[Xcelerate]

Wow, this was a very good article (I don't often say that).

I have long had a rather uncomfortable feeling about this idea that many chance mutations eventually produce some that are close to some objective function -- for essentially the same reason given in the article. There's just too many combinatorial possibilities to explore in the rather small time that the universe has been around (I mean, we're comparing like 10^50 to like 10^(10^(...)) here).

But this makes a lot of sense to me. Essentially what he's saying is that input sequences (vectors) pass through some kind of surjective mapping into a lower dimensional space of "networks" (I'll just call them a different space of vectors).

If I had to guess (and this is purely speculation here), this is probably due to some symmetry property.

[Retra]

A long time ago, I was trying to imagine how complex structures could evolve out of randomness, I came up with something I called "hinge evolution." Let the relevant behavior of a system be a function of the contact area between the two planes of the hinge, and the genetic variation be represented by the action of hinge. You can vary the hinge angle a lot while only causing minor shifts in the contact area, but if you close the hinge, there will be a giant discontinuous spike in the contact area. So I believed it wasn't absurd to think that things didn't have to evolve in continuous fashion: you could have 'silent' evolution happening in the background, but when certain conditions are met, a major shift in structure could happen all at once.

This article elucidates a much better explanation.