Perception and Intentionality

Dreyfus, in summing up the impossibility of creating an artificial mind, notes in passing that: "In this book I have only been concerned to argue that the current attempt to program computers with fully formed Athene-like intelligence runs into empirical difficulties and fundamental conceptual inconsistencies.... Could we then program computers to behave like children and bootstrap their way to intelligence? This question takes us beyond present psychological understanding and present computer techniques."(20) Artificial life attempts something of this nature but is bogged down by the nature of its assumptions and its fixation with mimicry. To create anything resembling a consciousness we would have to begin from scratch, or, more accurately, allow the system to begin from scratch. It would be necessary to accept from the beginning that we could not create a consciousness. We could only create an acceptable framework in which a consciousness might evolve.

What would constitute an acceptable framework?

First, the system would have to have access to the world through as many systems as possible, and it would have to be the "real" world. I consider it exceptionally unlikely that any sort of consciousness could evolve within a virtual reality. The level of detail and perceptual richness would almost certainly be inferior, and, more, any realistically hypothesized virtual reality would already have been filtered through the perceptions of its creators, not simply via what they considered worthy of being in it, but what they even remembered could go in it. How many creators, for example, would remember to put ultraviolet light, ultrasonic sounds, and other occurrences beyond the range of human perception into a virtual world? Each time we lessen the exposure of the system the full range of the chaotic universe in which we live, we are reducing the necessity for the system to evolve a sophisticated method (ie a consciousness) to deal with its complex perceptions.

Second, the system would have to have means to act upon its environment, and a reason to do so. A system which could only absorb information from its surroundings but could not act on that information would not become conscious. Similarly, a system which did not need something from the environment (or needed something only in very specialized situations) would be unlikely to develop any system for dealing with the world. Plants, for example, which need only freely available sunlight, have evolved only basic means for getting what they need and nothing which we recognize as a consciousness. Animals must hunt down or search out their food and have a considerably more sophisticated set of behaviors. Nor is it enough for a computer to have a programmed goal (like A.L.I.C.E's "become smarter than humans and immortal"). Dreyfus discusses this at length in Chapter Nine of What Computers Still Can't Do:

One difference... [is that] human concerns only need to be made as specific as the situation demands. This flexibility is closely connected with the human ability to recognize the generic in terms of purposes, and to extend the use of language in a regular but nonrulelike way. Moreover, man's ultimate concern is not just to achieve some goal which is the end of a series; rather, interest in the goal is present at each moment structuring the whole of experience and guiding our activity as we constantly select what is relevant in terms of its significance to the situation at hand. A machine table of objectives, on the other hand, has only an arbitrary relation to the alternatives before the machine, so that it must be explicitly appealed to at predetermined intervals to evaluate the machine's progress and direct its next choice.(21)
Again, this is what the artificial life programmers are trying to get at by programming their creations to survive. Their mistake is in making that the only objective of the system to the point that it blocked other, more complicated, behaviors. Artificial life have developed startlingly inventive ways to survive, reproduce, and kill each other, but none to my knowledge has developed anything identifiable as curiosity - which, considering their limited range of perception and sterile environment, is not particularly surprising.

I do not think there is any particular way to program, for example, curiosity into a system, or even the will to survive, without creating a false intentionality into the system. It would be interesting to see whether a system programmed with a simplistic knowledge that it required this or that substance, and given the means to recognize that substance, would begin to seek it out on its own. That would certainly be the optimal situation, and brings me to my next point:

Third, the system would have to have some method for learning. If the system truly is a "blank slate", or at the least a slate with some very faint writing on it, it will have to be capable of self-improvement. And here I run point-blank into one of the largest problems in artificial intelligence. Although there are computer programs capable of learning, there is nothing with anything near the ability to construct casual relationships in the way a human, or even an animal, brain does. Nor does it seem likely that human memory patterns could simply be pasted onto a computerized brain with any hope of success; that returns me to the AL problem of mimicry. The human and the computer memory are simply too different. A computer is perfectly capable of remembering everything in photographic detail. This may in fact be the problem; the system is too good at remembering, and floods itself with unnecessary detail, whereas a human remembers only a hazy picture and is able to respond much more quickly. One could, of course, "cripple" the computer memory to make it less perfect, but this seems a backward way of going about things. It seems more feasible (and elegant) to accept that the computer memory will not follow the pattern of the human memory, and try to create and program a method of learning that was more suited to the computer memory. Laying out any details of such a program is well beyond my level of expertise. Some research on humans with photographic memories might be done; I would guess that the computer learning would need to be more methodical and more structured than human learning, something that would be made up for by its greater accuracy and detail. This is certainly the point at which one hits Dreyfus's "present psychological understanding and present computer techniques." It certainly opens the question of how much a human, or how much a computer, is able to structure the world at "birth", and how much is a priori structuring innate to them. And if innate, is it innate to the "programming", or the genes, or is it innate to the physical structure?

Fourth, and finally, the system will have to be able to propagate itself. This is another sticky point. I consider it unlikely that a conscious species would not have some urge to extend its lifetime beyond the lifetime of its physical body. If it did not it would not experience much of the sympathy, empathy, and communal urge experienced by biological creatures, creating an extraordinarily selfish, quite possibly unpleasant, system. On the other hand, anything resembling "natural" biological reproduction is difficult to conceive. Replication via software strikes me as an unwholesome return to the premises of conciousness-as-program. It is possible that once the system reached a certain level of complexity (and, presumably, consciousness), it would be able to build a new, improved version of itself in the same way that it was once built by humans, but this seems a somewhat sterile and overcontrolled method of reproduction which would not introduce enough variability into the population. Again, I have no particular solution for this problem. It is worth noting that metal is considerably more durable than flesh, and so any computer intelligence would be likely to live considerably longer than a human one; reproduction would not be the violently urgent matter for them that it is for us.

I doubt that these four points provide in any way a conclusive blueprint for an artificial intelligence, but I believe they provide a viable ground on which to begin.




previous page

next page



Contents copyright © K. Feete, 2002. All rights reserved.