学术报告:Low-Rank Matrix Optimization: Theory and Algorithms


  8月20日:9:00—12:00 ,14:00—16:00

  8月22日:9:00—12:00 ,14:00—16:00

  8月23日:9:00—12:00 ,14:00—16:00


  报告标题问题Low-Rank Matrix Optimization: Theory and Algorithms

  报告佳宾 Qi Houduo传授



  One of the purposes in machine learning is to reveal and explore the structure among data. When data is put in arrays, they tend to be of low rank. Therefore, low-rank matrix optimization has become increasingly important in machine learning algorithms. This series of talks aims to provide a selective overview of the topic. We pay particular attention to efficient algorithms and try to elaborate on important optimization techniques that are uniquely related to low-rank optimization. The principle in guiding our selection of the material is the speed and solution quality of the resulting algorithms. After this course, one is expected to appreciate the difficulty of the problem, what works and what do not, and more importantly we are left for further exploration.

  We will focus on the following (selective) aspects of the topic:

  Low-rank matrix optimization: motivating examples (from Principle33 Component Analysis to Low-rank and sparse optimization, Low-rank Hankel matrix optimization for spectrally sparse optimization)

  Penalty methods (the need for regularization, proximal operators, DC penalties)

  Methods of Alternating Projections (global and linear convergence, and its penalized version)

  A show-case example: outlier detection via low-rank EDM (Euclidean Distance Matrix) optimization.


  Qi Houduo传授(主页:)为英国南安普敦大学传授,博士生导师。1990年毕业于北京大学统计学专业,1993年获曲阜师范大学硕士学位, 1996年中国迷信研讨院数学与体系迷信研讨院利用数学研讨所博士毕业。曾在香港理工大学、新南威尔士大学等做博士后研讨,获澳大利亚研讨委员会(ARC)帮助,和ARC和享有环球盛誉的Queen Elizabeth II Fellowship嘉奖。研讨标的目的有:束缚优化、矩阵优化、变分不等式、数值阐发等。在国际顶级期刊SIAM on Optimization, Mathematical Programming 等杂志颁发高程度研讨论文十余篇。