Everyone wants to make computers work more like the human brain these days.
There are more research projects than ever trying to unlock the mysteries of intelligence, ranging from government projects like the White House's brain mapping initiative to corporate projects like Google's ambitious artificial intelligence program to academic research like Stanford's attempts to help computers understand human languages.
But computer scientist and entrepreneur Jeff Hawkins, who is best known as the inventor of the Palm Pilot, already has developed a unified theory of the brain's inner workings and created algorithms for applying the theory to computer science. What's more, he has open-sourced his work so anyone can use the algorithms and software to build their own machine learning systems -- for free.
>'We're not just open sourcing a tool that we use to build our product, but the core of our product.'
Matthew Taylor
Hawkins was the co-founder of Palm and Handspring, which were behind several widely used mobile devices in the late 1990s, but artificial intelligence and neuroscience are his real passions. It's not a huge leap; Palm's handwriting recognition system was exactly the sort of thing that could benefit from a dose of artificial intelligence. After leaving Handspring, Hawkins founded the Redwood Center for Theoretical Neuroscience to study the brain full-time, and he co-authored On Intelligence with Sandra Blakeslee. In 2005, he co-founded Grok, originally known as Numenta, to turn his intelligence research into a marketable product.
But he wasn't content to keep the company's secrets to himself, so in addition to publishing a white paper outlining the theory and mathematics, the team has released NuPIC, an open source platform that includes the company's algorithms and a software framework for building prediction systems with them.
"We're not just open sourcing a tool that we use to build our product, but the core of our product," says Matthew Taylor, Grok's open source community manager. "Grok could not exist without NuPIC."
Grok makes a cloud-based service for monitoring IT infrastructure, such as servers and network gear. If something abnormal happens, or if the software predicts that a system is going to fail, it can notify your IT team. To do this, the software must be able to learn what normal looks like, spot anomalies, and make predictions based on the data. In other words, it needs to emulate the pattern-recognition systems of the human brain.
"Google's mission statement is to organize the world's information," says Taylor. "I like to think of our mission as understanding the world's data."
Grok is based on a set of machine learning algorithms created by Hawkins called cortical learning algorithms, or CLA. CLA attempts to realistically model the structure of the human brain, specifically the neocortex, which handles high-level cognitive functions like spatial reasoning and language processing. Physically, the neocortex consists of six layers; CLA mimics this hierarchy. It emulates only a tiny portion of the cortex now and it's far from being able to emulate the entire brain, but Taylor says it's a huge improvement over past work in machine learning.
It may seem obvious that computer scientists should look to the human brain to learn how to make machines more intelligent. But it's actually controversial.
"Many machine learning experts have told me that brains are, at best, good for inspiration but the details are not important," Hawkins wrote on the Numenta blog. "Of course, not everyone felt this way, but for many years I could count on one hand the people I knew who studied the neocortex in order to build intelligent machines."
NuPIC isn't the only collection of open source machine learning algorithms. Others include Apache Mahout, RapidMiner and Weka. But Taylor says NuPIC has a couple things that make it unique. One is what computer scientists call online learning. "A lot of these other algorithms you have a test data set that you train it on, and then go live," he says. "But if the patterns change you have to retrain. NuPIC doesn't operate like that, in the same way that your brain doesn't. As patterns change, it will forget the old patterns and remember the new patterns."
At first, the new patterns will appear to be anomalies, but CLA can adapt to new patterns much the same way your brain can adjust to change. "If one day you wake up and everything that you used to perceive as blue is now red, that will be really disturbing at first," he says. "But eventually you'll adjust."
The main audience for NuPIC will be academics and researchers who want to experiment with the CLA algorithms, but any company could use the platform to build their own products. "We don't have time to take this technology in all the different directions it could go," says Taylor. "We're building a product on top of it, but there's a great variety of ways it could be used."
Grok is using NuPIC for IT infrastructure monitoring, but it could be used for natural language processing, machine vision and robotics. Although no one but Grok has built a product on it yet, Taylor says companies like IBM and Seagate are looking at NuPIC. Meanwhile, outside developers are starting to contribute. And even people who have never contributed code or documentation are contributing ideas to the project's mailing list. "We're also having great discussions on the mailing list with people who know their neuroscience, know their computer science," he says.
