Is anyone else amused by the irony of using machine-learning-trained image generator in order to provide data to a machine-learning-trained image recognition program? I'm sure the researchers themselves and plenty of people here could come up with all sorts of logical reasons why this is fine, and very possibly given the right protocols it would be fine. But this sort of approach seems to lend itself toward increasing the risks of machine-learning. ie, you're doubling down on poor assumptions that are built-in to your training criteria or which creep into the neural net implicitly, because you are using the same potentially flawed assumptions on both ends of the process. Even if that's not the case, by using less real, accurately annotated data, you're far less likely to address true edge cases, and far more likely to overestimate the validity of the judgments of the final product compared to one with less synthetic training. And if there's one thing the machine learning community doesn't need any more of, it's overconfidence.

Edit: oops, turns out I mistakenly responded to the content of the paper instead of the fact that it exists and the form of its existence. Sorry.