[KongchangAI]
ConceptBayesian Methods / 贝叶斯神经网络 / MC Dropout

贝叶斯方法

从参数后验分布出发进行推断的统计方法,在深度学习中通过 MC Dropout、贝叶斯神经网络等技术实现不确定性估计,理论优雅但计算开销较大。

Timeline (last 90 days)

Oct 12

贝叶斯方法(如 MC Dropout、贝叶斯神经网络)从参数后验分布出发,通过多次随机前向传播采样来近似预测分布。

Unverified50%
Oct 12

分位数回归、集成学习和贝叶斯方法三种不确定性估计路径各有取舍:分位数回归计算最省,集成学习实现简单但训练成本高,贝叶斯方法理论最优雅但收敛慢。

Unverified50%
Oct 2

Methods for addressing measurement error include using multiple measurements or instrumental variables, and employing error-correction models or Bayesian methods to explicitly model the error.

Unverified50%
Oct 2

Bayesian methods can use historical failure rates as a prior and combine them with new observations to update the posterior distribution, conservatively estimating true risk when samples are scarce

Unverified50%
Aug 26

贝叶斯方法在20世纪90年代曾被神经网络热潮边缘化,但其理论价值在后来的深度学习不确定性量化中得到重新认可

Unverified50%

All Facts (5)

Source Articles