I saw a project on GitHub some time ago where someone trained a neural net with only a single, shared weight to play a driving game.

They used a gentic algorithm to evolve the network, its connections and activation functions (per node).

It got quite good.

Perhaps not that surprising that one can encode a lot in the graph structure itself, and given how DNA is quite good at dictating complex organs, I don't see why something similar couldn't be at work for insects and mammals alike. After all, nobody taught the newborn kangaroo babies how to crawl up into the pouch of their mom either.