ML PhD @GeorgiaTech 🚀 • previously MS @CMU_robotics • deep learning optimization and optimal control
Oct 3, 2023 • 4 tweets • 2 min read
📢Presenting "Mirror Diffusion Model"(#NeurIPS2023) --a diffusion model for fully tractable, simulation-free, convex constrained generation, with exciting applications to watermarked generation🤯
But how? ... by learning diffusions in dual space🙂
(1/n) arxiv.org/abs/2310.01236
Instead of learning approximate/implicit scores of the original (primal) constrained set, Mirror Diffusion Model (#MDM) applies #mirror #map and learns distributions in its #dual #space, which is a fully unconstrained Euclidean space!
