不确定性分解 - AWS 规范性指导

本文属于机器翻译版本。若本译文内容与英语原文存在差异,则一律以英文原文为准。

不确定性分解

贝叶斯神经网络 (BNNs) 产生预测分布 Mathematical formula showing conditional probability of y given x. ,该分布提供了一组不同的预测,您可以根据这些预测来估计方差 Mathematical symbol representing a function V with empty parentheses. ;即总的预测不确定性。 Mathematical square root symbol with variable x inside. 使用总方差定律,可以将总预测不确定性分为这两个不确定性组成部分:

总方差定律

给定输入 X icon, typically used to represent closing or canceling an action. 和指定 BNN 的随机参数 Theta symbol representing an angle or mathematical concept. ,目标变量 Predictive distribution 的预期值 Predictive distribution Mathematical expression showing expectation of y given x and theta. 由具有单个正向传播的 BNN 估算并表示为 Mathematical function f(x, θ) with x and θ as variables. 。给定输入和随机参数,目标 Mathematical formula showing nabla operator applied to vector y with respect to x and theta. 的方差也由 BNN 输出,并表示为 Mathematical formula showing s prime as a function of x and theta. 。因此,总预测不确定性是这两个数字的总和:

  • BNN 预测均值的方差 Mathematical notation showing the gradient of a function f with respect to theta. :认知不确定性

  • BNN 预测方差的平均值 Mathematical expression showing expectation of s squared, given theta. :随机不确定性

以下公式演示了如何根据(Kendall and Gal 2017)计算总不确定性。 BNNs 输入 X icon, typically used to represent closing or canceling an action. ,生成随机参数配置 Theta symbol representing an angle or mathematical concept. ,然后通过神经网络进行单次正向传播以输出均值 Mathematical function f(x, θ) with x and θ as variables. 和方差 Mathematical formula showing s prime as a function of x and theta. 。我们用 ~ 表示随机生成或模拟。使用固定的 X icon, typically used to represent closing or canceling an action. ,您可以多次重复这个过程 Lowercase letter "T" in a serif font against a white background. 来生成一个集合:

Mathematical formula showing calculation of total uncertainty using Bayesian Neural Networks.

这些 Lowercase letter "T" in a serif font against a white background. 许多样本 Mathematical formula showing a sequence of functions f and s with subscripts and superscripts. 为确定不确定性提供了必要的统计数据。要做到这一点,可以分别估计认知不确定性和随机不确定性,然后对求出它们的总和,如本部分第一个等式所示。