Google’s AI to save lives

Re­searcher David Sil­ver says health­care is the next fron­tier for Deep­Mind. Kather­ine Noyes re­ports

Tech Advisor - - Contents -

With more pos­si­ble board com­bi­na­tions than there are atoms in the uni­verse, Go has long been con­sid­ered the ul­ti­mate chal­lenge for AI re­searchers

Al­phaGo’s un­canny suc­cess at the game of Go was taken by many as a death knell for the dom­i­nance of the hu­man in­tel­lect, but Google re­searcher David Sil­ver doesn’t see it that way. In­stead, he sees a world of po­ten­tial ben­e­fits.

As one of the lead ar­chi­tects be­hind Google Deep­Mind’s Al­phaGo sys­tem, which de­feated South Korean Go champion Lee Se-dol 4 games to 1 in March (pic­tured above), Sil­ver be­lieves the tech­nol­ogy’s next role should be to help ad­vance hu­man health.

“We’d like to use these tech­nolo­gies to have a pos­i­tive im­pact in the real world,” he told an au­di­ence of AI re­searchers at the In­ter­na­tional Joint Con­fer­ence on Artificial In­tel­li­gence in New York.

With more pos­si­ble board com­bi­na­tions than there are atoms in the uni­verse, Go has long been con­sid­ered the ul­ti­mate chal­lenge for AI re­searchers. Al­phaGo was trained first on ex­pert hu­man moves, then on mil­lions of games of self-play. In its vic­tory against Se-dol, its moves were de­scribed by ex­perts as “cre­ative” in that they ob­vi­ously didn’t de­rive strictly from its train­ing ma­te­ri­als. Now, Deep­Mind is ap­ply­ing Al­phaGo’s deep-learn­ing smarts to ap­pli­ca­tions that in­clude health an­a­lyt­ics and health as­sis­tants for de­liv­er­ing per­son­alised medicine, Sil­ver said.

Ear­lier this year, UK-based Deep­Mind launched a health di­vi­sion. In July, it an­nounced a re­search part­ner­ship with Moor­fields Eye Hospi­tal that will fo­cus on ap­ply­ing ma­chine learn­ing to di­a­betic retinopa­thy and age-re­lated mac­u­lar de­gen­er­a­tion. It’s also been work­ing on sev­eral clin­i­cal mo­bile apps.

Many of the AI ad­vances that al­lowed Al­phaGo to achieve the suc­cess it did in Go could also help it ex­cel in health­care. Rather than find the best moves through sheer com­pu­ta­tional force the way IBM’s Deep­Blue did to beat chess champion Gary Kas­parov, Al­phaGo’s ap­proach is based on con­vo­lu­tional neu­ral net­works and re­in­force­ment learn­ing, al­low­ing it to essentiall­y teach it­self over time.

“Of course beat­ing Se-dol was ex­cit­ing, but for me, even more ex­cit­ing than the achieve­ment it­self was the man­ner in which Al­phaGo did it,” Sil­ver said. “It showed it can learn from data and self-play to fig­ure out knowl­edge for it­self.”

Re­in­force­ment learn­ing has long been con­sid­ered “a nice pipe dream,” he added. “Now it feels like these meth­ods re­ally work. That’s a real change. We can now look around at many, many dif­fer­ent do­mains. We’re by no means done with Al­phaGo.”

Al­phaGo played Lee Se-dol ear­lier this year

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.