虛擬數據生成於零樣本學習

Abstract

零樣本學習(zero shot learning)是指透過從已知類別的訓練資料中，將學習到的知識遷移到能辨識未知類別上。Li的方法[22]透過語義嵌入和視覺特徵當作訓練資料，用深度學習模型將影像特徵轉變為語義分類器，並與語義嵌入做線性/非線性組合，藉此來找出與該視覺特徵相符的類別。然而僅透過已知類別的資料當作訓練樣本，預測的成效可能有限，因為模型完全沒有對於未知類別的資訊。為了解決這個問題，本論文提出藉由隨機取樣的方式從已知類別的樣本中，產生出的新樣本，將其當作模擬未知類別的訓練樣本，也就是虛擬資料。讓模型在學習的時候，可以模擬未知類別的存在，使分類準確率得以提升。實驗於多個公開標準資料集驗證了所提出方法的可行性。

Zero-shot learning refers to the knowledge transfer to recognize unseen classes, with a model learned from seen classes. Li’s method [22] uses a deep learning method to transform visual embeddings into semantic classifiers, and performs linear / non-linear classification on semantic embeddings to predict the class.Nevertheless, solely using data of seen classes may have a limited prediction performance because the model does not have any information of unseen classes. In order to solve this problem, we propose to generate virtual training data by randomly combining seen classes, which simulate unseen data. With such virtual data, the model can simulate the situation of recognizing unseen classes during learning, and therefore the classification accuracy can be improved. Experiments on several benchmark datasets have verified the effectiveness of the proposed method.