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 2602.05657cs.LGmath.OC
收藏

非凸优化中(裁剪)SGD的严格长期尾衰减

Tight Long-Term Tail Decay of (Clipped) SGD in Non-Convex Optimization

École Polytechnique Fédérale de Lausanne(瑞士联邦理工学院洛桑分校);University of Novi Sad(诺维萨德大学);Carnegie Mellon University(卡内基梅隆大学)

Aleksandar Armacki, Dragana Bajović, Dušan Jakovetić, Soummya Kar, Ali H. Sayed

AI总结:

通过大偏差理论,研究非凸优化中SGD和裁剪SGD的长期尾衰减,给出梯度范数平方的指数级上界和下界,证明衰减率可达$e^{-t/\log(t)}$量级,比现有有限时间界快一个数量级。

摘要:

由于能够为算法的单次运行提供强保证,对SGD诱导过程的尾部行为的研究引起了广泛兴趣。虽然许多工作提供了高概率保证(量化固定概率阈值下的误差率),但缺乏直接研究失败概率的工作,即量化固定误差阈值下的尾部衰减率。此外,现有结果具有有限时间性质,限制了它们捕捉真实长期尾部衰减的能力,而后者对于现代学习模型(通常训练数百万次迭代)更具信息量。我们的工作通过大偏差理论的视角研究基于SGD的方法的长期尾部衰减,填补了这些空白,在此过程中建立了若干强结果。首先,对于非凸成本和有界噪声,我们给出了(普通)SGD产生的最佳迭代的梯度范数平方的尾部上界,长期衰减率为$e^{-t/\log(t)}$。接着,我们通过考虑在具有有界$p$阶矩($p \in (1,2]$)的重尾噪声下的裁剪SGD(c-SGD)来放宽噪声假设,证明了长期衰减率为$e^{-t^{\beta_p}/\log(t)}$的上界,其中当$p \in (1,2)$时$\beta_p = \frac{4(p-1)}{3p-2}$,当$p=2$时衰减率为$e^{-t/\log^2(t)}$。最后,我们给出了尾部衰减的下界,衰减率为$e^{-t}$,表明我们关于SGD和c-SGD的衰减率在多项式对数因子意义下是紧的。值得注意的是,我们的结果表明,与基于有限时间界的现有工作(分别显示SGD和c-SGD的衰减率为$e^{-\sqrt{t}}$和$e^{-t^{\beta_p/2}}$,$p \in (1,2]$)相比,长期尾部衰减快一个数量级。因此,我们揭示了尾部衰减比先前已知快得多的机制,为单次运行提供了更强的长期保证。

英文摘要:

The study of tail behaviour of SGD-induced processes has been attracting a lot of interest, due to offering strong guarantees with respect to individual runs of an algorithm. While many works provide high-probability guarantees, quantifying the error rate for a fixed probability threshold, there is a lack of work directly studying the probability of failure, i.e., quantifying the tail decay rate for a fixed error threshold. Moreover, existing results are of finite-time nature, limiting their ability to capture the true long-term tail decay which is more informative for modern learning models, typically trained for millions of iterations. Our work closes these gaps, by studying the long-term tail decay of SGD-based methods through the lens of large deviations theory, establishing several strong results in the process. First, we provide an upper bound on the tails of the gradient norm-squared of the best iterate produced by (vanilla) SGD, for non-convex costs and bounded noise, with long-term decay at rate $e^{-t/\log(t)}$. Next, we relax the noise assumption by considering clipped SGD (c-SGD) under heavy-tailed noise with bounded moment of order $p \in (1,2]$, showing an upper bound with long-term decay at rate $e^{-t^{β_p}/\log(t)}$, where $β_p = \frac{4(p-1)}{3p-2}$ for $p \in (1,2)$ and $e^{-t/\log^2(t)}$ for $p = 2$. Finally, we provide lower bounds on the tail decay, at rate $e^{-t}$, showing that our rates for both SGD and c-SGD are tight, up to poly-logarithmic factors. Notably, our results demonstrate an order of magnitude faster long-term tail decay compared to existing work based on finite-time bounds, which show rates $e^{-\sqrt{t}}$ and $e^{-t^{β_p/2}}$, $p \in (1,2]$, for SGD and c-SGD, respectively. As such, we uncover regimes where the tails decay much faster than previously known, providing stronger long-term guarantees for individual runs.

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。

↑