How to Build a Digital Brain

Palm founder Jeff Hawkins on neurology, big data, and the future of artificial intelligence

By Alexis C. Madrigal

For decades, Jeff Hawkins lived a double life, co-founding Palm, the first successful mobile-device company, and studying the brain in his spare time. Free from the traditional constraints of academia, he has been able to range across the field of neuroscience while applying his knowledge of (and funds from) technology endeavors to understanding cognition. He describes his latest venture, Grok (formerly Numenta), as the first step on a new quest for artificial intelligence—one that he hopes will, among other things, lead to the creation of software super-astronomers someday.

Alexis Madrigal: What is Grok?

Jeff Hawkins: Grok is software that helps companies take automated action from streaming data. It does this by finding complex patterns in machine-generated data, and making predictions. It might use smart-meter data to predict energy needs, or data from complex machinery to predict equipment failures. The underlying technology is based on the principles of human intelligence.

AM: So how does it actually work?

JH: Grok is self-learning—it finds patterns in data without human intervention. Feed Grok streams of data, and it automatically models the data the way a human analyst might—by understanding which data streams are useful, trying to represent the data, and tuning complex algorithm parameters to improve results. Because it’s automated, Grok is ideal for analyzing thousands of data streams. Grok also learns continuously. Unlike most other analytics techniques, Grok learns from every data point, versus having to be retrained. No analyst needs to make a decision about when to take models offline and update them.

AM: How do most people work with data now?

JH: What most people do today is put data in big databases and analyze the correlations. Say you have 1 billion users on Facebook, and you’re trying to figure out what advertisement to feed to 20 percent of them. You want one big model on all these data. What we do is different: Say someone has 10,000 smart meters, and they’re trying to figure out what energy consumption is going to be two hours from now. We build 10,000 models. You can’t have a data analyst doing that. If you want to model every machine in a factory or every windmill in a windmill farm, it’s all about automation. We build lots of little models—that’s the future of data.

Most advanced analytics require substantial human expertise and are done in batch fashion—data are gathered for some time period and then processed in big chunks. This process can be slow and is hard to apply to a broad range of problems as the world changes. It also means using huge databases, which are expensive—it’s complex to maintain and move large amounts of data around. Grok, like your brain, is a streaming system. Data pass through Grok, predictions are made—but Grok doesn’t need to store the data to function. With millions of devices generating billions of data streams, the ability to store only what’s critical can be a significant advantage.

In industrial applications, more-basic approaches—say, looking at a sensor to ensure a temperature does not exceed a certain value—can be used to monitor equipment and alarms. These approaches have limitations: By the time an alarm is triggered, it may be too late. Or what’s normal for one machine may not be normal for the next.

Grok lets you automate processes that previously required manual adjustment. Heating or cooling systems can be turned on or off intelligently. Applications can be migrated between servers based on load. Network traffic can be rerouted. Unusual behavior of heavy machinery can generate alerts that recommend specific action. Instead of reacting to problems, you can anticipate them.

AM: Who else might use your software?

JH: People who want to do anomaly prediction. Grok works as if it’s listening to very noisy melodies and going, “I recognize some of this. That sounds a little familiar.” And all of a sudden it says, “This sounds totally different. I’ve never seen this before,” so it goes beep, beep. There are a lot of people looking for anomalous behavior in credit-card and security applications. It turns out this might be even bigger than prediction. But of course, anomalies are the flip side of predictions—if I can’t predict well, then I have an anomaly.

AM: So how does this relate to your previous work on the brain?

JH: First of all, we have a very complex brain; it’s got all these different components. But we’re just talking neocortex here. Every mammal, from a human to a mouse to a dolphin, has one. What is the neocortex doing? It’s building a model of the world, of what we call sensory motor contingencies or sensory motor patterns: Why are you wearing glasses and what does that mean? Or: If I turn my head to the right, I have expectations about what I’m going to see. Most of what we learn about the world is how it behaves when we interact with it. The neocortex builds a model of what should happen in a particular context. A bigger neocortex lets you make a more complex model, and it lets you have more sensors. And that’s what intelligence is: it’s learning this model of the world.

AM: And Grok uses a similar principle?

JH: Here’s what we do inside Grok: we build this 60,000-neuron neural network that emulates a very small part of one layer of the neocortex. It’s about a thousandth the size of a mouse brain and a millionth the size of a human brain. So: not super-intelligent, but we’re using the principle by which the brain does all the inference and motor behavior. I’m very confident that this sequence memory we use is the core of how all intelligence works. The brain’s taking in streaming data, they’re noisy, they’re constantly changing, and it has to figure out what the patterns are and make predictions from them.

AM: Is this different from other artificial-intelligence research that’s going on these days?

JH: I’ve been observing the AI and AI-neural-network fields for years, and I’ve always been a bit of a contrarian. My view has been: let’s figure out how the neocortex works, and once we understand those principles, that will be the path of building machine intelligence. Classic AI says: forget the neuroscience; it’s a matter of programming and algorithms.

AM: I have to ask, why would you want to build super-intelligent machines?

JH: We can make the world more efficient, we can save energy, we can save resources, we can help detect diseases. When I ask myself, What’s the purpose of life?, I think a lot of it is figuring out how the world works. These machines will help us do that. Many, many years from now, we’ll be able to build machines that are super-physicists and super-mathematicians, and explore the universe. The idea that we could accelerate our accretion of knowledge is very exciting.

AM: But what are all the people going to do once there are all these super-intelligent machines?

JH: Take these models we’re building with Grok. No human is going to be displaced by these things. No one is doing this—it’s impossible. Take the telephone system, where electronic switching replaced all those operators. If we had to have an operator place every telephone call in the world, there would be a billion telephone operators. Did we lose a billion jobs? Not really. We lost a few jobs, and advanced the quality of life. It’s not some dystopian future where machines do everything and we sit around in lounge chairs.

AM: Actually, that sounds pretty good to me.


Read an extended interview at theatlantic.com/thefuture.

This article available online at:

http://www.theatlantic.com/magazine/archive/2013/07/the-intuition-machine/309392/