Jim,
Your letter proved to be very thought-provoking for me. I read it more than once and will peruse it even more. For now, the following statement you made worries me very much because it seems to contradict several things I know: > Much of our knowledge is based on non-causative relations. But algorithms are causal. Computers are causal, our brains are causal, a neuron fires only if some "preceding" neurons fire in turn. They use neural networks to simulate brain function, and they are causal. If our knowledge is non-causative, what are we doing representing it with causative means? It would help me if you gave an example or two. I have an example: a system of simultaneous equations. They must all be satisfied at once, so there is no "first" or "last" or any sort of causal relationship. However, the causation is not in the fact that they are simultaneous. The cause-effect relationship is "simultaneous equations ==>solution." I know that, if I have simultaneous equations, then I have a solution (or, in some cases, no solution, or many solutions) and this is what I use for my thinking. How do I know "simultaneous equations ==>solution?" Because I have studied all the methods for finding that solution. And all the methods, no exception, are causative. You select any arbitrary equation to be the "first" and process it in some way. Then you select the "next." Now next implies that there is another that precedes it, so you are forcing a cause-effect relationship. And so on. What is happening, is that I think of "simultaneous equations" as an object, and then I use causation at a higher level: if I have the equations, then I have a solution, omitting the intermediate steps. Sergio From: Jim Bromer [mailto:[email protected]] Sent: Thursday, June 21, 2012 2:15 PM To: AGI Subject: Re: [agi] Prediction Did Not Work (except in narrow ai.) On Thu, Jun 21, 2012 at 11:04 AM, Sergio Pissanetzky <[email protected]> wrote: Jim, thanks. I was thinking about how we use prediction for survival. Without prediction I would put my hand in the fire and leave it there, because I would not be able to predict that fire causes pain. Or that food is good for hunger. Just like a tree. Locomotion goes with prediction, without it I would be able to avoid pain, or seek food. Just like a tree. That's why we have a brain, to predict and to move. Sergio Yes, prediction is an important method of human thought. Perhaps I should have focused on saying that "prediction" as it has stood so far has not been reliable in producing higher intelligence. That seems like a strange idea since it is so useful in native intelligence. Much of our knowledge is based on non-causative relations. It is useful because we do not usually see the full scope of the causal relations. (The use of terms like, "full scope" become philosophically defeasible when we are talking about knowing because it is only by limiting the scope of what we are thinking about could we then say that we understand the full scope of that idea.) Similarly, much of our knowing is not based on hard edged prediction. But for the most part, if you can't get the airplane off the ground you cannot reliably discover advanced methods to improve the flight characteristics of the aircraft. What has happened is that we have discovered that our thinking is both more complicated then we imagined and more mysterious than we thought it should be at the beginning of the information age. On the other hand we can create extreme situations where the human mind fails just as our AGI programs have or would fail (for less extreme situations). For example, even if you could reliably pick out a number of objects in a scene, by reducing the light on the scene sufficiently, your analysis would fail just as miserably as most AGI programs would fail. This is an important thought experiment because it does reveal that the human mind is capable of effectively using a wider variety of methods in analyzing scenes than a computer program is. (This is a conclusion but it is a reasonable conclusion.) This then shows that theory behind AGI is not totally wrong. We can buttress this conclusion by pointing out that if the lighting of a scene (imagine an industrial setting) could be guaranteed to produce ideal lighting, many visual AI methods would succeed. If a researcher could establish what kinds of AI methods would work in the ideal situations, he could then systematically move to deal with individual variations that tend to produce worse results. And so on. Jim Bromer AGI | <https://www.listbox.com/member/archive/303/=now> Archives <https://www.listbox.com/member/archive/rss/303/18883996-f0d58d57> | <https://www.listbox.com/member/?&; ad2> Modify Your Subscription <http://www.listbox.com> ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
