交互關聯:在小樣本學習上改善RENet

Abstract

小樣本分類具有挑戰性，因為目標是在給予極少量有標籤樣本的情況下對有標籤樣本進行分類。交互關聯已經被證明在小樣本學習上可以產生更具有分辨性的特徵。在relational embedding network (RENet)中，交互關聯已經被用來提取注意力。然而，在RENet中只有利用一張影像本身之內所含有的資訊，如果我們利用兩張影像之間所含有的資訊而不是只有一張影像呢?這篇論文延伸了這個想法，並且提出了三個交互注意模組，分別是cross unscaled attention (CUA)、cross scaled attention (CSA) 和 cross aligned attention (CAA)。明確來說，CUA利用交互關聯幫助模型專注在重要特徵，CSA針對不同特徵圖進行縮放式它們更加完善配對，而CAA則是採用主成分分析使來自不同影像的特徵能夠進一步對齊。我們也為我們的模型發展了兩個end-to-end版本，這兩個版本比較簡單又有效率。實驗證實CUA、CSA和CAA三者全部都能夠在四個小樣本分類廣泛使用的標準數據集上對於最先進的方法取得改善，而CUA略為快速，CAA則是達成較高準確度

Few-shot classification is challenging since the goal is to classify unlabeled samples with very few labeled samples provided. It has been shown that cross correlation helps generate more discriminative features for few-shot learning. In the relational embedding network (RENet), cross correlation has been used to extract attention. However, Only the information contained within an image itself is exploited in RENet. What if we exploit the information contained between two images instead of only one image? This thesis extends the idea and proposes three cross attention modules, namely the cross unscaled attention (CUA), the cross scaled attention (CSA), and the cross aligned attention (CAA). Specifically, CUA exploits cross correlation to help the model focus on important features, and CSA scales different feature maps to make them better matched, and CAA adopts the principal component analysis to further align features from different images. We also develop two end-to-end versions for our model, which are simpler and more efficient. Experiments showed that all CUA, CSA, and CAA achieve consistent improvements over state-of-the-art methods on four widely used few-shot classification benchmark datasets, miniImageNet, tieredImageNet, CIFAR-FS, and CUB-200-2011, while CUA is slightly faster and CAA achieves higher accuracies.