Unverified50% confidenceSolutionExact time
Temperature scaling calibrates a model by dividing logit values by a learnable scalar parameter T before the softmax; T greater than 1 softens the probability distribution, T less than 1 sharpens the distribution, and the main model does not need to be retrained.
1
Sources
50%
Confidence
Long-term
Relevance
9/29/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSoftmax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69% similarUnverifiedSoftmax输出的概率值并不等同于真实的贝叶斯后验概率，而是被过度优化以拉开类别间距，导致预测概率系统性偏高68% similarUnverified温度超参数τ越小分布越集中，模型对难负样本越敏感，但τ过小会导致梯度对异常值敏感、训练不稳定67% similarUnverifiedHinton在原始论文中引入温度参数T来控制概率分布的柔软程度，T越高分布越平滑，类间相似关系暴露越充分，这些类间关系被称为暗知识（Dark Knowledge）66% similarVerifiedThe higher the temperature parameter, the flatter the probability distribution output by Softmax, and the more creative and diverse the output; the lower the temperature, the sharper the distribution and the more deterministic and conservative the output.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/958811API
curl https://kongchang.com/api/v1/knowledge/claims/958811MCP
get_claim(id=958811)