Lin, Zhouchen

Professor

Research Interests: Machine learning, computer vision

Office Phone: 86-10-6275 3313

Email: zlin@pku.edu.cn

Lin, Zhouchen is a professor with the Department of Machine Intelligence, School of EECS. He obtained his B.Sc. from Nankai University, Tianjin, in 1993, two M.Sc. from Peking University in 1995 and Hong Kong Polytechnic University in 1997, respectively, and Ph.D. from Peking University in 2000. His research interest include machine learning, computer vision, pattern recognition, image processing, and numerical optimization.

Dr. LIN has published more than 150 research papers, most of which are published in top-tier conferences and journals, such as IEEE TPAMI, IEEE TIT, IEEE TIP, IEEE TNNNLS, IJCV, ICCV, CVPR, ICML, NIPS, AAAI, and IJCAI. He holds more than 45 US patents. He is an associate editor of IEEE TPAMI and IJCV. He was an area chair of CVPR 2014, ICCV 2015, NIPS 2015, CVPR 2016, and a senior committee member of IJCAI 2017, and AAAI 2016/2017/2018. He was elected as a Distinguished Young Scholar of NSF China and a Fellow of International Association of Pattern Recognition in 2016.

Dr. LIN has run multiple research projects from NSF China and 973 programs. His major achievements are as follows:

1) He made seminal contributions to low-rank models. He is the first person to introduce low-rankness to subspace clustering (SC), the problem of clustering data into multiple subspaces and which is highly useful in signal and data processing. He further extended the matrix-based low-rank models to tensor-based ones, in which he solved a key problem of defining a computable rank for tensors. Thus he significantly contributed to the extension of compressed sensing theories from vectors to matrices and tensors. In particular, his low-rank representation has become a new hot research topic, with numerous variations and applications.

2) He also made original contributions to fast optimization algorithms for low-rank models. He is the first person who discovered that the augmented Lagrange multiplier method is particularly suitable for solving low-rank models, significantly promoted the boom of research on this type of algorithms, where he also made key contributions. He first proposed tricks like adaptive penalty, adoption of partial SVD, and linearization of the augmented Lagrangian function, which now have been the standard practice in the follow-up papers. He is also among the first persons who considered multi-block alternating direction method of multipliers and gave his unique algorithm with convergence rate. For non-convex problems, he made non-trivial extensions to the most popular algorithms for convex problems, such as accelerated proximal gradient, such that they also work for non-convex ones while maintaining the convergence rate.

3) His work on superresolution (SR) is ground-breaking. He is the first person who gives explicit limits of both reconstruction-based and learning-based SR. His analysis also reveals the factors that affect the effective magnification factor. His theories not only guide the selection of appropriate magnification factors, but also suggest that traditional imaging configurations should be changed in order to break the limits, thus also provide an indispensable guide to the imaging sensor and the display design.