Ubuyobozi bwa tekiniki

BYOL hamwe no Kutavuguruzanya Kwigenzura

BYOL (Bootstrap Own Latent) yiga amashusho yingirakamaro nta kirango na kimwe, kandi, igitangaje, nta ngero mbi.

Umva iki gitaboAI-byavuzwe · ~2 min

Kuriyi page2 min soma
  1. Incamake
  2. Kwibira cyane
  3. Gushyira mu bikorwa Isi
  4. Ejo hazaza ha BYOL no Kutavuguruzanya Kwigenzura
  5. Kubishyira mubikorwa
  6. Komeza Ubushakashatsi
  7. Ibibazo bikunze kubazwa

Incamake

It showed that self-supervised learning need not rely on pushing apart dissimilar images, sidestepping the need for huge batches of negatives.

Kwibira cyane

Uburyo bwinshi bwo kwiyobora hakiri kare bwari butandukanye: bakwega ibitekerezo bibiri byongeweho ishusho imwe hamwe mugihe basunikaga amashusho atandukanye, byasabye ingero nyinshi mbi kugirango birinde gusenyuka (aho umuyoboro usohora vector imwe kuri buri kintu). BYOL, kuva muri DeepMind muri 2020, yakuyeho ibibi rwose. Ikoresha imiyoboro ibiri: umuyoboro wa interineti numuyoboro ugamije. Ibintu bibiri byongerewe ishusho imwe inyura mumiyoboro ibiri; umuyoboro wa interineti wongeyeho umutwe wo guhanura kandi watojwe guhanura intego yibikorwa byerekana ubundi buryo bwo kureba. Icyitonderwa, ibipimo byurwego rwibipimo ntabwo byatozwa na gradient descente. Ahubwo ni impuzandengo yimuka (EMA) yuburemere kumurongo. Iyi asimmetrie wongeyeho intego ya EMA irinda gusenyuka kworoheje uburyo butandukanye bwo gutinya, guhuza cyangwa gukubita ibice bitandukanye kuri ImageNet.

Gushyira mu bikorwa Isi

Gutegura umugongo w'icyerekezo kuri miriyoni y'amafoto atashyizweho ikimenyetso, hanyuma ugahuza neza kuri dataset ntoya yanditseho ubuvuzi-bwerekana amashusho aho ibisobanuro byinzobere ari bike.

Kwiga robot yibiranga kuva kamera ya kamera mbisi nta kuranga intoki, kugabanya ikiguzi cyo kwigisha imirimo ya manipulation.

Kubaka amashusho-kugarura no kugabanya sisitemu ukoresheje BYOL yashyizwemo ayo mashusho asa nkaho nta kirango cyicyiciro.

Gutangiza icyogajuru cyangwa amashusho yikirere mububiko bunini butanditseho ububiko mbere yo guhuza neza imikoreshereze yubutaka cyangwa gutema amashyamba.

Ejo hazaza ha BYOL no Kutavuguruzanya Kwigenzura

Ibitekerezo bitavuguruzanya ubungubu ibyerekezo byinshi byo kwiyobora. SimSiam yambuye BYOL hasi, yerekana intego ya EMA ntabwo isabwa cyane niba guhagarika-gradient bikomeje, byimbitse gusobanukirwa nimpamvu yo gusenyuka birindwa. Tegereza ibyo bitabo bitarimo ibirango byo kwitegura kugirango ukomeze guhuza hamwe no kwerekana amashusho yerekana amashusho hamwe namahugurwa ya multimodal, no gukwirakwira kuri videwo, amajwi, amashusho yubuvuzi, hamwe na robo aho ibirango ari bike cyangwa bihenze, akenshi nkicyiciro cyo kwitegura mbere yumucyo ukurikiranwa neza.

Kubishyira mubikorwa

Impamvu ari ngombwa

  • Igiciro na bije. Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.
  • Ibyemezo bisobanutse. Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.
  • Kugenzura ubuziranenge. Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.

Witondere

  • Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.
  • Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.
  • Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.

Intambwe ikurikira

  1. Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.
  2. Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.
  3. Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.
  4. Tegura inzira yo gusubiza ibyabaye mbere yo gupima.

Komeza Ubushakashatsi

Ibaruwa y'amakuru y'ubuntu

Ibuka amakuru ya buri munsi ya AI

Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.

Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.

Isuzume

Take the BYOL and Non-Contrastive Self-Supervision quiz

Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.

Tangira ikibazo

Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano

Ibibazo bikunze kubazwa

What is BYOL and Non-Contrastive Self-Supervision?

BYOL (Bootstrap Own Latent) yiga amashusho yingirakamaro nta kirango na kimwe, kandi, igitangaje, nta ngero mbi. Yerekanye ko kwigira wenyine kwigira bidakeneye gushingira ku gutandukanya amashusho atandukanye, ukuraho ibikenewe byinshi bibi.

What makes BYOL non-contrastive?

The BYOL paper highlights learning representations without negative pairs.

What does the online network learn to predict?

BYOL trains the online branch to predict the target representation across two views.

How are target-network weights updated?

The paper describes updating the target with a slow-moving average.

Why does BYOL use an online predictor and stop-gradient on the target side?

The BYOL method uses asymmetry and a stop-gradient as parts of its design.

What happens to two views of the same image in BYOL?

The paper learns by predicting one view’s target representation from another.