人工知能の出現が目前に迫る？Jeff Hawkins hacks the human brainThis is a featured page

http://money.cnn.com/magazines/business2/business2_archive/2007/02/01/8398989/index.htm

The creator of the PalmPilot and the Treo isn't just making another gadget. He's attempting to fuse silicon and gray matter to produce the ultimate intelligent machine.

Erick Schonfeld, Business 2.0 Magazine editor-at-large
February 7 2007: 6:15 PM EST

Jeff Hawkins hacks the human brain

PalmPilot, Treoの作者Jeff HawkinsのNumenta社は人工知能を組み込んだ製品を18ヶ月後に発表の予定。
自ら学習し、判断するという脳の機能を組み込んだコンピュータはSFの悪夢を現実のものとするか
Jeff Hawkins
もともと、Hawkinsは脳の機能実現を長い間追求してきた。Palm社の設立文書にも本人が神経科学に一定の時間を割くことが謳ってあったほどである。
Hawkins自身も学習能力をコンピュータに理解させる方法については失敗を繰り返していたが、 Stanford grad studentのDileep Georgeが2003のある朝、研究所を訪問してきたことから、これは始まった。Georgeは脳の働きをアルゴリズムに変換してみせてHawkinsを驚かした。
はじめのうち、このアルゴリズムは早いが、スマートではなかったdirty.
Georgeは数理的処理mathの洗練を続け、ソフトウエアはどんどん改善されていった。
続いて、彼は視覚による認識の問題（画像認識？）でこのシステムを試してみた。
実際、子供は誰でも猫と犬を一目見て区別するが、コンピュータにはほとんど不可能に近い。
Georgeは犬、猫、蜂やヘリコプターの線書きを書き、新たにアルゴリズムを追加して、デジタルの描画を動かしてコンピュータを訓練した。
一旦、コンピュータが訓練を終わり、対象物を認識できるようになると、それまでに見たことのないそれらの図形の変形したものを見せた。
徐々にコンピュータは絵を正しいカテゴリーに入れ始め、どの程度、その答えに自信があるかの可能性も答え始めた。
まさにその時Jeffが考えていたことは単なる理論ではないことが最初に証明されたと関係者はいう。
Numentas社はこのことがあってから間もなく設立され、ソフトウエアはその後、大きく進化した。
同社はHawkinsが”階層型一時メモリhierarchical temporal memory(HTM)”システムと呼ぶものを開発中である。現在、HTMはいろいろなLinuxコンピュータ上でソフトウエアとして実行されているが、いずれ、シリコンの中に回路として組み込まれるものと思われる。
このシステムは大脳新皮質の構造を模倣する。HTMは人間のように、データから実際に学習して、一人前にならなければならないとNumentaの担当者はいう。
普通のコンピュータとHTMの決定的な違いは，HTMでは人間がプログラムする必要がないことである。それは自ら観察して学習する。
これはプログラマーとコンピュータの関係を根本的に変えることになる。
Machine Head
Numenta社の企業パートナー４社の一つはソフトウエアメーカーで、工場、製錬所、データセンターの電力システムの設計と分析を行っている。使用電力量の急激な増加その他の障害に極めて敏感である。このような急激な変動を避けることができれば、何百万ドルものチップウェーハーや薬品を廃棄しないで済むことになる。
もう一つの法人パートナーは自動車メーカーで名前を出すことを拒絶しているが，同社は車外に向けたセンサ（カメラ、赤外線、超音波）を装備し、HTMで訓練したシステムが交通や危険を車が理解できるかを研究中。
Numenta社は同時に石油企業からコンタクトを受けているが、地震や衛星データ上にHTMを展開して、石油探査に使えるかを相手は知りたがっている。
有るウェブ上の小売商はHTMモデルを使ってサイト上での消費者のクリックと購入行動から、より効果的な商品推奨エンジン構築について関心があるという。
Hawkinsにとっては、最終的なアプリケーションは量子メカニックスや生物学といった科学分野で新たな知識獲得を可能にすることのようだ。
現在、このような方向に向かってできるだけ多くの人の参加を実現する方法を探っているという。
もちろん、インテリジェントマシンのビジョンについては2001: A Space Odyssey からThe TerminatorまでSFの悪夢というジレンマが存在する。
尤もHawkinsはそれは考え過ぎだというが。Hawkinsが述べる次のビデオを参照。http://blogs.business2.com/business2blog/2007/02/jeff_hawkins_br.html

(Business 2.0) -- Jeff Hawkins was just another junior engineer at Intel in 1979 when he stumbled across an issue of Scientific American magazine that would illuminate a path to what would become his life's work.
It had nothing to do with the two great breakthroughs - the PalmPilot and the Treo - for which Hawkins would later become celebrated as one of the great technological and design geniuses of recent times. The issue was devoted to the human brain, and it featured an essay by DNA co-discoverer Francis Crick bemoaning the lack of a grand theory explaining how the roughly 3 pounds of gelatinous tissue each of us carries around in our skulls could possibly do all the fantastically complex tasks it does.

Jeff Hawkins wants to create the world's first truly intelligent computer.

Dubinsky, Hawkins's longtime business partner, runs operations so he can focus on brainstorming.


More from Business 2.0
NBC Universal's digital maven
Behind the (Jobs) music
Homebuilders feel the Phoenix pain
Fastest Growing Tech Companies
Current Issue
Subscribe to Business 2.0
Hawkins read it, put the magazine down, and thought to himself, "I have to work on this." Then and there, he set a goal of not just devising such a theory but using it to build a machine that, simply put, can think like a human.
People thought he was nuts. He tried to enroll in doctoral programs at MIT, then as now a hotbed of artificial intelligence; they wouldn't take him. He got into a biophysics doctoral program at the University of California at Berkeley but gave up after he was told he couldn't work on his brain-meets-machine interests because no one on the faculty would sponsor that kind of research.
Hawkins drifted back to the computer industry, but the brain obsession never waned. "In 1986, I laid out a plan, and it included making enough money to do what I wanted to do," Hawkins recalls. With the PalmPilot, first introduced in 1996, and the Treo, unveiled in 2001, Hawkins went about making his money.
Yet even as he was perfecting his groundbreaking inventions and trying to help manage the roller-coaster corporate fortunes of his companies, Hawkins quietly began puzzling out an overarching theory of how the brain works. Once in a while, he would pop up at some conference and hint that he was onto something epic - "the biggest idea I've ever had," as he once put it - but coyly refuse to give many details. There were whispers in Silicon Valley that Hawkins's project, whatever it was, was at best a distraction and quite possibly a technological white whale.
Now Hawkins is finally ready to open up about what he's been chasing. And what he says makes clear that his quest may well lead to a tremendous technical advance with far-ranging implications. Hawkins believes that his latest startup, called Numenta, is on its way to creating the first truly intelligent computer - a thinking machine that, in essence, learns the same way the human brain does.
Video: "We're not building humans," says Hawkins
Hawkins, now 49, founded Numenta in 2005 and brought in longtime business partner and Palm (Charts) veteran Donna Dubinsky as CEO. Numenta, Hawkins stresses, has nothing to do with the field known as artificial intelligence. What he has in mind is far more supple and elegant.
Rather than being inspired by biology, AI uses brute computing power and logic to make computers seem intelligent through their behavior. When IBM's (Charts) Deep Blue finally beat chess grand master Gary Kasparov a decade ago, it wasn't because it was smarter than he was. It was just faster.
Even today, computers don't have intuition. They have trouble recognizing images, understanding language, and dealing with ambiguous information. Humans have no trouble doing those things. We are intelligent, and computers are not.
Numenta's approach is radically different. Computers running Numenta software will not be programmed like regular computers. Rather, algorithms that Numenta has come up with allow machines to learn from observation, just as a child learns by observing the world around her.
Numenta is developing a new computer memory system that it says can remember the patterns of the world presented to it and use them, the way a human does, to make analogies and draw conclusions. If it works as Hawkins expects, the applications and business opportunities will be stunning. They could range from the mundane - helping radiologists or airport security officers to read X-ray images, predicting machine failures in factories, improving manufacturing yields at chip plants - to the mind-boggling: predicting tornadoes and stock prices, making smart cars, unraveling the mysteries of the cosmos. "I know this has to work because this is how the brain does it," Hawkins says.
Even with his track record, it's tempting to dismiss Hawkins's enthusiasm as overheated. "No one yet knows how human brains work," cautions Marvin Minsky, a venerated researcher who co-founded MIT's AI lab in 1959.
Nonetheless, some very impressive people have bought in. Bill Atkinson, one of the software engineers who designed the original user interface for the Mac computer, declares, "What Numenta is doing is more fundamentally important to society than the personal computer and the rise of the Internet." Atkinson pulled himself out of semiretirement to become one of the first outside developers of Numenta software.
More from Erick Schonfeld on The Next Net
The company also has four corporate partners developing their own applications: a major automaker, two defense contractors, and a software firm that monitors electrical surges in factories. Numenta plans to release a beta version of its software, along with a set of development tools and "learning algorithms," under a research license early this year.
Targeted at research scientists and hard-core programmers, the license will allow people to play around with the software for free until they're ready to create a commercial product. "We want as many people experimenting with the technology as possible," Hawkins says.
He knows that he needs to build a community of developers around Numenta with financial incentives to help his technology succeed, just as he did at Palm. He toyed with making Numenta a nonprofit, like his Redwood Neuroscience Institute from which the company sprang, but ultimately decided against it. "If you have a technology and people think they can make money from it, you will get thousands of people working on it," he explains. "I'm an impatient guy. I want to start seeing these machines working."
At his core, Jeff Hawkins is a designer in the grandest sense of the word. Whether it's a handheld computer or the human brain, he takes a big-picture, holistic approach. "When you look at the PalmPilot," he points out, "there was nothing new in it. Everything had existed in a prior product. The trick was to know what to include, what to exclude, and what we were trying to accomplish with it."
Hawkins is driven by his constant dissatisfaction with the way things are. Handwriting-recognition software was horrible, so he invented the Graffiti writing system used in the original Palm. An avid sailor, he recently commissioned a naval architect to build him a small, gaff-rigged schooner. The type of schooner he has in mind has two masts and, for the past 300 years, has required at least two people to operate it.
But Hawkins wants to be able to sail alone, so he designed a way to pull up the sails with one line instead of the usual two. "He can look at a problem and think about how to take all the constraints in the world and form the right thing out of them," Dubinsky says.
To understand how Numenta's software works, it helps to first understand Hawkins's concept of the brain. Hawkins is actually interested only in the neocortex, the outer, pink part of the brain where he believes intelligence resides. "Intelligence is about creating a model of the world and making predictions," he says.
He views the neocortex as a memory system that constantly adapts and reorganizes its connections to create that model. "When you come across new things every moment of your waking life," he says, "it looks at previously stored experience and predicts what will happen next."
The different regions of the neocortex all do pretty much the same thing, Hawkins believes. They store spatial and temporal patterns that can represent things like language, music, and vision. In Hawkins's view, all the human senses work the same way: Data from the world goes in as patterns of firing neurons, memories of the patterns are formed, and every piece of new information is matched to a stored sequence of patterns. In other words, there is one general brain algorithm that recognizes and interprets all those blips on the brain.
To many neuroscientists, this is a gross oversimplification. Hawkins is the first to admit that he has not done any firsthand neuroscience research. "People have been working on these problems for 50 years. There is hardly a new idea here," he concedes. "But what's missing is a concrete theory about how it all hangs together."
And not all experts think the big picture he paints is necessarily wrong. "His ideas are grounded strongly in neural anatomy and physiology," says Bob Knight, a neuroscientist at Berkeley (which, ironically, now houses Hawkins's research institute). The theory is sound, if a little lacking in specific details and experimental proof. Moreover, Knight concludes, "there is no doubt that if we knew how the brain worked, we would have the most unbelievable computer ever known to man."
Hawkins's early obsession with the brain was clear to anyone who knew where to look. A provision in Palm's founding documents, for instance, stated that Hawkins would be spending some of his time on neuroscience. But his rare utterances about Numenta - at PC Forum in 2005, he gave his first brief talk to a small crowd using a plastic brain model and a red dinner napkin as props - hid the seriousness with which he took the quest to create thinking machines.
Hawkins himself was largely stumped about how to convert his theory into something a computer could understand until Dileep George came to visit him at his brain research institute one morning in 2003. At the time, George was a Stanford grad student; like Hawkins, he was an electrical engineer who had gone back to school to learn about the brain. After being laughed at by other neuroscience grad students for suggesting that it might be possible to create a computational model of the neocortex, George started hanging out at the institute. "I very much liked the ideas Jeff proposed," George says. Then he did something that surprised even Hawkins. "I took the way the brain works," he says casually, "and converted it to algorithms."
The algorithms were quick and dirty at first. Most of the neuroscientists at the institute criticized George's approach as too simplistic. But Hawkins saw a kindred spirit. "This is great," he enthused to the grumbling group. "Dileep is taking my ideas seriously!"
George kept refining his math, and the software became better and better.
Then he cooked up a visual-recognition problem he calls Pictures to test the system.
While any child can identify a drawing of a cat or dog the first time it's seen, computers find the same task nearly impossible.
George made line drawings of simple objects such as a dog, a cat, a bee, and a helicopter. He loaded up his algorithm and trained the computer by animating the digital drawings.
Once the computer was trained and could identify the objects, George started to show it variations of the drawings it had never seen.
Slowly it started to put the drawings in the right categories and even gave a probability of how sure it was of its answer. Most computers today would find such a simple image-recognition problem unsolvable. Harry Saal, a Numenta board member and a founder of Network General, which was bought by McAfee (Charts), recalls the first time he saw George's Pictures.
"That this could be demoed on a stupid little laptop floored me," he says. "It was the first indication that what Jeff had was more than just a theory."
Numenta was founded soon thereafter, and the software has progressed considerably.
The company is developing what Hawkins calls a "hierarchical temporal memory" system. Today the HTM runs as software on a variety of Linux computers, but eventually it could be hardwired into silicon.
The system mimics the structure of the neocortex. "The HTM has to really learn from its data the way we learn growing up as children," explains Subutai Ahmad, Numenta's vice president for engineering.
Like the brain, it's a memory system arranged in a hierarchy of nodes where patterns and sequences of patterns are stored. These memory nodes pass information between levels as it comes streaming in over time from sensors connected to whatever is being observed - X-ray images, traffic patterns on highways or phone networks, engine or equipment performance. The idea is to first train an HTM by showing it enough data to create an accurate model and then set it loose to make predictions based on constantly changing new information from the real world.
When he first heard about Numenta, Atkinson, the former Apple (Charts) engineer, practically begged Hawkins to let him work there. In addition to his impeccable software credentials, Atkinson also knows a thing or two about the brain. Before Steve Jobs sidetracked him in the late 1970s, he spent 10 years pursuing a Ph.D. in neuroscience (which, like Hawkins, he never finished). At a meeting in Hawkins's office, Atkinson noticed the 1979 issue of Scientific American and pointed out his name in the credits: He had created the cover art, a 3-D computer graphics image of the brain. Atkinson was a shoo-in.
While not exactly an employee, Atkinson was one of the few early outside developers at Numenta and is given special access to weekly engineering meetings. Taking a page from his Apple days, he's developing simple demo applications to teach other programmers how to write software for a Numenta computer. One of his apps, for instance, can identify any of 15 languages based on a snippet of text. Another is training a computer to distinguish between the writings of authors such as Mark Twain and Nathaniel Hawthorne based solely on its knowledge of their previous works.
The key difference between an HTM and a regular computer is that you don't program an HTM. It learns by itself through observation.
This could fundamentally change the relationship between the programmer and the computer. "The programmer's job is no longer to tell it what to do," Atkinson notes. "An HTM can deliver more intelligence than the programmer has because it can learn things the programmer does not understand."
Ultimately, this simple fact could have profound implications. "As we build smart computers, will we feel dumb because we don't understand how this machine gets the answers right?" Saal asks. "I think that's about to happen."
Among Numenta's four corporate partners that are researching projects around the HTM is EDSA Micro, a San Diego-based maker of software that designs and analyzes electrical power systems in factories, refineries, and data centers, where equipment is often extremely sensitive to power surges and other disruptions. Avoiding such hiccups can mean not having to dump a million-dollar batch of chip wafers or drugs.
"Facilities might have 10,000 real data points coming in per second, but no way of interpreting that data," says Adib Nasle, EDSA's president. He's convinced that an HTM will be able to monitor all the data from electrical and other systems and find abnormalities that would otherwise be missed. For instance, a piece of equipment in a factory might experience a slight increase in current, some minor vibration, and a decrease in pressure in a valve. Any one of these signals might not be enough to set off an alarm, but taken together the pattern might be an early sign of equipment failure.
Nasle has been working with Numenta for about a year and is impressed with the progress he's seen. "Every month these guys put out a system that is 300 times faster than the one I had before," he says. He hopes to roll out an HTM-based product within two years.
Another partner is a car manufacturer, which declined to be identified, that wants to see if an HTM-trained system with outward-looking sensors (cameras, infrared, ultrasound) can help a car understand traffic and dangerous situations. If there's smoke coming from the car ahead, or if a red ball rolls out from the curb, we know to step on the brake. A car equipped with an HTM may one day know that too.
Numenta has also been approached by oil companies that want to unleash an HTM on their seismic and satellite data to find geologic patterns that could lead to new oil strikes. A chip company thinks it might be able to improve its manufacturing yields if a computer could learn to model all the different steps and variables that go into making a semiconductor.
A Web retailer has expressed interest in using an HTM to model consumers' clicking and purchasing behavior on its site so it could build a more effective product recommendation engine. "We talked to another Web-based company that has a problem with pornographic images being posted to its site," Dubinsky says. And a consulting firm thinks an HTM could help it to model companies. "Businesses are hierarchically organized," Hawkins says, "and there are a bunch of things you can measure - such as the flow of materials or the flow of dollars."
For Hawkins, the ultimate applications will be those that allow us to acquire new knowledge in areas of science such as quantum mechanics and biology. "What is exciting to me," he says, "is the prospect of building intelligent machines that sit comfortably in the realms of science where we have difficulty thinking. It will be like having a dedicated Einstein working around the clock on these problems."
That day, of course, is still far in the future. Even the earliest commercial Numenta applications are at least 18 months away. Hawkins is hoping the scheduled research release of the software in the first quarter of this year will help harness the collective brainpower of the scientific coder set and dramatically speed the advance of the technology.
"We're trying to figure out how to get as many people to work on this as possible," he says. Still, caution is in order. "This is a grand experiment that is being done with great hopes," Saal says. "But it is very speculative, and we are not yet in a position to say these applications will work." Some researchers take a dimmer view. Noted AI scientist Aaron Sloman labels Hawkins "overconfident" and charges him with "unwittingly ignoring most of the complexity" of the brain.
Moreover, there are deep moral dilemmas inherent in Hawkins's vision of intelligent machines, starting with the primal fears behind plots for everything from 2001: A Space Odyssey to The Terminator: Man makes machines smart, smart machines whale on man. While Numenta's machines aren't designed to self-replicate or have any desires, they could be taught to make independent decisions and direct other machines to act on them - even if those actions weren't in the interests of the humans they're supposed to serve.
Hawkins thinks such concerns are overblown. "People worried that the steam engine would lead to giant killer robots," he says. For him, the answers his thinking machines could provide far outweigh any hypothetical harm they might do, and he spends little mental energy on doomsday scenarios.
In fact, he's already homing in on his next challenge. As usual, it's big: the nature of time and the universe. "What's the universe really made of, and how did it get here?" he asks. Innumerable scientists are trying to answer the same questions, but Hawkins thinks they're going about it all wrong. He believes he can shed light on why the expansion of the universe seems to be accelerating. "I had an insight in my late 20s that I've been stewing on," Hawkins offers. "An expanding universe is the same as a universe where the rate of time is speeding up." Now, could a machine ever come up with such an insight?
Actually, Hawkins has no doubt that, someday, it could.


Threads for this page

Anonymous  (Get credit for your thread)

There are no threads for this page.  Be the first to start a new thread.