arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2512.00860cs.LG
收藏

NTKs的谱维度是常数:隐式正则化、有限宽度稳定性及可扩展估计的理论

The Spectral Dimension of NTKs is Constant: A Theory of Implicit Regularization, Finite-Width Stability, and Scalable Estimation

MBZUAI(穆罕默德·本·拉谢德人工智能研究院)

Praveen Anilkumar Shukla

AI总结:

本文提出通过NTK的有效秩理论,揭示深度网络隐式正则化和有限宽度稳定性,并设计可扩展估计器验证理论。

摘要:

现代深度网络通常具有高度的参数化程度,但往往能够很好地泛化,这表明了一种低内在复杂度的形式,这种复杂度并未反映在参数数量上。我们通过神经切线核(NTK)Gram矩阵的有效秩来研究这种复杂度,即r_eff(K) = (tr(K))^2 / ||K||_F^2。对于独立同分布的数据和无限宽度的NTK k,我们证明了一个极限定律:当n趋近于无穷大时,E[r_eff(K_n)]的极限为E[k(x,x)]^2 / E[k(x,x')^2] =: r_∞,具有子高斯集中趋势。我们进一步建立了有限宽度稳定性:如果有限宽度的NTK在算子范数上偏离O_p(m^{-1/2})(宽度m),则r_eff的变化为O_p(m^{-1/2})。我们设计了一个可扩展的估计器,使用随机输出探测和参数雅可比的CountSketch,并证明了其条件无偏性和一致性,具有显式的方差界。在CIFAR-10上使用ResNet-20/56(宽度16/32)在n∈{10^3, 5×10^3, 10^4, 2.5×10^4, 5×10^4}时,我们观察到r_eff≈1.0-1.3,且n的斜率≈0,与理论一致,且核矩预测与拟合常数紧密吻合。

英文摘要:

Modern deep networks are heavily overparameterized yet often generalize well, suggesting a form of low intrinsic complexity not reflected by parameter counts. We study this complexity at initialization through the effective rank of the Neural Tangent Kernel (NTK) Gram matrix, $r_{\text{eff}}(K) = (\text{tr}(K))^2/\|K\|_F^2$. For i.i.d. data and the infinite-width NTK $k$, we prove a constant-limit law $\lim_{n\to\infty} \mathbb{E}[r_{\text{eff}}(K_n)] = \mathbb{E}[k(x, x)]^2 / \mathbb{E}[k(x, x')^2] =: r_\infty$, with sub-Gaussian concentration. We further establish finite-width stability: if the finite-width NTK deviates in operator norm by $O_p(m^{-1/2})$ (width $m$), then $r_{\text{eff}}$ changes by $O_p(m^{-1/2})$. We design a scalable estimator using random output probes and a CountSketch of parameter Jacobians and prove conditional unbiasedness and consistency with explicit variance bounds. On CIFAR-10 with ResNet-20/56 (widths 16/32) across $n \in \{10^3, 5\times10^3, 10^4, 2.5\times10^4, 5\times10^4\}$, we observe $r_{\text{eff}} \approx 1.0\text{--}1.3$ and slopes $\approx 0$ in $n$, consistent with the theory, and the kernel-moment prediction closely matches fitted constants.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑