Thursday, December 13, 2007

The Everests of Artificial Intelligence

[updated after a new suggestion]

Here are some of the Everests for Computational Cognitive Modeling. Some people call them AI-complete. That might not be the best term, as it extends the notion of NP-Completeness, which is a precise, formal, mathematical notion, into a very blurry territory.

Anyway, I've put them from easier to harder...

Here are my feelings when delving into theory... (hat tip to her).

Do you have another problem that's missing from this list? I would appreciate additions and suggestions in the comments.


pfctdayelise said...

You think speech recognition is easier than MT?? I would guess it is harder.

What about face/picture recognition? (like a smart google images...)

Alexandre Linhares said...

I would classify face and image and object recognition in the same class as Bongard problems. I think that machine translation is daunting, because it involves culture, and cultural problems are always nasty. Peoples from Brazil and Argentina, like those from Australia and New Zealand, all have their "issues". But imagine the cultural gap between Arabs and Israelis... I think that there is something special about the first concepts one is forced down into one's mind. It is like these first concepts eventually mold all the other concepts in the space around them. Some concepts are central in some cultures, and they don't even exist in other cultures. This is why I think MT (of human-quality level) is so daunting. Speech recognition is obviously daunting, too. But I think that with a good theory, and a good database of podcasts and transcripts, a system might be able to learn to identify the words without having to learn the specifics of a culture. But this order is always, always, debatable, right? Just when we think we have nailed it the problem reappears, as a larger, farther off, one.

Thanks for the input!

Anonymous said...

Sense is originated in paradoxes, according to the philosopher Deleuze. Understanding is "making" (sic!) sense. Running a computer program means to calculate in finite time and space, hence it is a formally complete mapping, a formula, a tautology. There is nothing which has not been there. Ergo: first we have to implement a(t least one) paradox, then (artificially natural) intelligence. :))

(do you have this copycat-hofstadter stuff as delphi code, would be interested!)