CSE DSI Machine Learning Seminar - Ke Li (Simon Fraser University)

Next-Generation Generative AI with IMLE

Diffusion models and flow matching power the frontiers of generative AI, but they are slow, can generate off-manifold samples, and may not respect the conditioning input. These all are at their root a consequence of iterative denoising. Is iterative denoising truly necessary? In this talk, I will cover Implicit Maximum Likelihood Estimation (IMLE), a method we developed that does away with iterative denoising both at training and testing time. Instead, IMLE matches ground truth data to generated samples and trains the model with a simple regression-like objective. I will cover applications in image synthesis, 3D understanding, reinforcement learning, trajectory forecasting and robotics, which leverage IMLE's advantages of fast sample generation, sample efficiency, precision, mode coverage and adherence to the conditioning input. 

Ke Li is an Associate Professor of Computing Science at Simon Fraser University​ and a Canada CIFAR AI Chair at the Alberta Machine Intelligence Institute (Amii) and leads the APEX Lab. His research spans multiple areas of AI and computer science, including machine learning, computer vision, robotics, algorithms and systems. He is particularly passionate about tackling long-standing fundamental challenges that cannot be tackled with a straightforward application of conventional techniques. Previously, he was a Member of the Institute for Advanced Study in Princeton and a Research Scientist at Google, and received his Ph.D. from University of California, Berkeley and B.Sc. from the University of Toronto.