Tuesday, July 18, 2006

Confabulous robots

Today's NYTimes science section has an article about the state-of-the-art in AI, Brainy Robots Start Stepping Into Daily Life. I read it with an eye out for news about language-related developments. Tthe article was mostly a mishmash: computers imitating neural activtiy, artificial lifeguards, Turing tests, video games, Roombas, automatic phone-answering systems and the dream of putting a robot in every home. But sure enough, there was one guy working on an electronic butler that could "hold a coversation with its master, or order more pet food."

The butler researcher's name is Robert Hecht-Nielsen, and...
Last year he began speaking publicly about his theory of “confabulation,” a hypothesis about the way the brain makes decisions. At a recent I.B.M. symposium, Mr. Hecht-Nielsen showed off a model of confabulation, demonstrating how his software program could read two sentences from The Detroit Free Press and create a third sentence that both made sense and was a natural extension of the previous text.

For example, the program read: “He started his goodbyes with a morning audience with Queen Elizabeth II at Buckingham Palace, sharing coffee, tea, cookies and his desire for a golf rematch with her son, Prince Andrew. The visit came after Clinton made the rounds through Ireland and Northern Ireland to offer support for the flagging peace process there.”

The program then generated a sentence that read: “The two leaders also discussed bilateral cooperation in various fields.”
Pretty impressive, especially given the zeugma-inducing coordination in the first sentence ('sharing coffee, tea, cookies, and his desire for a golf rematch with her son, Prince Andrew"), and the backwards anaphora ('He' in the first sentence, 'Clinton' in the second.) This is confabulation? How's it working? And how's the sentence-generation working? Templates? Neural nets? Transitional probabilities? An actual PSG?

Judging from his bio, probably neural nets; he doesn't seem like he's much focussed on language. He's got a Darwinian view—or maybe it's a kind of optimality-theoretic view—well, neural-netty, anyway—of the solution to the abduction problem, which Jerry Fodor considers insoluble by current approaches to cognitive science. But it's not clear how the language-generation bit works.

His Wikipedia article ends with the following hardly believable quote:
"This new era, which as yet has no name, will be characterized by the eternal universal freedom from want provided by intelligent machines."
Wow! Those are some rose-colored glasses! Guess he hasn't seen Terminator. Or 2001. (In the article, his butler is explicitly compared to H.A.L., apparently without intending a negative connotation.) Or Westworld. Or the Matrix. Or I, Robot. Or... It's like he's chronoported1 here from a former life as a voiceover in a 1940s newsreel about Progress.

He probably grew up reading Asimov, where robots ultimately really do have the best interests of humanity at heart. Maybe it's productive to return to that kind of relentless optimism.

Maybe linguists can name this nameless era. I hereby throw the floor open to Nominations for the Name of the New Era. I always liked "The World of Tomorrow", but that'll hardy do as a name for an era. Deixis. Any thoughts?

Update: Thanks to Nikolas for providing a link to the following short paper outlining the confabulation idea: Cogent Confabulation. It's pretty interesting! The sentence-completion algorithm does in fact seem to be based on a transitional probability calculations. In short, it's a complicated example of linguifying -- the machine is computing what words in a sequence are likely to co-occur, given a big training set of natural English. It returns a null answer if the sequence is a nonsense string (and also sometimes if it isn't). Sadly the article doesn't explain how the complete-sentence cases were generated, but probably it's just more of the same, with the power cranked up an exponent or two. I think the machine isn't processing the sentences, or representing their meanings and drawing inferences, or anything similar.
But it is an exciting time to be doing this kind of thing. Maybe Asimov was right. The Foundation could be Microsoft's kinder, gentler offspring, the Bill & Melinda Gates Foundation, if it got seriously into this kind of thing.

1How about this coinage? I was fairly pleased with it. But Google tells me someone's already named some time-management software that, and it's appeared in a couple of sci fi stories with the 'time-travel' meaning. Not many hits, though, considering how obvious a form it is. Maybe it'll take off.

3 Comments:

Anonymous Anonymous said...

Is that a great post, or what.

I am very cautious of building too much 'knowledge' into our 'intelligent' machines and then sitting back and basking at their "deductive powers"...

2:23 AM  
Blogger Nikolas Coukouma said...

"Cogent confabulation" published in Neural Networks Letter is a paper by him about this confabulation idea. The example in it is phrase completion, not generation. A tech report discusses processing a bit. I haven't found anything detailed enough to satisfy me and the lack of benchmarks disturbs me.

4:46 AM  
Anonymous Anonymous said...

When I read this article the first thing that came to my mind was Asimov's sci-fi novel, Do Robots Dream of Electric Sheep? I think it's still along way before we see robots that are capable of making concise and logical human judgement. When that time comes the technology will probably be very advanced as compared with today's standard.

9:44 PM  

Post a Comment

<< Home