July 1999

That's Not How My Brain Works...

Continued from page 2

By Technology Review

TR: So we have fundamental assumptions about things that help us make sense of the world.
HAWKINS: Say I moved the doorknob on your front door up an inch. Now when you come home, you'd reach out for the doorknob, and it wouldn't be in the right spot. You'd notice that immediately, a misprediction. What if I made the doorknob a little wider or narrower? What if I made it stickier or heavier? I can think of a thousand changes I could make to your door and you'd notice them all. Now, the approach to this in traditional artificial intelligence (AI) research is to create a door database or door schema-a compilation of all the door's properties. Then the AI machine would test every one of those properties, one after another.

TR: And you're saying this is not how real brains work?
HAWKINS: I can guarantee you that. Your brain has no door database. We have to have a mechanism that tests all these door attributes at once. Autoassociative memories naturally make predictions about all their inputs. They are a great candidate mechanism. In a nutshell, intelligence is the ability of a system to make these low-level predictions about its input patterns. The more complex patterns you can predict over a longer time, the more you understand your environment and the more intelligent you are.

TR: How did these ideas lead to the PalmPilot?
HAWKINS: I was at Berkeley in the mid-1980s, which was just when neural networks were becoming fashionable again. A company called Nestor was trying to sell a neural-network pattern analyzer to do handwriting recognition-for $1 million. I thought, there have got to be easier, better ways of doing this. I took some of the math I was working on and designed a pattern classifier, which I received a patent on.

TR: What did you do with it?
HAWKINS: Just for fun, I built a hand-printed-character recognizer. Then I thought about building a computer that could use it. This started me down the path of building pen-based computers, first the GridPad and eventually the PalmPilot. The pattern recognizer in today's Palm products is based on the same recognition engine I created 12 years ago. It was inspired by the work I was doing in autoassociative memory.

TR: So the PalmPilot was just a byproduct, not a goal.
HAWKINS: Yes. I figured I could be successful building little computers that used my recognizer. It would give me some time to think about how I would get other people interested in autoassociative memories. Originally, I thought I would build portable computers for four or five years, make a name for myself, and then work full time on neurobiology.

TR: It has been almost 15 years.
HAWKINS: Yes, but that's still my intent. In the next couple of years I hope to start spending more time on autoassociative memories. If I get to my deathbed and I haven't made a significant contribution to the theory of how the brain works, I'll be disappointed.

TR: Meanwhile, might your ideas about brain function lead to other commercial possibilities?
HAWKINS: I wouldn't be surprised. One way to progress a science very rapidly is to find a commercial application for it. There is nothing like commercial success to get more people working on a problem.

TR: What sort of products do you imagine?
HAWKINS: Building autoassociative memories will be a very large business-some day more silicon will be consumed building such devices than for any other purpose. The amount of storage in a human brain is extremely large. It is impractical to use current memory technology to build memories anywhere near this capacity. Fortunately autoassociative memories are fundamentally different than the kinds of memories we put in computers. When you build memory chips, their capacity is limited by the physical size of the die. Since silicon will have a certain number of defects per square millimeter, if you start making the chips too big, you'll get a lot of chips with defects. Eventually the yield of good devices becomes unacceptably low-you have to throw away too many chips, driving the cost up.

TR: But this won't be true with autoassociative memories?
HAWKINS: Right-they are naturally fault-tolerant. If some percentage of the cells don't work properly, it doesn't really matter. Autoassociative memory chips will be very large and relatively cheap.

TR: What would they be used for?
HAWKINS: This is a little like asking in 1948 what the transistor would be used for. I believe autoassociative memories, like transistors, will be an enabling technology. The early applications will be modest. Ask what problems can benefit from a system that understands its environment, can predict what ought to be happening next and can recognize unexpected and undesirable events. Any human job that requires lots of attention to patterns and few motor skills is a candidate. Security surveillance could be an interesting market to start with.

TR: There are a lot of applications like that.
HAWKINS: How you get there in five steps, I really don't know. What drives me is my absolute certainty that this is the right approach to how brains are built.