BYOL da Kulawa da Kai ba Mai Ma'ana ba
BYOL (Bootstrap Your Own Latent) yana koyon fa'idodin hoto masu amfani ba tare da kowane lakabi ba kuma, abin mamaki, ba tare da misalan marasa kyau ba.
Dubawa
It showed that self-supervised learning need not rely on pushing apart dissimilar images, sidestepping the need for huge batches of negatives.
Zurfafa nutsewa
Yawancin hanyoyin kulawa da kansu na farko sun kasance masu bambanci: sun jawo ra'ayoyi guda biyu na hoto iri ɗaya tare yayin da suke tura hotuna daban-daban, wanda ke buƙatar samfurori marasa kyau da yawa don guje wa rugujewa (inda hanyar sadarwa ta fitar da nau'i ɗaya na kowane abu). BYOL, daga DeepMind a cikin 2020, ya cire mummunan gaba ɗaya. Yana amfani da cibiyoyin sadarwa guda biyu: cibiyar sadarwar kan layi da cibiyar sadarwar manufa. Ra'ayoyi guda biyu da aka haɓaka na hoto ɗaya suna tafiya ta hanyar cibiyoyin sadarwa guda biyu; hanyar sadarwa ta kan layi tana ƙara shugaban tsinkaya kuma an horar da ita don hasashen wakilcin cibiyar sadarwar da aka yi niyya na ɗayan ra'ayi. Mahimmanci, ma'aunin cibiyar sadarwa da aka yi niyya ba a horar da su ta hanyar zuriya ta gradient. Madadin haka su ne matsakaicin matsakaicin motsi (EMA) na ma'aunin kan layi. Wannan asymmetry da makasudin EMA yana hana ƙananan rugujewar hanyoyin banbance banbancen hanyoyin da ake tsoro, daidaitawa ko bugun ginshiƙan tushe akan ImageNet.
Fahimtar Fasaha
Sinadaran guda uku suna tsayawa ba tare da lahani ba: ƙarin MLP mai tsinkaya akan reshen kan layi, tsayawa-gradient akan reshen da aka yi niyya, da buƙatun sabunta EMA. Maƙasudin yana aiki azaman maƙasudin koma baya a hankali, don haka hanyar sadarwar kan layi tana korar tsayayye, maƙasudi mai rauni maimakon kwafin motsin kanta. Asymmetry na tsinkaya yana karya siffa wanda in ba haka ba zai bar rassan biyu su fita da yawa. Daidaita tsari a cikin na'ura kuma yana ba da gudummawar daidaitawa a fakaice.
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 BYOL da Kulawa da Kai da ba Ma'asumai ba
Ra'ayoyin da ba su saba da juna ba yanzu sun kafa hangen nesa mai kulawa da kai. SimSiam ya cire BYOL ƙasa, yana nuna makasudin EMA ba a buƙata sosai idan an kiyaye tasha-gradient, zurfafa fahimtar dalilin da yasa ake guje wa rugujewa. Yi tsammanin waɗannan girke-girke na horarwa na kyauta don ci gaba da haɗuwa tare da ƙirar hoto mai rufe fuska da horarwa na zamani, da kuma yada zuwa bidiyo, sauti, hoton likitanci, da na'ura mai ba da hanya tsakanin hanyoyin sadarwa inda alamun ba su da yawa ko tsada, galibi a matsayin matakin farko kafin a kula da lafiya mai nauyi.
Aiwatar da Gaskiyar Duniya
Gabatar da ƙashin bayan hangen nesa akan miliyoyin hotuna marasa lakabi, sannan daidaitawa akan ƙaramin ma'aunin bayanan hoto na likitanci inda bayanan ƙwararru ba su da yawa.
Koyan fasalulluka na tsinkayar mutum-mutumi daga rafukan kyamarori masu ɗanɗano ba tare da alamar hannu ba, rage farashin koyarwar ayyukan magudi.
Gina tsarin dawo da hoto da cirewa ta amfani da abubuwan haɗin BYOL waɗanda ke haɗa hotuna iri ɗaya na gani ba tare da kowane lakabin aji ba.
Ƙaddamar da samfurin tauraron dan adam ko na iska a kan ɗimbin ma'ajin da ba a lakafta su ba kafin daidaitawa don rarraba ƙasa ko saran gandun daji.
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 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
Jagora na gaba
Lakabin Lakabi da Koyarwar Kai
Tambayoyin da ake yawan yi
What is BYOL and Non-Contrastive Self-Supervision?
BYOL (Bootstrap Your Own Latent) yana koyon fa'idodin hoto masu amfani ba tare da kowane lakabi ba kuma, abin mamaki, ba tare da misalan marasa kyau ba. Ya nuna cewa koyo na kulawa da kai baya buƙatar dogaro da rarrabuwar hotuna daban-daban, tare da kawar da buƙatun manyan abubuwan da ba su dace ba.
Menene ya sa BYOL ya bambanta tsakanin hanyoyin kulawa da kai na zamaninsa?
BYOL ya nuna ƙaƙƙarfan koyon wakilci mai yuwuwa ba tare da nau'i-nau'i mara kyau ba, watsewa daga hanyoyin da suka bambanta.
Ta yaya ake sabunta ma'auni na cibiyar sadarwa a cikin BYOL?
Cibiyar sadarwar da aka yi niyya ita ce EMA (kwafin mai motsi a hankali) na hanyar sadarwar kan layi kuma ba a horar da su ta hanyar zuriya ta gradient.
Wace matsala ce mutane suka ji tsoron zai faru lokacin cire abubuwan da ba su da kyau, kuma me ya sa yake da mahimmanci?
Ba tare da munanan abubuwa ba, saitin butulci zai iya rugujewa zuwa fitarwa ta dindindin; Tsarin BYOL ya hana hakan.
Wanne bangare ne kawai aka ƙara zuwa reshen kan layi don karya siffa?
Reshen kan layi yana da ƙarin shugaban tsinkaya; asymmetry da yake haifarwa shine mabuɗin don gujewa rushewa.
Menene ainihin hanyar sadarwar kan layi ta horar da yin?
BYOL yana rage bambanci tsakanin tsinkayar kan layi da wakilcin manufa na ɗayan ra'ayi.