Keywords: sampling, data generation, denoising, sparse representations, diffusion score, transport.
Course Description
This course presents the state of the art in the generation of images, sounds, and scientific data using deep neural networks. We focus on sampling probability distributions obtained by transporting Gaussian white noise. After a review of the state of the art, we study transport via score diffusion, which performs progressive denoising to generate data (images, sounds, etc.). This requires estimating the score of the probability density using a deep neural network. The course introduces the mathematical and algorithmic foundations along with their applications.