Iisekethe ezidibeneyo ze-Photonic, okanye ii-chips optical, ezinokubonelela ngeenzuzo ezininzi ngaphezu koogxa babo be-elektroniki, njengokunciphisa ukusetyenziswa kwamandla kunye nokunciphisa latency ekubaleni. Yiyo loo nto abaphandi abaninzi bekholelwa ukuba banokusebenza ngokugqithisileyo ekufundeni koomatshini kunye nemisebenzi yobukrelekrele bokwenziwa (AI). I-Intel ikwabona amathemba amahle okusetyenziswa kwe-silicon photonics kweli cala. Iqela lakhe lophando kwi ubuchule obutsha obunenkcukacha ezinokuzisa uthungelwano lwe-optical neural inyathelo elisondele kwinyani.
Kutshanje , ezinikele ekufundeni komatshini, ichaza indlela uphando kwintsimi ye-optical neural networks yaqala ngayo. Uphando lukaDavid AB Miller kunye noMichael Reck lubonise ukuba uhlobo lwesekethe ye-photonic eyaziwa ngokuba yi-interferometer ye-Mach-Zehnder (MZI) inokuqwalaselwa ukuba yenze i-2 Γ 2 ukuphindaphinda kwe-matrix xa ibekwe i-MZI kwi-mesh engunxantathu yokuphindaphinda iimatriki ezinkulu, enye inako. fumana isekethe esebenzisa i-algorithm yokuphindaphinda i-matrix-vector, isibalo esisisiseko esisetyenziswa ekufundeni komatshini.
Uphando olutsha lwe-Intel lugxininise kwinto eyenzekayo xa iziphene ezahlukahlukeneyo ezinokuthi iitshiphusi ze-optical zichaphazeleke ngexesha lokuvelisa (ekubeni i-photonics yokubala i-analog kwindalo) ibangela ukungafani kokuchaneka kokubala phakathi kwee-chips ezahlukeneyo zohlobo olufanayo. Nangona uphando olufanayo luye lwenziwa, kwixesha elidlulileyo bagxininise ngakumbi kwi-post-fabrication optimization ukuphelisa ukungachaneki okunokwenzeka. Kodwa le ndlela ine-scalability engalunganga njengoko uthungelwano luba lukhulu, okukhokelela ekonyukeni kwamandla ekhompyutheni afunekayo ukuseta uthungelwano lwe-optical network. Endaweni yokulungiswa kwangemva kokwenziwa, i-Intel ithathele ingqalelo iitshiphusi zoqeqesho kanye phambi kokuba kwenziwe ngokusebenzisa uyilo olunyamezela ingxolo. Uthungelwano lwe-optical neural network lwaqeqeshwa kanye, emva koko iiparamitha zoqeqesho zaye zasasazwa kuzo zonke iimeko zothungelwano ezenziweyo ezinomahluko kumacandelo azo.
Iqela le-Intel liqwalasele izakhiwo ezimbini zokwakha iinkqubo zobuntlola ezenziweyo ezisekelwe kwi-MZI: GridNet kunye neFFTNet. I-GridNet ibeka i-MZIs kuthungelwano, ngelixa i-FFTNet izibeka kwiipateni zebhabhathane. Emva kokuqeqeshwa kokubini kukulinganisa kwidijithi ebhalwe ngesandla yokuqaphela umsebenzi onzulu wokulinganisa ukufunda (MNIST), abaphandi bafumanise ukuba iGridNet iphumelele ukuchaneka okuphezulu kuneFFTNet (98% vs. 95%), kodwa uyilo lweFFTNet "lomelele ngakumbi." Enyanisweni, ukusebenza kweGridNet kwehle ngaphantsi kwe-50% ngokongezwa kwengxolo eyenziweyo (uphazamiseko olulinganisa iziphene ezinokwenzeka kukwenziwa kwetshiphu zamehlo), ngelixa kwiFFTNet yahlala iphantse yatshintsha.
Izazinzulu zithi uphando lwazo lubeka isiseko seendlela zoqeqesho zobukrelekrele bokwenziwa ezinokuthi ziphelise imfuneko yokucoca iitships zamehlo emva kokuba zivelisiwe, okonga ixesha elixabisekileyo kunye nezixhobo.
Njengayo nayiphi na inkqubo yokuvelisa, iziphene ezithile ziya kwenzeka ezithetha ukuba kuya kubakho umahluko omncinci phakathi kweetshiphusi eziza kuba nefuthe ekuchanekeni kwezibalo, ubhala uCasimir Wierzynski, umlawuli omkhulu we-Intel AI Product Group. "Ukuba amaqumrhu e-optical neural azakuba yinxalenye ye-AI ye-hardware ecosystem, kuya kufuneka ukuba afudukele kwiitshiphusi ezinkulu kunye nobuchwepheshe bokwenza umzi-mveliso. Uphando lwethu lubonisa ukuba ukukhetha ulwakhiwo olufanelekileyo ngaphambili kunokonyusa kakhulu amathuba okuba iitshiphusi eziziphumo zifezekise umsebenzi ofunekayo, nokuba kukho imveliso eyahlukileyo.
Kwangelo xesha i-Intel ngokuyintloko iqhuba uphando, umgqatswa we-MIT PhD uYichen Shen waseka i-Boston-based startup Lightelligence, enyuse i-10,7 yezigidi zeedola kwinkxaso-mali kunye iprototype ye-optical chip yokufunda koomatshini ekhawuleza kangangezihlandlo ezili-100 kuneetshiphusi ze-elektroniki zanamhlanje kwaye ikwanciphisa ukusetyenziswa kwamandla ngomyalelo wobukhulu, obuphinda bubonise ngokucacileyo isithembiso sobuchwepheshe befotonic.
umthombo: 3dnews.ru