DINO Distilaasioŋ boppam
DINO anam la wuñuy saytu seen bopp, di tàggat ab transformatëru gis-gis ngir xam nataal yu amul benn etiketu, ci fexe ba reso bi jàngal boppam.
Résumé
It produces features so clean that object boundaries emerge for free in the attention maps.
Plongeur bu xóot
DINO, di distillation boppam bu amul etiket, Meta AI (ci jamono jooju Facebook AI) moo ko siiwal ci 2021. Dafay jëfandikoo ñaari kopi benn reso - ndongo ak jàngalekat - ba noppi di leen dundal ay xeeti mbay yu wuute yu benn nataal. The student tries to match the teacher's output distribution, even though the teacher only sees a different view. Li gëna am solo mooy jàngalekat bi duñu ko tàggat ci saasi; its weights are an exponential moving average of the student's, slowly trailing behind. Ngir moytu reso bi daanu ci benn tontu budul soppeeku, DINO dafay center ak sharpen outputs jàngalekat bi. Resultaa bi yéeme mooy, kàrtu self-attention yi ci segment transformateur vision biy génne ci object yi te kenn musu leen wax lu object bi doon.
Gis-gis xarala
Both networks output a high-dimensional probability distribution after a softmax. Ndongo dafay gis ay mbay yu ndaw yu bawoo ci dëkk bi boole ci gis-gis àdduna bi, waaye jàngalekat bi dafay gis gis-gis àdduna bi kese — pexe mbay yu bari yuy push mbootaayu dëkk bi ak àdduna bi. The loss is cross-entropy between teacher and student distributions, with gradients flowing only through the student. Ñaari pexe ñooy tere mabb: centering dafay dindi moyenne biy dawal ci logits jàngalekat yi, ba noppi tàngoor wu woyof daf leen di ñaw, ñuy ekilibre seen biir suko defee génne yi di wuute.
njeextalu pexe
Gaawaay ak yaatuwaay
Visual IA mën na otomatise saytu, gis ak etiketu liggéey ci eskaal.
Tabax tànneef
Ekipu kreatif yi mën nañu defar konsept yu gëna gaaw te duñu def lu bari ci loxo.
Ekip ak def liggéey
Liggéeyukaay yi mën nañu jëfandikoo siñaal nataal wala wideo yu jafewoon lool ci liggéey.
Ëlëgu DINO distilaasioŋ boppam
DINO dafa tàmbali liggéey bu am solo. DINOv2 (2023) dafa yokk rëset bi ba lu ëpp benn miliyaar ciy nataal yuñ tànn, ba noppi joxe ci anam yu bari yu ñuy xoole, yuy xëccoo ak model yiñ yor ci xayma xóotaayu, segmentation, ak retrieval - mën nañu ko jëfandikoo te amul benn fine-tuning. Xaarandil distillation boppam ngir des ci digg bi ndax barab bi dafa topp xeetu fondasioŋ bu amul etiket ngir gis-gis, robotik, ak sistem multimodal, fu annotation seer lool. Moomeelu segmentation biy feeñ itam mingi wéy di gëna dooleel gëstu ci gis-gis buñ mëna tekki, ubbeeku ci vocabulaire.
Doxal ci àdduna dëgg
Segmentation objet buñu dul saytu, fu kàrtu DINO di fësal mbir yi te amul benn etiketu mask
Seetlu nataal ak gis koppi, jëfandikoo man-mani DINO ngir gis nataal yu nuru ñaar wala yu nuru
DINOv2 dafa melni yaxu ndigg luy gelé ngir xayma xóotaayu liggéey yi ak wax luy am
Tàggat bu njëkk ci wàllu medsin wala xeetu xoolum satelit fu done yuñ etikete jafe wala yu seer
Risk yi ak balustrade yi
Yelleefi nataal ak nangu mën na nekk risku yoon sudee fi ñu bawoo leerul.
Performance model bi mën na wuute ci leeraay bi, demographie bi ak environmaa bi.
Njuumteg positive yi mën nañu dem te kenn duko seetlu fileek xool wuñu buntu wóolu sa bopp.
Roadmap ngir samp gi
Mandargal kritërium nangug njub, woowaat ak njëgu njuumte.
Saytu ak done yu méngoo ak anam yi ñuy liggéeyee dëgg.
Yokk jàngat nit ngir xam fu wóorul dara wala am njeexital yu rëy.
Toppal model drift bi nga baaxal ko ginaaw bi kamera bi wala done yi soppeekoo.
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 DINO Self-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
DreamFusion ak jël misaalu distilaasioŋ
Laaj yi ñuy faral di laaj
What is DINO Self-Distillation?
DINO anam la wuñuy saytu seen bopp, di tàggat ab transformatëru gis-gis ngir xam nataal yu amul benn etiketu, ci fexe ba reso bi jàngal boppam. It produces features so clean that object boundaries emerge for free in the attention maps.
Luy 'Déedet' ci DINO?
DINO mooy distillation boppam te amul benn etiket — dafay saytu boppam, te soxlawul ay tegtal yu nit.
naka lañuy yeesal diisaayu reso jàngalekat yi ci DINO?
Jàngalekat bi EMA ndongo la, kon dafay topp ndongo ci anam wu jaar yoon, du ñu koy tàggat ci saasi.
Ban jafe-jafe la centering ak sharpening ci génnug jàngalekat bi di tere?
Centre ak sharpening dañuy boole seen biir ngir jàngalekat yi mëna am njariñ yu bari, moytu saafara yu amul benn njariñ.
Ci pexem DINO ci mbay mi, ban xalaat la jàngalekat bi am?
Jàngalekat bi dafay gis mbay mi ci àdduna bi kese, ndongo daal di gis mbay mi ci dëkk bi, loolu dafay ñaax nit ñi ñu baña bay ci dëkk bi ak ci àdduna bi.
Lan mooy màndarga yu yéeme yi feeñ ci kàrtu bàyyi xel ci transformatëru gis-gis biñ tàggat ci DINO?
DINO's self-attention dafay wane bu baax ay pexe yu mbir yi di daje, di def segmentation doonte musul gis ay mask.