Unverified50% confidenceFactExact time
训练指标的周期性波动往往与训练数据的遍历方式有关，若每个epoch数据顺序固定或困难样本固定位置出现会导致规律性起伏
1
Sources
50%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
Related Claims
Unverified批次归一化在训练模式使用当前批次统计量，在推理模式使用训练过程中累积的移动平均值78% similarUnverifiedTraditional large model training relies heavily on manually annotated data and external corpus collection, which are typically the bottlenecks constraining iteration speed.75% similarUnverifiedBehavioral cloning suffers from the distribution shift problem: during training the policy only encounters the state distribution covered by expert trajectories, and during deployment the accumulation of small deviations causes the actual state distribution visited to gradually diverge from the training distribution.73% similarUnverified数据增强通过对训练样本施加随机变换（如旋转、翻转、缩放、裁剪、颜色抖动等）来人工扩大训练集规模，其理论基础是正则化72% similarUnverified一个高质量的路由器往往需要数千到数万条标注样本来训练，且决策边界会随底层模型更新而失效需持续重训72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/692748API
curl https://kongchang.com/api/v1/knowledge/claims/692748MCP
get_claim(id=692748)