Amasekhethi ahlanganisiwe ezithombe, noma ama-optical chips, anamandla okunikeza izinzuzo eziningi ngaphezu kozakwabo be-electronic, njengokusebenzisa amandla okuphansi kanye nokubambezeleka okuncishisiwe kukhompyutha. Kungakho abacwaningi abaningi bekholelwa ukuthi bangaphumelela ngokwedlulele emisebenzini yokufunda komshini nokudala ubuhlakani bokwenziwa (AI). I-Intel futhi ibona amathemba amahle okusetshenziswa kwe-silicon photonics kule ndawo. Iqembu labacwaningi ku kuchazwe ngokuningiliziwe izindlela ezintsha ezingaletha amanethiwekhi okusebenza kwe-optical neural isinyathelo esisodwa eduze neqiniso.
Kwakamuva , enikelwe ekufundeni komshini, itshela ukuthi ucwaningo emkhakheni wamanethiwekhi we-optical neural lwaqala kanjani. Umsebenzi wesayensi ka-David AB Miller no-Michael Reck ubonise ukuthi uhlobo lwesekhethi ye-photonic eyaziwa ngokuthi i-Mach-Zehnder interferometer (MZI) ingalungiselelwa ukwenza ukuphindaphinda kwe-matrix engu-2 Γ 2, lapho, uma ibekwe i-MZI kugridi engunxantathu ukuze iphindaphindeke. omatikuletsheni amakhulu, ungathola isekethe esebenzisa i-algorithm yokuphindaphinda ye-matrix-vector - ukubala okuyinhloko okusetshenziswe ekufundeni komshini.
Ucwaningo olusha lwe-Intel lugxile kulokho okwenzekayo lapho ukukhubazeka okuhlukahlukene okuvezwa ama-optical chips ngesikhathi sokukhiqiza (ngoba i-photonics yekhompyutha iyi-analogi ngokwemvelo) kubangela umehluko ekunembeni kwekhompyutha phakathi kwama-chips ahlukene ohlobo olufanayo. Nakuba izifundo ezifanayo sezenziwe kakade, esikhathini esidlule bezigxile kakhulu ekuthuthukisweni kwangemva kokukhiqiza ukuze kuqedwe ukunemba okungenzeka. Kodwa le ndlela inokulinganisa okungekuhle njengoba amanethiwekhi eba makhudlwana futhi abe makhudlwana, okuholela emandleni okucubungula engeziwe adingekayo ukuze kumiswe amanethiwekhi okukhanya. Esikhundleni sokwenza kahle kwangemva kokukhiqiza, i-Intel ibheke ukuthi kungenzeka yini ukuqeqeshwa ngesikhathi esisodwa kwama-chips ngaphambi kokwenziwa kusetshenziswa isakhiwo esibekezelela umsindo. Inethiwekhi ye-optical neural eyireferensi yaqeqeshwa kanye, okwathi ngemva kwalokho imingcele yokuqeqeshwa yasatshalaliswa ezenzakalweni ezimbalwa zenethiwekhi eziqanjiwe ezinomehluko ezingxenyeni zazo.
Ithimba le-Intel licubungule izakhiwo ezimbili zokwakha amasistimu obuhlakani bokwenziwa asuselwe ku-MZI: GridNet kanye ne-FFTNet. I-GridNet ibeka ama-MZI kugridi ebikezelwayo, kuyilapho i-FFTNet iwabeka kuphethini yovemvane. Ngemva kokuqeqeshwa kokubili ekwenzeni amadijithi okuqashelwa okujulile kwenkomba yomsebenzi (MNIST), abacwaningi bathole ukuthi i-GridNet ithole ukunemba okuphakeme kune-FFTNet (98% uma iqhathaniswa nama-95%), kodwa ukwakheka kwe-FFTNet "kwakuqine kakhulu." Eqinisweni, ukusebenza kwe-GridNet kwehle ngaphansi kuka-50% ngokungezwa komsindo wokwenziwa (ukuphazamiseka okulingisa ukukhubazeka okungenzeka ekukhiqizeni ama-optical chips), kuyilapho ku-FFTNet kwahlala kucishe kube njalo.
Ososayensi bathi ucwaningo lwabo lubeka isisekelo sezindlela zokuqeqesha ze-AI ezingaqeda isidingo sokulungisa kahle ama-chip optical ngemva kokwakhiwa, okonga isikhathi nezinsiza ezibalulekile.
βNjenganoma iyiphi inqubo yokukhiqiza, kuzokwenzeka amaphutha athile, okusho ukuthi kuzoba nomehluko omncane phakathi kwama-chips, futhi lokhu kuzothinta ukunemba kwezibalo,β kubhala uCasimir Wierzynski, umqondisi omkhulu we-Intel AI Product Group. "Uma ama-optical neural essences eba yingxenye esebenzayo ye-AI ββhardware ecosystem, kuzodingeka athuthele kuma-chips amakhulu kanye nobuchwepheshe bokukhiqiza izimboni. Ucwaningo lwethu lukhombisa ukuthi ukukhetha i-architecture efanele ngaphambili kungakhuphula kakhulu amathuba okuthi ama-chips azoba umphumela athole ukusebenza okufunayo, noma kukhona ukuhlukahluka kokukhiqiza. "
Ngenkathi i-Intel ngokuyinhloko yenza ucwaningo, i-MIT Ph.D. Yichen Shen yasungula i-Lightelligence esekelwe eBoston, ekhulise amaRandi ayizigidi ezingu-10,7 ngoxhaso lwemali yebhizinisi kanye i-prototype optical chip yokufunda ngomshini, eshesha ngokuphindwe izikhathi eziyi-100 kunama-chip esimanje kagesi, futhi inciphisa ukusetshenziswa kwamandla ngokuhleleka kobukhulu, okuphinde kubonise ngokucacile isithembiso sobuchwepheshe bezithombe.
Source: 3dnews.ru