larry@JPL-VLSI.ARPA.UUCP (12/17/86)
I don't think we need a practioners' and a philosophers' AI discussion list, but more effort to bring the two types of discussion together. This is such an effort. There seems to me to be little gain from giving a Turing Test, which measures intelligence on a single dimension with a binary scale. Further, it's only useful after one's work has produced a (hopefully) intelligent machine, giving little help in the creation of the machine. More useful would be a test that treated intelligence as a multi-dimensional activity, somewhat like the various clinical IQ tests but considerably expanded, perhaps with social or emotional dimensions. I'd also like to see more microscopic measures, based on my belief that "higher" intellectual capabilities are composites of essentially independent capacities. Memory and emotion, for instance, seem to depend upon quite different mechanisms. (Though in an organism as complex as a human there may not be any measures that are completely orthogonal.) Consciousness might be one of those higher capacities, but my guess is that it is not essential for intelligence. On the other hand, I doubt that it is an epiphenomenon having no effect on intelligent systems. Perhaps it serves to integrate what would otherwise be disparate parts working against their individual and collective survival--in other words, consciousness insures that there are no orthogonal measures of intelligence! Before we can investigate (and duplicate) consciousness we first must investigate the functions on which it depends. One of them is memory, which seems to come in many varieties. Perhaps the most crucial dimension of memory (for the study of consciousness) is its volatility. The most volatile is very short term (a half to one-and-a-half seconds) and seems to be primarily electrical in nature. Short term memory (15-30 minutes) may be primarily a chemical phenomenon. Longer term memory seems more related to biological mechanisms, and seems to come in two types, which I call mid-term (half-hour to about a day) and long-term. The transfer between mid- and long-term memory apparently occurs during sleep or sleep-like phenomena. To relate this to consciousness, I would guess that consciousness is primarily a function of very-short-term memory but depends in successively lesser ways on the other volatile memory banks. So to duplicate conscious- ness we might have to utilize some kind of multi-buffered pipeline memory. Free will is another of those nebulous ideas that may seem not to relate to AI practice. I would first say that the connection between freedom and willing may be spurious. I see others, including machines, making decisions all the time, so will is obviously a real phenomenon and probably an indispensable one for intelligence (unlike consciousness). But at least in machines most decisions are based on information and rules stored in some kind of memory (with the remaining decisions the result of error). I surmise that human decisions are similarly determined. Secondly, some psych research indicates that decisions are rarely (or never) consciously made. Instead we seem to subconsciously perform a very rapid vector summation of many conflicting motives (some "rational," some emotional). Then we decide on motion along the vector (either in a positive or negative direction), and then create a publicly acceptable reason for our decision which finally pops up into the conscious. (And most of us are so quick and skilled at subconscious rationalization that it seems to us as if the "reason" preceded the decision.) To duplicate/emulate this form of decision-making analog computation may be more efficient than symbolic computation. Larry @ jpl-vlsi.arpa