# Deep Unsupervised Learning using Nonequilibrium Thermodynamics **Authors**: Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, Surya Ganguli **Year**: 2015 **Source**: arXiv:1503.03585 **URL**: https://arxiv.org/abs/1503.03585 ## Summary The paper that first formalized **diffusion-based generative models** for machine learning. Drawing from non-equilibrium statistical physics, the authors propose systematically destroying structure in a data distribution through an iterative forward diffusion process, then training a neural network to learn the reverse — restoring structure from noise. This became the conceptual foundation for all modern [[diffusion_models|diffusion models]].