Smaller generalization error derived for deep compared to shallow residual neural networks

Abstract

Estimates of the generalization error are proved for a residual neural network with $L$ random Fourier features layers $\bar z_{\ell+1}=\bar z_\ell + \text{Re}\sum_{k=1}^K\bar b_{\ell k}e^{{\rm i}\omega_{\ell k}\bar z_\ell}+ \text{Re}\sum_{k=1}^K\bar c_{\ell k}e^{{\rm i}\omega'_{\ell k}\cdot x}$. An optimal distribution for the frequencies $(\omega_{\ell k},\omega'_{\ell k})$ of the random Fourier features $e^{{\rm i}\omega_{\ell k}\bar z_\ell}$ and $e^{{\rm i}\omega'_{\ell k}\cdot x}$ is derived. The derivation is based on the corresponding generalization error to approximate function values $f(x)$. The generalization error turns out to be smaller than the estimate ${\|\hat f\|^2_{L^1(\mathbb{R}^d)}}/{(LK)}$ of the generalization error for random Fourier features with one hidden layer and the same total number of nodes $LK$, in the case the $L^\infty$-norm of $f$ is much less than the $L^1$-norm of its Fourier transform $\hat f$. This understanding of an optimal distribution for random features is used to construct a new training method for a deep residual network that shows promising results.