Continuous Normalizing Flows

Slides: https://bayesgroup.github.io/bmml_sem... My presentation will be a continuation of Victor’s talk about Neural ODE. I’ll explain how this idea can be applied to normalizing flows, making them more flexible and computable. Furthermore, we will talk about FFJORD — an unbiased stochastic estimator of the likelihood based on continuous normalizing flows. As authors of the paper state, this approach allows creating reversible generative models with completely unrestricted architectures.