Abacwaningi beMicrosoft benze uhlelo lobuhlakani bokwenziwa olukwazi ukukhiqiza amagama-ncazo wezithombe, ezimweni eziningi, avele anembe kakhulu kunezincazelo ezenziwe abantu. Lokhu kuphumelela kuphawule ingqopha-mlando ekuzibophezeleni kweMicrosoft ekwenzeni imikhiqizo namasevisi ayo ahlanganiswe futhi afinyeleleke kubo bonke abasebenzisi.
"Incazelo yesithombe ingenye yemisebenzi eyinhloko yombono wekhompiyutha, okwenza izinsizakalo eziningi zenzeke," kusho u-Xuedong Huang (), Isikhulu Sezobuchwepheshe se-Microsoft kanye ne-CTO ye-Azure AI Cognitive Services e-Redmond, eWashington.
Imodeli entsha isiyatholakala kubathengi ngeComputer Vision ku , eyingxenye ye-Azure AI, futhi ivumela onjiniyela ukuthi basebenzise lesi sici ukuze bathuthukise ukutholakala kwezinsizakalo zabo. Iphinde ifakwe ohlelweni lokusebenza lwe-Seing AI futhi izotholakala ngokuhamba kwesikhathi kulo nyaka ku-Microsoft Word ne-Outlook ye-Windows ne-Mac, kanye ne-PowerPoint ye-Windows, Mac kanye nakuwebhu.
Ukuchazwa Okuzenzakalelayo kusiza abasebenzisi ukuthi bafinyelele okuqukethwe okubalulekile kwanoma yisiphi isithombe, kungakhathaliseki ukuthi isithombe esibuyiselwe emiphumeleni yosesho noma umdwebo wokwethulwa.
"Ukusetshenziswa kwamagama-ncazo achaza okuqukethwe kwezithombe (okuthiwa omunye umbhalo noma omunye umbhalo) emakhasini ewebhu kanye nemibhalo kubaluleke kakhulu kubantu abayizimpumputhe noma abangaboni kahle," kusho uSaqib Sheikh (), Umphathi Wesoftware e-Microsoft's AI Platform Group e-Redmond.
Isibonelo, ithimba lakhe lisebenzisa isici esithuthukisiwe sencazelo yesithombe kuhlelo lokusebenza lwabantu abayizimpumputhe nabangaboni kahle. , ebona lokho ikhamera ethwebulayo bese ikutshela ngakho. Uhlelo lokusebenza lusebenzisa amagama-ncazo akhiqiziwe ukuchaza izithombe, okuhlanganisa nenkundla yezokuxhumana.
“Ngokufanelekile, wonke umuntu kufanele engeze umbhalo ohlukile kuzo zonke izithombe kumadokhumenti, kuwebhu, ezinkundleni zokuxhumana, njengoba lokhu kuvumela abantu abangaboni ukuthi bafinyelele okuqukethwe futhi babambe iqhaza engxoxweni. Kodwa, maye, abantu abakwenzi lokhu, ”kusho uSheikh. "Kodwa-ke, kunezinhlelo zokusebenza ezimbalwa ezisebenzisa isici sencazelo yesithombe ukwengeza omunye umbhalo uma ungekho."
U-Liruan Wang, umphathi jikelele wocwaningo e-Microsoft's Redmond Lab, uhole ithimba labacwaningi elazuza futhi ladlula imiphumela yabantu. Isithombe: Dan DeLong.
Incazelo yezinto ezintsha
"Ukuchaza izithombe kungomunye wemisebenzi eyinhloko yombono wekhompiyutha, okudinga uhlelo lobuhlakani bokwenziwa ukuze luqonde futhi luchaze okuqukethwe okuyinhloko noma isenzo esivezwe esithombeni," kuchaza uLiruan Wang (), umphathi jikelele wocwaningo elebhu yakwa-Microsoft Redmond.
“Kufanele uqonde ukuthi kwenzekani, uthole ukuthi buyini ubudlelwano phakathi kwezinto nezenzo, bese ufingqa futhi ukuchaze konke ngomusho ngolimi olufundeka ngabantu,” kusho yena.
U-Wang uhole ithimba labacwaningi, ekulinganiseni izilinganiso (amagama-ncazo wento inoveli esikalini, incazelo enkulu yezinto ezintsha) uthole umphumela ofana nowomuntu, futhi wawudlula. Lokhu kuhlola kukuvumela ukuthi uhlole ukuthi amasistimu e-AI akhiqiza kahle kangakanani izincazelo zezinto ezibonisiwe ezingafakiwe kusethi yedatha lapho imodeli yaqeqeshwa khona.
Ngokuvamile, amasistimu encazelo yesithombe aqeqeshwa kumasethi edatha aqukethe izithombe ezihambisana nencazelo yombhalo yalezi zithombe, okungukuthi, kumasethi ezithombe ezisayiniwe.
"Ukuhlolwa kwe-nocaps kubonisa ukuthi uhlelo lukwazi kanjani ukuchaza izinto ezintsha ezingatholakali kudatha yokuqeqeshwa," kusho u-Wang.
Ukuze kuxazululwe le nkinga, ithimba le-Microsoft liqeqeshe kusengaphambili imodeli ye-AI enkulu kudathasethi enkulu equkethe izithombe ezimakiwe ngamagama, ngayinye ehlotshaniswa nento ethile esithombeni.
Amasethi ezithombe anamathegi egama esikhundleni samagama-ncazo aphelele asebenza kahle kakhulu ukuze adalwe, okuvumela ithimba lika-Wang ukuthi lifake idatha eningi kumodeli yalo. Le ndlela yokwenza inikeze imodeli lokho iqembu elikubiza ngokuthi silulumagama esibonakalayo.
Njengoba u-Huang echaza, indlela yokufundisa ngaphambi kokufunda ngokusebenzisa isilulumagama esibukwayo ifana nokulungiselela izingane ukufunda: okokuqala, incwadi yezithombe isetshenziswa lapho amagama ngamanye ahlotshaniswa nezithombe, isibonelo, ngaphansi kwesithombe se-apula kubhalwe ukuthi "i-apula" futhi ngaphansi kwesithombe sekati kunegama elithi "ikati".
“Lokhu kuqeqeshwa kwangaphambili ngesilulumagama esibonwayo, empeleni, kuyimfundo yokuqala edingekayo ukuqeqesha lolu hlelo. Yile ndlela esizama ngayo ukuthuthukisa uhlobo lwenkumbulo yemoto, ”kusho uHuang.
Imodeli eqeqeshwe ngaphambilini ibe isicwengwa ngedathasethi okuhlanganisa nezithombe ezinelebula. Kulesi sigaba sokuqeqeshwa, imodeli ifunda ukwenza imisho. Uma isithombe esiqukethe izinto ezintsha sivela, isistimu ye-AI isebenzisa isichazamazwi esibonakalayo ukuze idale izincazelo ezinembile.
"Ukuze usebenze ngezinto ezintsha ngesikhathi sokuhlolwa, uhlelo luhlanganisa lokho olukufundile ngesikhathi sokuqeqeshwa kwangaphambili nangesikhathi sokulungiswa okulandelayo," kusho u-Wang.
Ngokwemiphumela , lapho ihlolwa ezivivinyweni ze-nocaps, uhlelo lwe-AI lukhiqize izincazelo ezizwakalayo nezinembe kakhulu kunabantu abazenzela izithombe ezifanayo.
Ushintsho olusheshayo oluya endaweni yokusebenza
Phakathi kwezinye izinto, uhlelo olusha lokuchaza izithombe luhle ngokuphindwe kabili kunemodeli esetshenziswe emikhiqizweni namasevisi e-Microsoft kusukela ngo-2015, uma luqhathaniswa nelinye ibhentshimakhi yemboni.
Ngokucabangela izinzuzo ezizotholwa yibo bonke abasebenzisi bemikhiqizo namasevisi e-Microsoft kulokhu kuthuthukiswa, u-Huang washeshisa ukuhlanganiswa kwemodeli entsha endaweni yokusebenza ye-Azure.
"Sithatha lobu buchwepheshe be-AI obuphazamisayo siye e-Azure njengenkundla yokusebenzela uhla olubanzi lwamakhasimende," esho. “Futhi lokhu akulona nje impumelelo ocwaningweni. Isikhathi esisithathile ukufaka le mpumelelo endaweni yokukhiqiza i-Azure nakho kwaba impumelelo. ”
U-Huang wengeze ngokuthi ukuzuza imiphumela efana neyomuntu kuyaqhubeka nokuthambekela osekuvele kusungulwe ezinhlelweni zobuhlakani bokuqonda ze-Microsoft.
“Kule minyaka emihlanu edlule, sithole imiphumela efana neyomuntu emikhakheni emihlanu emikhulu: ekuqapheliseni inkulumo, ekuhumusheni ngomshini, ekuphenduleni imibuzo, ekufundeni ngomshini nasekuqondeni umbhalo, nango-2020, naphezu kwe-COVID-19, encazelweni yesithombe. ' kusho uJuan.
Kulesi sihloko
- Vakashela isayithi ukuze ufunde kabanzi mayelana nombono wekhompyutha ()
- Funda isihloko
Qhathanisa imiphumela yencazelo yezithombe uhlelo olunikeze ngaphambili namanje lisebenzisa i-AI
Isithombe sihlonishwa yi-Getty Images. Incazelo yangaphambilini: Ukusondela kwendoda elungisa i-hot dog ebhodini lokusika. Incazelo entsha: Umuntu wenza isinkwa.
Isithombe sihlonishwa yi-Getty Images. Incazelo edlule: Indoda ihlezi ekushoneni kwelanga. Incazelo entsha: I-Bonfire ogwini.
Isithombe sihlonishwa yi-Getty Images. Incazelo yangaphambilini: Indoda egqoke ihembe eliluhlaza okwesibhakabhaka. Incazelo entsha: Abantu abambalwa abagqoke imaski yokuhlinzwa.
Isithombe sihlonishwa yi-Getty Images. Incazelo yangaphambilini: Indoda ekwi-skateboard indizela phezulu odongeni. Incazelo entsha: Umdlali we-baseball ubamba ibhola.
Source: www.habr.com