Universal Approximation of Input-Output Maps by Temporal Convolutional Nets

被引:0
|
作者
Hanson, Joshua [1 ]
Raginsky, Maxim [1 ]
机构
[1] Univ Illinois, Urbana, IL 61801 USA
来源
ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019) | 2019年 / 32卷
基金
美国国家科学基金会;
关键词
STABILITY;
D O I
暂无
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
There has been a recent shift in sequence-to-sequence modeling from recurrent network architectures to convolutional network architectures due to computational advantages in training and operation while still achieving competitive performance. For systems having limited long-term temporal dependencies, the approximation capability of recurrent networks is essentially equivalent to that of temporal convolutional nets (TCNs). We prove that TCNs can approximate a large class of input-output maps having approximately finite memory to arbitrary error tolerance. Furthermore, we derive quantitative approximation rates for deep ReLU TCNs in terms of the width and depth of the network and modulus of continuity of the original input-output map, and apply these results to input-output maps of systems that admit finite-dimensional state-space realizations (i.e., recurrent models).
引用
收藏
页数:11
相关论文
共 50 条