9
$\begingroup$

If IQ were used as a measure of the intelligence of machines, as in humans, at this point in time what would be the IQ of our most intelligent AI systems? If not IQ, then how best to compare our intelligence to a machine, or one machine to another?

This question is not asking if we can measure the IQ of a machine, but if IQ is the most preferred, or general, method of measuring intelligence then how does artificial intelligence compare to our most accepted method of measuring intelligence in humans. Many people may not understand the relevance of a Turing Test as to how intelligent their new car is, or other types of intelligent machines.

$\endgroup$
6
$\begingroup$

It depends on how the IQ test is presented:

  1. If as for humans (effectively, as a video of the book containing the test questions being opened etc), then all AI programs would score zero.

  2. If presented as the test set of a supervised learning problem (e.g. as for Bongard Problems) then one might imagine that a number of ML rule induction techniques (e.g. Learning Classifier Systems, Genetic Programming) might achieve some limited success.

So all current AI programs require the problem to be 'framed' in a suitable fashion. It doesn't take too much thought to see that removing the need for such 'framing' is actually the core problem in AI, and (despite some of the claims about Deep Learning), eliminating framing remains a distant goal.

More generally (just as with the Turing test), in order for an IQ test to be a really meaningful test of intelligence, it should be possible as a side effect of the program's capabilities, and not the specific purpose for which humans have designed it.

Interestingly, there is only one program that I'm aware of that sits between 1. and 2.:

Phaeaco (developed by Harry Foundalis at Douglas Hofstadter's research group) takes noisy photographic images of Bongard problems as input and (using a variant of Hofstadter's 'Fluid Concepts' architecture) successfully deduces the required rule in many cases.

$\endgroup$
4
$\begingroup$

at this point in time, what would be the IQ of our most intelligent AI systems?

Zero.

There are many different kinds of IQ tests including written, visual, and verbal assessments, but the majority of questions are based on abstract-reasoning problems that involve creative thinking and true intelligence.

In other words, the computer would have to exhibit something that does not yet exist… "strong AI".

The intelligent computers of science fiction do not exist. At all. We are not even close. We have absolutely NO IDEA how to bridge the gap between what we can do now and what is depicted in pop-culture films. Even with cars that drive themselves and computers that play 'Go' — an underachieving mosquito possesses more cognitive intelligence than all the world's super computers …combined!

…or possibly "disqualified" for cheating.

Even if we could pre-format the questions in a style and delivery system it understands, what does memorization, attention, or speed mean in the context of a computer? I'm not even sure if a standardized IQ test makes sense in this context. It might be like asking how a computer would do in a spelling bee.

In human terms, we're not allowed to bring along reference materials to look up an answer; but how do you rectify that when reference-lookup is innate to a computer's existence? How do you measure memory when storage is non-volatile? This gets into an existential question about the nature of learning and knowledge vs. just taking a lot of notes.

Still, how do you even teach a computer what is meant by "which animal is least like the other four?" Did the computer really figure out what was being asked out of general intelligence, or is the computer simply designed to parse out IQ-style questions specifically? If you designed something with a foreknowledge of what would likely be asked, the computers of today might simply be able to "recognize" it as question-style 496.527b and plug in the variables.

But that's not general intelligence by any definition we use or understand. It's just a specialized, slick interpreter designed to parse out a specific type of standardized question. Ask it a style of question which it is not expecting, and you'll see the computer is exhibiting no innate intelligence at all.

Until we create strong AI, a computer has effectively no IQ.

$\endgroup$
  • $\begingroup$ "how do you even teach a computer what is meant by "which animals is least like the other four?"". Like this: foundalis.com/res/diss_research.html $\endgroup$ – NietzscheanAI Sep 14 '16 at 16:50
  • $\begingroup$ @NietzscheanAI Yes, but I'm talking about understanding the nature of the question itself. That's the kind kind of general intelligence I'm describing in the context of this post. $\endgroup$ – Robert Cartaino Sep 14 '16 at 17:00
  • $\begingroup$ Looking at our respective answers, we seem to be in agreement about the core issue: transfer learning is what's missing from AI. What I'm saying is that (of all current AI architectures) the 'Fluid Concepts' approach is closest in spirit to achieving that, by virtue of the flexibility with which it manipulates/analogises domain knowledge. $\endgroup$ – NietzscheanAI Sep 14 '16 at 17:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.