具注意力機制之類別層級映射模型用於無監督式域適應

Abstract

無監督式域適應的目標是訓練一個泛化的分類器來去預測目標資料的類別，來源資料是有被標註的，但是目標資料卻沒有被標註，且來源資料跟目標資料的分布有一定的落差。先前的論文已經在對齊來源跟目標資料取得一定的進步。然後這些論文的表現依然會受到一定程度的限制，因為他們少考慮了類別層級的問題，而那些問題會影響分類器的邊界效果。因此最近有研究開始考慮類別層級的問題並且給目標資料假標記來解決問題。在這篇論文中，我們提出了一個創新的架構-具注意力機制之類別層級映射模型，去更進一步的改進先前的研究。我們利用了源分類器將兩個域的不同類別的特徵分開，再提出了一個嶄新的目標函數-軟性類別層級差異去將兩個域的相同類別的特徵作對齊，但由於目標域沒有標注，所以我們同時利用源域目標域做K-means集群分析拿到中心點，對於每筆目標域資料看它離哪個中心點最近去給相對的假標注。另外，先前有用假標注的研究可能會有錯誤標注的問題，錯誤的標注將會造成對齊的錯誤進而影響分類。因此我們考慮了假標注的可靠性，我們不僅使用了假標註並且對於每個假標註我們會給一個信心值來決定對齊程度的大小。此外，對於圖片而言，並不是每個區域都適合做域適應，因此我們設計了一個注意力機制去找出適合做域適應的區域。根據實驗結果，我們的模型取得了一定的成功並且贏了許多先進的研究。

Unsupervised Domain Adaptation (UDA) aims to train a classifier to predict the category of unlabeled data on target domain given labeled data on source domain. Previous methods make remarkable progress in learning aligned domain-invariant representations. However, the performance of these methods is limited because they ignored categorical information which results in misleading decision boundaries. Therefore, recent researches take the issue into consideration and some of them try to assign pseudo labels to target domain data. In this paper, we propose a novel model architecture named Attentive Class-Aware Mapping Network (ACMN) to improve the class-aware alignment. We use a domain classifier separate data representations from different domains which empirically help the class-aware alignment. Then we employs novel discrepancy objective function \textit{Soft-Class Discrepancy (SCD)} to align representations within the identical categories through pseudo labels from spherical K-means clustering. Note that previous approaches may run the risk of misleading domain alignment due to the unreliability of pseudo labels. In view of this, we consider the reliability of pseudo labels and assign the confidence values to pseudo labels. Moreover, because not all the regions are suitable for domain alignment in an image, we design a novel attention mechanism to find proper transferable regions. Experimental results on several UDA benchmarks show that our method consistently outperforms the state-of-the-art approaches.