Reverse Markov Learning for high-dimensional generative models with distributional shifts

Modelling distributional shifts in high-dimensions is challenging for at least two reasons. There is a computational challenge: most generative models like diffusion-based models quickly become very expensive in a high-dimensional context. There is also a statistical challenge on how to model the distributional shifts in a robust way, especially if (in a causal context) interventions will extend to previously unseen support. I want to show in this talk how using Reverse Markov Learning.(building on earlier work of Engression) can address these two challenges and provide a lightweight generative model in this setting. I will show as a case study the emulation of global circulation models under various forcing regimes.

Joint work with Xinwei Shen, Malte Meinshausen, and Tong Zhang.