Title：In Pursuit of Deciphering ReLU Networks and Beyond
Speaker: Fenglei Fan (Fudan University）
Time: Monday, April 24, 2023, 09:30-10:30
Location:Mingde Building B201-1
Abstract:Deep learning, represented by deep artificial neural networks, has been dominating numerous important research fields in the past decade. Although deep learning performs excellently in many tasks, it is notoriously a black box model. A neural network with the widely used ReLU activation is a piecewise linear function over polytopes, which has a simple functional structure. To enhance the interpretability of deep learning, it is of great importance to figure out the properties of polytopes and decipher the functional structure of a ReLU network. In collaboration with leading peer groups from Harbin Institute of Technology, Cornell University, BIGAI, and RIKEN AIP, we dedicate ourselves to answering the following fundamental questions: what kind of functions does a ReLU network learn? and how can we leverage the enhanced understanding to build novel machine learning models? I will share with you our work and discuss issues and opportunities in the field. We welcome interaction with prospective students, postdoctoral fellows, and new collaborators.
