Autoencodeurs masqués
Masked Autoencoders (MAE) xeetu xam-xam la buñuy saytu seen bopp, di jàngal modelu gis-gis bi ñu defaraat nataal yi ginaaw bi ñu nëbbee li ëpp ci nataal yi.
Résumé
By learning to fill in the blanks, the model builds rich visual understanding without any human labels.
Plongeur bu xóot
Autoencoders yu maskeer, yu Kaiming He ak ay naataango ci Meta AI dugal ci 2021, dañuy jël benn nataal, xaaj ko ay pàcc yu ndaw, ba noppi nëbb ci lu bari ci ñoom, lu bari ci 75%. Encoder Vision Transformer dafay liggéey ci patch yiñ mëna gis, waaye decoder bu woyof bi dafay jéema defaraat pixel yi fi ñàkk. Ndax barina luñu nëbb, model bi mënul kopie pixel yi ci wetam rek, te dafa wara jàng structure yu am solo, lu melni jëmm ak cër yi. Encoder bi dafay sànni ay patch yuñ maske, dafay tax tàggat yaram bi gaaw, ba noppi sakkanal mémoire bi. Ginaaw ñu ko tàggatee, dañuy sànni dekodeer bi, enkodeer bi daal di koy toxal bu baax ci wàllu xaaj, gis ak xaaj.
Gis-gis xarala
Li gëna am solo mooy asymétrie: encoder bu diis bi dafay gis 25% ci patch yi maskewul, waaye decoder bu ndaw bi dafay defaraat leneen li. Patch yi dañu leen platre, dugal leen ci ligne, ba noppi ñu jox leen encodage positionnel. Perte ci tabaxaat bi mooy njuumte kaare moyen bi ñuy xayma ci patch yuñ maske rek, rawatina ci valeur pixel yuñ normalisee. Rasio masking yu kawe yi dañuy forse njàngum semantik moo gën interpolation bu woyof, ba noppi sànni token masked yi ci encoder bi dafay dagg ordinatër bi bu baax moo gën defar nataal bi yépp.
njeextalu pexe
Gaawaay ak yaatuwaay
Visual IA mën na otomatise saytu, gis ak etiketu liggéey ci eskaal.
Tabax tànneef
Ekipu kreatif yi mën nañu defar konsept yu gëna gaaw te duñu def lu bari ci loxo.
Ekip ak def liggéey
Liggéeyukaay yi mën nañu jëfandikoo siñaal nataal wala wideo yu jafewoon lool ci liggéey.
Ëlëgu autoencodeur yu maske
Tabaxwaat maskeer-style MAE mingi nekk rëset bu ñuy njëkka tàggat ci anam yépp. Gëstukat yi dañu koy yokk ci wideo (di nëbb ay cube espace-temps), spectrogram audio, scanner medikaal, ak nataali satelit, ndax foofu etiketu yi bariwul te seen njëg jafe lool. Xaarandi boole bu gëna dëgër ak làkk ngir xeetu fondasioŋ multimodal, dekodeer yu gëna am njariñ, ak maskeer buy méngoo ak gox yi am xibaar. Bi ordinatër di màgg, masked pretraining ci kaw nataal yu bari yu amul etiketu dañu wara wéy di gëna dëgaral njubte gi ci wàll wi ci noonu lañuy wàññi koolute ci annotation nit yu seer yi.
Doxal ci àdduna dëgg
Taggat ab Transformatëru gis-gis ci ay milioŋ ciy nataal yuñu etiketewul, ba noppi nga defar ko bu baax ngir mëna tànnal ImageNet ci anam wu jaar yoon
Jàng man-mani yi ci scanner medikaal yi amul etiketu (rayon X, MRI) fu annotation eksper yi seer lool te bariwul
Defar anam wi ci wideo ci maskeer ay patch espace-temps ngir tàggat xeetu xàmmee jëf (VideoMAE)
Tàggat bu njëkk ci satelit ak nataali jawwu ji ngir jàppale jëfandikoo kàrtu suuf ak gis coppite te du am etiketu loxo
Risk yi ak balustrade yi
Yelleefi nataal ak nangu mën na nekk risku yoon sudee fi ñu bawoo leerul.
Performance model bi mën na wuute ci leeraay bi, demographie bi ak environmaa bi.
Njuumteg positive yi mën nañu dem te kenn duko seetlu fileek xool wuñu buntu wóolu sa bopp.
Roadmap ngir samp gi
Mandargal kritërium nangug njub, woowaat ak njëgu njuumte.
Saytu ak done yu méngoo ak anam yi ñuy liggéeyee dëgg.
Yokk jàngat nit ngir xam fu wóorul dara wala am njeexital yu rëy.
Toppal model drift bi nga baaxal ko ginaaw bi kamera bi wala done yi soppeekoo.
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 Masked Autoencoders 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
Muse Masked Nataal Generatif
Laaj yi ñuy faral di laaj
What is Masked Autoencoders?
Masked Autoencoders (MAE) xeetu xam-xam la buñuy saytu seen bopp, di jàngal modelu gis-gis bi ñu defaraat nataal yi ginaaw bi ñu nëbbee li ëpp ci nataal yi. Su model bi jàngee joxe leeral ci barab yu amul dara, dafay tabax xam-xam bu xóot te amul benn etiketu nit.
Lan mooy liggéey bi gëna am solo bi ñuy saytu seen bopp ci Autoencoder Masked?
MAE dafay nëbb patch nataal yu bari ba noppi di tàggat model bi ngir defaraat pixel yi ñàkk, te soxlawul etiketu nit.
Lu tollu ci ban paccu nataal la MAE bu njëkk bi di maskeer?
MAE bu njëkk bi dafay mask luy tollu ci 75% ci patch yi, muy ratio bu yéeg buy forse model bi mu jàng structure bu am solo moo gën koppi dëkkandoo yi.
Lan moo waral encoder MAE bi du def ludul ci patch yiñ mëna gis (ñu maskewul)?
Soo sànnee ay token yuñ maske ci encoder bi dafay wàññi bu baax ordinatër bi ak mémoire bi, ndax dafay jëfandikoo tuuti ci patch yi.
Lan mooy dall dekodeer bi ginaaw bi MAE defee tàggat yaram?
Decodeur bu woyof bi soxla nañu ko ngir tabaxaat ci diiru tàggat yaram; Ginaaw loolu encoder bi jàng toxal ci liggéey yu melni gis.
Ban fonction perte la MAE di faral di jëfandikoo ngir tabaxaat?
MAE dafay wàññi njuumte kaare moyen bi diggante valeur pixel yiñ séentu ak yiñ dëgg, ñu xayma ko ci patch yuñ maskeer yi kese.