When: Thursday 4th of November, 1pm AEDT
Where: This seminar will be presented online via Zoom, RSVP here.
Speaker: Max Revay
Title: Thesis presentation: A Behavioral Approach to Robust Machine Learning
Deep learning has demonstrated impressive results on a variety of problems, however, many of these results have not translated to applications in physical systems, partly due to the cost of system failure and partly due to the difficulty of ensuring reliable and robust model behavior. In this talk, we present new approaches to training models that have built-in Stability and robustness guarantees. These properties are important when reliability is a key consideration, e.g. in safety-critical systems. Our approach to ensuring stability and robustness of the models trained is distinct from prior methods; where prior methods learn a model and then attempt to verify robustness/stability, we directly optimize over sets of models where the necessary properties are known to hold. The methods developed are relevant to a variety of domains, including supervised learning, system identification, and learning for control. More specifically, we apply methods from robust and nonlinear control to the analysis and synthesis of recurrent neural networks, equilibrium neural networks, and recurrent equilibrium neural networks. The techniques developed allow us to enforce properties such as incremental stability, incremental passivity, and incremental l2gain bounds / Lipschitz bounds. A central consideration in the development of our model sets is the difficulty of fitting models. All models can be placed in the image of a convex set, or even R^N, allowing useful properties to be easily imposed during the training procedure via simple interior point methods, penalty methods, or unconstrained optimization.
Max Revay received his Bachelor’s degree in Mechatronics and Applied Mathematics at the University of Sydney in 2017. Since then, he has been a Ph.D. student at the Australian Center for Field Robotics working on Robust Machine Learning. His research interests include robust machine learning, non-linear control, and networked systems.