Den Beste and Jay Currie have a couple of interesting posts on Artificial Intelligence, which happens to be one of my favorite areas of computer science (and the one in which I got my best marks in in university). I do not share Jay’s confidence in genetic programming as (let’s say) a cornerstone for future progress for AI. Like symbolic logic manipulation, neural networks and expert systems, I believe that this will be a dead-end — a useful technique, but not framework for the “solution”.
My belief is that the future of AI will come from work done in Douglas Hofstadter’s Center for Research on Concepts and Cognition group, particularly the work involving “fluid analogies” and “parallel terraced scans”.
Consider vision, which Den Beste presents as a kind of pipeline of recognition, from the simple (individual pixels and their surrounding) to the complex (I’m looking at a picture of Queen Street). This is good and correct (as far as we know), but it’s the back channel that makes it all work. [Warning: all of the following is highly IMHO, though there's reasonable backing for what I'm saying here. This also should not be read as a critism of SDB's post, but an as attempt to build upon it]. Each stage of the pipeline is telling not just the next stages (the higher levels) what to look for, but also the previous stages. Thus, for example, when one stage of the system recognizes the concept of a light pole, it tells the previous stages to “look for more light poles, look for lights on those poles, look for flyers pasted to the light poles”, etc.. In fact, there’s a lot of guessing going on to make this work: a higher level stage of processing might try and “guess” that the light pole is a fishing pole, a sword, a ruler, a candy cane or what have you. It then will look for “supporting evidence” that this is true, by “tuning” the previous stages to look for related information. If the supporting information is not present, the guess will be downgraded in importance and other guesses come into play. Eventually some of the guess will be right (this is all happening very quickly because of parallelism) and a complete and self-consistent “low energy” mental picture will be built up of what is actually being seen.
Think about this: once your brain knows where the street is, it no longer needs to devote much effort into looking for cars in the sky: it knows where the cars should be. You can see that through this mechanism, the hard, massively open-ended problem of vision becomes simpler and simpler as more and more information is learned during the vision process. If there is a car in the sky, say one being carried by a helicopter for some sort of car commercial, your brain is jolted because it just doesn’t fit your previous model of how the world is expected to work.
You have the concrete experience that this model of vision is on the right track. Have you ever seen a picture of something you can’t recognize until your told what it is, and then suddenly everything falls into place? Have you seen the Crone/Young Woman illusion? These are instances of the highest levels of your vision processing system re-tuning the lower levels to recognize something they didn’t originally. Check out Hofstadter’s book Fluid Concepts & Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought for more about these concepts.
I also recommend the book “What Computers Still Can’t Do” by Hubert Dreyfus for a skeptic’s look at AI. There’s a great quote in the book that “the first man to climb to the top of a tree didn’t get us any closer to putting a man on the moon.”