Unverified50% confidenceFactExact time
缩放点积将点积除以√d（d为嵌入维度），使点积方差保持在合理范围，让Softmax输出更平滑的概率分布以稳定训练
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
Unverified计算 softmax 函数时指数运算可能导致浮点数溢出，标准做法是先减去输入向量的最大值再取指数68% similarUnverifiedIn Naive Bayes, logarithmic probability space must be used to avoid underflow caused by multiplying together small probability values.67% similarVerifiedThe reason for dividing by √d_k in the Attention formula is to prevent excessively large dot products from causing the softmax gradient to vanish.66% similarUnverifiedFor small targets, large images should be tiled into overlapping 1008 px patches for separate inference and then merged; never resize directly, as this will produce confident but blurry boundaries or destroy small targets.65% similarUnverifiedSoftmax函数具有平移不变性，任意放大所有logit的尺度不改变预测类别，却会将输出概率压向极端值，是其天然的过度自信倾向65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/704267API
curl https://kongchang.com/api/v1/knowledge/claims/704267MCP
get_claim(id=704267)