在現有的機器學習方法中,大多假設所有處理的資料為相互獨立且具有相同分佈的情況。然而,在實際的生活應用上,我們從網路或其他方式所蒐集到的資料並不一定符合這種假設,使得一般機器學習方法在針對不同資料庫的情況下做處理時會得到較差的效果,此類問題稱為跨資料庫情況下的問題。近年來,傳遞學習(transfer learning)方法被認為能有效的解決跨資料庫所造成的問題,因此在此篇論文中,我們將利用傳遞學習,並考慮在低維度空間上表現出的具有鑑別力的資訊,建立一個新的傳遞學習架構。最後實驗部分以跨資料庫情況下的表情辨識以及年齡分類來驗證我們所提出之方法的成果,由結果可看出我們所提出的架構不論在表情辨識或年齡分類的實驗下,皆比未處理跨資料庫問題的方法好,也優於現有的傳遞子空間學習法(transfer subspace learning)。
Most machine learning methods assumed training and test sets are independent and identically distributed, and may have degraded performance in practical applications since training and test sets are usually not independent and identically distributed. This is called cross-database problem. The ability of transfer learning has been identified to be helpful to solve the cross-database problem. In this thesis, we propose a novel transfer learning framework to utilize transfer learning and consider the discriminate information presented in the low-dimensional feature space. In experiment, we evaluate the effectiveness of our method on cross-database facial expression recognition and age classification. Our experiment shows that, our proposed framework outperforms conventional non-transferred subspace learning methods and most existing transfer subspace learning methods in both facial expression recognition and age classification.