January 04, 2007

Neural Nets in toyland

Apparently there's some sort of handheld toy that plays 20 questions. It's supposed to be some kind of neural net, and you can play with it (i.e., train it for them free of charge) online.

It's always illuminating to get to play around with some kind of AI technology. As good as your ideas might sound, and as impressive as the results might be in a controlled environment, nothing beats unleashing a few million curious people on it for seeing just how useful it really is. It's when people are trying to have fun with something that flaws tend to be the most glaring.

I've heard people say the results are incredible, but I'm not that impressed. You can easily get a clear impression of the program's typical trainer in the questions it asks. Tell it you are thinking of a person involved in music (Monteverdi, let's say) and the program repeatedly asks things like "were you popular in the sixties?" long after you've already told it that "you" lived before 1900.

It also likes to ask things like "do you have blond hair?" when you've already made it clear that you're thinking of someone who is not an entertainer and who lived centuries ago (Gengis Khan, for example).

In short, it repeatedly asks questions which 1) have already been answered in an indirect way and 2) which would be good questions to ask assuming you're thinking of Britney Spears.

These would seem to be artifacts of its training, and maybe it will get better in time. But it also seems fairly easy to break. On my Monteverdi round it got really confused: questions 20 and 21 (right after "Were you popular in the 90s?") were "I guessed that it was a sample?" (it was told right away that it was a person), followed up by the fairly left-field "Are you larger than a pound of butter?", which I am going to use every chance I get in future games of 20 questions.

To its credit, though, it did get Gengis Khan fairly quickly, and though on my Stanley Kubrick round it kept asking me whether the person in question was skinny, it does know about Kubrick -- at the end if it can't guess it gives you a list and asks you to give it the correct answer.

Posted by Alan Hogue at January 4, 2007 10:29 AM
Comments

Ben's gadget-crazy dad got us one for Christmas. I've been pretty pleased on the "thing" side -- it got Christmas tree and piano, but guessed pumpkin for squash.

More than the tech, though, I'm impressed with the marketing. We ended up with ours because it was featured in the IEEE magazine, Spectrum. That's a pretty niche publication to bother with...

Posted by: Sara Brumfield at January 5, 2007 08:34 AM

Reminds me of a family story about a high school classmate of mine. She was born prematurely, weighing two pounds. A friend of her mother's couldn't accept the idea of a child that size. Finally she got two pounds of butter out of the refrigerator, placed them on the counter, and looked at them.

Posted by: Martha Bridegam at January 6, 2007 08:32 PM