New research by CDS Faculty Fellow Florentin Guth, CDS PhD alum Zahra Kadkhodaie, and CDS Professor Eero Simoncelli shows that the probability of natural images varies by factors up to 10^14,000 — far more than previously recognized. Accepted to this year's NeurIPS, their paper, “Learning normalized image densities via dual score matching,” offers a method, "dual score matching," that adapts diffusion models to directly estimate probability, offering a clearer window into how neural networks lea