UGoogle AI uqikelela imozulu

UGoogle AI uqikelela imozulu

Uqikelelo lwemozulu lwanamhlanje lusekelwe kwiimodeli ezintsonkothileyo ezibandakanya imithetho elawula ukushukuma kweatmosfera kunye neelwandle, kwaye le mizekelo isebenza kwezinye zezona khompyutha zinamandla zikhoyo. Nangona kunjalo, i-Alphabet (inkampani engumzali kaGoogle) ikwazile ukuqikelela iimeko zemozulu zehlabathi kwiintsuku ezili-10 ezizayo ngomzuzu nje omnye isebenzisa umatshini omnye olingana nekhompyuter yakho, ngenxa yobukrelekrele bokwenziwa obuphuhliswe yi-DeepMind. I UGoogle AI uqikelela imozulu kwaye oku kusandula ukuqalisa.

Kweli nqaku siza kukuxelela indlela iGoogle AI eqikelela ngayo imozulu kunye nendlela obu buchwepheshe buvele ngayo.

UGoogle AI uqikelela imozulu

imodeli yoqikelelo lwemozulu

Okumangalisayo kukuba, le nkqubo ye-AI igqwesa uninzi lweenkqubo zokuqikelela imozulu yanamhlanje phantse kuzo zonke iinkalo. Okubangel’ umdla kukuba, kubonakala ukuba ngeli xesha ubukrelekrele bokwenziwa busebenza njengokuphelelisa ubukrelekrele bomntu kunokuba bubuthabathel’ indawo.

I-European Centre for Medium-Range Weather Forecasts (ECMWF) inesistim eqhubela phambili ngokumangalisayo eye yaphuculwa kakhulu kunyaka ophelileyo, iphucula amandla ayo okuxela kwangaphambili. Ibanjwe kwiindawo zayo e-Bologna, eItali, Kukho ikhompyutha enkulu exhotyiswe malunga neeprosesa ezisisigidi (ngokuchaseneyo nezimbini okanye ezine ezifunyenwe kwikhompyuter yakho) kunye namandla angaqhelekanga ekhompyuter angama-30 epetaflops, alingana nezibalo ezimangalisayo ezingama-30.000 zetriliyoni ngomzuzwana.

Esi sixhobo sikhulu sokubala siyimfuneko kwesinye sezixhobo zayo, i-High Resolution Forecasting (HRES), eqikelela ngokuchanekileyo iipatheni zemozulu yehlabathi yexesha eliphakathi, ethi Ngokuqhelekileyo bathatha iintsuku ezili-10, kunye nesisombululo esichukumisayo sesithuba seekhilomitha ezilithoba. Ezi ngqikelelo zisebenza njengesiseko soqikelelo lwemozulu olunikelwa zizazi ngemozulu kwihlabathi jikelele. Kutshanje, iGraphCast, ubukrelekrele bokwenziwa obuphuhliswe nguGoogle DeepMind, busetyenziselwe ukulinganisa amandla ale nkqubo yoyikekayo yokuqikelela imozulu.

Iziphumo zezifundo ze-AI

igraphcast

Iziphumo zokuthelekisa, ezipapashwe ngoLwesibini kwiphephancwadi iSayensi, zityhila ukuba iGraphCast igqwesa i-HRES ngokuqikelela izinto ezininzi zemozulu. Ngokophononongo, Umatshini kaGoogle ugqwesa iECMWF's kwi-90,3% ye-1.380 yeemetrics ezivavanyiweyo.

Xa kugxilwe kuphela kwi-troposphere, umaleko we-atmospheric apho iziganeko ezininzi zemozulu zenzeka khona, kwaye ngaphandle kwedatha evela kwi-stratosphere, emalunga ne-6 ukuya kwi-8 yeekhilomitha ngaphezu komphezulu woMhlaba, ubukrelekrele bokwenziwa (A.I.) ) bugqwesa ii-supercomputers ezibekwe esweni ngabantu kwi-99,7% iimeko. ezahlukeneyo zicazululwe. Okumangalisayo kukuba, le mpumelelo yaphunyezwa kusetyenziswa umatshini ofana ngokusondeleyo nekhompyutha yomntu eyaziwa ngokuba yiyunithi yokulungisa i-tensor okanye i-TPU.

Ngokuka-Álvaro Sánchez González, umphandi kuGoogle DeepMind, iiTPU zizixhobo ezikhethekileyo ezibonelela ngoqeqesho olusebenzayo kunye nokwenziwa kwesoftware yobukrelekrele bokwenziwa xa kuthelekiswa nePC eqhelekileyo, ngelixa igcina ubungakanani obufanayo. Kanye njengokuba ikhadi lemizobo lekhompyuter ligxile ekunikezeni imifanekiso, iiTPU ziyilelwe ukuba zigqwese kwiimveliso zematrix. Kuqeqesho lweGraphCast, sasebenzisa ii-TPU ezingama-32 ekuhambeni kweeveki ezininzi. Nangona kunjalo, xa uqeqesho lugqityiwe, iTPU enye inokuvelisa uqikelelo ngaphantsi komzuzu, njengoko kuchazwe nguSánchez González, omnye wabadali besixhobo.

Iinkqubo zeGraphCast kunye nokuqikelela

google AI iqikelela imozulu

Ukwahlula okuphawulekayo phakathi kweGraphCast kunye neenkqubo ezikhoyo zokuqikelela ukukwazi ukubandakanya idatha yembali. Abadali baqeqeshe inkqubo besebenzisa idatha yemozulu ephuma kugcino lwe-ECMWF olususela ngo-1979. Olu lwazi lubanzi lubandakanya Imvula eSantiago kunye nezitshingitshane ezichaphazele iAcapulco kwisithuba seminyaka engama-40. Emva koqeqesho oluninzi, iGraphCast inekhono elimangalisayo lokuvelisa uqikelelo lwemozulu oluchanekileyo.

Ifuna kuphela ulwazi lweemeko zemozulu kwiiyure ezintandathu ngaphambi nangemva koqikelelo lwakho ukuqikelela ngokuchanekileyo imozulu kwiiyure ezintandathu ukusuka ngoku. Uqikelelo luxhomekeke kwelinye kwaye uqikelelo olutsha ngalunye lwazisa olwangaphambili. UFerran Alet, umyili oncedisana nalo matshini unomtsalane we-DeepMind, ucacisa ukusebenza kwawo kwangaphakathi: «Inethiwekhi yethu ye-neural ilindele iimeko zemozulu kwiiyure ezintandathu kwangaphambili. Ukuqikelela imozulu kwiiyure ezingama-24, sivavanya nje imodeli amaxesha amane. Kungenjalo, besinokuqeqesha imifuziselo eyahlukeneyo yamaxesha ahlukeneyo, anje ngeeyure ezintandathu enye ngeeyure ezingama-24. Nangona kunjalo, "Siyaqonda ukuba imigaqo esisiseko elawula imozulu ihlala ihambelana kwisithuba seeyure ezintandathu."

"Ngoko ke, ukuba sinokufumanisa imodeli efanelekileyo yeeyure ze-6 kwaye sisebenzise izibikezelo zayo njengegalelo, sinokuyichaza ngokuchanekileyo imozulu kwiiyure ze-12 ezizayo kwaye siphinda le nkqubo rhoqo kwiiyure ezintandathu." Ngokuka-Alet, le ndlela ibonelela ngedatha eninzi yemodeli enye, okukhokelela kuqeqesho olusebenzayo.

Ukuza kuthi ga ngoku, uqikelelo lwemozulu lusekwe kuqikelelo lwemozulu lwamanani, olusebenzisa iiequations zenzululwazi eziphuhliswe kwimbali yonke ukunika ingxelo ngobunzima obahlukeneyo bokuguquguquka kwemozulu. Iziphumo zabaphandi ziseka iseti yee-algorithms zemathematika ukuba ii-supercomputers kufuneka ibaleke ukwenza uqikelelo lweeyure ezimbalwa ezizayo, iintsuku, okanye iiveki (nangona ukuthembeka kuncipha kakhulu ngaphaya kweentsuku ezili-15). Nangona kunjalo, ukwenza lo msebenzi kufuna i-supercomputer ephezulu kakhulu, ebandakanya iindleko ezinkulu kunye nemizamo yobunjineli ebanzi.

Imodeli yeGoogle AI iqikelela imozulu

Eyona nto iphawuleka ngakumbi kukuba ezi nkqubo abazisebenzisi iimeko zemozulu zosuku olungaphambili okanye kunyaka ophelileyo, nangona zisenzeka kwindawo enye nangexesha elinye.

Ngokuchasene noko, isondela kumsebenzi kwi-angle eyahlukileyo, phantse ngokuchaseneyo. Ngesakhono sayo sokufunda nzulu esiphucukileyo, isebenzisa uvimba obanzi wedatha yemozulu yangaphambili ukuze ifumane ulwazi olubanzi lwendlela entsonkothileyo yesizathu kunye nesiphumo esilawula ukuqhubeka kwemozulu yoMhlaba.

Ngokutsho kukaJosé Luis Casado, umkhulumeli we-Meteorological Agency yaseSpain (AEMET), idatha yembali ayithathelwa ingqalelo kwimodeli ye-atmospheric. I-Casado icacisa ukuba le modeli isekelwe kwimigqaliselo ekhoyo kunye noqikelelo lwamva nje olwenziwe yimodeli ngokwayo. Ngokuyiqonda ngokuchanekileyo imeko yangoku ye-atmosfera, kunokwenzeka ukuqikelela ukuqhubela phambili kwayo kwixesha elizayo. Ngokungafaniyo neendlela zokufunda koomatshini, le ndlela ayisebenzisi idatha yembali okanye izibikezelo.

Ndiyathemba ukuba ngolu lwazi unokufunda ngakumbi malunga neGoogle ye-AI eqikelela imozulu kunye neempawu zayo.


Shiya uluvo lwakho

Idilesi yakho ye email aziyi kupapashwa. ezidingekayo ziphawulwe *

*

*

  1. Uxanduva lwedatha: UMiguel Ángel Gatón
  2. Injongo yedatha: Ulawulo lwe-SPAM, ulawulo lwezimvo.
  3. Umthetho: Imvume yakho
  4. Unxibelelwano lwedatha: Idatha ayizukuhanjiswa kubantu besithathu ngaphandle koxanduva lomthetho.
  5. Ukugcinwa kweenkcukacha
  6. Amalungelo: Ngalo naliphi na ixesha unganciphisa, uphinde uphinde ucime ulwazi lwakho.