So basically everything starts as noise in the neural net, which may develop to an idea
That can be tested.
If I'm reading Imagination Engines text heavy web site clearly. FWIW it's a notion I agree with.
In principle all human knowledge can be processed by a set of very complex neural networks, since that's what a human brain is.
But IRL humans cannot program this system directly (and did not know it existed till the invention of the microscope)
That level is only dealt with directly when we learn to do manual tasks like walk, or learning a language solely by matching sounds in one language against sounds (and their meanings) in another. Even this assumes we know a language already (What if the person was deaf since birth?)
For everything else we operate with higher level abstractions. Words on paper --> language conventions --> concepts --> restructure thinking --> change weights within the NNs.
multilayer neural nets deep learning be easier if we acknowledged that?