MGD: Moment Guided Diffusion for Maximum Entropy Generation
New method solves slow sampling in high dimensions, tested on financial and cosmic data.
Deep Dive
A team led by Etienne Lempereur and Stéphane Mallat introduced Moment Guided Diffusion (MGD). It combines diffusion models with maximum entropy principles to generate samples from limited data. MGD uses a stochastic differential equation to guide moments to target values, avoiding the exponential slowdown of traditional MCMC methods. It was applied to estimate negentropy in complex systems like financial time series and turbulent flows using wavelet scattering moments.
Why It Matters
Enables more reliable, theoretically-grounded generation of data in fields like finance and science where information is scarce.