Qeexid
Gradient descent si joogto ah ayuu u hagaajiyaa parameter-yada jihada ugu badan yaraysa loss-ka. Noocyada sida SGD, Adam, iyo AdaGrad waxay bedelaan sida tallaabooyinka loo cabbiro loona simo.
Ku dhawaad dhammaan tababarka shabakadaha neerongyada casriga ah waxay ku tiirsan yihiin nooc ka mid ah gradient descent.
Si fudud
Ka fiirso in aad ceeryaamo dhex socoto oo aad hoos u socoto: tallaabo kasta waxaad dareemaysaa jihada dhulku ugu qalloocsan yahay, taasoo aad tallaabo yar ku qaadato ilaa aad gaadho dooxo (loss yar).
Halka aad ka aragto
- Tababarka classifier-yada, transformers, iyo moodeelada diffusion.
- Fine-tuning LLM-yada iyadoo la isticmaalayo optimizer-yada AdamW.
Sida ay u shaqeyso
1.Xisaabi loss-ka
Cabbir inta saadaalinta hadda jirta khaldan tahay.
2.Xisaabi gradient-yada
Ogow sida miisaan kasta u saameeyo loss-ka.
3.Cusboonaysii miisaannada
Ka jar learning-rate × gradient.
4.Ku celceli
Tallaabo badan (ama epoch) ilaa la gaaro isku-dheelitir (convergence).
Maxay muhiim u tahay
- Fahamka gradient descent wuxuu fudaydinayaa sida shabakadaha neerongyadu run ahaantii u "bartaan".
Inta badan la khaldo
Gradient descent had iyo jeer wuxuu helayaa xalka ugu wanaagsan ee caalamiga ah.
Loss-yada non-convex waxay leeyihiin dhowr dooxo; dhaqanka wuxuu ku qanacsanaadaa local minima ku filan.