Seminar Series - Machine Learning and Physics: Bridging the Gap. Professor Stephan Mandt, I&CS

Tuesday, June 9 at 2 p.m. via Zoom


Physicists and machine learners have much in common, but speak different languages. Both communities deal with high dimensional state spaces and build sophisticated approximation schemes, e.g., of intractable integrals or dynamic processes, to come up with tractable models of the world. In this lecture, I will start with an introduction to Bayesian machine learning and the popular training paradigm of variational inference. I will then discuss several physics-inspired innovations, including identifying and fixing slow convergence problems in time series models by artificial gauge fields, using perturbation theory to improve variational bounds, and using Langevin diffusions to explore parameter uncertainties in deep learning. I will also present several use cases of variational methods, in particular from the domain of neural image and video compression. As an example of how machine learning can be used in physics, I will discuss my group’s recent work on predicting thermodynamic properties of chemical mixtures.