Jeff Hawkins
Will Computers Think Like Humans?
December 2007 / January 2008
Volume 5 Number 6
Jeff Hawkins is the co-founder of two prominent Silicon Valley companies, Palm and Handspring. But he wants to be remembered for his contributions to the study of how the brain works and his latest startup, Numenta, which is trying to create intelligent machines that can mimic the way the brain works. Hawkins wanted to work on the brain as far back as the late 1970s, but he ran into a wall of skepticism. He created well-known portable computers for a living and revisited brain research once he became wealthy. In 2002, he founded the Redwood Neuroscience Institute and in 2005 he wrote with Sandra Blakeslee On Intelligence, a book about the brain that has been translated into 15 languages.
He argues that the brain is a memory-prediction system that forms the basis of intelligence, perception, creativity and consciousness. Based on this theory, he says, we can finally build intelligent machines, ones that will likely exceed human ability in surprising and useful ways.
INNOVATION: If we get into a conversation with you about the brain, what are some terms we need to know?
HAWKINS: The neocortex is something we focus on. It's about 60 percent of the volume of the human brain. It's the outer layer that wraps around everything else. It's where all high-level vision, touch, language, hearing and thought occur. Anything you can tell me or everything you learned in school—€”that is all stored in the neocortex.
When it comes to intelligence, that is the most important area?
Yes, the neocortex is the part of the brain dedicated to intelligence. Other parts of the brain are closely related to it like the thalamus and hippocampus. The thalamus is an egg-shaped thing at the center of your brain used mostly for routing information and it is intimately connected to the neocortex.
How do you describe the brain's hardware, so to speak? That's part of the fun of your book. You say there are 30 billion cells in the brain.
The neocortex is a sheet of cells, about three millimeters thick. It's the size of a larger dinner napkin. It has a similar structure everywhere you look. But different areas do different things. There are language areas and hearing areas and vision areas and areas that light up when you play chess. The miraculous thing —€”speculated many years ago —€” is all of the regions are basically doing the same thing. What makes language into language is how the regions are connected together. The different regions are highly connected together. One word that describes it is it is hierarchy. In that hierarchical memory is where everything is stored.
How did you get interested in the brain?
I can't imagine why everyone isn't fascinated with the brain. It's the center of everything in our lives. More concretely, in 1979 I was fresh out of college. I started my first job at Intel in the summer of that year. In September of that year, an issue of Scientific American was dedicated to the brain. One article was written by Francis Crick, who was new to brain science at the time. He said people think they know how it works but we really don't know anything. He said what is lacking is a theoretical framework. I thought wow. This is a great thing to work on.
You pitched it to then-Intel Chairman Gordon Moore himself?
I did. (laughs). I suggested that Intel let me work on this. He passed it on to Ted Hoff, who had some background in early neural networks. They turned it down. They probably made the right call because it did take so long to come to fruition. I pitched it to Wang Laboratories because I was living in Boston at the time. I pitched it to MIT's artificial intelligence lab. They weren't interested in brains, only computers.
You felt computers were heading down the wrong path?
I felt the field of AI was heading down the wrong path. A lot of people were trying to build intelligent machines. No one was looking at the brain. They said we didn't need to know how brains worked. I felt that was wrong. I felt confident they would not succeed in many of the things they wanted to do until they understood the brain. The neuroscience world wasn't wrong. It's just they weren't doing much theory. They were collecting lots of facts. There were very few people working on overall theories of how the brain works. It was very difficult for them to do that and get funding.
You weren't studying this just to understand it? You wanted to build an intelligent machine?
First, I wanted to come up with a theory of how the neocortex works. Going back to Francis Crick's plea back in 1979, I knew we needed a theoretical framework. My first task was to come up with a theoretical framework for the neocortex, which is 60 percent of the brain. And, I realized even back in 1979 that if you could do that, if you could figure out how brains work, you could build a machine that works on the same principles.
Why is it appealing to do a machine that mimics the brain, based on what a computer can do today?
Well, there are a lot of things computers can't do today. Just totally can't do. There is no computer today that you can show a picture to and have it tell you what it is. It seems like it's easy. It doesn't exist. We don't have computers that can understand language. We don't have machines that can help us think through problems in science and math. We don't have computers that can drive cars reliably. With any new technology, it's really difficult to know where it's going to turn out. When they built the first computer, they couldn't anticipate satellites and cell phones. This is a fundamental technology. It's hard for us to anticipate how it will benefit society. It's clear it will.
Can you talk about Vernon Mountcastle and his importance in neuroscience?
He was a neurophysiologist at Johns Hopkins University. He is retired now. He did many important things in neuroscience. He was the first one to propose that there is a common algorithm underlying everything in the neocortex. Up until then, people assumed that vision ran on vision algorithms. Hearing ran on hearing algorithms. Speech runs on speech algorithms. He said that that is not what the neuroscience tells us. The anatomy says it's all the same. It's an amazing insight when you think about it. Vision and language and hearing all seem like different things. But it turns out it was just the same basic problem and task. It was an incredibly liberating insight.
How does the brain work?
What we have done is put together a bunch of pieces of knowledge that other people already had and integrated it in a more cohesive way. There was no magic here. It's an analogy to when I designed the PalmPilot. There was nothing new in that product. It was a matter of picking the most important things and putting them in the right order. That is what we have done with this theory. We said let's take what is important from these fields of research and put them in a cohesive whole. Having said that, if I were to describe it in one sentence, the neocortex is a large memory system. It is structured like a hierarchy, which is not like a linear memory system in a computer. It's fundamentally different in the way information is stored. The hierarchy is self training. You feed in sensory data to the bottom of the hierarchy. We have proposed that each region of the hierarchy is doing the same thing as Vernon Mountcastle suggested. And it's a fairly simple operation. It is looking for common sequences or patterns. It is a hierarchical memory that stores sequences in space and time.
If you relate this theory to how we do things, why can we remember songs or music very easily?
We can remember a tremendous amount about our lives. We don't remember everything. We are not like a photo album. The way the memory works is associative. If I give you part of it, you can fill out the rest. I play a melody and the rest of it comes back. If I say I remember the last time I saw you, other memories will come back. It is part of the way the memory is structured in the brain. A computer doesn't do that. You give a new piece of data to a computer and it doesn't say this is like something else I stored.
You started the Redwood Neuroscience Institute and the book in 2005. What is the reaction you got?
Interesting. When I started the institute, I worried if it would be respected. What kind of scientist would come here? Very quickly, that was not a problem. We had 120 visitors over three years I managed it. They were very well known and almost every one of them said it was great. The science was good. There wasn't a problem having it being accepted. The book is a separate thing. In it, I propose these theories and it has developed since then. Like anything it is mixed. Some say it is amazing, historic. I could point out a couple of statistics. It has been translated to 15 languages. It is being used in 12 courses but it's not written as a textbook. I am speaking as the only invited speech at the Society for Neuroscience conference. There are 30,000 members of it and it is only the third time they've hosted a talk from outside their community.
Then you started Numenta?
We started Numenta two and a half years ago because Dileep George, one of the scientists working at the institute, got excited and showed you could implement this theory in software. I predicted that in the book. I didn't know how to do it. Dileep showed me how to do it. He showed me the math. I said this is great.
Do you want to start a technology business around this? From my perspective, the technology way of Silicon Valley is a great way to get a lot of people working on a problem very rapidly. Academia moves at a slow pace. It's a good system and very thorough, but it's slow. In the technology world, people are much more interested in trying new things and experimenting. I said if we create a company, we can get thousands of people working on this. That's the goal behind Numenta, to build a technology platform and get all of these people excited about it.
When you look at what Numenta has come up with, how is it brain-like?
It's brain-like in the sense that you can map the technology onto parts of the brain. It is not like a neural network or brain emulator. Essentially, we figured out how hierarchical memories in the brain store knowledge about the world. Now we have built software in a system for building hierarchical memory.
The algorithms can be mapped onto neurons but that is not the point. You can use our technology without knowing anything about neuroscience. I imagine the applications will divide at some point in time. As we figure out better ways of doing this, we don't have to emulate the neuroscience. It's a hierarchical memory system. You train it by taking sensory data into it in raw form and it learns on its own. That's brain like. It builds a model of the world on its own. It is self-training. Then it can do inference and prediction, things that computers are not good at.
In your lifetime, what do you want to get done?
In all the years I was working at Palm, I said to my partner Donna Dubinsky, "This is all great, but neuroscience is what I really care about and that is what I want to be remembered for." Besides raising a family and other personal things, the major goal in my life is to come up for a theoretical framework just like Francis Crick asked. We'll see if we've done that. That's what I hope to accomplish in my life. And we're making progress.
Dean Takahashi reports from Silicon Valley for Innovation.
Copyright © 2012 | Innovation America