Numenta Is Imitating Your Brain

A startup's approach to advanced pattern recognition could trump rivals' in the hunt to capitalize on finding trends in large streams of data by Stephen Baker

Through his development of the Palm (PALM) Pilot and Treo smartphone, Jeff Hawkins helped change the way people access information computers. Now he may have come up with a way to alter how we comb through that data.

Hawkins' software startup Numenta is trying to replicate the thinking patterns of the human brain in an effort to recognize subtle patterns in immense streams of data. Researchers, of course, say such tools could lead to advances in data-rich fields from drug discovery to law enforcement.

Help for Data Centers

But perhaps the first industry to benefit from Numenta will be computing itself. In an interview in early July, Hawkins points to one of his company's research projects, an engagement with EDSA, a San Diego-based company that monitors and fine-tunes electricity usage in a host of industries, including data centers. These centers, the heart of data-gobbling companies like Google (GOOG), consume enormous quantities of energy. A U.S. government study predicts that power consumption at data centers, which reached $4.5 billion in 2006, could hit twice that figure by 2011—and far higher if energy prices continue to rise.

These massive brains of the world need ever brainier software to make better use of their power consumption. That's where Numenta comes in. Numenta's software, which is in research trials with partners like EDSA, is based on the brain's neocortex. That's gray matter associated with the higher levels of thinking. The software is designed to pick up the "deeper structures" in the patterns of electrical flows in data centers. In other words, instead of focusing on simple input and output of voltage, it might study the electrical path of each query into a data center.

The way Hawkins describes it, the Numenta system establishes a hierarchy of information. At its lowest level, it feeds on streams of data coming in from sensors. It recognizes patterns in those streams, just as we do when we grapple with the sounds and images that pour into our brains, which we learn to recognize as language or faces. The equivalent to sensory input in the data center would be a crazy quilt of electrical signals streaming among thousands of computers, many of them often working on the same jobs at the same time.

Numenta's system looks for the patterns, Hawkins says, "over space and time." It might see, for example, tiny surges in electrical demand associated with certain types of jobs or communications between specific machines. This analysis is what EDSA Chief Executive Mark Ascolese calls "advanced education."

More Efficient Operations

This, Ascolese hopes, will help EDSA create more sophisticated mathematical models of data center operations, adding more details about the tiniest patterns. EDSA's tools first optimize the flow of electricity in the model and then benchmark it against the live operation. This involves reading the flow "several million times per second, to detect even the slightest deviation and assess downstream implications."

The better companies can analyze their data, the more efficiently they can run energy-hogging data centers. Competing Web giants are out to stockpile as much of the world's information as possible in their data centers. This includes not just Web pages, but also video, maps, blog posts, music—in short, the digital universe. Storing these petabytes of data requires electricity. And plowing through these mountains to find and serve up Web pages or music eats up more of it.

No wonder companies are building data centers near the cut-rate hydropowered electricity in the Pacific Northwest. Microsoft (MSFT) has even been looking into building data centers in Siberia, where fuel is cheap and air conditioning—a major expense—is a nonissue nine months of the year. According to Yahoo's (YHOO) tech research chief, Prahbakar Raghavan, the race between Microsoft, Yahoo, Google, and IBM (IBM) hinges in part on which company can "turn electricity most efficiently into computing power." What's more, according to EDSA, an hour of downtime at a data center can cost as much as $6 million.

Boundless Market

Numenta is hardly alone in the market for analyzing data patterns. Vast sectors of the economy, from scientific research to marketing, are focused on finding meaningful trends in rivers of data. This is the heart of Google's business. It's crucial to the digital hunt for terrorists at the National Security Agency. Medical researchers are burrowing through patterns of genetic and health data in their hunt to conquer diseases and create new drugs. Since the basic pattern-finding challenges are similar, a breakthrough in any one of these areas could spread quickly into other industries. This means that startups like Numenta face rivals in many industries. At the same time, though, a breakthrough could open wide opportunities.

One direct rival is at IBM. Stream technology, a major initiative at IBM Research, uses supercomputers to analyze the flows of real-time data. Early customers are banks and brokerages, which are looking for patterns in financial transactions. Some might highlight inefficiencies. Others could signal changing market dynamics, giving them a chance to gird for, say, a commodity crash or a dollar movement before it occurs.

Numenta also has company in its attempts to mimic the brain. Robert Hecht-Nielsen, a pioneer in neural networking and now a vice-president for research at Fair Isaac (FIC), has designed a computing platform based upon his understanding of the neurons in the brain. He's harnessing this into an electronic butler called "Chancellor," which he says will soon be able to understand voice commands and carry out shopping and home management tasks.

If these projects take off, the market for advanced pattern recognition may be boundless. Hawkins says a horse trainer recently approached him about feeding sensors from horse's hooves into the Numenta software. Shifting patterns of the horse's weight distribution and muscle use, he says, could signal a leg injury before the animal starts to limp. That may sound like a small niche, but if using the software on hooves can prove it works, trainers could be a crucial market.