Tuesday, May 26, 2009

Your brain is an embodied computer

I was thinking this morning (in the shower, if you must know) about brains and neural networks. Let me first just say that I believe the brain is nothing more than a network of neurons exchanging little electrical impulses. In particular, I reject the notion of "the ghost in the machine", or any other explanation of behaviour or consciousness that invokes a non-material entity like a soul or cosmic will. As far as I'm concerned, good ol' physics and chemistry is enough to 'splain it all.

By extension, then, a computer-modeled neural network could also be used to model the brain. We call these programs "artificial neural networks".

It struck me that human brains take years and years to develop. An infant has a neural network (brain), but it's not wired in a way that enables the baby to be a self-sufficient contributing member of society. Not yet, anyway. So the baby plays with blocks, blankets, food, and sometimes plastic bags... all as constructive input to train their neural networks. A neural network -- as implemented in the human and many other animals, at least -- is a pattern-searching-and-predicting machine. It's function is to incorporate experience into some sort of internal model, and use that model to predict the outcome of events and help its owner to survive and thrive. This training happens over a long period of time. It goes like gangbusters in the first few years of life, then slows down somewhat. But a human isn't even trusted to survive on their own until the age of 16 (in Canada, anyway). That suggests that 16 years of training is the minimum for a productive human being.

So, if we hope to produce useful neural-network computers, will we have to train them for 16 years? No, I don't think so. In fact, training on the order of days seems more palatable.

Why is there such a huge difference in training time? I put it down to two reasons:
  1. (uninteresting) An artificial neural network would probably only be trained on a specific task, while a human is faced with understanding "reality as we know it". That's a much larger and more complex task, so naturally takes longer.
  2. (interesting) An artificial neural network makes adjustments (learns) by changing numbers in the computer's memory. Real brains, on the other hand, have to actually alter their wiring.
Number 2 is the one that struck me in the shower today. Though I'm not a neuro-scientist (as you'd probably guessed by now), I suspect that changing the wiring in the brain takes longer than changing a number in computer memory. The benefit, though, is that the brain is a fully parallel-processing machine, while a computer can really only model the behaviour of one neuron at a time. So, when it comes to making snap decisions, like how to move to catch a ball without spilling your coffee, a real brain kicks ass because many parts of the neural circuits are doing their part at the same time. A standard computer would have to operate the artificial neurons one-by-one. This would be too slow, and the computer would probably take the ball in the crotch and spill the coffee on its keyboard.

My suspiscion (without knowing the current evidence, at this point) is that sleeping is the slow process of incorporating the day's lessons into the wiring of the brain. And I like the word "incorporate" here, because it derives from the Latin word corpus meaning "body". Hence, incorporating knowledge is the activity of representing that knowledge into the physical body.

This post is from the blog

1 comment: