English

Learning Hierarchical Polynomials of Multiple Nonlinear Features with Three-Layer Networks

Machine Learning 2024-11-27 v1 Artificial Intelligence Statistics Theory Machine Learning Statistics Theory

Abstract

In deep learning theory, a critical question is to understand how neural networks learn hierarchical features. In this work, we study the learning of hierarchical polynomials of \textit{multiple nonlinear features} using three-layer neural networks. We examine a broad class of functions of the form f=g\bpf^{\star}=g^{\star}\circ \bp, where \bp:RdRr\bp:\mathbb{R}^{d} \rightarrow \mathbb{R}^{r} represents multiple quadratic features with rdr \ll d and g:RrRg^{\star}:\mathbb{R}^{r}\rightarrow \mathbb{R} is a polynomial of degree pp. This can be viewed as a nonlinear generalization of the multi-index model \citep{damian2022neural}, and also an expansion upon previous work that focused only on a single nonlinear feature, i.e. r=1r = 1 \citep{nichani2023provable,wang2023learning}. Our primary contribution shows that a three-layer neural network trained via layerwise gradient descent suffices for \begin{itemize}\item complete recovery of the space spanned by the nonlinear features \item efficient learning of the target function f=g\bpf^{\star}=g^{\star}\circ \bp or transfer learning of f=g\bpf=g\circ \bp with a different link function \end{itemize} within \cO~(d4)\widetilde{\cO}(d^4) samples and polynomial time. For such hierarchical targets, our result substantially improves the sample complexity Θ(d2p){\Theta}(d^{2p}) of the kernel methods, demonstrating the power of efficient feature learning. It is important to highlight that{ our results leverage novel techniques and thus manage to go beyond all prior settings} such as single-index and multi-index models as well as models depending just on one nonlinear feature, contributing to a more comprehensive understanding of feature learning in deep learning.

Keywords

Cite

@article{arxiv.2411.17201,
  title  = {Learning Hierarchical Polynomials of Multiple Nonlinear Features with Three-Layer Networks},
  author = {Hengyu Fu and Zihao Wang and Eshaan Nichani and Jason D. Lee},
  journal= {arXiv preprint arXiv:2411.17201},
  year   = {2024}
}

Comments

78 pages, 4 figures

R2 v1 2026-06-28T20:12:47.915Z