arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2103.01447cs.LGcs.DCmath.OC
收藏

ZeroSARAH:零全梯度计算的高效非凸最优算法

ZeroSARAH: Efficient Nonconvex Finite-Sum Optimization with Zero Full Gradient Computation

King Abdullah University of Science Technology(阿卜杜拉国王科技大学)

Zhize Li, Slavomír Hanzely, Peter Richtárik

摘要:

我们提出了ZeroSARAH --方差缩减方法SARAH的一种新变体(Nguyen等人,2017)--用于最小化大量非凸函数的平均值$\frac{1}{n}\sum_{i=1}^{n}f_i(x)$。据我们所知,在这个非凸有限和制度,所有现有的方差减少的方法,包括SARAH,SVRG,SAGA和他们的变种,需要计算在初始点$x^0$的所有$n$数据样本的全梯度,然后定期计算全梯度每隔几次迭代(SVRG,SARAH和他们的变种)。请注意,SVRG、SAGA及其变体通常比SARAH的变体实现更弱的收敛结果:$n^{2/3}/ε^2$与$n^{1/2}/ε^2$。因此,我们专注于莎拉的变体。提出的ZeroSARAH及其分布式变体D-ZeroSARAH是\emph{first}方差缩减算法,其\emph{do not require any full gradient computations}甚至不针对初始点。此外,对于标准和分布式设置,我们证明了ZeroSARAH和D-ZeroSARAH获得了新的最先进的收敛结果,这可以改进以前最著名的结果(例如,SPIDER、SARAH和PAGE)。避免任何完整的梯度计算(这是耗时的步骤)在许多应用中是重要的,因为数据样本$n$的数量通常非常大。特别是在分布式设置中,对所有数据样本的全梯度的周期性计算需要周期性地同步所有客户端/设备/机器,这可能是不可能的或负担不起的。因此,我们预计ZeroSARAH/D-ZeroSARAH将对分布式和联邦学习产生实际影响,因为在这些学习中,完全设备参与是不切实际的。

英文摘要:

We propose ZeroSARAH -- a novel variant of the variance-reduced method SARAH (Nguyen et al., 2017) -- for minimizing the average of a large number of nonconvex functions $\frac{1}{n}\sum_{i=1}^{n}f_i(x)$. To the best of our knowledge, in this nonconvex finite-sum regime, all existing variance-reduced methods, including SARAH, SVRG, SAGA and their variants, need to compute the full gradient over all $n$ data samples at the initial point $x^0$, and then periodically compute the full gradient once every few iterations (for SVRG, SARAH and their variants). Note that SVRG, SAGA and their variants typically achieve weaker convergence results than variants of SARAH: $n^{2/3}/ε^2$ vs. $n^{1/2}/ε^2$. Thus we focus on the variant of SARAH. The proposed ZeroSARAH and its distributed variant D-ZeroSARAH are the \emph{first} variance-reduced algorithms which \emph{do not require any full gradient computations}, not even for the initial point. Moreover, for both standard and distributed settings, we show that ZeroSARAH and D-ZeroSARAH obtain new state-of-the-art convergence results, which can improve the previous best-known result (given by e.g., SPIDER, SARAH, and PAGE) in certain regimes. Avoiding any full gradient computations (which are time-consuming steps) is important in many applications as the number of data samples $n$ usually is very large. Especially in the distributed setting, periodic computation of full gradient over all data samples needs to periodically synchronize all clients/devices/machines, which may be impossible or unaffordable. Thus, we expect that ZeroSARAH/D-ZeroSARAH will have a practical impact in distributed and federated learning where full device participation is impractical.

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。

↑