> A good estimation of p(x) makes it possible to efficiently complete many downstream tasks: sample unobserved but realistic new data points (data generation), predict the rareness of future events (density estimation), infer latent variables, fill in incomplete data samples, etc.
To give an example, we use generative models for model-based reinforcement learning. Basically, RL algorithms have relatively poor sample efficiency. They require a lot more data than supervised learning for a number of reasons. Rather than train an agent against real data, which is limited, we train a generative model using supervised learning and use the model to simulate real data—either using the generated outputs or the latent representations. This covers several of the author's described use cases.
man people are lazy - it's literally right there in the intro:
"Flow-based deep generative models conquer this hard problem with the help of normalizing flows, a powerful statistics tool for density estimation. A good estimation of p(x)p(x) makes it possible to efficiently complete many downstream tasks: sample unobserved but realistic new data points (data generation), predict the rareness of future events (density estimation), infer latent variables, fill in incomplete data samples, etc."
