Distilaasioŋ Xam-xam
Distilaasioŋ xam-xam dafay tàggat xeetu 'ndongo' bu ndaw ngir toppandoo xeetu 'jàngalekat' bu mag te jaar yoon.
Résumé
It matters because it shrinks powerful models so they run cheaply on phones and servers while keeping much of the accuracy.
Plongeur bu xóot
Modèle yu mag yi dañu jaar yoon waaye dañu yeex te seer lool ci jëfandikoo. Distilaasioŋ xam-xam dafay toxal seen kàttan ci xeetu model bu dëgër, ndax dafay tax ndongo yi jàngee ci li jàngalekat bi génne, te baña jàngee ci etiket yu dëgër yi kese. Li gëna am solo ci Hinton ak ay naataangoom mooy ni seddaleb probabilite bu jàngalekat bi dafay yor 'xam-xam bu lëndëm': doonte dafay wax luy waaja am ci 'xaj', probabilite relatif yi ci 'luu' ak 'oto' dañuy wane ni jàngalekat bi di gisee nuru yi. Noofal probabilite yooyu ak tàngoor dafay wane structure boobu, ba noppi ñu tàggat ndongo yi ñu méngale ko, lu bari ci wetu etiketu dëgg yi. Lépp soo ko boolee mu nekk model bu gëna ndaw, gëna gaaw, gëna mëna jëfandikoo lu ëpp benn buñ tàggat ci etiket yi kese. DistilBERT ak TinyBERT xeetu làkk yuñ siiw lañu.
Gis-gis xarala
Perte classique bi dafay boole ab terme distillation (difference KL ci digganté probabilite yu woyof yu ndongo yi ak jàngalekat yi) ak entropie croix buñ miin ci etiket yu dëggu yi. Softening dafay jëfandikoo tàngoor T ci softmax bi: T bu gëna kawe dafay dakkal séddale bi suko defee nuru yu ndaw yi am ci diggante klaas yi nekk siñaal yuñ mëna jàng; gradient distilaasioŋ bi dañu koy jëfandikoo ci T-kare. Variante yi dem nañu lu ëpp li ñuy génne: distillation bu sukkandiko ci màndarga yi méngoo ak diggante yi nëbbu, distillation bu sukkandiko ci lëkkaloo gi méngoo ak lëkkaloo gi am ci diggante misaal yi.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu Xam-xam Distillation
Distillation leegi jéego buñ miin la ci yobbu ay model yu baax te nekk na lu am solo ci vague tay bu model yu ndaw te mëna ubbeeku. Tendens buy gaawa màgg mooy distillation ci niveau sequence ci model lakk yu mag, fu model bu dëgër di defar ay done di tàggat wala ay tombi xalaat (bokk ci seen xalaat) ngir jàngal ndonngo yu ndaw yi, di dindi liiñ bi ak ay done synthetik. Xaarandi lëkkaloo bu gëna dëgër ak kantite ak dagg, gëna bari ci aparey yi, ak waxtaan buy wéy ci wàllu lisence ak kalite suñuy distil ci model yuñ moom seen génne nekk siñaal tàggat concurrent.
Doxal ci àdduna dëgg
DistilBERT dafay tënk BERT ci lu tollu ci 40% ci paramet yu néew, ci noonu muy tëye li gëna bari ci xam-xam làkk wi ngir gëna gaaw ci jël doggal.
Wàññi ab xeetu gis-gis bu yaatu suko defee ab xaaj nataal mëna dox ci saa si ci aplikaasioŋu kamera telefon bu xarañ.
Distill model bu mag ci chaîne de pensée ci model bu gëna ndaw ngir mëna tontu laaji math wala codage ci anam wu gëna xéewale.
Dajale ensemble model ci benn ndongo suko defee njëgu liggéey bi ak latency bi wàññeeku te du ñàkk lu bari ci njub.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
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 Knowledge Distillation 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
Xam-xam IA
Laaj yi ñuy faral di laaj
What is Knowledge Distillation?
Distilaasioŋ xam-xam dafay tàggat xeetu 'ndongo' bu ndaw ngir toppandoo xeetu 'jàngalekat' bu mag te jaar yoon. Dafa am solo ndax dafay wàññi model yu am doole yi, ñu mëna dox ci telefon yi ak serveur yi ci anam wu yomb, boole ci baña gaawa jeex.
Luy mébetu distilaasioŋ xam-xam?
Distillation dafay toxal kàttanu jàngalekat bu mag ci modelu ndongo bu gëna ndaw te gëna gaaw.
Luy 'xam-xam bu lëndëm' bu jàngalekat bi?
Xam-xam bu lëndëm mooy jumtukaay bi jàngalekat bi di séddale ci anam wu leer, lu melni ni 'luu' ak 'xaj' nuru, du tontu bi gëna kawe rek.
Ban cër la tàngoor (T) di def ci distilaasioŋ?
Tàngoor wu gëna rëy dafay dakkal séddaleb probabilite yi, di wane nuru yi ndongo yi wara jàng.
Ban ñaari mbir ñoo boole perte distillation bu yàgg bi?
Ndongo dafa wara méngoo ak séddaleb jàngalekat bi (terme KL) ci noonu mu méngoo ak etiketu dëgg gi ci suuf (entropie cross).
Lan mooy misaalu xeetu làkk buñ distiye?
DistilBERT xeetu model bu siiw la buñ distiye ci BERT, di tëye performance yu bari ak ay paramet yu néew.