Re: why not have the machine take iq tests
even iq tests for children. there are enough differnt types of tests making the sytem robust
Hi,
That's certainly a natural question to ask...
I think one could engineer a system to do really well on current written IQ tests, with a couple years' focused work by a small team...
IQ tests are hard for people, because people do them using the same abilities they use to do other, more complex and useful tasks....
But just as folks have engineered specialized systems for chess and Jeopardy, one could engineer a special system for IQ tests...
For example, with rather modest effort, this group engineered a system that does well on specific portions of the standard IQ test
http://www.kurzweilai.net/computer-programs-that-ace-iq-tests
I believe a similar approach could be taken to each of the other portions of the IQ test.... Putting them together, one would then have a high-IQ AI system -- without much capability to do anything else...
The tests used to measure intelligence of preschoolers are a bit more interesting to me, in the sense that they tend to involve intelligent engagement in the world, which is harder to "fake" ...
The bottom line is, though, IQ tests are moderately effective for humans because they are designed to exploit particular features of the human mind. It's hard for humans to pass IQ questions without being pretty generally intelligent, but the same wouldn't necessarily be true for AI minds...
Yeah, I have looked at this test before:
http://en.wikipedia.org/wiki/Wechsler_Preschool_and_Primary_Scale_
of Intelligence
Note that the test items often involve vision and action, not just
abstract puzzle solving...
Making an OpenCog system do well on this test would be a great
objective for a research project....
Or, even if that's not the explicit objective of the research, I
would imagine that after some more development is done, it will be
able to pass those tests automatically...
If an Al system first deals with these preschool/primary IQ
problems, and then moves on to the more advanced IQ problems --
THEN it will probably be able to deal with the more advanced IQ
problems in a more human-like, AGI-ish way, rather than just
it seems to me to be a more explicit well defined goal, you always
ask
what can i show you or others. I thinkg the ability to pass a
number of iq tests for children, would provide focus if not
insight
put toghetr a proposal
Yeah... the child IQ tests would be a reasonable research focus....
I can see that making an AGI pass these tests would be
attention-grabbing in the media, and also scientifically compelling
to a certain subset of the community.
I'd have to find a child psychologist willing to actually share some
specific test questions; they are cagey about supplying them online
or in books for some reason... I guess they don't want parents to
coach their tots on the specific questions ;p .... But that wouldn't
be too hard.
Selmer Bringsjord, at RPI, has done some work making robots pass
vision and movement oriented IQ tests
http://www.rpi.edu/dept/NewsComm/Magazine/dec02/atrengselaer/atrengsse
laer5.html
[Selmer is an interesting guy btw.... He is a rather good AI
researcher, mathematician and cognitive scientist -- yet also a
pretty hard-core Christian, who believes human-level AGI is
impossible because computers and robots don't have souls !! The
capability of the human mind to maintain fundamental contradictions,
yet still reason logically in closely related areas, is pretty
impressive...]
But Selmer is taking an approach of especially engineering and tuning
a system just to pass each specific test.... To me, that is not very
interesting.... It's like taking a really dumb kid and spending
hundreds of hours coaching the kid on specific sorts of IQ questions
that you know he is going to be asked. The child IQ tests could be a
good objective for an AGI research project, but only if pursued in
the right spirit...
Sure thing -- will do tomorrow or Wednesday, thanks ;)