Live Gradient Compensation for Evading Stragglers in Distributed Learning2021-05-100 PRE AI DOI:10.1109/infocom42981.2021.9488815原文链接分享收藏摘要 En