I think really hits the nail on the head. It also may be the foundation for why we find things like neural networks more usable for things like vision than directly formalizing the approach. In other words, you could use an approach like coq to verify the neural net software calculates correctly, but you couldn't use to formalize what it calculates.