kp@uts.amdahl.com (Ken Presting) (03/17/90)
In article <1990Mar16.104707.29360@hellgate.utah.edu> kbreinho%ug.utah.edu@cs.utah.edu (Keith Breinholt) writes: > >WHEN we have the algorithms, input devices and other hardware so that machines >can learn an arbitrary environment "out of the box", THEN AND ONLY THEN will >I believe that machines are sentient. > This position seems to me to be too strong. Human beings have very complex sensory and motor functions, and the use of these functions is a large part of human intelligence. But we don't count ourselves as lacking in intelligence because we can't see radio waves, and can't bend steel beams with our bare hands. Our sensorimotor functions are arbitrarily limited by natural selection to a class which are useful and reliably reproducible. I'd like to suggest that a machine with radically limited facilities for interaction with its environment could still exhibit category formation, concept learning, and behavioral complexity more than sufficient to merit the term "intelligent" (whatever we eventually decide that means). One thing about Keith's claim that I agree with is that the environment itself must be the real world, with real people, places, and things. Furthermore, I would go beyond Keith's statement to say that what the machine does learn must include all the things that people learn - how to use the sensory and motor facilities they have, and how to participate in the activities of other people. Human beings engage in a number of behaviors which are generally agreed to involve intellect - negotiating contracts, writing legal opinions, arguing about science and philosophy, for example. It is easy to overlook the role of perception in these activities, but it is important that when people engage in them, observations and premises based on observation are used regularly. None of these intellectual activities is a pure calculation. A machine which had very limited inputs and outputs (perhaps no more than a keyboard and printer) could participate in arguments or negotiations with a few practical handicaps, but with no intellectual handicaps. For example, suppose that a computer were acting as a judge in a courtroom, with the court stenographer typing on its keyboard, and a bailiff to announce the computer's directions and rulings. The machine would be dependent on the cooperation of able-bodied helpers, but that does not seem to be an intellectual handicap. The machine would have all the information available (eg) to an appellate court. The question then is what can a printer-keyboard machine learn, and how? The central observation here is that the input to the machine, which appears to be symbolic from our point of view, is no different from our own afferent nerve impulses, from another point of view. If the inputs are categorized according to programmed operations, then the machine does not "learn" categories in the human sense of learning. Another question is how the machine could exhibit any behavior that would reasonably be called "motivated" as opposed to "reflex". It seems natural to suppose that a printer-keyboard machine could have *curiosity*, but perhaps not much else. But if what we're after is behavior that resembles what we call "intelligent" in humans, that may be sufficient. I used the courtroom example mainly for variety. I think that arguing about science and philosophy is actually the more suggestive case. The main issue is, since the human interface to its environment is already a subset of the possible interactions, why should another subset be inadequate to support the development of intelligence, simply because it's a smaller subset? Ken Presting