A new arXival describes work to improve the computational efficiency of GANs (generative adversarial networks) for emulating cosmological N-body simulators. It seems to me that we’re putting the cart before the horse here. The notion is that since cosmological N-body simulations are expensive we’ll take a relatively small set of these and learn how to produce more of them with a GAN since GANs have proven so effective at big data tasks like in-painting for missing photographic image segments. So, (ignoring the discrepancy in the number of unique training instances in each scenario and) assuming that GANs will be great at emulating from a small set of cosmological N-body simulations, then what are we going to do with them. Presumably we don’t just want them to generate pretty pictures; presumably we want to do some kind of inference? In that case we’re in the regime of ‘likelihood-free inference’ and ‘Approximate Bayesian Computation’. Here GANs have been studied but there’s no clear advantages to using them over other likelihood-free methods in general and there’s not a great deal of theory to tell us how the GAN posterior might relate to the posterior of the model we’re trying to emulate (as Xi’an points out). In the era of so-called ‘precision cosmology’ we surely want to be moving beyond O(1) posterior approximations.
All that to say, there’s no doubt that GAN projects are a good way to take a slice of the funding pie for machine learning, and getting students to work on them will make for some students who are readily employable outside of academia, but I’m yet to see them answer a meaningful science question. They’re a fun idea though, so I’d be happy to eventually be proved wrong on this.