Logical Methods for Testing the Effectiveness of Artificial Intelligence Prototypes
The efforts of producing a computer that is able to imitate basic human abilities, like thought and reasoning, have been increasing progressively since the invention of the first computing machine.
Soon after came the scientific and philosophical dilemma of finding an appropriate method for determining whether the hypothetical machine would or not be considered as artificial intelligence. This goes back at one of the basic questions of human kind- “What is human nature? What makes a man human?”- Because in order to name a machine as capable of human abilities we should know which are the general characteristics that define a human being.
Perhaps the most famous test of artificial intelligence is the Turing test, elaborated in 1950. As part of his argument Turing put forward the idea of an 'imitation game', in which a human being and a computer would be interrogated under conditions where the interrogator would not know which was which, the communication being entirely by textual messages. Turing argued that if the interrogator could not distinguish them by questioning, then it would be unreasonable not to call the computer intelligent.
There are several ways of interpreting and discussing the Turing test, which in effect all lead to the discussing of some key philosophical views on the mind. Considering the mind as an input-output device with the exterior world, the four main metaphysical takes on the mind would have four different views of how the decision of choosing an appropriate output to a given input is mediated. The dualistic view on the mind would see as responsible for this mediation the right conscious experiences, the mind-brain identity theory would name the right neurophysiological mechanisms, functionalism the right procedures while the behaviorist would not identify any. Accordingly, these four views would view the mental processes as phenomenological, neurophysiological, computational, while the behaviorist would see the mind as lacking any processes but just having qualities resulted from behavioral dispositions.
What these four views state effectively about the way the mind works can be explained in an easier way, as follows. The dualistic view implies a certain awareness of the self, of the past experiences and is probably the closest to imply a freedom of choice. The mind-brain identity sees the mind and the brain as a whole and as the only defining element of humans. All the processes of thinking can be explained through chemical and biological reactions. Functionalism is the closest to the Turing test and to the most basic way we perceive computers. It states that the mind acts as an axiomatic system and all the outputs are given as results of the use of the axioms and theorems present in the brain. Behaviorism sees every decision made by the mind as the choice of one alternative over another, where the preferred output is the response that has repeatedly been followed by a reinforcing stimulus.
We can see now that a method that would test the effectiveness of such a machine would be entirely based on a chosen interpretation of the mind. Moreover, not only the test should be based on this interpretation, but also the whole way the machine is built. It is obvious that the Turing test is entirely based on functionalism because the way computers were built, especially in the beginning was based on functionalism.
Nevertheless, the Turing test has been severely criticized, especially by logicians. There has even been invented a counter-test, or experiment to prove its lack of validity. This test, the Chinese Room Experiment, developed by John Searle has the main quality that assumes the premises of the Turing test, namely that the human mind works according to theory that it has built in. Shortly, the test says that an English native speaker that has no knowledge of Chinese, is locked in a room and given a set of instructions and then two batches of Chinese writings. When the third one is given, he is able to correlate one set of formal symbols to another set of formal symbols and thus, using the rules, he is able to give back certain sorts of Chinese symbols with certain sorts of shapes in response to certain sorts of shapes given him in the third batch. Thus, the person would be able to communicate in Chinese, and convince the interviewer of his abilities, without actually knowing any word of Chinese. He produces the answers by manipulating uninterpreted formal symbols, but as far as the Chinese is concerned, he simply behaves like a computer; he performs computational operations on formally specified elements. For the purposes of Chinese, he is simply an instantiation of the computer program. Searle’s conclusion is that computers can perform a program according to how they have been programmed but never understand what is actually that they are doing. Thus, there can be no discussion of whether the computer actually has intelligence or not.
Coming back to the idea of creating a new test, based on one or more philosophical interpretations of the mind, the one that seems most probable of designing is the one based on behaviorism. Indeed, the behaviorist theory supposes that most basic responses are learnt and integrated into complex patterns by a process of conditioning. According to the behaviorists, one of the main ways of conditioning – “operant conditioning” supposed that a response that has repeatedly followed by a reinforcing stimulus will occur with greater frequency and will thus be “selected” over other possible responses.
A behaviorist machine would suppose that it is programmed in such a way that it identifies similar situations, meaning, it has introduced in it several types of causalities, conditionings and other similar types of relationships between one or several objects. Thus the computer will be able to relate the data he receives to these kinds of relationships. Later, the computer will be able to identify the kind of association he is asked to operate with when he will be stimulated. Now, suppose we introduce the computer to several events. The computer will judge the events described according to his pre-existent knowledge of relationships and will thus be able to identify which kind of relationship is addressed. He will then associate to every event a certain kind of reaction to certain kind of action. The one that occurs most often will then be judged as the correct one and thus the only one according to which the computer should act in the future.
Concretely, we have a computer, which we teach relationships like “cause-effect”, for instance. We provide the computer with some information, say, some stories from a newspaper. He will be able to identify among these stories the ones in which the relationship of “cause-effect” occurs. According to the comments in the articles, the computer will be able to identify which kind of cause generates which kind of effect. By systematizing the events, the machine will then be able to formulate a general situation that will allow him to say “the appropriate reaction to this kind of event is this one”. Therefore, when given a concrete event, based on his “cause-effect” resolution, the computer will be able to give a reply to that event.
To make the whole case even clearer, let us consider the case of ethics. The computer is presented several articles from the newspaper that describe crimes, the way they were committed and what happened to the delinquents. Even if some of these articles would say that the criminals were not yet found by the police, the computer will be able to judge that, for instance violence-which has been the theme in most of those articles- brings the incarceration of the criminals. The machine is now presented with a situation from an interviewer that presents a situation and asks whether he should act in a violent way or not. The computer will undoubtedly answer “No”.
Thus, the behavioristic computer acts exactly as the behavioristic person- by reacting to certain stimuli. The whole question of a behavioristic machine is not if we are able or not to build such a computer but indeed if the mind and the human being is conditioned to such a simplistic functioning. Here, the question of free will occurs. No doubt that even human beings can be convinced using certain stimuli that hey should act in a certain way in different situations. In fact, this is the main theory manipulation works on. Still, it can be argued that as long as there is free will, man is not a purely behavioristic being.
Summarizing, my essay started off by stating that if we want to create artificial intelligence we should first define what we understand through the human mind, and identify a certain mechanism through which it works. Then, I presented four different such mechanisms that describe the way the mind works, corresponding to four philosophical views. I conjectured that any kind of test would presuppose that the machine was created bearing in mind one of these concepts of the mind. After presenting and discussing one previous such test, I have proposed an alternative one, based on the behaviorist interpretation of the mind.
Bibliography:
http://members.aol.com/lshauser/turingho.html
http://users.ox.ac.uk/~jrlucas/mmg.html
posted by IH @ 10:33 AM
1 comments
