Over the weekend I read this interesting post from Scott Adams:
My Crackpot Theory of Intelligence
http://dilbert.com/blog/entry/my_crackpot_theory_of_intelligence/
In his post Scott basically says that intelligence is the same as pattern recognition. At first it sounds rather idiotic (I think Scott likes to start off sounding idiotic just to get people’s dander up (and what the hell is ‘dander’ and why would you not like it ‘up’?)), but as I read it sounded increasingly plausible. I have favored the idea of evolutionary programming techniques (most famous is genetic algorithms) for a long time (that is how I got my start programming), largely I think because one of the core elements of my personality is being lazy as hell. I favor the idea of writing a master program which then does the real programming (its first task might be to rewrite it, thus hopefully removing all my human errors) in a much faster and vastly less error prone way, but writing that master program is the challenge.
If, though, we take on the idea of pattern matching as intelligence, then I can see some ways around this. Pattern matching can be recursive and like fractals, operates at each level of resolution. The basic issue I see, though, is seeding the initial database. Simply making patterns match isn’t intelligence (would anyone care to consider ‘grep‘ intelligent, no matter how useful it is? (btw, I found this amusing quote at the wiki page: “You can’t grep dead trees”)), you need to have a database of conclusions based on locating (or not locating) patterns. For humans part of that is hard-wired in, but a large amount of it is learned through experience (I love Scott’s analogy of dreams: your brain is just bubble sorting new memories into its database), but for a machine it might be better to hard-code some basic information before we set it loose to start its own learning. Of course, the massive upside to machine learning is it only needs to be successful once, then that instance can be replicated an infinite time.
Regarding the multiple senses involved in the pattern matching, that is something else that rings true to me. Sometimes isolated bits of information are meaningless no matter how many ways they are presented. If your sensory tools cannot provide you with any meaningful context then the information contained in the data is lost. Additional sensory tools provide additional context enable a more rich extraction of information from the available data. The addition of more (different) sensors might work in a geometric fashion, so perhaps just the few that Scott describes would be enough.
When I was much younger (before I learned anything about computers or programming) I was very interested in AI. As I learned about programming I decided to pick up some books on AI and see how I could contribute but found the books to be so dense in math as to be incomprehensible (sort of like my love/hate relationship with image analysis). Also, since I very early on became a ‘speed freak’ when it comes to computing, the idea of using LISP which appeared to be something that had to be interpreted at run time (as opposed to compiled), not to mention totally impossible for me to read, sucked the life out of that idea straight away. I have tried my hand at genetic programming using C++ (something that can be compiled and optimized for the hardware architecture) from time to time, but I haven’t taken anything to completion enough to test for utility. It seems I am too lazy to write my master program!
