In this talk, we will discuss the relationship between linear finite element methods (FEM) and deep neural networks (DNN) using ReLU as activation function. Using this relationship, we discuss and compare approximation properties of both FEM and DNN, give some theoretical justification of using low bit-width DNN models in applications and using DNN for numerical solution of partial differential equations.

许进超，男，数学家，国际计算数学领域权威学者 ，美国宾州大学终身教授，计算数学与应用研究中心主任，湘潭大学计算与应用数学研究所所长，北京国际数学研究中心计算方法与应用实验室主任，美国数学学会会士、美国工业与应用数学学会会士 ，北京大学长江学者讲座教授，海外杰出青年。

海报