稀缺区间(\(N\))中防止过拟合。
Evaluation across eight analytic functions, a damped harmonic oscillator physics-informed neural network (PINN)—where ER-KAN achieves a 4.2× lower solution MSE than standard Multi-Layer Perceptrons (MLPs)—and Burgers' equation demonstrates the model's superior reliability. The author also introduces the noise degradation ratio as a standardized metric for efficient-KAN research.
通过对八个解析函数、一个阻尼谐振子物理信息神经网络(PINN,其中 ER-KAN 的求解 MSE 比标准多层感知机低 4.2倍)以及伯格斯方程的评估,证明了该模型卓越的可靠性。作者还引入了噪声退化率作为高效 KAN 研究的标准化指标。
📌 Document Overview & Metadata
- arXiv Identifier: arXiv:2608.14773 [cs.LG] (v2)
- Title: ER-KANs: Efficient and Robust Kolmogorov-Arnold Networks for Data-Scarce Scientific Machine Learning
- Author: Harshil Lodhiya
- Primary Subject: Machine Learning (
cs.LG), Artificial Intelligence (cs.AI) - MSC Classes: 68T05, 41A30, 65D15
- ACM Classes: I.2.6; G.1.2; I.5.1
- Submitted: August 14, 2026; Last revised: August 25, 2026
- Resources & Code: GitHub Repository
- License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (View License:
)
- arXiv 标识符: arXiv:2608.14773 [cs.LG] (v2)
- 标题: ER-KANs: Efficient and Robust Kolmogorov-Arnold Networks for Data-Scarce Scientific Machine Learning
- 作者: Harshil Lodhiya
- 主要学科: 机器学习 (
cs.LG)、人工智能 (cs.AI)- MSC 分类: 68T05, 41A30, 65D15
- ACM 分类: I.2.6; G.1.2; I.5.1
- 提交时间: 2026年8月14日;最后修订: 2026年8月25日
- 资源与代码: GitHub 仓库
- 许可证: 知识共享 署名-非商业性使用-禁止演绎 4.0 国际版 (查看许可证:
)
🔬 Abstract & Key Findings
Recent efficient-KAN literature has heavily prioritized clean-data benchmarking, masking severe architectural vulnerabilities to noise. When evaluated against clean ground truth under corrupted training data (\(\sigma = 0.1\)): * ChebyKAN: Test MSE increases by 10.6× * Vanilla KAN: Test MSE increases by 7.9× * Standard MLP: Test MSE increases by 1.7× * ER-KAN (Proposed): Test MSE increases by just 1.4×
近期关于高效 KAN 的文献过度关注干净数据的基准测试,掩盖了其对噪声的严重架构脆弱性。在受到污染的训练数据(\(\sigma = 0.1\))下与干净的真实值进行对比评估时: * ChebyKAN: 测试 MSE 增加了 10.6倍 * Vanilla KAN: 测试 MSE 增加了 7.9倍 * 标准 MLP: 测试 MSE 增加了 1.7倍 * ER-KAN(本研究提出): 测试 MSE 仅增加 1.4倍
Architectural Innovations of ER-KAN
- Shared Gaussian RBF Bases: Implemented across all edges within a layer, optimizing locality and keeping parameter counts exceptionally low (e.g., matching MLP accuracy using a compact 595-parameter footprint).
- Curriculum Noise Injection: Trains the network progressively against structured noise to build resilience.
- Entropy-Weighted Adaptive Regularization: Intelligently restrains model complexity to protect against severe overfitting when sample size \(N\) is small.
ER-KAN 的架构创新
- 共享高斯 RBF 基: 在层内的所有边缘上实现,优化了局部性并使参数数量保持在极低水平(例如,使用仅 595 个参数的紧凑占用空间即可匹配 MLP 的准确率)。
- 课程噪声注入: 针对结构化噪声逐步训练网络,以建立鲁棒性。
- 熵加权自适应正则化: 当样本量 \(N\) 较小时,智能抑制模型复杂度,防止严重过拟合。
📊 Experimental Evaluation
ER-KAN was rigorously tested across multiple complex scientific scenarios: 1. Analytic Functions: Evaluated across eight different functions with varying sample sizes (\(N \in \{50, 200, 500\}\)) and noise levels (\(\sigma \in \{0, 0.03, 0.1\}\)). 2. Damped Harmonic Oscillator (PINN): ER-KAN achieved a 4.2× lower solution MSE compared to a standard MLP. 3. Burgers' Equation (PINN): Transparently reported as a failure case where all evaluated models struggled to converge, highlighting actual architectural limitations rather than hiding them.
ER-KAN 在多个复杂的科学场景中进行了严格测试: 1. 解析函数: 在具有不同样本大小(\(N \in \{50, 200, 500\}\))和噪声水平(\(\sigma \in \{0, 0.03, 0.1\}\))的八种不同函数上进行评估。 2. 阻尼谐振子(PINN): 相比于标准 MLP,ER-KAN 实现了低 4.2倍 的求解 MSE。 3. 伯格斯方程(PINN): 透明地报告为一个失败案例,所有评估的模型在该案例中均难以收敛,凸显了真实的架构局限性而非刻意隐瞒。
🔗 Navigation & External Links
- Full-Text Options: View PDF | HTML Version | TeX Source
- Citations & References: NASA ADS | Google Scholar | Semantic Scholar
- Bookmarks: BibSonomy | Reddit