The design of convolutional neural networks (CNNs) has undergone two phases: manual design at the early stage, which requires much engineering insights,  and the automatic search at the current stage, which heavily relies on computing power. Whether there is an underlying theory for designing good CNNs becomes a crucial research problem. In this talk, I will illustrate our efforts on pursuing this goal. Although I haven’t found a unified principle that can result in all the effective CNNs, I do find multiple principles that can help design CNNs from various aspects. 

Zhouchen Lin received the PhD degree in applied mathematics from Peking University in 2000. He is currently a professor with the Key Laboratory of Machine Perception, School of Electronics Engineering and Computer Science, Peking University. His research interests include machine learning, computer vision, image processing, pattern recognition, and numerical optimization. He is an area chair of CVPR, ICCV, NIPS/NeurIPS, AAAI, IJCAI, ICLR and ICML for multiple times. He is a fellow of the IAPR and the IEEE and a recipient of NSF funding for distinguished young scholars.