There is a very interesting paper on getting neural networks to design their own internal structure.
https://arxiv.org/abs/1911.11691
That is kind of an awkward paper for people experimenting with NEAT or HyperNEAT neural networks. I suppose though they can just pull out the relevant structure afterwards and then it is just a way of designing NEAT networks.
https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_topologies
4 Likes