That paper doesn't mention game theory or the Nash equilibrium at all.
Perhaps you mean [2], which does. However, the article seems to misinterpret this: the fact that that system approximates the Nash equilibrium should be expected. Any system (including humans) that performs well in those types of games will, almost by definition.
It doesn't mention Nash equilibrium, but the whole approach they are discussing comes from the idea of pitting two networks against each other in a "minimax two-player game".
"We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models ... This framework corresponds to a minimax two-player game."
Edit: Regarding the paper you linked to. The fact that a Nash equilibrium is expected is an important part of why it's interesting to structure things this way: it's an optimization process guaranteed to stabilize (whereas it's problematic that other approaches will diverge).
"When applied to Leduc poker, Neural
Fictitious Self-Play (NFSP) approached a Nash equilibrium, whereas common
reinforcement learning methods diverged"
The articles claims (based on the paper I linked, and others) that:
What we see in these 3 players are 3 different ways game theory plays in Deep Learning. (1) As a means of describing and analyzing new DL architectures. (2) As a way to construct a learning strategy and (3) A way to predict behavior of human participants. The last application can make your skin crawl!
Of these claims, I think all are wrong.
One can certainly make the case that it is interesting to measure the performance of a system compared to the Nash equilibrium (when possible), but the author seems to think that the designers of the system are somehow using game theory to design the system.
They are in one of the cases pointed out, not the other two. The novel idea is to have two networks compete against each other in a two-player minimax game, developing progressively better strategies until something like Nash equilibrium is reached--at which point you have one well-trained classifying network and one well-trained generating network.
Perhaps you mean [2], which does. However, the article seems to misinterpret this: the fact that that system approximates the Nash equilibrium should be expected. Any system (including humans) that performs well in those types of games will, almost by definition.
[2] https://arxiv.org/pdf/1603.01121.pdf