Firms will be able to build sys­tems even with­out ex­ten­sive ex­per­tise

The East African - - OUTLOOK - By CADE METZ New York Times News Ser­vice

A new breath test for di­ag­nos­ing the dis­ease.

They are a dream of re­searchers but per­haps a night­mare for highly skilled com­puter pro­gram­mers: Ar­ti­fi­cially in­tel­li­gent ma­chines that can build other ar­ti­fi­cially in­tel­li­gent ma­chines.

With re­cent speeches in both Sil­i­con Val­ley and China, Jeff Dean, one of Google’s lead­ing engineers, spot­lighted a Google project called Au­toml. ML is short for ma­chine learn­ing, re­fer­ring to com­puter al­go­rithms that can learn to per­form par­tic­u­lar tasks on their own by analysing data. Au­toml, in turn, is a ma­chine-learn­ing al­go­rithm that learns to build other ma­chine­learn­ing al­go­rithms.

With it, Google may soon find a way to create AI tech­nol­ogy that can partly take the hu­mans out of build­ing the AI sys­tems that many be­lieve are the fu­ture of the tech­nol­ogy in­dus­try.

The project is part of a much larger ef­fort to bring the lat­est and great­est AI tech­niques to a wider col­lec­tion of com­pa­nies and soft­ware de­vel­op­ers.

The tech in­dus­try is promis­ing ev­ery­thing from smart­phone apps that can recog­nise faces to cars that can drive on their own. But by some es­ti­mates, only 10,000 peo­ple world­wide have the ed­u­ca­tion, ex­pe­ri­ence and tal­ent needed to build the com­plex and some­times mys­te­ri­ous math­e­mat­i­cal al­go­rithms that will drive this new breed of ar­ti­fi­cial in­tel­li­gence.

The world’s largest tech busi­nesses, in­clud­ing Google, Face­book and Mi­crosoft, some­times pay mil­lions of dol­lars a year to AI ex­perts, ef­fec­tively cor­ner­ing the mar­ket for this hard-to-find tal­ent. The short­age isn’t go­ing away any­time soon, just be­cause mas­ter­ing these skills takes years of work.

The in­dus­try is not will­ing to wait. Com­pa­nies are de­vel­op­ing all sorts of tools that will make it eas­ier for any op­er­a­tion to build its own AI soft­ware.

“We are fol­low­ing the same path that com­puter sci­ence has followed with ev­ery new type of tech­nol­ogy,” said Joseph Sirosh, a vice pres­i­dent at Mi­crosoft, which re­cently un­veiled a tool to help coders build deep neu­ral net­works, a type of com­puter al­go­rithm that is driv­ing much of the re­cent progress in the AI field. “We are elim­i­nat­ing a lot of

We want to go from thou­sands of or­gan­i­sa­tions solv­ing ma­chine learn­ing prob­lems to mil­lions.” Jeff Dean, en­gi­neer at Google

the heavy lift­ing.”

This is not al­tru­ism. Re­searchers like Dean be­lieve if more peo­ple and com­pa­nies are work­ing on ar­ti­fi­cial in­tel­li­gence, it will pro­pel their own re­search. At the same time, com­pa­nies like Google, Ama­zon and Mi­crosoft see se­ri­ous money in the trend that Sirosh de­scribed. All of them are sell­ing cloud-com­put­ing ser­vices that can help other busi­nesses and de­vel­op­ers build AI.

“There is a real de­mand for this,” said Matt Scott, a co-founder and the chief tech­ni­cal of­fi­cer of Ma­long, a startup in China that of­fers sim­i­lar ser­vices. “And the tools are not yet sat­is­fy­ing all the de­mand.”

This is most likely what Google has in mind for Au­toml, as the com­pany con­tin­ues to hail the project’s progress. Google’s chief ex­ec­u­tive, Sun­dar Pichai, boasted about Au­toml last month while un­veil­ing a new An­droid smart­phone. Even­tu­ally, the Google project will help com­pa­nies build sys­tems with ar­ti­fi­cial in­tel­li­gence even if they don’t have ex­ten­sive ex­per­tise, Dean said. Today, he es­ti­mated, no more than a few thou­sand com­pa­nies have the right tal­ent for build­ing AI, but many more have the nec­es­sary data.

“We want to go from thou­sands of or­gan­i­sa­tions solv­ing ma­chine learn­ing prob­lems to mil­lions,” he said.

Google is in­vest­ing heav­ily in cloud-com­put­ing ser­vices — ser­vices that help other busi­nesses build and run soft­ware — which it ex­pects to be one of its pri­mary eco­nomic en­gines in the years to come. And after snap­ping up such a large por­tion of the world’s top AI re­searchers, it has a means of jump-start­ing this en­gine.

Neu­ral net­works

Neu­ral net­works are rapidly ac­cel­er­at­ing the de­vel­op­ment of AI. Rather than build­ing an im­age-recog­ni­tion ser­vice or a lan­guage trans­la­tion app by hand, one line of code at a time, engineers can much more quickly build an al­go­rithm that learns tasks on its own.

By analysing the sounds in a vast col­lec­tion of old tech­ni­cal sup­port calls, for in­stance, a ma­chine-learn­ing al­go­rithm can learn to recog­nise spo­ken words. But build­ing a neu­ral net­work is not like build­ing a web­site or some run-of-the-mill smart­phone app. It re­quires sig­nif­i­cant math­e­mat­i­cal skills, ex­treme trial and er­ror, and a fair amount of in­tu­ition. Jean-françois Gagné, chief ex­ec­u­tive of an in­de­pen­dent ma­chine-learn­ing lab called El­e­ment AI, refers to the process as “a new kind of com­puter pro­gram­ming.”

In build­ing a neu­ral net­work, re­searchers run dozens or even hun­dreds of ex­per­i­ments across a vast net­work of ma­chines, test­ing how well an al­go­rithm can learn a task like recog­nis­ing an im­age or trans­lat­ing from one lan­guage to an­other. Then they ad­just par­tic­u­lar parts of the al­go­rithm over and over again, un­til they set­tle on some­thing that works. Some call it a “dark art,” just be­cause re­searchers find it dif­fi­cult to ex­plain why they make par­tic­u­lar ad­just­ments.

But with Au­toml, Google is try­ing to au­to­mate this process. It is build­ing al­go­rithms that an­a­lyse the de­vel­op­ment of other al­go­rithms, learn­ing which meth­ods are suc­cess­ful and which are not. Even­tu­ally, they learn to build more ef­fec­tive ma­chine learn­ing. Google said Au­toml could now build al­go­rithms that, in some cases, iden­ti­fied ob­jects in pho­tos more ac­cu­rately than ser­vices built solely by hu­man ex­perts.

Bar­ret Zoph, one of the Google re­searchers be­hind the project, be­lieves that the same method will even­tu­ally work well for other tasks, like speech recog­ni­tion or ma­chine trans­la­tion.

Newspapers in English

Newspapers from Kenya

© PressReader. All rights reserved.