Unverified50% confidenceFactExact time
困惑度等于测试集交叉熵损失的指数：PPL = exp(loss)，loss=1.77对应PPL≈5.86
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Verified分类任务中的交叉熵损失本质上是对数似然函数的负值，最小化交叉熵等价于最大化模型对训练数据的对数似然76% similarUnverified交叉熵损失函数 H(p,q) = -Σp(x)log q(x) 的最小值在q=p时取到，等于真实分布的熵75% similarUnverified线性回归的最小二乘损失与逻辑回归的交叉熵损失分别对应高斯噪声假设与伯努利分布假设下的最大似然估计72% similarUnverified逻辑回归的损失函数（交叉熵）本质上就是负对数似然70% similarUnverified当模型对某个训练样本的预测概率接近1时，该样本对应的交叉熵损失趋近于零，模型几乎无法获得梯度更新信号69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/912650API
curl https://kongchang.com/api/v1/knowledge/claims/912650MCP
get_claim(id=912650)