Wàcci degrade stokastik ak saa
Momentum ab coppite la ci wàccinu degrade biy dajale ab moyenne buy daw ci degrade yi weesu, loolu mooy tax optimisation bi gëna gaaw ci vale yi te dakkal oscillations yi.
Résumé
It is one of the most widely used training tricks in deep learning.
Plongeur bu xóot
Wàcci degrade stochastic bu leer (SGD) dafay yeesal parametre yi ci jéego ci wàll wi jàkkarloo ak degrade mini-batch bi fi nekk. Ci peyisaas yu am formu ravin yu gudd te sew, lii dafay zig-zag ci kaw miir yu gaaw yi, di raam ci dër bu woyof bi. Momentum, bi Polyak siiwal, ginaaw ga Rumelhart ak ay naataangoom, dafay saafara lii ci tëye vecteur gaawaay: jéego bu nekk dafay boole gradient bu bees bi ak benn wàll (koefisieŋu momentum bi, lu bari ci 0.9) ci gaawaay bi njëkk. Jubluwaayi degrade yu dëppoo dañuy gëna dooleel ak gaaw, ci noonu la komposant yiy osciller di fomm wàll wi. Analogie physique bi mooy benn bal bu diis buy wàcci: dafay tabax gaawaay ci yoon yu dëgër te du gëna néew luñu koy def ci ay dos d'argent yu bari, di joxe convergence bu gëna gaaw, gëna nooy gëna vanille SGD.
Gis-gis xarala
Coppite bi dafay tëye gaawaay v bi ñuy yeesal ni v = beta * v + gradient, suko defee parametre yi di toxu ci wàññi tolluwaayu jàng bi v. Ak koeffisientu momentum beta, jéego bi am solo ci yoon wu méngoo dañu koy yokk lu tollu ci facteur 1/(1 - beta); ci beta = 0.9 lu tollu ci fukki yoon. Ci wàllu math, lii ab moyenne mouvement pondéré exponentiel la ci degrade yi, di dindi bruit yu ndaw yi boole ci baña bàyyi yoon wi gëna am solo ci wàcci gi.
njeextalu pexe
dogal yu gëna leer
Daf lay jàppale nga tàqale kàddu yu leer ci wàllu xarala ak làkku fësal njaay.
Njëgg ak budget
Mën nga laaj laaj yu gëna baax ci samp gi balaa ngay dugal xaalis wala sa jotu liggéey.
Ekip ak def liggéey
Ekip yi bokk xam-xam ñoo gëna mëna jël yenn dogal ci wàllu produit, politik ak jàng.
Ëlëgu wàccinu gradient stokastik ak saa
Momentum mingi wéy di nekk fondaasioŋ: optimisateurs adaptifs yu melni Adam ak ay variantam dañuy dugal benn estimation bu njëkk ci momentum, te SGD ak momentum ba leegi nekk na baseline bu dëgër buy faral di generalise gëna xéewale ci xeetu gis-gis yu mag. Gëstu baa ngi wéy ci wàllu oraaru momentum, decoupled weight decay, ak jaxasoo bi ak tàggat yaram bu bari lool. Xaarandil ni momentum bi dina des ci benn cër bu am solo ginaaw bi optimisatër yi di jëm kanam ci model yu gëna mag.
Doxal ci àdduna dëgg
Taggat reso yu xóot yu melni ResNet, fu SGD ak momentum 0.9 nekk rëset buñ miin.
Nooyal xayma gradient bu bari bruit sooy jëfandikoo ay lote yu ndaw.
Daw ci platoo yu xóotul yi ci yóbbu gaawaay ci gox yu dalal yi.
Dafay nekk kàddu momentum ci biir optimisatër yiy méngoo ak anam yu melni Adam ak RMSprop.
Risk yi ak balustrade yi
Ekip yu bari mën nañu jëfandikoo benn baat ci anam wu wuute, kon teela leeral yaatuwaayam.
Benchmark yi mën nañu nuru lu am doole waaye performance yi ci àdduna bi duñu tolloo.
Bëgg kalite done ak palaŋu jàngat dafay faral di jur njariñ yu yomba dagg.
Roadmap ngir samp gi
Tàmbaleel ci joxe leeral ci làkk wu leer ci njariñ li nga soxla.
Tannal benn metric bu baax ak benn anam bu baaxul balaa ngay saytu.
Doxal ab pilote bu ndaw ak ay done yu representatif, du ab demo bu leer.
Dokumenteer fi wàccinu gradient stochastic ak momentum di jàppale ak fi pexe yu gëna yomba gëna baax.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Stochastic Gradient Descent with Momentum quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Wàccig degrade
Laaj yi ñuy faral di laaj
What is Stochastic Gradient Descent with Momentum?
Momentum ab coppite la ci wàccinu degrade biy dajale ab moyenne buy daw ci degrade yi weesu, loolu mooy tax optimisation bi gëna gaaw ci vale yi te dakkal oscillations yi. Benn la ci pexe tàggat yaram yiñ gëna jëfandikoo ci jàng bu xóot.
Lan la momentum bi di dajale ci diiru tàggat yaram?
Momentum dafay tëye vecteur gaawaay bi nekk moyenne moyenne pondérée exponentielle ci gradient yu bees yi, di nooyloo yoonu yeesal bi.
Ci ravine bu gudd te sew, ban jafe-jafe la SGD bu leer di dundu te momentum bi dafay jàppale ci saafara?
Su momentum amul, SGD dafay osciller ci yoon yu gaaw yi ci ravin bi. Momentum dafay fomm oscillation yooyu ba noppi gaaw ci suufu vale bi.
Ban analogie physique lañuy gëna jëfandikoo ngir fësal momentum?
Momentum dañu koy méngale ak bal bu diis buy tabax gaawaay ci yoon wu méngoo te baña dellu ginaaw ci ay dos d'état yu bari bruit.
Lu tollu ci ñaata la momentum bi di yokk jéego bu am solo bi ci yoon wu méngoo sudee beta = 0.9?
Facteur amplification mingi tollu ci 1/(1-beta), ak 1/(1-0.9) = 10, kon yoon yu méngoo dañuy gaaw lu tollu ci fukki yoon.
Ban optimisatër bu bees moo boole ci xayma saa yu njëkk yu melni momentum?
Adam dafa boole xayma gradient yi ci saa yu njëkk yi (moyenne) ak xayma ci ñaareelu saa yi (variance) ngir mëna ànd ak eskaal bi.