Controllable Invariance through Adversarial Feature Learning 1705.11122.pdf
Given paired observations <x,y>, we are interested in the task of predicting the target y based on thvalue of x using a discriminative approach, i.e. directly modeling the conditional distribution p(y|x).As the input x can have highly complicated structure, we employ a dedicated model or algorithm toextract an expressive representation h from x. In addition, we have access to some intrinsic attribute sof x as well as a prior belief that the prediction result should be invariant to s. Thus, when we extractthe representation h from x, we want the representation h to preserve variations that are necessary topredict y while eliminating information of s.
給定成對(duì)觀測(cè)值<x,y>份殿，我們感興趣的是使用判別方法基于x的值來(lái)預(yù)測(cè)目標(biāo)y的任務(wù)汉操，即直接建模條件分布p(y|x)。由于輸入x可以具有高度復(fù)雜的結(jié)構(gòu) 我們采用一種專用的模型或算法從x中提取一個(gè)表達(dá)式h。 此外,我們可以訪問(wèn)一些x的內(nèi)在屬性以及先驗(yàn)的觀點(diǎn)匹耕，即預(yù)測(cè)結(jié)果應(yīng)該是不變的。 因此瀑志，當(dāng)我們從x中提取表示h時(shí)小槐，我們希望表示h在消除s的信息時(shí)保留必要的預(yù)測(cè)y的變化。
To achieve the aforementioned goal, we employ a deterministic encoder E to obtain the representation by encoding x and s into h, namely, h = E(x, s). It should be noted here that we are using s as an additional input. Intuitively, this can inform and guide the encoder to remove information about undesired variations within the representation. For example, if we want to learn a representation of image x that is invariant to the lighting condition s, the model can learn to “brighten” the representation if it knows the original picture is dark, and vice versa.
為了實(shí)現(xiàn)上述目標(biāo)拍屑，我們采用確定性編碼器E來(lái)通過(guò)將x和s編碼為h來(lái)獲得表示途戒，即h = E（x，s）僵驰。 這里應(yīng)該注意喷斋，我們使用s作為附加輸入。 直觀地蒜茴，這可以通知和指導(dǎo)編碼器去除表示內(nèi)的不期望的變化的信息星爪。 例如，如果我們要學(xué)習(xí)對(duì)照明條件s不變的圖像x的表示粉私，則如果知道原始圖像較暗顽腾，模型可以學(xué)習(xí)“增亮”表示，反之亦然诺核。
