Sharp Generalization for Nonparametric Regression in Interpolation Space by Over-Parameterized Neural Networks Trained with Preconditioned Gradient Descent and Early Stopping

TL;DR

This paper demonstrates that over-parameterized neural networks trained with preconditioned gradient descent and early stopping can achieve sharp nonparametric regression rates, surpassing standard kernel regression and NTK regime results.

Contribution

The authors introduce a novel analysis framework for neural network training, showing improved generalization rates via a new kernel decomposition and local Rademacher complexity control.

Findings

Achieves regression rate of O(n^{-2αs'/(2αs'+1)}) for target functions in interpolation space.

Surpasses nearly-optimal and standard NTK regression rates.

Provides theoretical evidence that PGD enables neural networks to escape the NTK regime.

Abstract

We study nonparametric regression using an over-parameterized two-layer neural networks trained with algorithmic guarantees in this paper. We consider the setting where the training features are drawn uniformly from the unit sphere in $\RR^d$, and the target function lies in an interpolation space commonly studied in statistical learning theory. We demonstrate that training the neural network with a novel Preconditioned Gradient Descent (PGD) algorithm, equipped with early stopping, achieves a sharp regression rate of $\cO(n^{-\frac{2\alpha s'}{2\alpha s'+1}})$ when the target function is in the interpolation space $\bth{\cH_K}^{s'}$ with $s' \ge 3$. This rate is even sharper than the currently known nearly-optimal rate of $\cO(n^{-\frac{2\alpha s'}{2\alpha s'+1}})\log^2(1/\delta)$~\citep{Li2024-edr-general-domain}, where $n$ is the size of the training data and $\delta \in (0,1)$ is a small probability. This rate is also sharper than the standard kernel regression rate of $\cO(n^{-\frac{2\alpha}{2\alpha+1}})$ obtained under the regular Neural Tangent Kernel (NTK) regime when training the neural network with the vanilla gradient descent (GD), where $2\alpha = d/(d-1)$. Our analysis is based on two key technical contributions. First, we present a principled decomposition of the network output at each PGD step into a function in the reproducing kernel Hilbert space (RKHS) of a newly induced integral kernel, and a residual function with small $L^{\infty}$-norm. Second, leveraging this decomposition, we apply local Rademacher complexity theory to tightly control the complexity of the function class comprising all the neural network functions obtained in the PGD iterates. Our results further suggest that PGD enables the neural network to escape the linear NTK regime and achieve improved generalization.