Post by tkorrovi on Mar 14, 2008 21:37:27 GMT -5
Turing may be right that the "imitation game" might be enough for determining intelligence, but not every imitation game and any interrogator's impression is good enough. There should be a criterion of how to decide whether the result is positive. This is easier to find when we know something about the system, ie when we know whether some particular behavior is pre-programmed or not. This enables objective test, which is not Turing test, and is quite easy when the system has almost no behavior pre-programmed in the beginning. The test is much more difficult using an "imitation game", but not impossible when using an appropriate criterion. The test should consist of three phases, the first phase would be for finding out whether the system has a certain behavior pre-programmed, the second phase would be for training some particular behavior, and the third phase would be to find out whether it learned that particular behavior, the phases two and three can be combined. The result of such test is positive only when the result of the first phase is negative, and the result of the third phase is positive. Now let's see what kind of tests Turing proposed for determining intelligence. In his paper "Computing Machinery and Intelligence" (Turing, A.M., 1950) which is the basic document describing the Turing test, Turing proposed the following test, which should determine whether the system has intelligence:
Interrogator: In the first line of your sonnet which reads "Shall I compare thee to a summer's day," would not "a spring day" do as well or better?
Witness: It wouldn't scan.
Interrogator: How about "a winter's day," That would scan all right.
Witness: Yes, but nobody wants to be compared to a winter's day.
Interrogator: Would you say Mr. Pickwick reminded you of Christmas?
Witness: In a way.
Interrogator: Yet Christmas is a winter's day, and I do not think Mr. Pickwick would mind the comparison.
Witness: I don't think you're serious. By a winter's day one means a typical winter's day, rather than a special one like Christmas.
I think Turing was wrong here, this test would not prove the presence of intelligence, even if the interrogator would be convinced after that thest, that the machine has intelligence. The reason is that the system could have all the functionality necessary for that test pre-programmed, like the way to determine whether a phrase scans or not, what means winter's day, and what people generally think about it. This is also not so unlikely, as many things can be pre-programmed, and nothing much more than a simple if-then logic would be capable to pass such test. The test which could prove the presence of intelligence, should first determine that the system has no knowledge about the particular things which it should deal with during the test. Like it should begin with the question, whether a certain phrase is scanned or not, and the system should give either wrong answers, or then say that it doesn't know. It should also show no knowledge about what a winter day, spring day or summer day means. Then it should be taught in a way or another, in the form of conversation, what these things mean. And when after that indeed the system gives the same answers as in the conversation provided in the Turing paper, then the result of the test is positive, and a few such tests can indeed prove the presence of intelligence. But not the particular test that Turing proposed. Therefore we cannot even exactly say that Turing test can be used to determine intelligence, a certain kind of Turing test, yes, but not the Turing test the way which Turing proposed it. Such test would really prove almost nothing, and i don't know about Artificial Intelligence, as it it not very clear what it means, but at least such test cannot be used to test an Artificial Consciousness, or a true AI.
Interrogator: In the first line of your sonnet which reads "Shall I compare thee to a summer's day," would not "a spring day" do as well or better?
Witness: It wouldn't scan.
Interrogator: How about "a winter's day," That would scan all right.
Witness: Yes, but nobody wants to be compared to a winter's day.
Interrogator: Would you say Mr. Pickwick reminded you of Christmas?
Witness: In a way.
Interrogator: Yet Christmas is a winter's day, and I do not think Mr. Pickwick would mind the comparison.
Witness: I don't think you're serious. By a winter's day one means a typical winter's day, rather than a special one like Christmas.
I think Turing was wrong here, this test would not prove the presence of intelligence, even if the interrogator would be convinced after that thest, that the machine has intelligence. The reason is that the system could have all the functionality necessary for that test pre-programmed, like the way to determine whether a phrase scans or not, what means winter's day, and what people generally think about it. This is also not so unlikely, as many things can be pre-programmed, and nothing much more than a simple if-then logic would be capable to pass such test. The test which could prove the presence of intelligence, should first determine that the system has no knowledge about the particular things which it should deal with during the test. Like it should begin with the question, whether a certain phrase is scanned or not, and the system should give either wrong answers, or then say that it doesn't know. It should also show no knowledge about what a winter day, spring day or summer day means. Then it should be taught in a way or another, in the form of conversation, what these things mean. And when after that indeed the system gives the same answers as in the conversation provided in the Turing paper, then the result of the test is positive, and a few such tests can indeed prove the presence of intelligence. But not the particular test that Turing proposed. Therefore we cannot even exactly say that Turing test can be used to determine intelligence, a certain kind of Turing test, yes, but not the Turing test the way which Turing proposed it. Such test would really prove almost nothing, and i don't know about Artificial Intelligence, as it it not very clear what it means, but at least such test cannot be used to test an Artificial Consciousness, or a true AI.