Bridging Graphics and Machine Learning: differentiable simulators as inductive biases
Derek Nowrouzezahrai - McGill University
April 4, 2025, 2:30 p.m. - April 4, 2025, 2:30 p.m.
MAASS 217
Hosted by: Paul Kry
After briefly touring the increasing interplay between computer graphics and machine learning, I will highlight one recent path of growing intersection between these two communities: the development and application of differentiable simulators. Here, I will support the claim that leveraging the expertise and advances developed in the computational physics, computational statistics, numerical methods, and — ultimately — computer graphics communities can lead to powerful inductive biases for end-to-end physics-oriented learning tasks. I will survey the many reasons — motivated in part by a recent case study — for my excitement in maintaining and a research agenda that combines computer graphics-oriented methodologies with physics-based machine learning.
Derek Nowrouzezahrai is a Full Professor at McGill University and a Core Faculty Member of the Quebec Institute for Artificial Intelligence (Mila). He is the Co-director of the the McGill Graphics Lab and Director of McGill University's Centre for Intelligent Machines, a grouping of over a dozen labs in the Faculty of Engineering and the Faculty of Science. Derek completed a Post-Doc at Disney Research Zurich after his graduate work at the University of Toronto. Derek’s group works on devising new mathematical models of visual phenomena and dynamics, developing efficient and differentiable numerical approaches to solve forward and inverse physics-based problems. These include problems in generative content creation, light transport, fluid dynamics and control, robotics, augmented reality, digital manufacturing, computational optics and imaging, and image and geometry processing. Derek's research has been adopted in feature films, video games, amusement parks, and consumer products.
