顶尖高手过招，为何有时“笨办法”反而更有效？

[LG] Is your batch size the problem? Revisiting the Adam-SGD gap in language modeling  

[Max Planck Institute for Intelligent Systems]  

arxiv.org