Le mardi 28 Novembre 2006 20:06, tOM Trottier a écrit :
> FYI,
> 5. TEACHING COMPUTERS TO RECOGNIZE IMAGES
> http://www.eurekalert.org/pub_releases/2006-11/ps-rtc110106.php
Thank you for the link, it's been a long time since I hadn't laughed that
much ; I tested a range of my personal images on the site, and the guesswork
of the AI is about as good as a seer. Of course, with categories as broad as
'transports', 'sports', 'people', there's at least one match out of 15. I can
do that much without even looking at a picture.
I'm totaly unconvinced, but it's fun nonetheless. Makes me think aout AI
pranks of the old times, like the fake psychanalyst and such. (comp : "Tell
me about your problem" user : "I don't like eating soup" comp "tell me more
about soup eating" etc... for hours).
But somehow, I think someday there will be something out of it ; it's just
that today technology can't cope with a single picture in less than 2
seconds. We human beings need years of training, and we're able to see our
samples in volume ; it's much easier to infer a flat projection from a know
3D sample than recognize an arbitrary matrix of points. The day we start
traigning computers with 3D models, then we might get on something out of 2D
pictures. But the amount of needed computing power is scary.
--
Manuel Viet
==============================================
List usage info: http://www.zuikoholic.com
List nannies: olympusadmin@xxxxxxxxxx
==============================================
|