Random hidden layers construct a nonlinear representation of the input, while only the final readout is trained. The paper connects this construction to approximation in the reproducing kernel Hilbert space of the NNGP kernel. Research source
About this paper
Connects neural network approximation guarantees to the RKHS of the neural network Gaussian process kernel.
A random multilayer representation, followed by training only its final layer, provides constructive approximation guarantees. The required network width depends on the RKHS norm, and the analysis is compared with Barron-type bounds and numerical experiments.
Citation
BibTeX
@inproceedings{takhanov2024multilayerrandomfeatures,
title = {Multi-layer random features and the approximation power of neural networks},
author = {Takhanov, Rustem},
year = {2024},
booktitle = {UAI 2024},
url = {https://proceedings.mlr.press/v244/takhanov24a.html}
}