题 目：代数多层网格法与深度神经网络
主讲人：许进超 教授 美国宾州大学
摘要：In this talk, I will present a general framework for the design and analysis of Algebraic or Abstract Multi-Grid (AMG) methods. Given a smoother, such as Gauss-Seidel or Jacobi, we provide a general approach to the construction of a quasi-optimal coarse space and we prove that under appropriate assumptions the resulting two-level AMG method for the underlying linear system converges uniformly with respect to the size of the problem, the coefficient variation, and the anisotropy. Our theory applies to most existing multigrid methods, including the standard geometric multigrid method, the classic AMG, energy-minimization AMG, unsmoothed and smoothed aggregation AMG, and spectral AMGe. These results are summarized in a recent survey article entitled “Algebraic Multigrid Methods” in Acta Numerica (Vol 26). Finally we will discuss our ongoing investigation on relationship and cross-application between algebraic multigrid techniques and deep neural networks.
时间：2017年6月15日下午4:00开始
地点：南山一阶
欢迎感兴趣的同学和老师参加！
伟德BETVLCTOR1946教务处
伟德BETVLCTOR1946数学与计算科学学院
科学工程计算与数值仿真湖南省重点实验室
