Title: Memorization in Diffusion-Based Generative Modeling
Abstract: The aim of generative modeling is to create new samples from a probability measure, given only samples from that measure. In recent years diffusion models have emerged as a powerful framework for generative modeling, especially in the context of (conditional) image generation. However these diffusion models often suffer from memorization, a phenomenon in which the newly generated samples simply reproduce the given samples. The talk describes the underlying dynamical systems structure that leads to memorization; and it includes theoretical and numerical studies of regularization techniques that can be used to ameliorate it.