LGRT 1621
The past few years has seen extraordinary advancements in generative modeling. Modern generative models such as diffusion models and continuous normalizing flows have been able to surpass the performance of previous state of the art models like GANs and VAEs due to advances in deep learning and the development of new training algorithms. In this talk I will give a primer on how these models are formulated and trained via matching algorithms. These methods construct a time indexed probability distributions between a prior and data that is used as a target to train a parametric model. This paradigm enables efficient training of diffusion models and continuous normalizing flows with algorithms such as score matching and flow matching.