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 2502.02431cs.LGcs.AI
收藏

无调度优化器、AdEMAMix 与加速 SGD 变体之间的联系

Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD Variants

Kempner Institute, Harvard University(哈佛大学肯普纳研究所);Department of Computer Science, Harvard University(哈佛大学计算机科学系)

Depen Morwani, Nikhil Vyas, Hanlin Zhang, Sham Kakade

AI总结:

本文揭示了无调度优化器、AdEMAMix 等新算法与噪声主导情形下加速 SGD 理论之间的联系,并基于 AdEMAMix 提出 Simplified-AdEMAMix,在保持性能的同时简化了动量结构。

摘要:

深度学习优化领域的最新进展引入了一些新算法,例如无调度(Schedule-Free)优化器、AdEMAMix、MARS 和 Lion,它们对传统动量机制进行了修改。在另一条研究路线中,随机梯度下降(SGD)在噪声主导情形下的理论加速已经通过将动量系数与当前梯度的权重解耦而实现。在本文中,我们在上述两条研究路线之间建立了明确的联系。我们通过一个 1.5 亿参数规模的语言建模任务上的初步实验来验证理论发现。我们发现,AdEMAMix 最接近随机梯度下降的加速版本,并表现出更优的性能。基于这些认识,我们对 AdEMAMix 进行了修改,提出了 Simplified-AdEMAMix,它在大批量和小批量设置下均保持与 AdEMAMix 相同的性能,同时不再需要两个不同的动量项。Simplified-AdEMAMix 的代码可在以下仓库获取:https://github.com/DepenM/Simplified-AdEMAMix/。

英文摘要:

Recent advancements in deep learning optimization have introduced new algorithms, such as Schedule-Free optimizers, AdEMAMix, MARS and Lion which modify traditional momentum mechanisms. In a separate line of work, theoretical acceleration of stochastic gradient descent (SGD) in noise-dominated regime has been achieved by decoupling the momentum coefficient from the current gradient's weight. In this paper, we establish explicit connections between these two lines of work. We substantiate our theoretical findings with preliminary experiments on a 150m language modeling task. We find that AdEMAMix, which most closely resembles accelerated versions of stochastic gradient descent, exhibits superior performance. Building on these insights, we introduce a modification to AdEMAMix, termed Simplified-AdEMAMix, which maintains the same performance as AdEMAMix across both large and small batch-size settings while eliminating the need for two different momentum terms. The code for Simplified-AdEMAMix is available on the repository: https://github.com/DepenM/Simplified-AdEMAMix/.

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。

↑