Wednesday, August 17, 2005

Tip of the Tongue...

You know that feeling of irritated frustration when something is on the tip of your tongue? "What's the word? Damnit, it's..." and then, thirty seconds later, "Oh yeah, 'fish'!" The same kind of frustration you get from being almost able to beat a game - that feeling that you nearly understand, you nearly have it... but not quite.

Well, I can't speak for the population at large, but that is a sensation I feel quite often. A tickling in my brain when I encounter something I almost understand. Almost, but not quite. Personally, I think it's a good thing, since it's led me down many interesting intellectual paths. It also makes me a giant geek. You heard it here first: Craig is a giant geek.

When I feel this sensation, I am almost inevitably forced to investigate. It doesn't matter what the object is. Today I was curious about narwhals. I knew some stuff, but not very much. Just enough that when I heard them mentioned, my brain said, "oh, yeah: narwhals. I wonder..."

So I looked up narwhals. Google's image search is... pretty random. You'd think it would tend to pick, I dunno, pictures of narwhals. Whatever, I learned a lot.

Unfortunately, it is not always something as simple as narwhals. For example, I've started feeling that way about music composition. So, I suppose I'll be reading up on that, too.

I'm not here just to talk about myself, though: I'm also here to talk about my second favorite subject: pattern recognition.

I don't mean, this once, my "pattern adaptation control" theory. I'm talking about the whole field of "pattern recognition".

You see, we can do all sorts of froody stuff with pattern recognition. Whether you use symbolic logic, or neural nets (or whatever they call them these days - I haven't studied them in three years), or genetic simulation, or any of the dozens of other kooky methods... no matter which method you use, you can usually get the patterns to be recognized.

The problem is the "the". We want something that doesn't require extensive training or seeding. Something we can just plug in and watch it learn. Or, something we can train in the basics and have it come up with the complex stuff on the fly.

The problem is that these systems don't get that feeling. That, "I almost understand, I can figure this out if I try" feeling. They can't, because they can't figure it out. If you create a speech recognition system, it is precisely as good as you make it to be. If you want to make it use a syntax assessor to help it, you need to make an entirely different system that does that and plug it in. The same for any other kind of help.

Maybe our brain simply has ten million custom systems. Certainly, much of the brain defaults to a given purpose. But brains are highly adaptable: if you lose a chunk of it, especially in your youth, another chunk of it will take over. It may not do quite as good a job, but it will certainly do better than any computer could hope to do when their syntax assessor crashes.

That's the part we can't seem to figure out. How to make a system which can learn to recognize arbitrary patterns and meaningfully mate them to other parts of the system. We can't get a computer to say, "I almost understand how this works... and if I study, I will!"

Just in case you'd never heard it explained before.

No comments: