Unverified50% confidenceFactExact time
点估计置信度在分布偏移下会失效，而MC Dropout等贝叶斯式不确定性度量能真正识别信号失效
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMC Dropout通过在推理时保持Dropout激活并多次前向传播来近似贝叶斯推断，从预测方差中量化认知不确定性77% similarUnverified当一组采样结果全部错误时，组相对优势会坍缩为零，导致高不确定性场景下模型拿不到有效的梯度反馈73% similarUnverified在分布外（out-of-distribution）输入上，高置信度的错误判断相当常见72% similarVerified双重下降现象表明，当模型参数量远超数据量时，测试误差反而可能再次下降71% similarUnverified当某次发布导致评估指标明显下滑时，这是触发回滚的客观信号70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/986750API
curl https://kongchang.com/api/v1/knowledge/claims/986750MCP
get_claim(id=986750)