arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2409.08770cs.LGmath.OC
收藏

同时增加批量大小和学习率可加速随机梯度下降

Increasing Both Batch Size and Learning Rate Accelerates Stochastic Gradient Descent

Meiji University(明治大学)

Hikaru Umeda, Hideaki Iiduka

AI总结:

本文理论分析四种批量与学习率调度器,证明同时增大批量和学习率(含预热衰减)可稳定最小化全梯度范数并加速小批量 SGD,数值实验验证了结论。

摘要:

小批量随机梯度下降(SGD)的性能在很大程度上取决于如何设置批量大小和学习率,以在训练深度神经网络时最小化经验损失。本文对配备四种调度器的小批量 SGD 进行了理论分析:(i)恒定批量大小与衰减学习率调度器,(ii)增加批量大小与衰减学习率调度器,(iii)增加批量大小与增加学习率调度器,以及(iv)增加批量大小与预热衰减学习率调度器。我们证明,使用调度器(i)的小批量 SGD 并不总能最小化经验损失的全梯度范数的期望,而使用调度器(ii)、(iii)和(iv)中的任意一种都可以做到。此外,调度器(iii)和(iv)能够加速小批量 SGD。本文还提供了支持这些分析的数值结果,表明使用调度器(iii)或(iv)比使用调度器(i)或(ii)能更快地最小化经验损失的全梯度范数。

英文摘要:

The performance of mini-batch stochastic gradient descent (SGD) strongly depends on setting the batch size and learning rate to minimize the empirical loss in training the deep neural network. In this paper, we present theoretical analyses of mini-batch SGD with four schedulers: (i) constant batch size and decaying learning rate scheduler, (ii) increasing batch size and decaying learning rate scheduler, (iii) increasing batch size and increasing learning rate scheduler, and (iv) increasing batch size and warm-up decaying learning rate scheduler. We show that mini-batch SGD using scheduler (i) does not always minimize the expectation of the full gradient norm of the empirical loss, whereas it does using any of schedulers (ii), (iii), and (iv). Furthermore, schedulers (iii) and (iv) accelerate mini-batch SGD. The paper also provides numerical results of supporting analyses showing that using scheduler (iii) or (iv) minimizes the full gradient norm of the empirical loss faster than using scheduler (i) or (ii).

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。

↑