InfoNCE ak SimCLR
InfoNCE mooy ñàkka wuute giy jàngal benn model ñu boole ñaari mbir yu méngoo, ba noppi tàqale ñaari mbir yu méngoowul ci espace embedding.
Résumé
SimCLR is a landmark framework that used this loss to learn powerful image representations from unlabeled data, rivaling supervised pretraining.
Plongeur bu xóot
InfoNCE (Xayma buy wuutale bruit ngir leeral yuñ bokk) dafay tàggat encoder suko defee ab laaj ak limu baax dëgg am poñ yu gëna rëy ci nuru laaj bi ak ay negatif yu bari. Ci yoon sax entropi cross-max la ci kaw poñ yu nuróo: ngir ancre, lu baax moo wara jël ndam li ci lu bonn li. SimCLR (2020) dafa jëfandikoo lii ci nataal yi: jëlal benn nataal, jëfandikoo ñaari yokkute yu bari ngir sos peer bu baax, daw ñoom ñaar ci encoder buñ bokk boole ci boppu projection, ba noppi jëfandikoo entropy cross-scale bu tàngoor (NT-Xent, benn InfoNCE variant) suko defee ñeneen ñi yokk nataal yépp. SimCLR dafa wane ni yokkute ci done yu am doole, boppu projection bu amul rëdd, tolluwaayu lots yu mag, ak tàngoor wuñ boole dañu may model yi ñuy saytu seen bopp ñu méngoo ak yi ñuy saytu ci ImageNet - te amul benn etiket ci diiru tàggat yaram.
Gis-gis xarala
NT-Xent dafay xayma nuru cosinus ci digganté L2-normalisé embeddings, xaaj ko ak tàngoor τ, ba noppi jëfandikoo softmax cross-entropy di jëfandikoo bu baax bi ni klaas bu dëggu ci misaal yi ci biir lots yépp. τ bu wàññeeku dafay gëna ñaw séddale bi ba noppi di gëna yar negatif yu dëgër yi. Boppu projection bu SimCLR (MLP) dañu koy jëfandikoo ci diiru tàggat yaram rek ba noppi sànni ko ginaaw ga - muy wane balaa ñuy toxal bopp bi gëna baax. Lote yu bari dañu am solo ndax dañuy joxe negatif yu bari ci benn jéego.
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 InfoNCE ak SimCLR
Mébet yu wuute yi dañu tasaaroo fu sori ginaaw SimCLR: CLIP dafay méngale nataal yi ak mbind yi ci jëfandikoo InfoNCE ci anam yu bari, te benn ñàkk bi mooy dawal audio, wideo, ak xeetu seet. Gëstu leegi wàññi na dependence ci lots yu bari ak negatif yu bari jaaraleko ci bànku mémoire (MoCo), wala dindi negatif yu leer yépp (BYOL, SimSiam, DINO). Xaarandi wéyal jaxaso bu wuute, distilaasioŋ, ak tàggat modelu maske, ak njubte multimodal (bind, nataal, audio) muy frontiere bi gëna am solo ci modelu fondasioŋ yi.
Doxal ci àdduna dëgg
SimCLR dafay njëkka tàggat encodeur nataal ci kaw nataal yu amul etiket, ba noppi ñu defar ko ci benn set bu ndaw bu am etiket ngir xaaj.
CLIP dafay jëfandikoo ab mébetu InfoNCE ngir méngale nataal yi ak seeni mbind, loolu mooy tax ñu mëna tànnal nataal yi amul benn kapsioŋ.
Tabax seetlu/jëfandikoo nataal fu nataal yu nuróo toog jegewaale ci barabu samp giñ jàng.
Tàggat sa bopp ngir nataali medsin wala satelit fu etiket yi néew waaye done yu bees yi bari nañu.
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 InfoNCE and SimCLR Objectives 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
Lookahead ak gaynde yiy gëna xéewale
Laaj yi ñuy faral di laaj
What is InfoNCE and SimCLR Objectives?
InfoNCE mooy ñàkka wuute giy jàngal benn model ñu boole ñaari mbir yu méngoo, ba noppi tàqale ñaari mbir yu méngoowul ci espace embedding. SimCLR benn kaada la bu jëfandikoo ñàkk bii ngir jàng misaali nataal yu am doole ci done yuñ etiketewul, di xëcc tàggat buñ saytu.
Lan la mébetu InfoNCE di ñaax ab model mu def?
InfoNCE softmax la ci kaw nuru yi, di neexal nuru gu rëy ci dëgg lu baax ak nuru gu néew ci lu baaxul.
Ci SimCLR, naka lañuy defaree ñaar ñu baax?
SimCLR dafay defar peer bu baax ci ñaari gis-gis yuñ yokk ci benn nataalu balluwaay; yeneen nataal yi dañuy nekk negatif.
Ban cër la tàngoor τ def ci NT-Xent / InfoNCE?
Séddale nuru ak τ balaa softmax di saytu ni perte bi di wuutale lu baax ak lu bon lu dëgër.
Lan mooy dall boppu projection bu SimCLR ginaaw biñu ko njëkkee tàggat?
SimCLR gisna ni màndarga yi laata boppu projection bi moo gëna mëna toxal, moo tax bopp bi dañu koy sànni ginaaw bi ñu ko tàggatee.
Lu tax SimCLR di yéem ci dayo yu bari?
Ci SimCLR yeneen nataal yi ci batch bi dañuy nekk negatif, kon batch yu gëna mag dañuy joxe negatif yu bari ak jàng bu gëna am doole.