intelegence is easy, it's emulating stupidity that is the hard bit....a rare few of us do after all hopefully learn from out mistakes.
Also wouldn't you want a AI that's less fickle than a human.
It's also intersting to note that a lot of the problems solved appear to be of the visual type e.g. the word 'cat' had to be provided and that 'blank slate' theory has been disproven, though that's not an issue if the computer algorythms have long enough to evolve.
I agree with your IO stuff, that bares strong relation to neurology.
Personally I'm working on linguistics modeling and the senses, which is based on neurology I won't go into until I have something publishable, but you can find it out if you look for neurology in that area... you won't find anything in linquistics in that area though... it seems to be a hard problem even for humans.
My my key problem was seeding, so I may take a look at deap learning to see what it has to offer, but I think a few lightly ranked examples (who ranking can be changed by the algorythm) would probably be most benifiial.... at least to do some primary set reduction on the data.