Jàng mu wuutale
Njàngale mu wuute dafay jàngal xeetu nit ñi ñu boole mbir yu nuróo, ba noppi tàqale mbir yu wuute yi ci barab buñ boole.
Résumé
It matters because it lets AI learn powerful representations from mostly unlabeled data, powering image search, recommendations, and multimodal models.
Plongeur bu xóot
Duñu wax luy waaja am ci etiket bi, jàngat bu wuute dafay jàng ci méngale: ñu jox benn mbir bu ancre, model bi dafay tàggat ci anam wu méngoo 'positif' mu wàcci ci wetam ci espace vecteur bi fekk 'negatif' yi méngoowul ñu wàcci fu sori. Benn rëset buñuy saytu boppam (lu melni SimCLR) dafay defar lu baax ci jël ñaari yokkute yu bari ci benn nataal (dagg, melo jitter, blur); leneen lu nekk ci batch bi negatif la. Modèle bi dafay màndargaal dugal yi ci vecteur yi, perte bi dafay neexal nuru bu rëy ci ñaar ñi ak nuru bu néew ci ñeneen ñi. Loolu dafay defar ay embeddings fu distance bi di fësal lu muy tekki, kon liggéey biy wàcci soxla etiket yu néew lool. CLIP dafay jëfandikoo benn xalaat ci anam yu bari, di méngale nataal yi ak seeni mbind.
Gis-gis xarala
Workhorse perte mooy InfoNCE (softmax ci kaw poñ yu nuru), lu bari ak nuru cosine xaaj ak tàngoor wuy saytu ni sharply positive yi di taamu. Li gëna am solo mooy, performance dafay gëna baax ak negatif yu bari, moo tax ay lots yu bari wala bànku mémoire / queue (ni ci MoCo) ñoo leen di jox. Yenn pexe yu melni BYOL ak SimSiam dañuy bàyyi negatif yu leer, lu moy loolu dañuy jëfandikoo momentum wala reso target-gradient ngir moytu dagg, fu embeddings yépp di nuru.
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 jàng bu wuute
Njàngale mu wuute mingi boole ak saytu sa bopp bu maskeer ak defar ci ay mébet yu wuute yuy jàpp nuru àdduna bi ak detay yu ndaw yépp. Li gëna am solo mooy multimodal: nataal-tekst yu wuute (ak leegi audio ak wideo) embeddings yi dañuy jàppale seetlu, jëmmal-yokkum generation, ak zero-shot classification, te emprent bi dina màgg. Xaarandil liggéey bu gëna bari ci wàññi bëgg-bëggu lots yu mag, ci yokk bu baax ak pexe mine yu baaxul, ak ci yaatal jegewaale ci domen yu melni nataali medsin ak time-series fu etiket yi néew te seer.
Doxal ci àdduna dëgg
CLIP jàng barabu nataal-bind buñ bokk suko defee nga mëna seetee ci bibliotek foto ak ab mbind buñ bind melni 'xaj ci kaw skateboard'.
Tàggat yaxu ndigg li ñuy gis ak SimCLR ci kaw nataal yu amul etiketu, ba noppi defar ko bu baax ngir gis feebar bi ak benn set bu ndaw bu am etiketu.
Tabax produit wala way recommandés fu embedding yi mbir yi jëfandikukat bi bëgg toog jegewaale ngir am dëkkandoo bi gëna jege.
Sistem yiy saytu kanam yi di tàggat ay nataali benn nit ñu jegewaale, nit ñu wuute ñu sori seen biir.
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 Contrastive Learning 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
Jàng bu boole
Laaj yi ñuy faral di laaj
What is Contrastive Learning?
Njàngale mu wuute dafay jàngal xeetu nit ñi ñu boole mbir yu nuróo, ba noppi tàqale mbir yu wuute yi ci barab buñ boole. Dafa am solo ndax dafay may IA mu jàng misaal yu am doole ci done yu bari te amul etiket, gëna dooleel seetlu nataal, xalaat ak xeetu multimodal.
Lan mooy mébetu jàngat bu wuute?
Njàngale mu wuute dafay forme barab buy tënk, suko defee ñaar ñuy méngoo ñu jegewaale, ñaar ñi méngoowul ñu sori seen biir.
Ci anamu nataal buñuy saytu sa bopp bu melni SimCLR, naka lañuy sos peer bu baax?
SimCLR dafay forme ay positif ci ñaari gis-gis yuñ yokk ci benn nataal, ak yeneen nataal ci batch bi di nekk negatif.
Ban fonction perte moo gëna méngoo ak jàng bu wuute?
InfoNCE, softmax ci kaw poñ yu nuru ak tàngoor, mooy mébet buñ miin buy wuutale.
Lan moo waral pexe yu melni MoCo di jëfandikoo bànku mémoire wala rang?
Njàngale mu wuute dafay am njariñ ci mbir yu bari yu baaxul; ab raŋ moo leen di jox ci noonu lañuy tëye dayo batch bi.
Ban jafe-jafe la BYOL ak SimSiam di moytu te duñu jëfandikoo ay negatif yu leer?
Su negatif yi amul, model bi mën na xayma lépp ci benn vecteur bi; stop-gradient ak momentum yiñ bëgga tere dañuy tere mabb gi.