Unverified50% confidenceFactExact time
Loss Scaling最初由NVIDIA在2018年提出，用于解决FP16训练中梯度下溢问题
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
NVFP4强化学习训练：4位量化稳定性挑战与工程实践
Related Claims
UnverifiedFocal Loss was proposed by Kaiming He et al. in the 2017 RetinaNet paper to address the problem of extreme foreground-background sample imbalance in object detection.68% similarUnverified应对损失尖峰的工程策略包括梯度裁剪（将梯度范数限制在1.0以内）、从尖峰前检查点回滚重训以及使用QK Normalization65% similarUnverifiedDuring training, MRL simultaneously optimizes loss functions at multiple dimensional levels, with the total loss being a weighted sum of losses at each granularity.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/510047API
curl https://kongchang.com/api/v1/knowledge/claims/510047MCP
get_claim(id=510047)