BYOL ak saytu sa bopp bu amul wuute
BYOL (Bootstrap Your Own Latent) dafay jàng nataal yu am njariñ te amul benn etiket, te lu yéeme mooy amul misaal yu baaxul.
Résumé
It showed that self-supervised learning need not rely on pushing apart dissimilar images, sidestepping the need for huge batches of negatives.
Plongeur bu xóot
Pexe yi njëkka am ngir saytu seen bopp dañu doon wuute: dañu boole ñaari gis-gis yu gëna yokk ci benn nataal, boole ci tàqale nataal yu wuute, loolu dafay laaj misaal yu bari yu baaxul ngir moytu daanu (fu reso bi di génne benn vecteur ngir lépp). BYOL, ci DeepMind ci 2020, dindi na lépp lu baaxul. Dafay jëfandikoo ñaari reso: reso bu nekk ci net bi ak reso biñ bëgga jëfandikoo. Ñaari gis-gis yuñ yokk ci benn nataal dañuy jaar ci ñaari reso yi; reso bi ci net bi dafay yokk benn boppu wax luy waaja am, ñu tàggat ko mu wax luy waaja am ci reso bi ñuy wut ci beneen gis-gis bi. Li gëna doywaat mooy diisaayu reso biñ bëgga tàggat nekkul ci wàccinu gradient bi. Lu moy loolu dañuy nekk moyenne exponentielle (EMA) ci poid yi ci net bi. asymétrie bi boole ci EMA target bi dafay tere xeetu contraste yu ñuy ragal, méngoo wala daan ligne yu wuute ci ImageNet.
Gis-gis xarala
Ñatti ingredient ñooy taxawal mabb te amul lu baaxul: benn MLP buy wax luy waaja am ci bànxaas bi ci net bi, benn stop-gradient ci bànxaas biñ bëgga am, ak benn mébet buñ yeesal ci EMA. Bindkat bi dafay melni mébetu regression buy toxu ndànk, moo tax reso bi ci net bi dafay topp mébet bu dëgër, bu dellu ginaaw, du nekk kopi buy toxu ci boppam. Asymétrie predicteur bi dafay dog symétrie bi nga xamni luko moy dina may ñaari car yi ñu génne ci anam wu amul benn njariñ benn constant. Normalisasioŋ ci lots ci projectër bi itam dafay jàppale ci yamale gi.
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 BYOL ak saytu sa bopp bu amul benn wuute
Leegi xalaat yu wuutewul ñooy ancre lu bari ci gis-gis bi ñuy saytu seen bopp. SimSiam dindi na BYOL gëna wàcci, wane ni EMA target bi amul benn njariñ sudee dañuy tëye stop-gradient, gëna xóotal xam-xam bi waral ñuy moytu mabb. Xaarandil yii rëset yu amul etiket ngir wéy di boole ak modeling nataal bu maskeer ak tàggat multimodal, ba noppi tasaaroo ci wideo, audio, nataali medsin, ak robotik fu etiket yi néew wala seer, lu bari ci etapu tàggat yaram balaa lightweight supervised fine-tuning.
Doxal ci àdduna dëgg
Taggat ab yaxu ndigg ci gis-gis ci ay milioŋ ciy nataal yuñ etiketewul, ba noppi defar bu baax ci ab done bu ndaw buñu etikete ci nataali medsin, fu annotation yu kàngam yi néew.
Jàng man-mani gis-gis robot ci ay stream kamera yu ñor te du am etiketu loxo, wàññi njëgu jàngale liggéey yuñ jaay xel.
Tabax sistemu seet nataal ak deduplication jëfandikoo BYOL embeddings yuy boole nataal yu nuru ci gis-gis te amul benn etiketu klaas.
Tàmbali modelu satelit wala nataali jawwu ji ci kaw archive yu yaatu yu amul etiket balaa ñuy gëna defar ngir jëfandikoo suuf wala dagg garab.
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 BYOL and Non-Contrastive Self-Supervision 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
Pseudo-Etiketu ak Tàggat Bopp
Laaj yi ñuy faral di laaj
What is BYOL and Non-Contrastive Self-Supervision?
BYOL (Bootstrap Your Own Latent) dafay jàng nataal yu am njariñ te amul benn etiket, te lu yéeme mooy amul misaal yu baaxul. Mu wane ni jàng bi ñuy saytu sa bopp warul aju ci tàqale nataal yu wuute, di moytu soxla ay xeeti negatif yu bari.
Lan moo wuutale BYOL ci anam yi ñuy saytoo seen bopp ci jamonoom?
BYOL dafa wane ni njàngum representation bu dëgër mën na am te du am ñaari mbir yu baaxul, baña bàyyi gis-gis yu wuute.
naka lañuy yeesal poid yi ci reso bi ñuy wut ci BYOL?
Reseau biñuy wut mooy EMA (kopi buy dem ndànk) ci reso bi ci net bi te tàggatu ñu ci wàccinu gradient.
Ban jafe-jafe la nit ñi ragaloon ni dina am suñu dindie mbir yu baaxul yi, te lu tax mu am solo?
Su negatiif yi amul, tabb bu yomb bi mën na wàcci ba nekk génne gu dul soppeeku; Designu BYOL dafa tere loolu.
Ban composant lañuy yokk ci branch bi ci net bi ngir dog symétrie bi?
Sucursal bi ci net bi amna beneen bopp buy wax luy waaja am; asymétrie bi muy defar mooy caabi ngir moytu dagg.
Lan lañu tàggat ci reso bi ci net bi?
BYOL dafay wàññi wuute gi nekk ci digganté li ñuy wax ci net bi ak li ñuy wax ci beneen gis-gis bi.