AI pioneer says field needs more hu­mans

Austin American-Statesman Sunday - - TECH SUNDAY - By Ryan Nakashima

Andrew Ng has PALO ALTO, CALIF. — led teams at Google and Baidu that have gone on to cre­ate self-learn­ing com­puter pro­grams used by hun­dreds of mil­lions of peo­ple, in­clud­ing email spam fil­ters and touch-screen key­boards that make typ­ing eas­ier by pre­dict­ing what you might want to say next.

As a way to get ma­chines to learn without su­per­vi­sion, he has trained them to rec­og­nize cats in YouTube videos without be­ing told what cats were. And he rev­o­lu­tion­ized this field, known as ar­ti­fi­cial in­tel­li­gence, by adopt­ing graph­ics chips meant for video games.

To push the bound­aries of ar­ti­fi­cial in­tel­li­gence fur­ther, one of the world’s most renowned re­searchers in the field says many more hu­mans need to get in­volved. So his fo­cus now is on teach­ing the next gen­er­a­tion of AI spe­cial­ists to teach the ma­chines.

Nearly 2 mil­lion peo­ple around the globe have taken Ng’s on­line course on ma­chine learn­ing. In his videos, the lanky, 6-foot-1 Bri­ton of Hong Kong and Sin­ga­porean up­bring­ing speaks with a dif­fi­cult-to-place ac­cent. He often tries to get stu­dents com­fort­able with mind-bog­gling con­cepts by ac­knowl­edg­ing up front, in essence, that “hey, this stuff is tough.”

Ng sees AI as a way to “free hu­man­ity from repet­i­tive men­tal drudgery.” He has said he sees AI chang­ing vir­tu­ally every in­dus­try, and any task that takes less than a second of thought will even­tu­ally be done by ma­chines. He once said fa­mously that the only job that might not be changed is his hair­dresser’s — to which a friend of his re­sponded that in fact, she could get a robot to do his hair.

At the end of a 90-minute in­ter­view in his sparse of­fice in Palo Alto, Calif., he re­veals what’s par­tially be­hind his am­bi­tion.

“Life is shock­ingly short,” the 41-year-old com­puter sci­en­tist says, swivel­ing his lap­top into view. He’s cal­cu­lated in a

Chrome browser win­dow how many days we have from birth to death: a lit­tle more than 27,000. “I don’t want to waste that many days.”

An early start

An up­start pro­gram­mer by age 6, Ng learned cod­ing early from his fa­ther, a med­i­cal doc­tor who tried to pro­gram a com­puter to di­ag­nose pa­tients us­ing data. “At his urg­ing,” Ng says, he fid­dled with these con­cepts on his home com­puter. At age 16, he wrote a pro­gram to cal­cu­late trigono­met­ric func­tions like sine and co­sine us­ing a “neu­ral net­work” — the core com­put­ing en­gine of ar­ti­fi­cial in­tel­li­gence mod­eled on the hu­man brain.

“It seemed re­ally amaz­ing that you could write a few lines of code and have it learn to do in­ter­est­ing things,” he said.

After grad­u­at­ing high school from Sin­ga­pore’s Raf­fles In­sti­tu­tion, Ng made the rounds of Carnegie Mel­lon, MIT and Berke­ley be­fore tak­ing up res­i­dence as a pro­fes­sor at Stanford University.

There, he taught ro­botic he­li­copters to do ae­rial ac­ro­bat­ics after be­ing trained by an ex­pert pi­lot. The work was “in­spir­ing and ex­cit­ing,” re­calls Pi­eter Abbeel, then one of Ng’s doc­toral stu­dents and now a com­puter sci­en­tist at Berke­ley.

Abbeel says he once crashed a $10,000 he­li­copter drone, but Ng brushed it off. “Andrew was al­ways like, ‘If these things are too sim­ple, ev­ery­body else could do them.’”

The GPU

Ng’s stand­out AI work in­volved find­ing a new way to su­per­charge neu­ral net­works us­ing chips most often found in video-game ma­chines.

Un­til then, com­puter sci­en­tists had mostly re­lied on gen­eral-pur­pose pro­ces­sors — like the In­tel chips that still run many PCs. Such chips can han­dle only a few com­put­ing tasks si­mul­ta­ne­ously, but make up for it with blaz­ing speed.

Neu­ral net­works, how­ever, work much bet­ter if they can run thou­sands of cal­cu­la­tions si­mul­ta­ne­ously. That turned out to be a task em­i­nently suited for a dif­fer­ent class of chips called graph­ics pro­cess­ing units, or GPUs.

So when graph­ics chip maker Nvidia opened up its GPUs for gen­eral pur­poses beyond video games in 2007, Ng jumped on the tech­nol­ogy. His Stanford team be­gan pub­lish­ing pa­pers on the tech­nique a year later, speed­ing up ma­chine learn­ing by as much as 70 times.

Geoffrey Hin­ton, whose University of Toronto team wowed peers by us­ing a neu­ral net­work to win the pres­ti­gious ImageNet com­pe­ti­tion in 2012, cred­its Ng with per­suad­ing him to use the tech­nique.

That win spawned a flurry of copy­cats, giv­ing birth to the rise of mod­ern AI.

“Sev­eral dif­fer­ent peo­ple sug­gested us­ing GPUs,” Hin­ton says by email. But the work by Ng’s team, he says, “was what con­vinced me.”

Let it go

Ng’s fas­ci­na­tion with AI was par­al­leled by a de­sire to share his knowl­edge with stu­dents. As on­line ed­u­ca­tion took off ear­lier this decade, Ng dis­cov­ered a nat­u­ral out­let.

His “Ma­chine Learn­ing” course, which kicked off Stanford’s on­line learn­ing pro­gram along­side two other cour­ses in 2011, im­me­di­ately signed up 100,000 peo­ple without any mar­ket­ing ef­fort.

A year later, he co-founded the on­line-learn­ing startup Cours­era. More re­cently, he left his high-pro­file job at Baidu to launch deeplearn­ing.ai, a startup that pro­duces AI-train­ing cour­ses.

Every time he’s started some­thing big, whether it’s Cours­era, the Google Brain deep learn­ing unit, or Baidu’s AI lab, he has left once he felt the teams he has built can carry on without him.

“Then you go, ‘Great. It’s thriv­ing with or without me,’” says Ng, who con­tin­ues to teach at Stanford while work­ing in pri­vate in­dus­try.

For Ng, one of his next chal­lenges might in­clude hav­ing a child with his roboti­cist wife, Carol Rei­ley.

“I wish we knew how chil­dren (or even a pet dog) learns,” Ng says in an email fol­low-up. “None of us to­day know how to get com­put­ers to learn with the speed and flex­i­bil­ity of a child.”

Andrew Ng, seen in his Palo Alto, Calif., of­fice last month, is one of the best-known ar­ti­fi­cial­in­tel­li­gence re­searchers in the world. He sees AI as a way to “free hu­man­ity from repet­i­tive men­tal drudgery,” and wants to ramp up the teach­ing of ma­chines by AI spe­cial­ists.

ERIC RISBERG / AS­SO­CI­ATED PRESS

Com­puter sci­en­tist Andrew Ng (right) works with stu­dents last month in his of­fice at Stanford University in Palo Alto, Calif. Ng’s work has in­cluded the de­vel­op­ment of self­learn­ing tech­nolo­gies such as spam fil­ters.

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.