Discrete Generative Model

Discrete generative models aim to learn and generate data represented by discrete variables, offering advantages in interpretability and efficiency compared to continuous counterparts. Current research focuses on developing novel training algorithms, such as flow-matching methods on statistical manifolds and improved gradient estimators like the gapped straight-through estimator, to overcome challenges in training and sampling from these models. These advancements are improving the performance of discrete generative models across various applications, including image generation, protein design, and anomaly detection, particularly in scenarios requiring high controllability or interpretability of the generated data. The development of more robust and efficient discrete generative models holds significant potential for advancing diverse fields.