Juncai He, Penn State University
In this talk, I will first show the connections between ReLU DNNs and linear finite element functions. In addition, I will talk about the approximation properties for ReLU DNNs with a new understanding from the perspective of hierarchical basis. I will then present a constrained linear model that extracts features from images for classifications. Based on such a model, I will finally demonstrate how a new type of CNN, known as MgNet, can be derived by making minor modifications of a classic geometric multigrid method for partial differential equations and then discuss the theoretical and practical potentials of MgNet.
Meeting ID: 535 409 777 (no password)