Graphics can be replaced Winograd Turing test to define a human-level AI?

Getty Images: details
Earlier this year, chatbot called Eugene Goostman "won the" Turing test for artificial intelligence as part of a competition organized by the University of the United Kingdom. Almost immediately, it became clear that instead of proving that software has reached the level of human intelligence, all this, especially in this tournament showed that it was a piece of software obtained a little clever to fool people into thinking they were talking about other rights, which is very different from a measure of the ability to "think". (In fact, some observers do not believe that the bot is very smart in everything.)

Obviously, it is necessary to test the best, and we may be one, as a sort of questions called Winograd system that is easy for humans to respond, but a serious challenge for a computer.

The problem with the Turing test is that it is not really test if the program is able to think of artificial intelligence: This is a test to see if the AI program can deceive man . And humans are really, really stupid. Soak for all sorts of stuff that the AI ​​is programmed well and can be used to convince us that we are talking about real people who can think.

For example, the chatbot Eugene Goostman pretends to be a boy at the age of 13 years, because boys aged 13 years and often irregular idiots (I was one), and that would relieve many of the circumstances in which failure of Amnesty International simply. So really, not a bot clever cat at all, it's just really good to make you overlook the times when it is stupid, focusing on interactions league when the algorithm knows how to answer questions you ask him.

In theory, the Turing test is still valid, but we need a better process to test the process of artificial intelligence. Competition for the new Amnesty International, sponsored by Nuance Communications and CommonsenseReasoning.org and price offers U.S. $ 25,000 AI that can successfully respond Winograd said plans, named after Terry Winograd, a science teacher computer at the University of Stanford.

Here is an example:

Do not get into the bag cup brown because it is too large. What is too big?

Cup, of course. But it is not clear. It is clear to us, because we know everything about prices and bags. We do not even have to "think" about it, it almost intuitive. But for a computer program, it is not clear what "it" refers to. To succeed in response to a question like this, it must be artificial intelligence knowledge base and the ability to think.

Here is another:

Jim relieved Kevin because he was very angry. , Who was upset?

These are the rules of graphics Winograd must follow:

1. Mentioned two parties in a sentence with noun phrases. It can be two men, two women and two objects or groups of persons or inanimate things.

2. Uses possessive pronoun or a recipe in a sentence referring to a party, but it is also the right kind of party. In the case of man is "he / him / her", females, it is, "it / him / her"; inanimate object is "he / she / it". "They / themselves "Groups is

3. Includes the issue of determining the reference awareness or jealous recipe. 0 answer is always the first part mentioned in the sentence (repeated but for clarity sentence), and the answer is 1 second part.

4. The word there (called a special word) that appears in the wholesale and perhaps this question. When it is replaced by another word (called another word), while still makes perfect sense, but it changes the answer.

For more details (including examples of some of the means that can include graphics Winograd clues that Amnesty International could use), this paper is easy to understand and worth reading. In fact, it is so nice to read, I'll fly the finish and then here:

Such as Turing, and we believe that getting the behavior concern main reason for the development of an artificially intelligent system. We also agree that the understanding of the English language in the broadest sense is an excellent indicator of intelligent behavior. When we have a slight disagreement with Turing is whether the free form conversation in English is the appropriate vehicle. WS our [graphics Winograd] challenge does not allow the object to hide behind the pretext of verbal tricks, fun, or standardized answers. Assuming that the subject is willing to take a WS test at all, much will be learned completely unequivocal about it in a few minutes. What we have proposed here is certainly less demanding intelligent conversation about sonnets (for example), designed by Turing, He, however, offer a challenge test, which is less susceptible to abuse.



It is important to emphasize that we are a little skeptical that you can really "test" level of human AI that way. Test with highly structured with questions and answers that suit you specific unambiguous or bad, and there is great potential for smart (but do not think) Amnesty International to find ways to exploit it.

Question, then, is whether the "intelligence" is just a technological system that is complex enough to correctly answer a series of questions that a bit more complicated biological system (we) decided arbitrarily what constitutes a measure requires thought.

It seems inevitable that at some point we have to say that intelligence is a real feeling and thinking, and "Blade" is the road