真實世界中的電腦視覺應用問題經常具有小樣本資料的特性，也就是需要辨識的物件種類數量繁多，但是每個物件種類所能收集到的標註影像資料卻僅有少數樣本可以使用。 本群體計畫將結合四個子計畫協力解決標註資源不足時的深度學習技術，網路設計 ，及實務應用，包含下列應用情境：(1) 有巨量的樣本，但僅有少量被標註，或是 標註資訊不完整；(2) 每個類別或事件，僅有少量之標註樣本；(3) 標註資訊有雜訊而不完全正確，期能在低標註資源的情況下，發展有效的機器學方法及網路。

This project will focus on deep learning with a small (or no) training dataset, and thus expects to bring the following contributions to academia and industry: (1) Developing efficient and accurate unsupervised/semi-supervised image clustering/classification algorithms; (2) Developing data augmentation techniques for computer vision applications without enough training samples; (3) Developing efficient techniques for , (4) Collaborating with local companies to apply our developed technologies to their real applications.