Ubuhlakani bokwenziwa abulandeli umqondo wenqubekelaphambili yesayensi
Sibhale izikhathi eziningi ku-MT mayelana nabacwaningi nezingcweti ezimemezela izinhlelo zokufunda zomshini “njengamabhokisi amnyama” (1) ngisho nakulabo abawakhayo. Lokhu kwenza kube nzima ukuhlola imiphumela nokusebenzisa kabusha ama-algorithms avelayo.
Amanethiwekhi e-Neural - inqubo esinikeza ama-bots ahlakaniphile aguqulayo kanye namajeneretha ombhalo ahlakaniphile angakha ngisho nezinkondlo - ahlala eyimfihlakalo engaqondakali ezibukelini zangaphandle.
Ziya ziba zinkulu futhi ziba yinkimbinkimbi, zibamba amasethi edatha amakhulu, futhi zisebenzisa ama-comute arrays amakhulu. Lokhu kwenza ukuphindaphinda nokuhlaziya amamodeli atholiwe kubize futhi kwesinye isikhathi kungenzeki kwabanye abacwaningi, ngaphandle kwezikhungo ezinkulu ezinesabelomali esikhulu.
Ososayensi abaningi bayazi kahle le nkinga. Phakathi kwabo kukhona uJoel Pino (2), usihlalo we-NeurIPS, ingqungquthela eyinhloko yokukhiqiza kabusha. Ochwepheshe abangaphansi kobuholi bakhe bafuna ukudala "uhlu lokuhlola ukukhiqizwa kabusha".
Umbono, ngokusho kuka-Pino, owokukhuthaza abacwaningi ukuthi banikeze abanye imephu yomgwaqo ukuze bakwazi ukuphinda badale futhi basebenzise umsebenzi osuwenzile. Ungamangala ngobugagu bejeneretha yombhalo entsha noma amandla angaphezu kwawomuntu erobhothi lomdlalo wevidiyo, kodwa ngisho nochwepheshe abangcono kakhulu abazi ukuthi lezi zimangaliso zisebenza kanjani. Ngakho-ke, ukukhiqizwa kabusha kwamamodeli we-AI kubalulekile hhayi kuphela ekuhlonzeni amagoli amasha nezikhombisi-ndlela zocwaningo, kodwa futhi njengomhlahlandlela osebenzayo ongawusebenzisa.
Abanye bazama ukuxazulula le nkinga. Abacwaningi be-Google banikeze "amakhadi emodeli" ukuchaza kabanzi ukuthi amasistimu ahlolwe kanjani, okuhlanganisa imiphumela ekhomba iziphazamisi ezingaba khona. Abacwaningi e-Allen Institute for Artificial Intelligence (AI2) bashicilele iphepha elihlose ukunweba uhlu lokuhlola lokukhiqiza kabusha lwe-Pinot kwezinye izinyathelo zenqubo yokuhlola. “Bonisa umsebenzi wakho,” bayanxusa.
Kwesinye isikhathi ulwazi oluyisisekelo luyashoda ngenxa yokuthi iphrojekthi yocwaningo iphethwe, ikakhulukazi amalabhorethri asebenzela inkampani. Nokho, imvamisa, kuwuphawu lokungakwazi ukuchaza izindlela zocwaningo eziguqukayo neziya ngokuba nzima. Amanethiwekhi e-Neural ayindawo eyinkimbinkimbi kakhulu. Ukuze uthole imiphumela engcono kakhulu, ukushuna kahle kwezinkulungwane “zamafindo nezinkinobho” ngokuvamile kuyadingeka, abanye abakubiza ngokuthi “umlingo omnyama”. Ukukhethwa kwemodeli elungile kuvame ukuhlotshaniswa nenani elikhulu lokuhlola. Umlingo uba eqolo.
Isibonelo, ngenkathi i-Facebook izama ukuphinda umsebenzi we-AlphaGo, uhlelo olwakhiwe yi-DeepMind Alphabet, umsebenzi ubonakale unzima kakhulu. Izidingo ezinkulu zekhompiyutha, izigidi zokuhlolwa ezinkulungwaneni zamadivayisi ezinsukwini eziningi, kuhlanganiswe nokuntuleka kwekhodi, kwenze uhlelo "lube nzima kakhulu, noma lungenzeki, ukuphinda ludale, luvivinye, luthuthuke futhi lunwebe," ngokusho kwabasebenzi be-Facebook.
Inkinga ibonakala ikhethekile. Kodwa-ke, uma sicabanga ngokuqhubekayo, ukwenzeka kwezinkinga zokuphindaphindeka kwemiphumela nemisebenzi phakathi kwethimba elilodwa locwaningo nelinye kubukela phansi yonke indlela enengqondo yokusebenza kwesayensi nezinqubo zocwaningo ezaziwa yithi. Njengomthetho, imiphumela yocwaningo lwangaphambilini ingasetshenziswa njengesisekelo socwaningo olwengeziwe olukhuthaza ukuthuthukiswa kolwazi, ubuchwepheshe kanye nenqubekelaphambili jikelele.
