Abstract:

This talk presents progress towards developing structure or architecture-aware mathematical frameworks for reasoning about inference and learning in deep neural networks. By being architecture-aware, we can extract concrete analytical insights about training neural networks that are somewhat opaque from the perspective of standard tools such as Lipschitz constants. Yisong will present two research thrusts along this direction.  The first thrust is a majorize-minimize framework that develops a novel architecture-aware trust region for deep learning optimization, which we call "Deep Relative Trust".  The second thrust is a control-theoretic treatment of neural ODEs (and related architectures), leading to new algorithms that can enforce desirable properties such as stability, adversarial robustness, and forward invariance.

Bio: 

Yisong is a professor of Computing and Mathematical Sciences at the California Institute of Technology. His research interests lie primarily in machine learning, and span the entire theory-to-application spectrum from foundational advances all the way to deployment in real systems. He work closely with domain experts to understand the frontier challenges in applied machine learning, distill those challenges into mathematically precise formulations, and develop novel methods to tackle them.

More info: http://www.yisongyue.com/