IGoogle namuhla ikhiphe imodeli ye-TimesFM-3, okuwumkhiqizo wesizukulwane sesithathu ochungechungeni lwayo lwamamodeli ayisisekelo ochungechunge lwesikhathi. Inezici ezihlanganisa ukubikezela okuguquguqukayo okuningi okwakhelwe ngaphakathi, indlela yokunaka ekabili, kanye nokufihla amapheshana aqhubekayo, futhi ingenza ukubikezela kochungechunge lwesikhathi oluguquguqukayo oluningi nge-zero-shot.
IT之家 ithole olwazini olusemthethweni lwe-Google ukuthi i-TimesFM-3 inamapharamitha ayizigidi ezingu-330, futhi yaqeqeshwa kusengaphambili eqoqweni ledatha lochungechunge lwesikhathi lomhlaba wangempela nolwenziwe ngokwenziwa, oluqukethe amaphuzu esikhathi angaphezu kuka-1 trillion. I-TimesFM-3 izuza ekusebenzeni kahle kweka-previous generation kanye nekhono lokwenza generalization nge-zero-shot, futhi ithuthukisa ukusekelwa kwezimo eziyinkimbinkimbi eziguquguqukayo ngezindlela ze-zero-shot.
Ngesikhathi esifanayo, le modeli ingabikezela ndawonye uchungechunge lwesikhathi oluningi oluthuthuka ngokuhlanganyela, ibambe ubudlelwano phakathi kwalo ukuze ithuthukise ukunemba kukonke, ngaphandle kokulungiswa okuqondene nomsebenzi othile. Izici ezisekelwa yimodeli ngokwemvelo zihlanganisa lezi:
Ukubikezela okuqondiswe ezinhlosweni eziningi: ukubikezela ngesikhathi esisodwa uchungechunge lwesikhathi oluningi oluhlobene, njengokubikezela ndawonye u-ayisikhilimu wemikhiqizo ehlukene. Le modeli isekela ukubikezela kwamaphuzu nokubikezela kwama-quantile kuzo zonke izinhloso.
Ama-covariate esikhathi esidlule: ukufaka izici ezaziwa kuphela ngokubheka umlando, njengokugcwala kwabantu kwangaphambilini.
Ama-covariate aguquguqukayo esikhathi esidlule nesizayo: ukusebenzisa izehlakalo zesikhathi esizayo ezaziwayo ukuqondisa ukubikezela, njengemikhankaso yokuphromotha ehleliwe noma isibikezelo sezulu.
Inkomba:
TimesFM-3: A zero-shot foundation model for multivariate forecasting