- 學位論文
「基於類級的最大均值差異之無監督域適應深度網路」
Unsupervised Domain Adaptation Deep Network Based on Class-Wise MMD
摘要
本篇論文旨在非監督域適應（unsupervised domain adaptation, UDA），即對不具標籤的目標域資料，從具有標籤的源域資料中學習域不變性特徵（domain-invariant feature）。透過最小化兩域樣本的最大均值差異（maximum mean discrepancy, MMD）已被證明可以有效地拉近兩域樣本在特徵空間的分布，進而學習到域不變性特徵；然而在最小化兩域資料的MMD的過程中，不保證能對齊各類別資料。Long等人提出類級MMD（class-wise MMD）來解決這個問題，不過利用最小化類級MMD來拉近同一類的兩域資料會同時最大化類內距離（intra-class distance）造成特徵可辨性降低。Wang等人提出調整類級MMD裡面隱含的類內距離之權重來減輕此問題，不過該方法是在線性轉換的特徵空間中計算兩域資料均值之歐式距離來定義的MMD，這樣的定義並不符合Gretton等人在雙樣本檢定所定義之MMD性質。本篇論文將改進Wang等人所提的方法，將兩域樣本透過卷積網路（CNN）的非線性轉換映射到特徵空間，提出了在一個再現核希爾伯特空間計算的具基於類級的可辨性MMD，此MMD可以用核計巧（kernel trick）簡單計算投影空間中的向量內積。使得相同類別的兩域樣本在特徵空間的分布能夠有效對齊，同時減低特徵可辨性降低的風險，因而增強了目標域資料分類明確性，進而達到域適應的目的。
並列摘要
This paper aims to address unsupervised domain adaptation (UDA), specifically learning domain invariant features from labeled source domain data for unlabeled target domain data. By minimizing the Maximum Mean Discrepancy (MMD) between the two domains, it has been demonstrated that the distributions of samples in the feature space can effectively be brought closer together, thus facilitating the learning of domain invariant features. However, when minimizing the MMD between the two domains, there is no guar antee that data from different classes will be aligned properly. Long et al. proposed class wise MMD to tackle this issue, but minimizing class wise MMD to bring together samples from the same class in both domains simultaneously maximizes the intra class distance, which leads to a reduction in feature discriminability. Wang et al. proposed adjusting the weights of the implicit intra class distances within class wise MMD to mitigate this problem. However, their method calculates the MMD in a linearly transformed feature space, defining MMD as the Euclidean distance between the mean of two domains, which does not adhere to the properties of MMD as defined in the two sample test by Gretton et al. This paper improves upon Wang et al.'s method by non linearly transforming samples from both domains into a feature space using Convolutional Neural Networks (CNNs). It introduces a class wise discriminative MMD computed in a Reproducing Kernel Hilbert Space (RKHS). This MMD can b e easily calculated using the kernel trick, simplifying the computation of vector inner products in the projection space. This approach effectively aligns the distributions of samples from the same class in the feature space while reducing the risk of feat ure discriminability reduction. As a result, it enhances the clarity of target domain data classification, thus achieving the goal of domain adaptation.