中文

Lipschitz算子的深度算子网络逼近速率

数值分析 2023-07-20 v1 机器学习 数值分析

摘要

我们建立了一类神经深度算子网络(DON)的普适性与表达速率界,用于仿真可分的Hilbert空间X\mathcal XY\mathcal Y(的子集)之间的Lipschitz(或Hölder)连续映射G:XY\mathcal G:\mathcal X\to\mathcal Y。所考虑的DON架构通过X\mathcal XY\mathcal Y的(双正交)Riesz基使用线性编码器E\mathcal E和解码器D\mathcal D,以及对序列空间2(N)\ell^2(\mathbb N)上Lipschitz连续的无限维参数坐标映射的逼近网络。与先前工作([Herrmann, Schwab and Zech: Neural and Spectral operator surrogates: construction and expression rate bounds, SAM Report, 2022]、[Marcati and Schwab: Exponential Convergence of Deep Operator Networks for Elliptic Partial Differential Equations, SAM Report, 2022])要求例如G\mathcal G全纯不同,本表达速率结果仅要求G\mathcal G的Lipschitz(或Hölder)连续性。当前表达速率界证明的关键在于使用受Kolmogorov叠加定理启发的超表达激活函数(例如[Yarotski: Elementary superexpressive activations, Int. Conf. on ML, 2021]、[Shen, Yang and Zhang: Neural network approximation: Three hidden layers are enough, Neural Networks, 2021]及其参考文献),或使用最近在[Zhang, Shen and Yang: Neural Network Architecture Beyond Width and Depth, Adv. in Neural Inf. Proc. Sys., 2022]中提出的具有标准(ReLU)激活的非标准NN架构。我们通过以下仿真逼近速率界说明了抽象结果:a)参数椭圆变分不等式的解算子,以及b)Hilbert-Schmidt算子的Lipschitz映射。

关键词

引用

@article{arxiv.2307.09835,
  title  = {Deep Operator Network Approximation Rates for Lipschitz Operators},
  author = {Christoph Schwab and Andreas Stein and Jakob Zech},
  journal= {arXiv preprint arXiv:2307.09835},
  year   = {2023}
}

备注

31 pages