From Optimal Score Matching to Optimal Sampling

Music City Center 

Score-based generative algorithms, particularly those leveraging score matching and denoising diffusion techniques, have achieved state-of-the-art performance in generating high-quality samples from complex, structured probability distributions. These methods are exceptionally versatile, demonstrating success across a variety of modalities, including natural images and audio. Notable implementations, such as OpenAI's SORA, showcase their power and flexibility. We will discuss some recent theoretical advances of those algorithms.

score estimation

diffusion models

density estimation