Pedro Domingos

never thought I'd ever hear a journalist use the phrase "generative models" enthusiastically. They ain't what they use to be.

did journalists use to hold them in contempt?

They just weren't very sexy outside of academia. And the fact that generative models could "generate" was sort of secondary.

In most instances people create "generative models" as a way of specifying some set of simplifying inductive assumptions.

e.g. latent Dirichlet allocation, where we assume documents are generated by sampling words independently at random.

That's of course absurd and not going to give you realistic documents, but it becomes useful when you turn the inference crank.

This formal probabilistic specification is quite different from "sampler boxes" like GANs, but we're... using the same name. :)

illuminating, thanks.