I have been thinking about this for a while now and decided to share it with you and get your opinions. The classic Turing test was intended as a way of conveying the strong AI viewpoint that when an entity displays intelligence behaviorally indistinguishable from a human then we have in fact created intelligence. These days we see the use of image based security in most account registration pages. Basically when you create an account on most public mail servers, etc., you may have to go through the step of keying in the characters you see in a provided image. The purpose behind this is to avoid robots, or automated programs, from creating accounts. Clearly then, this is a test that is intended to distinguish humans from non-intelligent machines. It is a form of a Turing test, although less encompassing. To me it appears like a really nice challenge to actually develop a program that can in fact recognize the characters in such images. Sure it is a tough problem, but it has the one nice feature. The image is expected to be unambiguously recognizable to any human creating an account. So I believe that must be a limit to which the characters in the image can be distorted or noise added. So, if we start off with a program that can handle images of a certain class (let us say images generated on a particular website), and progressively improve it to work for more and more web sites, then we might eventually get to the limit where no other image classes exist that are beyond the scope of our programs which are still readable by any human creating an account. If we reach such a point, then mail server companies might change to some other tests, but which still need to be mini Turing tests. We could then take that as the next challenge and continue.
No comments:
Post a Comment