InfoNCE da Makasudin SimCLR
InfoNCE ita ce hasara mai ma'ana wacce ke koya wa abin ƙira don jawo madaidaicin nau'i-nau'i tare da tura madaidaicin nau'i-nau'i a cikin haɗa sararin samaniya.
Dubawa
SimCLR is a landmark framework that used this loss to learn powerful image representations from unlabeled data, rivaling supervised pretraining.
Zurfafa nutsewa
InfoNCE (Kimanin Haɓaka Harutu don bayanin juna) yana horar da encoder ta yadda tambaya da tabbataccen sa na gaskiya su sami mafi girman makin kamanni fiye da tambayar da yawancin munanan abubuwa. Yana da gaske a softmax giciye-entropy akan maki kamanceceniya: don anga, mai kyau yakamata yayi nasara akan abubuwan da ba su dace ba. SimCLR (2020) ya aiwatar da wannan don hotuna: ɗaukar hoto ɗaya, yi amfani da haɓaka bazuwar guda biyu don ƙirƙirar ingantacciyar guda biyu, gudanar da duka ta hanyar encoder mai raba tare da kan tsinkaya, kuma amfani da daidaitaccen ma'aunin zafin jiki na giciye (NT-Xent, bambance-bambancen InfoNCE) don haka ra'ayoyin biyun da aka haɓaka suna jan hankali yayin da duk sauran hotuna a cikin tsari mara kyau. SimCLR ya nuna cewa ƙaƙƙarfan haɓaka bayanai, shugaban tsinkaya mara tushe, manyan nau'ikan tsari, da yanayin zafin jiki tare suna barin samfuran kulawa da kai suyi daidai da waɗanda ake kulawa akan ImageNet - ba tare da wani lakabi ba yayin horo.
Fahimtar Fasaha
NT-Xent yana ƙididdige kamanceceniyar cosine tsakanin abubuwan haɗin L2-na al'ada, yana rarraba ta zazzabi τ, kuma yana amfani da softmax giciye-entropy yana kula da tabbatacce a matsayin daidaitaccen aji tsakanin duk misalan cikin-tsari. Ƙananan τ yana ƙaddamar da rarrabawa kuma yana ƙulla mummunan mummunan rauni. Ana amfani da shugaban tsinkayar SimCLR (MLP) ne kawai yayin horarwa kuma a watsar da shi daga baya - wakilci kafin canja wurin kai da kyau. Manyan batches suna da mahimmanci saboda suna samar da abubuwa da yawa a cikin mataki ɗaya.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar InfoNCE da Makasudin SimCLR
Maƙasudai daban-daban sun bazu fiye da SimCLR: CLIP yana daidaita hotuna tare da rubutu ta amfani da InfoNCE ta hanyoyin, kuma asara iri ɗaya tana tafiyar da sauti, bidiyo, da samfura na dawo da su. Bincike a yanzu yana rage dogaro ga manyan batches da yawa mara kyau ta bankunan ƙwaƙwalwar ajiya (MoCo), ko kuma cire abubuwan da ba su da kyau gaba ɗaya (BYOL, SimSiam, DINO). Yi tsammanin ci gaba da haɗawa da bambance-bambance, distillation, da pretraining-masked-modeling, tare da daidaitawa na zamani (rubutu, hoto, sauti) a matsayin babban yanki don ƙirar tushe.
Aiwatar da Gaskiyar Duniya
SimCLR yana horar da mai rikodin hoto akan hotuna marasa lakabi, sannan daidaitawa akan ƙaramin saiti mai lakabi don rarrabuwa.
CLIP ta amfani da manufar InfoNCE don daidaita hotuna tare da taken su, yana ba da damar rarraba hoto mai harbi.
Gina bincike na gani/dawowa inda hotuna masu kama da juna ke zama kusa da juna a cikin sararin da aka koya.
Horon da ake sa ido kan kai don ɗaukar hoto na likita ko tauraron dan adam inda alamun ba su da yawa amma ɗanyen bayanai suna da yawa.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
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
Jagora na gaba
Lookhead and Lion Optimizers
Tambayoyin da ake yawan yi
What is InfoNCE and SimCLR Objectives?
InfoNCE ita ce hasara mai ma'ana wacce ke koya wa abin ƙira don jawo madaidaicin nau'i-nau'i tare da tura madaidaicin nau'i-nau'i a cikin haɗa sararin samaniya. SimCLR babban tsari ne na ƙasa wanda yayi amfani da wannan asarar don koyan ƙaƙƙarfan wakilcin hoto daga bayanan da ba a lakabi ba, yana fafatawa da horarwar da aka sa gaba.
Menene manufar InfoNCE ke ƙarfafa abin ƙira ya yi?
InfoNCE shine softmax akan kamanceceniya wanda ke ba da lada mai girma kamance don tabbataccen gaskiya da ƙarancin kamanni don rashin kyau.
A cikin SimCLR, ta yaya aka ƙirƙiri tabbataccen biyu?
SimCLR yana haifar da ingantattun nau'i-nau'i daga ra'ayoyi guda biyu da aka haɓaka na hoton tushe ɗaya; sauran hotuna suna aiki azaman marasa kyau.
Wace rawa zazzabi τ ke takawa a cikin NT-Xent/InfoNCE?
Rarraba kamanceceniya ta τ kafin softmax yana sarrafa yadda hasarar ta ke bambanta tabbatacce daga mummunan mummunan.
Me zai faru da shugaban tsinkayar SimCLR bayan pretraining?
SimCLR ya gano cewa fasalulluka kafin canjin kai na tsinkaya mafi kyau, don haka ana jefar da kai bayan an riga an gama horo.
Me yasa SimCLR ya dogara da manyan batch?
A cikin SimCLR sauran hotunan da ke cikin tsari suna aiki azaman mara kyau, don haka manyan batches suna ba da ƙarin rashin ƙarfi da ƙarfi koyo.