This paper introduces a novel deep unsupervised learning algorithm that leverages non-equilibrium thermodynamics to model complex datasets. The core idea involves a forward diffusion process that systematically degrades data structure, and then learning a reverse diffusion process to reconstruct it, creating a flexible generative model. This approach allows for efficient learning, sampling, and probability evaluation in deep generative models, even with numerous layers. The authors demonstrate the method's efficacy on various datasets, including images, highlighting its ability to handle tasks like denoising and inpainting by multiplying distributions.

Ver todos los episodios