摘要：
Generative models such as GANs, diffusion models, flow matching, and related approaches have demonstrated remarkable success in generation tasks across diverse data modalities. Several works have sought to establish a unified framework for these generative models. Generator Matching is a new modality-agnostic framework for generative modeling. Building upon the iterative, step-wise nature of existing generative models, it leverages generators that characterize the infinitesimal evolution of Markov processes. Furthermore, it expands the design space to previously unexplored Markov processes and enables the rigorous construction of superpositions of Markov generative models.
In this talk, we will briefly introduce the background of generative modeling, and then focus on Generator Matching as a novel paradigm. We will detail its methodology and show how it unifies existing generative models. Additionally, we will present new models and superposition techniques within this framework, along with practical applications.

论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。