Inethiwekhi ye-neural ye-Nvidia ishintsha imidwebo elula ibe yizindawo ezinhle

Impophoma yomuntu obhemayo kanye nempophoma yomuntu onempilo

Sonke siyazi ukudweba isikhova. Okokuqala udinga ukudweba okusaqanda, bese kuba omunye umbuthano, bese uthola isikhova langa. Yiqiniso, lokhu kuyihlaya, futhi endala kakhulu, kodwa Nvidia onjiniyela bazama ukwenza iphupho libe ngokoqobo.

Intuthuko entsha, ebizwa ngokuthi i-GauGAN, idala izindawo ezinhle zomhlaba ezivela emidwebeni elula kakhulu (ilula ngempela - imibuthano, imigqa nakho konke). Kunjalo, lokhu kuthuthukiswa kusekelwe kubuchwepheshe besimanje - okungukuthi amanethiwekhi e-neural adversarial akhiqizayo.

I-GauGAN ikuvumela ukuthi udale imihlaba ebonakalayo enemibala - hhayi ezokuzijabulisa kuphela, kodwa nomsebenzi. Ngakho-ke, abaklami bezakhiwo, abaklami bendawo, abathuthukisi begeyimu - bonke bangafunda okuthile okuwusizo. I-Artificial intelligence ngokushesha "iyaqonda" ukuthi yini umuntu ayifunayo futhi igcwalise umqondo wangempela ngenani elikhulu lemininingwane.

"Ukucabanga ngomklamo kulula kakhulu usebenzisa i-GauGAN ngoba ibhulashi elihlakaniphile lingathuthukisa umdwebo wangempela ngezithombe ezisezingeni eliphezulu," kusho omunye wabathuthukisi be-GauGAN.

Abasebenzisi baleli thuluzi bangashintsha umbono wangempela, balungise isimo sezwe noma esinye isithombe, bengeze isibhakabhaka, isihlabathi, ulwandle, njll. Noma ngabe yini inhliziyo yakho ekufisayo, futhi ukungeza kuthatha imizuzwana embalwa.

Inethiwekhi ye-neural yaqeqeshwa kusetshenziswa isizindalwazi sezigidi zezithombe. Ngenxa yalokhu, uhlelo lungakwazi ukuqonda ukuthi yini umuntu ayifunayo nendlela yokufeza lokho akufunayo. Ngaphezu kwalokho, inethiwekhi ye-neural ayikhohlwa ngemininingwane emincane kakhulu. Ngakho-ke, uma udweba umdwebo womdwebo wechibi kanye nezihlahla ezithile eduze kwalo, khona-ke ngemva kokuba indawo iphile, zonke izinto eziseduze zizobonakala esibukweni samanzi echibini.

Uhlelo lungatshelwa ukuthi indawo ebonakalayo kufanele ibe yini - ingambozwa ngotshani, iqhwa, amanzi noma isihlabathi. Konke lokhu kungashintshwa ngomzuzwana, ukuze iqhwa libe isihlabathi futhi esikhundleni sezwe eligcwele iqhwa, umdwebi uthola indawo ewugwadule.

Kufana nencwadi enemibala ekutshela ukuthi usibeke kuphi isihlahla, ukuthi ilanga likuphi, nokuthi isibhakabhaka sikuphi. Bese, ngemva komsebenzi wokuqala, inethiwekhi ye-neural iphila isithombe, yengeze imininingwane edingekayo kanye nokuthungwa, futhi idwebe ukubonakaliswa. “Konke lokhu kusekelwe ezithombeni zangempela,” kusho omunye wabathuthukisi.

Nakuba uhlelo lungenakho "ukuqonda" komhlaba wangempela, uhlelo lukhiqiza izindawo ezihlaba umxhwele. Lokhu kungenxa yokuthi amanethiwekhi amabili e-neural asetshenziswa lapha, ijeneretha kanye ne-criminator. Ijeneretha idala isithombe bese isikhombisa kumbandlululi. Yena, ngokusekelwe ezigidini zezithombe ezibonwe ngaphambilini, ukhetha izinketho ezingokoqobo kakhulu.

Yingakho ijeneretha “yazi” lapho ukuboniswa kufanele kube khona. Kuyaphawuleka ukuthi ithuluzi liguquguquka kakhulu futhi lifakwe inombolo enkulu yezilungiselelo. Ngakho-ke, ngosizo lwayo, ungapenda izithombe, uvumelane nesitayela somculi othile, noma uvele udabule ngokungeza ngokushesha ukuphuma noma ukushona kwelanga.

Abathuthukisi bathi isistimu ayivele ithathe izithombe kwenye indawo, ingeze ndawonye bese ithola umphumela. Cha, zonke "izithombe" eziwumphumela ziyakhiqizwa. Okusho ukuthi, inethiwekhi ye-neural "idala" njengomculi wangempela (noma okungcono nakakhulu).

Okwamanje, uhlelo alutholakali mahhala, kodwa maduze luzokwazi ukuluzama. Lokhu kungenziwa ku-GPU Technology Conference 2019, eyenzekayo njengamanje eCalifornia. Abanenhlanhla abakwazi ukuvakashela umbukiso sebengakwazi kakade ukuhlola i-GauGAN.

Amanethiwekhi e-Neural sekuyisikhathi eside efundiswa ukubamba iqhaza enqubweni yokudala. Ngokwesibonelo, ngonyaka odlule, abanye babo ingakha amamodeli e-3D. Ngaphezu kwalokho, abathuthukisi abavela ku-DeepMind baqeqeshe inethiwekhi ye-neural ukuze yakhe kabusha izikhala ezinezinhlangothi ezintathu nezinto ezivela emidwebo, izithombe, nemidwebo. Ukuze udale kabusha umfanekiso olula, inethiwekhi ye-neural idinga isithombe esisodwa; ukwakha izinto eziyinkimbinkimbi kakhulu, izithombe ezinhlanu ziyadingeka ukuze "ukuqeqeshwa".

Ngokuqondene ne-GauGAN, leli thuluzi lizothola ngokucacile isicelo esifanelekayo sokuhweba - izindawo eziningi zebhizinisi nesayensi zinesidingo salezo zinsizakalo.

Source: www.habr.com