arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2206.04030stat.MLcs.LGmath.PRmath.STstat.TH
收藏

SGD的高维极限定理:有效动力学与临界标度

High-dimensional limit theorems for SGD: Effective dynamics and critical scaling

Courant Institute, New York University(纽约大学柯朗研究所);Northwestern University(西北大学);University of Waterloo(滑铁卢大学)

Gerard Ben Arous, Reza Gheissari, Aukosh Jagannath

AI总结:

本文研究高维恒定步长SGD的标度极限,提出可按统计量、初始化与步长选择获得ODE/SDE极限的定理框架,揭示临界步长修正及过参数化降低次优收敛概率的作用。

摘要:

我们研究高维情形下恒定步长随机梯度下降(SGD)的标度极限。我们证明了当维度趋于无穷大时,SGD的摘要统计量(即有限维函数)轨迹的极限定理。我们的方法允许选择所追踪的摘要统计量、初始化方式和步长。它同时给出弹道式(ODE)和扩散式(SDE)极限,且极限显著依赖于前述选择。我们展示了步长的一个临界标度区间:低于该标度时,有效弹道动力学与总体损失的梯度流相匹配;但在该临界标度处,会出现一个新的修正项,改变相图。围绕该有效动力学的不动点,相应的扩散极限可能相当复杂,甚至是退化的。我们在若干常用例子上演示了该方法,包括尖峰矩阵和张量模型的估计,以及通过两层网络对二分类和XOR型高斯混合模型进行分类。这些例子展现出令人惊讶的现象,包括收敛的多模态时间尺度,以及从随机(例如高斯)初始化出发,以远离零的概率收敛到次优解。同时,我们通过证明当第二层宽度增大时,后一概率趋于零,展示了过参数化的益处。

英文摘要:

We study the scaling limits of stochastic gradient descent (SGD) with constant step-size in the high-dimensional regime. We prove limit theorems for the trajectories of summary statistics (i.e., finite-dimensional functions) of SGD as the dimension goes to infinity. Our approach allows one to choose the summary statistics that are tracked, the initialization, and the step-size. It yields both ballistic (ODE) and diffusive (SDE) limits, with the limit depending dramatically on the former choices. We show a critical scaling regime for the step-size, below which the effective ballistic dynamics matches gradient flow for the population loss, but at which, a new correction term appears which changes the phase diagram. About the fixed points of this effective dynamics, the corresponding diffusive limits can be quite complex and even degenerate. We demonstrate our approach on popular examples including estimation for spiked matrix and tensor models and classification via two-layer networks for binary and XOR-type Gaussian mixture models. These examples exhibit surprising phenomena including multimodal timescales to convergence as well as convergence to sub-optimal solutions with probability bounded away from zero from random (e.g., Gaussian) initializations. At the same time, we demonstrate the benefit of overparametrization by showing that the latter probability goes to zero as the second layer width grows.

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。

↑