Transfer Learning with Scarce Annotations
Zhirong Wu
Microsoft Research Asia
Abstract: In the past few years, progresses have been shown that much of the visual problems (e.g.classification, detection, segmentation) can be approached by collecting a large amountlabeled data, and training a huge neural net. This draws scalability limitations to the openworld where new objects may constantly appear, and labeling costs are often hard tomanage.
I will show our recent progress on recognition with unsupervised learning and learning withscarce annotations. Our key insights are two folds: 1) A simple unsupervised learningalgorithm by discriminating instances could achieve the state-of-the-art performance. 2)Advances in unsupervised learning can directly translate to few-shot recognition. 3) Simplelabel propagation algorithms can create abundance of labeled data reliably by propagatinglabels to the unlabeled data. We demonstrate state-of-the-art performance for severalimage and video recognition problems under the constraints that labeled data is scarce.
Bio: Zhirong Wu is currently a Researcher in the visual computing group at Microsoft ResearchAsia. Previously, he obtained his B.Eng from the department of automation at TsinghuaUniversity, and his Ph.D from the Chinese University of Hong Kong advised by Prof. XiaoouTang. He was a visiting PhD student at Princeton Vision Group advised by Prof. JianxiongXiao. He was Post-doctoral scholar at UC Berkeley with Dr. Stella Yu