專利授權區國立清華大學國際產學營運總中心 Operations Center for Industry Collaboration

搜尋專利授權區

關鍵字

» 新增關鍵字

選單

專利授權區

專利授權區
專利名稱(中)	基於機器學習從輸入特徵中過濾資訊的方法與系統
專利名稱(英)	METHOD AND SYSTEM FOR FILTERING INFORMATION FROM INPUT FEATURES BASED ON MACHINE LEARNING
專利家族	中華民國：(公開號)
專利權人	國立清華大學 100.00%
發明人	李祈均,陳金博,許暐彤
技術領域	通信傳輸,資訊工程,電子電機

專利摘要(中)
本發明提出一種計算機實施方法，基於原始訊號和資訊消除模型生成不包括第一資訊的特徵，並基於此特徵和機器學習模型來執行指定任務，其中第一資訊使第一屬性可辨識。資訊消除模型的訓練包括提供第一和第二對抗網路以及最小化損失函數以訓練資訊消除模型。第一和第二對抗網路的輸入層是基於資訊消除模型的輸出層以及輸入特徵來產生。第一對抗網路的生成器和第二對抗網路的判別器用以執行指定任務，並且第一對抗網路的判別器和第二對抗網路的生成器用以辨識第一屬性。損失函數關聯於第一和第二對抗網路的輸入層的解耦損失以及各個生成器和判別器的損失。

專利摘要(中)

本發明提出一種計算機實施方法，基於原始訊號和資訊消除模型生成不包括第一資訊的特徵，並基於此特徵和機器學習模型來執行指定任務，其中第一資訊使第一屬性可辨識。資訊消除模型的訓練包括提供第一和第二對抗網路以及最小化損失函數以訓練資訊消除模型。第一和第二對抗網路的輸入層是基於資訊消除模型的輸出層以及輸入特徵來產生。第一對抗網路的生成器和第二對抗網路的判別器用以執行指定任務，並且第一對抗網路的判別器和第二對抗網路的生成器用以辨識第一屬性。損失函數關聯於第一和第二對抗網路的輸入層的解耦損失以及各個生成器和判別器的損失。

專利摘要(英)
A computer-implemented method is provided. According to this computer-implemented method, based on an original signal and an information elimination model, a feature not including first information is generated, the first information allowing a first attribute to be identifiable. A specified task is then performed based on the generated feature and a machine learning model. The training of the information elimination model includes providing first and second adversarial networks and minimizing a loss function to train the information elimination model. Input layers of the first and second adversarial networks are generated based on an output layer and an input feature of the information elimination model. The generator of the first adversarial network and the discriminator of the second adversarial network are configured to perform the specified task, while the discriminator of the first adversarial network and the generator of the second adversarial network are configured to identify the first attribute. The loss function is associated with a disentangling loss of the input layers of the first and second adversarial networks, as well as the losses of each generator and discriminator.

專利摘要(英)

A computer-implemented method is provided. According to this computer-implemented method, based on an original signal and an information elimination model, a feature not including first information is generated, the first information allowing a first attribute to be identifiable. A specified task is then performed based on the generated feature and a machine learning model. The training of the information elimination model includes providing first and second adversarial networks and minimizing a loss function to train the information elimination model. Input layers of the first and second adversarial networks are generated based on an output layer and an input feature of the information elimination model. The generator of the first adversarial network and the discriminator of the second adversarial network are configured to perform the specified task, while the discriminator of the first adversarial network and the generator of the second adversarial network are configured to identify the first attribute. The loss function is associated with a disentangling loss of the input layers of the first and second adversarial networks, as well as the losses of each generator and discriminator.

聯絡資訊
承辦人姓名	李曉琪
承辦人電話	03-5715131 #31061
承辦人Email	hsiaochi@mx.nthu.edu.tw