HWM (Malaysia) - - GEAR - By James Lu

Google is one, if not the big­gest, con­trib­u­tor to AI de­vel­op­ment today, and you can find AI in al­most ev­ery Google prod­uct, from the Smart Re­ply fea­ture in Gmail, to au­to­com­plete in Google Search, to the next sug­gested video in YouTube, to the Por­trait mode fea­ture on the Pixel 2 XL smart­phone, to Google As­sis­tant in your Google Home. In fact, AI is now the key fo­cus for the com­pany, with CEO Sun­dar Pichai stat­ing at Google I/O ear­lier this year his be­lief that we are “shift­ing from a mo­bile-first world to an AI-first world.”

But what else has Google been us­ing AI for?

More ac­cu­rate trans­la­tions

You may not know this, but in Novem­ber 2016, Google Trans­late switched from its old sen­tence-based trans­la­tion sys­tem to a new one based on neu­ral net­works.

Google’s Neu­ral Ma­chine Trans­la­tion sys­tem trans­lates whole sen­tences at a time, rather than just piece by piece. It uses this broader con­text to help it fig­ure out the most rel­e­vant trans­la­tion, which it then re­ar­ranges and ad­justs to be more like a hu­man speak­ing with proper gram­mar. Since it’s easier to un­der­stand each sen­tence, trans­lated para­graphs and ar­ti­cles are a lot smoother and easier to read. Ad­di­tion­ally, the net­work learns over time, mean­ing that ev­ery time it trans­lates some­thing, it gets bet­ter and bet­ter.

Be­come the undis­puted best Go player in the world in 40 days

Ev­ery­one’s fa­mil­iar with Google’s world cham­pion-beat­ing Al­phaGo AI, but much more im­pres­sive is Google’s new Al­phaGo Zero AI. Pre­vi­ous ver­sions of Al­phaGo learned the game through an­a­lyz­ing thou­sands of pro­fes­sional games. Al­phaGo Zero skipped this step en­tirely, and was just taught the ba­sic rules of the game, and im­proved only by play­ing matches against it­self. Af­ter just three days of train­ing, it had sur­passed the ver­sion of Al­phaGo that beat Go leg­end Lee Se-dol in 2016, and by

21 days, it had reached the level of Al­phaGo Master, the ver­sion of Go that beat the num­ber one ranked player in the world, Ke Jie, ear­lier this year. By day 40, Al­phaGo Zero had sur­passed all pre­vi­ous ver­sions of Al­phaGo and was able to beat Al­phaGo Master 100 games to 0.

Al­phaGo Zero is able to do this by us­ing a novel form of re­in­force­ment learn­ing. The sys­tem starts off with a neu­ral net­work that knows noth­ing ex­cept the rules of Go. It then plays games against it­self, by com­bin­ing this neu­ral net­work with a pow­er­ful search al­go­rithm. As it plays, the neu­ral net­work is tuned and up­dated to pre­dict moves, as well as the even­tual win­ner of the games. This up­dated neu­ral net­work is then re­com­bined with the search al­go­rithm to cre­ate a new, stronger ver­sion of Al­phaGo Zero, and the process be­gins again. This tech­nique is more pow­er­ful than pre­vi­ous ver­sions of Al­phaGo be­cause it is no longer con­strained by the lim­its of hu­man knowl­edge. In­stead, it is able to learn from the strong­est player in the world: Al­phaGo it­self.

Di­ag­nose dis­eases bet­ter than hu­man doctors

Ac­cord­ing to Google, di­a­betic retinopa­thy is the fastest grow­ing cause of blind­ness, with nearly 415 mil­lion di­a­betic pa­tients at risk world­wide. If caught early, the dis­ease can be treated; if not, it can lead to ir­re­versible blind­ness. Un­for­tu­nately, doctors ca­pa­ble of de­tect­ing the dis­ease are not avail­able in many parts of the world, such as In­dia, where di­a­betes is preva­lent.

How­ever, Google may have a so­lu­tion. One of the most com­mon ways to de­tect di­a­betic eye dis­ease is to have a spe­cial­ist ex­am­ine pic­tures of the back of the eye and rate them for dis­ease pres­ence and sever­ity. Sever­ity is de­ter­mined by the type of le­sions present, which are in­dica­tive of bleed­ing and fluid leak­age in the eye. Work­ing closely with doctors both in In­dia and the U.S., Google cre­ated a de­vel­op­ment data set of 128,000 images that were each eval­u­ated by three to seven oph­thal­mol­o­gists from a panel of 54 oph­thal­mol­o­gists. This dataset was used to train an AI neu­ral net­work to de­tect di­a­betic retinopa­thy.

The re­search re­sults were very promis­ing, with the AI’s per­for­mance on-par with that of trained oph­thal­mol­o­gists. In fact, Google’s AI had an F-score (com­bined sen­si­tiv­ity and speci­ficity met­ric, with max=1) of 0.95, which was ac­tu­ally slightly bet­ter than the me­dian F-score of the hu­man oph­thal­mol­o­gists (mea­sured at 0.91).

Cre­ate an AI that makes other AIs

Au­toML is an AI Google has cre­ated to help it cre­ate other AIs. In Google’s words, “What if we could au­to­mate the process of ma­chine learn­ing?”

Cur­rently, most AIs are built around neu­ral net­works that sim­u­late the way the hu­man brain works in or­der to learn. Neu­ral net­works can be trained to rec­og­nize pat­terns in in­for­ma­tion, such as speech, text, and vis­ual images. But to train them re­quires large data sets and hu­man in­put to make sure the AI is learn­ing cor­rectly.

How­ever, Google has re­vealed that not only has Au­toML suc­cess­fully cre­ated its own neu­ral net­work AI with­out hu­man in­put, but it’s also vastly more pow­er­ful and ef­fi­cient than the top per­form­ing hu­man-de­signed sys­tems.

In a com­par­i­son of neu­ral net­works built for im­age recog­ni­tion, Au­toML’s im­age recog­ni­tion AI man­aged an in­cred­i­ble 82-per­cent ac­cu­racy, which is higher than any hu­man-de­signed AI yet.

Google has re­vealed that not only has Au­toML suc­cess­fully cre­ated its own neu­ral net­work AI with­out hu­man in­put, but it’s also vastly more pow­er­ful...

Newspapers in English

Newspapers from Malaysia

© PressReader. All rights reserved.