Ragewar Kaifi-Arewa
Sharpness-Aware Minimization (SAM) hanya ce ta ingantawa wacce ke neman ba kawai asara ba amma ƙarancin asara a duk faɗin unguwar ma'aunin nauyi - mafi ƙanƙanta.
Dubawa
Flatter minima tend to generalize better, so SAM often improves test accuracy and robustness without changing the model architecture.
Zurfafa nutsewa
Daidaitaccen horo yana rage hasara a wuri ɗaya a cikin sararin nauyi, amma mafita guda biyu tare da asarar horo iri ɗaya na iya zama daban-daban: mafi ƙarancin 'kaifi' yana zaune a cikin kunkuntar kwari inda ƙananan ƙarancin nauyi ke haɓaka asarar, yayin da mafi ƙarancin 'lebur' yana jure ɓarna kuma yawanci yana haɓaka mafi kyawun bayanan gaibu. SAM, wanda Google masu bincike suka gabatar a cikin 2020, ya bayyana wannan a sarari. A kowane mataki ya fara gano ma'aunin nauyi na kusa (a cikin ƙaramin radius rho) wanda ke haɓaka asarar - maƙwabcin mafi munin yanayi - sannan ya sabunta ma'aunin asali na asali don rage asarar a waccan maƙasudin. Wannan makasudin min-max yana tura haɓakawa zuwa yankuna waɗanda ba su da ƙanƙanta iri ɗaya, suna ba da cikakkiyar fa'ida akan rarrabuwar hoto da ƙari.
Fahimtar Fasaha
Kowane matakin SAM wucewa biyu ne. Da farko, lissafta gradient a ma'aunin halin yanzu kuma ɗauki matakin 'hankali' na girman rho a cikin alƙawarin gradient don isa wurin mafi muni a kusa. Na biyu, lissafta gradient a waccan wuri mai ruɗani kuma amfani da shi don sabunta ma'aunin asali. Radius rho yana sarrafa girman girman unguwar da kuke kariya. Kudin yana kusan wucewa biyu gaba-baya kowane mataki, wanda ke ninka lissafin - babban koma baya mai amfani.
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 Ragewar Kaifi-Aware
SAM ya haifar da dangin bin diddigin da nufin babban rauninsa, ƙididdige ninki biyu: ingantattun bambance-bambancen kamar ESAM, LookSAM, da hanyoyin da ke dagula ma'aunin nauyi kawai ko amfani da SAM kowane ƴan matakai. Adaptive SAM (ASAM) yana sake daidaita radius don zama marar bambanci. Masu bincike sun ci gaba da yin muhawara daidai dalilin da yasa flatness ke taimakawa da yadda za a auna shi, kuma ra'ayoyi masu fa'ida suna yaduwa zuwa kyakkyawan daidaita manyan nau'ikan harshe da haɓaka ƙarfi ga canjin rarraba.
Aiwatar da Gaskiyar Duniya
Haɓaka Mai Canjin hangen nesa da daidaiton ResNet akan ImageNet ta horo tare da SAM maimakon SGD bayyananne.
Haɓaka ƙarfi don yin lakabin amo, tun da ƙananan ƙananan ba su da yuwuwar haddar gurɓatattun takalmi.
Kyawawan tsarin ƙirar harshe da aka riga aka horar da su tare da SAM don samun ingantacciyar fahimta akan ƙananan bayanan bayanan ƙasa.
Amfani da bambance-bambancen ESAM ko LookSAM lokacin da aka ninka lissafin kuɗin vanilla SAM yayi tsada sosai.
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 Sharpness-Aware Minimization 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
DenseNet da Haɗin Haɗin
Tambayoyin da ake yawan yi
What is Sharpness-Aware Minimization?
Sharpness-Aware Minimization (SAM) hanya ce ta ingantawa wacce ke neman ba kawai asara ba amma ƙarancin asara a duk faɗin unguwar ma'aunin nauyi - mafi ƙanƙanta. Flatter minima yakan zama mafi kyawu, don haka SAM sau da yawa yana haɓaka daidaiton gwaji da ƙarfi ba tare da canza ƙirar ƙirar ƙira ba.
Wane irin ƙarami ne SAM ke ƙoƙarin nema?
SAM ya yi niyya ga ƙaramin ƙarami saboda asarar da ta rage ƙasa ƙarƙashin ƙananan ruɗar nauyi tana ƙoƙarin haɓaka mafi kyawun bayanan gaibu.
Fassara nawa gaba-baya nawa daidaitaccen matakin SAM ke buƙata?
SAM yana ƙididdige gradient don nemo mafi munin yanayi a kusa, sannan wani gradient a can don sabuntawa - kusan ninki biyu farashin da aka saba.
Menene radius hyperparameter rho ke sarrafawa a cikin SAM?
Rho ya saita nisan matakin ' hawan' don nemo makwabcin mafi munin yanayi, yana bayyana girman girman yanki na SAM.
Menene farkon matakai biyu na SAM a kowace maimaitawa?
SAM da farko yana dagula ma'aunin nauyi a cikin radius rho a cikin alkiblar da ta fi girma asara, sannan ta sauko daga ainihin ma'ana ta amfani da gradient da aka lissafta a wurin.
Me yasa SAM sau da yawa yana inganta ƙarfi don yin lakabin amo?
Ƙwaƙwalwar labulen hayaniya yawanci yana buƙatar kaifi, kunkuntar ƙarami; ta hanyar fifita yankuna masu fa'ida, SAM yana ƙin cewa wuce gona da iri.