Journal de Mathématiques Pures et Appliquées ( IF 2.3 ) Pub Date : 2021-07-16 , DOI: 10.1016/j.matpur.2021.07.009 Zuowei Shen 1 , Haizhao Yang 2 , Shijun Zhang 1
This paper concentrates on the approximation power of deep feed-forward neural networks in terms of width and depth. It is proved by construction that ReLU networks with width and depth can approximate a Hölder continuous function on with an approximation rate , where and are Hölder order and constant, respectively. Such a rate is optimal up to a constant in terms of width and depth separately, while existing results are only nearly optimal without the logarithmic factor in the approximation rate. More generally, for an arbitrary continuous function f on , the approximation rate becomes , where is the modulus of continuity. We also extend our analysis to any continuous function f on a bounded set. Particularly, if ReLU networks with depth 31 and width are used to approximate one-dimensional Lipschitz continuous functions on with a Lipschitz constant , the approximation rate in terms of the total number of parameters, , becomes , which has not been discovered in the literature for fixed-depth ReLU networks.
中文翻译:
ReLU 网络在宽度和深度方面的最佳逼近率
本文重点研究深度前馈神经网络在宽度和深度方面的逼近能力。构造证明,具有宽度的 ReLU 网络 和深度 可以近似一个 Hölder 连续函数 以近似率 , 在哪里 和 分别是 Hölder 阶和常数。这样的速率在宽度和深度方面分别达到常数时是最佳的,而现有的结果只是在近似速率中没有对数因子的情况下接近最佳。更一般地,对于任意连续函数f on,近似率变为 , 在哪里 是连续性的模数。我们还将我们的分析扩展到有界集合上的任何连续函数f。特别是,如果深度为 31 且宽度为 31 的 ReLU 网络 用于逼近一维 Lipschitz 连续函数 与 Lipschitz 常数 ,就参数总数而言的近似率, ,变成 ,在固定深度 ReLU 网络的文献中尚未发现。