From the colorful pictures of brain coral that hang on the walls to the promotional key chains sporting little plastic cerebral cortices, you'd have to be gray-matterless not to notice the decorative theme at this nonprofit scientific research organization. Though it may all seem a little over the top to the average visitor, such brain mania seems excusable for someone who's spent about 25 years studying the workings of this most thoughtful of organs.
The institute's director, Jeff Hawkins, was interested in the brain even before he helped spawn an industry with his most famous invention, the PalmPilot. In his spare time, he learned the sciences behind brain research, and after becoming versed in them he developed his own theory, which is contrary to some of the established ideas. In his first book, "On Intelligence," Hawkins explains his theory and how it can be used to build truly smart machines--a question others have tackled, through the study of and neural networks, but haven't resolved.
Hawkins says the main difference between his idea and others is that the other methods try to copy human behavior using the wrong notion of how the brain works. The brain doesn't produce an output for every input, Hawkins says. Instead, it stores experiences and sequences and makes predictions based on those memories. Using that realization about intelligence as a starting point, scientists and inventors can create new and smarter machines, he says.
And as if it weren't enough to be designing future handheld and phone devices atwhile running a research institute, Hawkins is now debating whether he wants to head a start-up devoted to creating such intelligent machines. He recently spoke with CNET News.com about his book and his brain theory, and about how long we might have to wait for the appearance of computers based on that theory.
Q: This book has been about two years in the writing. What was its genesis?
A: I have been working on this theory for a while. I've been going out and giving talks about it, but in a limited time period I was trying to present the theoretical framework and the detailed biology behind it. I found I just could not cover all the basics in a talk.
Someone suggested to me, "You know, you need to put this in book form, because that is the only way you're going to be able to get it all down and have people sit down and read it." I realized that they were right.
How would a machine that worked more like the brain do a better job?
Current computers just don't understand what is being done, and they . The problem with something like is that computers are trying to just recognize speech. They take some pattern and try to match it against some template. We understand speech, but with current systems, there is no understanding. So when you have real data coming in that is messy for the most part, you can't match it.
In your book, you talk a lot about the cortex. What is it, and why is it so significant?
If you look at a human brain, you can essentially divide it into two pieces. You have got this big thing on top, which is the cortex, and you have everything else stuck up in the middle. It looks like a little post and that thing in the middle is the old brain. It's what every other animal has, but only mammals have the cortex.
The cortex is a thin sheet like a dinner napkin, and it's about as thick as six business cards stacked flat on top of one another, about six millimeters thick. It is important because it was determined many years ago that this is where all intelligence lies. It is the location for language, map, music, art, programming culture--everything that we think about (as) humans. This is where all the things that we think (of) as higher-level thought perception occur. The key to understanding what intelligence is, is in understanding the cortex.
So if I want to build intelligent machines, I'm not going to base them on the old brain. I want to base them on the rational part of the human experience. Fortunately, the cortex is this extremely uniform structure.
You talk about the brain as always predicting things. Humans act on those predictions, and experiences provide sensory input that's sent back to the brain, which develops new predictions. A computer is mostly computing its most recent thing and involves very little prediction. Elaborate on that difference.
Well, our brains work on a completely different principle than computers. It doesn't mean you can't emulate a brain on a computer, but you have to understand what the brain is doing first. The failings of (artificial intelligence) come from the idea that you have some input and then you have some output. You feed in some information, and the output you get determines the success of the system.
They didn't have a concept of what thinking is or what perception is or what it means to understand something. The biggest conceptual difference between computers and brains is the ability to predict. Brains have this input, and their output is this internal prediction mechanism. It's basically saying, "Hey, before I act, before I do anything, I need to check. Do I understand what's going on?" Success is not whether you have the right behavior; it's whether you actually compute (with the future in mind) and you can see what's going to happen next.
Assuming your theories are correct and there is a common algorithm that allows the brain to essentially process different sensory inputs in a similar manner, how close are we to understanding that algorithm?
There are some things that I don't understand, but for the most part I think I've got the basics of it down. It's not like years away. I have a graduate student here who is building this stuff now. There are also a couple of computer science departments working on this. Computer scientists love this kind of stuff; biologists are very receptive to this; so I've had some really great feedback.
Machines that can
perform tasks usually
reserved for humans
are finally becoming
How long before this work moves out of academia and into commercial uses?
I would predict within a year there will be start-ups working on this. I am debating whether I want to do that myself.
What do you see as the challenges to making this into a commercially appealing product?
It's just time and effort. The idea is there; the technology is there; it's good enough. The only hesitation I would have about doing this is that I am still involved heavily in PalmOne. I also run this institute...Right at the moment I am in more of the mood of, "Let's see if I can get 1,000 other people working on this."
But it takes a long time.
Two years now, no problem: It will be happening. You know, there will be businesses started on this, and people will be working on it.
Is that project and your work at PalmOne at all complementary? Or are they really two different things at this point?
They are complementary, and I am interested in both projects because I think they represent the future of computers. In , it's the future of personal computing where devices that fit in your pocket are going to have superfast wireless connections. The work here is like the future of computing in general.
Do you feel like you're still able to contribute in the ways that you want at PalmOne? In the early days, obviously, you were crafting a whole direction.
In the early days I came up with the direction and then I would craft the nuts and bolts. What I am doing now at PalmOne is I have a few really big ideas that I am pursuing and I can't be as involved in the details. But the conceptual things, what do you do after --I can write a position paper, go and meet with the team, get a manager to explain what these products have to do. But I can't be there day to day when someone comes back and says, "You know, the button is sticking."
Do you think this book writing effort is a one-time thing? Or are you going to come out with another one?
Oh boy--I hope so. This is hard work. I found I could only really write if I had huge blocks of time, and I only have those on weekends.