Rayadurgam Srikant, University of Illinois Urbana-Champaign
On the loss surface of neural networks for binary classification

Oct 22, 2018, 2:00pm; EEB 132


Deep neural networks used for classification problems are trained so that their parameters achieve a local minimum of an appropriate loss function. The loss function is typically intended to approximate the classification error in training samples. In this talk, we will show that approximations of widely-used neural network architectures have the property that every local minimum of a surrogate loss function is a global minimum, and further achieves the global minimum of the training error.

Joint work with Shiyu Liang, Ruoyu Sun, and Jason Lee.


R. Srikant is the Fredric G. and Elizabeth H. Nearing Endowed Professor of Electrical and Computer Engineering and the Coordinated Science Lab at the University of Illinois at Urbana-Champaign. His research interests include communication networks, cloud computing, stochastic control, and machine learning. He is a recipient of the IEEE INFOCOM Achievement Award and the IEEE Koji Kobayashi Computers and Communications Award. His best paper awards include an INFOCOM Best Paper Award and an Applied Probability Best Publication Award. He is a past Editor-in-Chief of the IEEE/ACM Transactions on Networking, and is a Fellow of the IEEE.