Engines of Our Ingenuity

No. 164:
NEURAL NETWORKS

by John H. H. Lienhard

Click here for audio of Episode 164.

Today, we wonder how humans and computers think. The University of Houston's College of Engineering presents this series about the machines that make our civilization run, and the people whose ingenuity created them.

As the blast-front of the computer revolution sweeps outward, a question nags us. "How will the fully-evolved computer differ from our own brain -- what'll be left for us?" We haven't found the answer, but Ann Finkbeiner gives a progress report. She tells us about two ways computer engineers try to copy human thinking.

The first way is to use what's called artificial intelligence. That means trying to convert human thought into rules. Suppose we want a computer to play chess. We teach it the rules for moving chessmen and for anticipating the consequences of moves. We tell it not to put its knight there, because the queen'll take it. Next we go to chess-masters and ask them to tell us how they make decisions. We reduce their answers to rules -- as best we can -- and feed them into the computer, too. When we're through, we have a program that can play pretty well against the best chess-players.

That works because chess is fairly logical. And the computer is way ahead of us in manipulating logic. On the other hand, we have an edge over computer chess programs because we ultimately use something beyond logic. We can generalize and associate ideas; and that's quite beyond an artificial-intelligence machine. In simple terms, the machine and the human act out the balance between Captain Kirk and Mr. Spock.

Now people are looking at a second method for replicating human thought. It's called the neural network. The idea is to copy the actual machinery of the human brain -- not just to try to write rules for thinking. The brain is made up of building blocks called neurons. They're fairly simple, but they have an important feature. Their logical action is adaptive. Neurons can learn.

Experiments with machines using neural elements are beginning to give some startling results. What one can do is find an answer that's good enough very quickly. Asked to multiply 7 by 6, one might tell you the answer is about 40. That's not impressive at face value, but it's what we do when -- say -- we recognize a face. It'd take enormous digital-computer capacity to do that precisely. But we don't do it precisely. We don't compare every feature with the ones stored in our memory. We recognize friends instantly -- but approximately -- by comparing a few key features. A digital computer can't think that way, but a neural network machine can.

The jury is still out on these new machines. They do hold a lot of promise in tasks like reading fingerprints and translating languages. More than that, they help us understand human thinking. Most fascinating: they tell us that our ability to tolerate imperfection is what gives us an edge over the digital computer.

I'm John Lienhard, at the University of Houston, where we're interested in the way inventive minds work.

(Theme music)


The Engines of Our Ingenuity is Copyright © 1988-1997 by John H. Lienhard.

Previous Episode | Search Episodes | Index | Home | Next Episode