Aadam iyo Hagaajinta La Qabsiga
Aadam waa farsameeyaha shaqada faraska ka dambeeya inta badan shabakadaha neerfaha ee casriga ah, isaga oo si toos ah u hagaajinaya heerka waxbarasho ee u gaarka ah cabbir kasta.
Dulmar
It matters because it makes training deep models faster and far less finicky than plain gradient descent.
quusid qoto dheer
Adam (Qiyaasta La Qabsiga La Qabsiga), oo ay soo bandhigeen Kingma iyo Ba 2014, waxay isku daraan laba fikradood. Marka hore, xawli: waxa ay xajisaa celcelis ahaan qudhuntay ee gradients hore (daqiiqadii ugu horaysay) si ay cusboonaysiinta u dhisto xawaaraha jihooyin joogto ah. Midda labaad, halbeeg-beeg-beegista: waxay daba-galaysaa celceliska gradients labajibbaaran (daqiiqada labaad) waxayna u qaybisaa tillaabo kasta xididka labajibbaaran ee qiimahaas, sidaas darteed cabbirada leh jaan-gooyooyin waaweyn oo buuq badan ayaa qaadaya tillaabooyin yar yar kuwa dhif ah-cusbooneysiina waxay qaadaan tillaabooyin waaweyn. La qabsigan macneheedu waxa weeye in aad inta badan isticmaali karto hal heer wax-barasho oo dhan shabakada oo dhan. Kala duwanaanshuhu, AdamW, wuxuu qurxiyaa qudhunka miisaanka cusboonaysiinta gradient-ka wuxuuna noqday meesha ugu habboon ee lagu tababaro transformers waaweyn iyo moodooyinka luqadda.
Aragtida Farsamada
Aadam waxa uu hayaa laba celcelis orod halkiibeeg: m (gradients) iyo v ( gradients afar gees ah), oo lagu cusboonaysiiyay heerka qudhunka beta1 (caadi ahaan 0.9) iyo beta2 (caadi ahaan 0.999). Sababtoo ah labaduba waxay ku bilowdaan eber, waxaa lagu saxaa eex iyadoo loo qaybinayo (1 - beta ^t). Cusboonaysiinta waa theta = theta - lr * m_hat / (sqrt(v_hat) + epsilon), halkaasoo epsilon (qiyaastii 1e-8) ay ka hortagto u qaybinta eber. Tani waa sababta Aadan uu ugu baahan yahay hagaajin yar oo heerka waxbarashada ah marka loo eego SGD cad.
Saamaynta Istiraatijiyadeed
Qiimaha iyo miisaaniyada
Go'aamada qaab-dhismeedku waxay horseedaan waxqabadka iyo kharashka hawlgalka sannadaha.
Go'aamo cad
Waxbarashada farsamada waxay ka caawisaa kooxaha inay doortaan xidhmo sax ah, ma aha oo kaliya kan ugu cusub.
Xakamaynta tayada
Doorashooyinka injineernimada ee wanaagsan waxay yareeyaan shilalka la isku halleyn karo ee wax soo saarka.
Mustaqbalka Aadan iyo La-qabsiga Hagaajinta
Adam iyo AdamW ayaa weli ah kan ugu sarreeya, laakiin cilmi-baaristu waxay riixaysaa hufnaanta moodooyinka trillion-parameters, halkaasoo kaydinta laba qiime oo dheeraad ah miisaankiiba ay qaali tahay. Kala duwanaanshaha iftiinka xusuusta sida Adafactor, 8-bit Adam, iyo wanaajiyayaal cusub sida Lion (kaas oo isticmaala kaliya calamad ku salaysan) iyo Sophia waxay hiigsanaysaa inay la jaanqaado tayada Adam oo leh xasuus yar ama isku xidhid degdeg ah. Filo hagaajinta la qabsiga oo si gaar ah loogu habeeyey si loo qaybiyo, tababar sax ah oo hooseeya si loo sii horumariyo.
Dhaqangelinta Adduunka-dhabta ah
Tababarka moodooyinka luqadaha waaweyn sida GPT iyo Llama, kuwaas oo u adeegsada AdamW sida hagaajinta caadiga ah.
Hagaajinta sawirka hore loo tababaray (tusaale, ResNet) ee xog-ururin leh oo leh heerka waxbarashada Adam ee caadiga ah.
Tababarka moodooyinka faafinta ee ka dambeeya sawir-dhaliyeyaasha sida Stable Diffusion.
Ku socodsiinta 8-bit Adam ee maktabadaha sida bitsandbytes si ay ugu habboonaato dawladaha hagaajinta xusuusta GPU xaddidan.
Khatarta & Dariiqyada Ilaalada
Hagaajinta hal bartilmaameed waxay qarin kartaa daciifnimada nidaamka ballaaran.
Kaabayaasha dhaqaalaha iyo dayactirka inta badan waa la dhayalsadaa.
Nabadgelyada iyo daldaloolada u fiirsashada ayaa kori kara marka nidaamyadu noqdaan kuwo aad u adag.
Qorshe Hawleedka Dhaqangelinta
Qeex daahida, tayada, iyo bartilmaameedyada qiimaha ka hor inta aan la hirgelin.
Benchmark marka la eego culeyska dhabta ah iyo xaaladaha xogta.
La socodka qalabka khaladaadka, leexashada, iyo saamaynta isticmaalaha.
U diyaari dib-u-noqoshada iyo dariiqyada jawaab-celinta dhacdada ka hor inta aanad miisaan.
Sii wad Sahaminta
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 Adam and Adaptive Optimizers 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
Hagaha xiga
ZeRO iyo Optimizers Sharded
Su'aalaha soo noqnoqda
What is Adam and Adaptive Optimizers?
Aadam waa farsameeyaha shaqada faraska ka dambeeya inta badan shabakadaha neerfaha ee casriga ah, isaga oo si toos ah u hagaajinaya heerka waxbarasho ee u gaarka ah cabbir kasta. Waa arrin sababtoo ah waxay ka dhigaysaa tababarka moodooyinka qoto dheer si ka dhakhso badan oo aad uga liita kuwa ka soo jeeda jaan-goynta cad.
Waa maxay labada qiyaasood ee uu Aadan ku raad-joogo halbeeg kasta?
Aadam waxa uu hayaa celceliska jajaban ee gradients (daqiiqada ugu horeysa) iyo jaangooyooyinka labajibbaaran (daqiiqad labaad) ee cabbir kasta.
Muxuu Aadam ugu dabaqayaa sixitaanka eexda qiyaasihiisa xilligan?
M iyo v labaduba waxay ku bilowdeen eber, sidaas darteed qiyaasaha hore waa eex hooseeya; qaybinta (1 - beta^t) ayaa saxaysa tan.
Waa maxay faraqa ugu weyn ee u dhexeeya Aadan iyo AdamW?
AdamW waxa uu si toos ah ugu dabaqaa qudhunka miisaanka miisaanka halkii uu ku qasi lahaa ereyga isbedbeddelka, kaas oo wanaajinaya guud ahaan beddelka.
Doorkee ayuu ereyga yar ee epsilon ka ciyaartaa qaanuunka cusboonaysiinta Aadan?
Epsilon (qiyaastii 1e-8) ayaa lagu daraa hooseeyaha si halbeegyada ku dhow-eber-jibbaaran-jibbaaran aysan u keenin qarax.
Waa maxay sababta Aadan inta badan loogu tilmaamo 'laqabsi'?
Marka loo qaybiyo xididka labajibbaaran ee halbeeg kasta oo isku celcelis-jibbaaran, Aadan waxa uu cabbiraa cabbirka cabbirka halbeeggii halkii uu isticmaali lahaa hal qiime caalami ah.