The article introduces Marginalization Models (MAMs), a new type of generative model for high-dimensional discrete data. MAMs offer scalable and flexible generative modeling with tractable likelihoods. They facilitate fast evaluation of marginal probabilities, overcoming a significant limitation of methods with exact marginal inference. The authors propose scalable methods for learning the marginals and demonstrate the effectiveness of MAMs on a variety of discrete data distributions. MAMs achieve significant speedup in evaluating the marginal probabilities and enable high-dimensional generative modeling beyond previous methods’ capabilities.
Publication date: 20 Oct 2023
Project Page: https://github.com/PrincetonLIPS/MaM
Paper: https://arxiv.org/pdf/2310.12920
