[FOM] AI-completeness

Timothy Y. Chow tchow at alum.mit.edu
Thu May 3 11:23:05 EDT 2007


"Mitch Harris" <maharri at gmail.com> wrote:
> See the article in the New Yorker
> 
>   http://www.newyorker.com/archive/2006/10/16/061016fa_fact6
> 
> about a neural network that has better-than-human performance in
> predicting income of movies based on plot.

That's a fascinating article...thanks for the pointer.  However, your 
summary is a little misleading.  Two or three human beings have to read 
through the screenplay or watch the movie and decide what features to 
extract and plug into the neural network.  Plus, it sounds like the humans 
have to make judgments like, "emotional connection between lead character 
and little child in scene 4 rates an 8 out of 10."  That's a huge amount 
of human intervention.

On the other hand, the article also mentions a hit-song analyzer that 
sounds like it requires very little human intervention (though it's hard 
to tell for sure from the article).  That's pretty impressive.

> To me, it seems like cheating to allow a very large neural network as
> a proposed solution to  an AI-complete problem.

Perhaps, but (artificial) neural networks haven't yet proved to be a 
silver bullet, so I see no reason to rule them out.  Neural nets can't 
play a decent game of go or translate Shakespeare into good Chinese.

Tim


More information about the FOM mailing list