In this talk, I use three examples to illustrate how datasets affect the design of deep learning algorithms. For image classification problems with a large label space, we designed hierarchical deep convolutional neural networks capable of learning a category hierarchy from data with uneven between-category distinctions. For fine-grained classification problems with small training sets, we designed a novel fine-tuning scheme called Selective Joint Fine-Tuning, which performs multitask learning by sampling a subset of most relevant data from a large source domain. For object detection or semantic segmentation with image level labels only, we designed a weakly supervised learning algorithm capable of inferring object level and pixel level labels from image level labels for the training images.

Yizhou Yu is a professor in the Department of Computer Science at the University of Hong Kong. He was a professor at University of Illinois, Urbana-Champaign (UIUC) from 2000 to 2010. He received his PhD degree in computer science from University of California, Berkeley. Professor Yu has made significant contributions to AI and visual computing, including deep learning, computer vision, and computer graphics. He is an ACM Distinguished Member and IEEE Fellow. His current research interests also include biomedical data analysis, computational visual media, and geometric processing. Professor Yu has served as an associate editor of many international journals, including IET Computer Vision, IEEE TVCG, The Visual Computer and International Journal of Software and Informatics. He has also served on the program committee of many leading international conferences, including SIGGRAPH, SIGGRAPH Asia, and International Conference on Computer Vision.