Abstract

It is known that recursive training from generative models can lead to the socalled `collapse' of the simulated probability distribution. This note showsthat one in fact gets two different asymptotic behaviours depending on whetheran external source, howsoever minor, is also contributing samples.

 

Quick Read (beta)