arXiv · Machine Learning Theory· Xin Teng, Muxiao Li, Hongyi Wen·· 4 hr agoAI score31
分布鲁棒 Mixture-of-Experts 训练
Distributionally Robust Mixture-of-Experts Training
AI brief
研究者提出 Distributionally Robust MoE Training（DRMoET），把逐层专家视为内生鲁棒组，优化高损失路由结果而非仅均衡流量。
Source: arXiv · Machine Learning Theory · arxiv.org