分解不確定性 - AWS 方案指引

本文為英文版的機器翻譯版本,如內容有任何歧義或不一致之處,概以英文版為準。

分解不確定性

貝葉斯神經網路 (BNNs) 會產生預測分佈 Mathematical formula showing conditional probability of y given x. ,提供一組不同的預測,您可以從中估計變異數 Mathematical symbol representing a function V with empty parentheses. ;也就是總預測不確定性 Mathematical square root symbol with variable x inside. 。總預測不確定性可以使用總差異法,分成這兩個不確定因素:

總變異數法

目標變數 Predictive distribution 的預期值 Predictive distribution ,指定指定 BNN Theta symbol representing an angle or mathematical concept. 的輸入 X icon, typically used to represent closing or canceling an action. 和隨機參數 Mathematical expression showing expectation of y given x and theta. ,是由具有單一轉送傳播的 BNN 預估,並表示為 Mathematical function f(x, θ) with x and θ as variables. 。指定輸入和隨機參數的目標變異 Mathematical formula showing nabla operator applied to vector y with respect to x and theta. ,也會由 BNN 輸出,並表示為 Mathematical formula showing s prime as a function of x and theta. 。因此,總預測不確定性是這兩個數字的總和:

  • BNN 的預測平均值差異 Mathematical notation showing the gradient of a function f with respect to theta. – 流行不確定性

  • BNN 預測變異數的平均值 Mathematical expression showing expectation of s squared, given theta. :正式不確定性

下列公式示範如何根據 (Kendall and Gal 2017) 計算總不確定性。BNNs輸入 X icon, typically used to represent closing or canceling an action. 、產生隨機參數組態 Theta symbol representing an angle or mathematical concept. ,並透過神經網路進行單一轉送傳播,以輸出平均值 Mathematical function f(x, θ) with x and θ as variables. 和變異數 Mathematical formula showing s prime as a function of x and theta. 。我們以 ~ 表示隨機產生或模擬。使用固定的 X icon, typically used to represent closing or canceling an action. ,您可以重複此程序 Lowercase letter "T" in a serif font against a white background. 多次,以產生集合:

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

這些 Lowercase letter "T" in a serif font against a white background. 許多範例 Mathematical formula showing a sequence of functions f and s with subscripts and superscripts. 提供必要的統計資料來確定不確定性。您可以分別估計流行不確定性和暫時性不確定性,然後採取其總和,如本節中第一個方程式所示。