How do we know until we've tried it?
I mean, at a certain level of abstraction, you can just view humans as a bunch of atoms that have been placed in a heat bath for a very long time. And yet intelligence seems to have spontaneously self-organized from it.
They are nowhere close to the complexity of a brain, and certainly do not have enough capacity to be a realistic simulation of the universe.
> How do we know until we've tried it? Tried what, is the question? What's an ant-level intelligence? We have algorithms that can detect images/speech at super-human level (at least in some conditions). Worthy questions, but unrelated to these algorithms.
I am thinking, in particular, of an experiment done by Richard Feynman where he got ants walking across small sheets of glass. After shuffling the sheets around so the chemical trail became a loop, he was able to get the ants tirelessly marching around in a circle (which they would presumably do until they dropped).
Heck, you probably don't even need an ANN at all to do that, much less an evolutionary one.
They didn't? I'm not sure what you're saying here. Are you arguing intelligent design or something? Because the standard evolutionary theories postulate that humans did, in fact, spontaneously arise from a heat bath (the heat coming from the Sun).
That's not what I typically think of when I hear the word "spontaneous". Spontaneous emergence sounds to me like lightning hitting a swamp over a period of time, and then by a pure chance arrangement of molecules, a fully formed swamp man comes out of it.
That's way different than an evolutionary process.
> Do you think creating a 3D universe (like that being worked on by open.ai), putting the evolvable agents in it and trying to make them evolve learning and intelligence is a good idea
Basically simulating a universe, seeded with some things we think are already somewhere along the road to intelligence, then seeing if progress is made just by simulating for a long time.
The question is, what is that function?
Not necessarily. Everything points to the fact that the humain brain (at least, I don't know much about ants) does not work in any way similar to neural networks, and as such there's no guarantee that you can represent its behaviour by the current algorithms.
For example, real neurons have no supervision signal. Memory is also a big issue (see the work being done by DeepMind and FAIR on differentiable neural computers, and memory networks) and Reinforcement Learning still struggles with long-term planing, switching strategies, etc.
Yes, the brain is most likely not simply modelling one giant ANN, but it's much more plausible that ANN-like components have a role in it.
Doesn't the universal function approximation theorem provide just such a guarantee? It doesn't guarantee any algorithm will converge on that representation, but the capability to represent it is there.
An ant’s fitness function is different from its ancestor’s fitness function, which is different from its own ancestor’s fitness function, which is different from...
You don’t get from zero to complexity with a single fitness function.
The question we should ask is: How do we vary the fitness function over time in order to evolve something complex?
I have no idea what the answer is, but...
“If you know the question, you know half.”
-Herb Boyer, geneticist
These are mathematical models that optimise a specific fitness function, they borrow some biological concepts but they are not going to "evolve into ant-level intelligence".