Cibiyoyin sadarwa na Matsi-da-Haɗa kai
Tubalan Squeeze-and-Excitation (SE) suna barin cibiyar sadarwa mai jujjuyawa ta koyi nawa za ta auna nauyin kowane tasha, ta sake daidaita su bisa yanayin duniya.
Dubawa
This cheap attention-like mechanism won the 2017 ImageNet competition and became a standard CNN building block.
Zurfafa nutsewa
Hu, Shen, da Sun ne suka gabatar da su a cikin 2017, toshe SE yana ƙara hankalin tashar ta musamman ga CNN. Yana aiki a matakai biyu. 'Matsi' yana amfani da matsakaita taɗi na duniya don ruguje kowane taswira (tsawo x faɗin) zuwa lamba ɗaya, yana samar da mai siffa guda ɗaya akan kowane tashoshi wanda ke taƙaita kunnawar sa a duniya. The 'excitation' yana ciyar da cewa vector ta hanyar ƙananan yadudduka masu cikakken haɗin gwiwa tare da ƙwanƙwasa (a ReLU sannan sigmoid) don samar da nauyin kowane tashar tsakanin 0 da 1. Waɗancan ma'aunin nauyi suna ninka taswirar fasalin asali na asali, haɓaka tashoshi masu amfani da damping waɗanda ba su da mahimmanci. SENet ta sami nasarar ƙalubalen rarraba ILSVRC 2017, yanke kuskuren saman-5 zuwa kusan 2.25%. Toshe yana ƙara ƙarin sigogi kaɗan kawai da ƙididdigewa, da ramummuka cikin ResNet, Inception, ko MobileNet tare da ƙaramin canji.
Fahimtar Fasaha
Matsi yana samar da vector z mai tsayin C inda z_c shine matsakaicin sarari na tashar c. Ƙaddamarwa mai ban sha'awa s = sigmoid (W2 * ReLU (W1 * z)), inda W1 ya rage girman ta hanyar raguwar rabo r (yawanci 16) kuma W2 yana mayar da shi, yana kiyaye ƙarin farashi. Fitowar ita ce taswirar shigar da sikelin taswirar tasha-hikima ta s. Wani nau'i ne na gating kai: hanyar sadarwa ta yanke shawara, daga kididdigar duniya, wace tashoshi ke da mahimmanci ga wannan takamaiman shigarwar.
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 Cibiyoyin Sadarwar Matsi-da-Farawa
SE tubalan suna rayuwa a cikin ingantattun gine-gine: EfficientNet da MobileNetV3 sun sanya su cikin tubalan ginin su. Ra'ayin ya haifar da dangin kulawa da kayayyaki, CBAM yana ƙara kulawar sararin samaniya, ECA-Net ya maye gurbin kwalban tare da juyin juya halin 1D mai arha, kuma waɗannan dabaru na sake fasalin nauyi a yanzu suna bayyana a cikin ganowa, rarrabuwa, har ma da wasu nau'ikan masu canza hangen nesa. Yi tsammanin kulawar tashar zai kasance madaidaicin madaidaicin farashi mai rahusa a duk inda rikice-rikice ya ci gaba.
Aiwatar da Gaskiyar Duniya
SENet ta sami nasarar ƙalubalen rarraba na ImageNet ILSVRC 2017 ta ƙara abubuwan SE zuwa kashin baya na ResNeXt.
EfficientNet da MobileNetV3 sun haɗa samfuran SE a cikin kowane toshe don haɓaka daidaito akan na'urorin hannu
Abubuwan gano abubuwa da ƙirar rarrabuwa suna saka shingen SE don jaddada tashoshi fasalin fasali
ECA-Net da CBAM suna ƙaddamar da ra'ayin SE tare da sake fasalin tashar mai rahusa ko sane
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 Squeeze-and-Excitation Networks 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
Cibiyoyin sadarwa na capsule
Tambayoyin da ake yawan yi
What is Squeeze-and-Excitation Networks?
Tubalan Squeeze-and-Excitation (SE) suna barin cibiyar sadarwa mai jujjuyawa ta koyi nawa za ta auna nauyin kowane tasha, ta sake daidaita su bisa yanayin duniya. Wannan tsari mai arha mai kama da hankali ya lashe gasar ImageNet ta 2017 kuma ya zama daidaitaccen ginin CNN.
Menene matakin 'matsi' a cikin toshe SE yake yi?
Matsin yana shafi matsakaitan haɗaɗɗiyar duniya, yana rushe kowace taswira zuwa ma'auni ɗaya a kowane tashoshi.
Menene fitowar matakin 'excitation'?
Farin ciki yana wuce matsewar vector ta cikin yadudduka FC guda biyu suna ƙarewa cikin sigmoid, yana samar da ma'aunin ma'aunin tashoshi.
Ta yaya ake amfani da ma'aunin tashin hankali ga taswirar fasalin asali?
Ana ninka taswirar fasalin kowane tashar ta hanyar koyon nauyin da aka koya, haɓakawa ko murkushe tashar.
Me yasa tashin hankali yayi amfani da kwalabe tare da raguwar raguwa (sau da yawa 16)?
Ƙaƙƙarfan kwalban yana raguwa sannan ya dawo da girman tashar, don haka shingen SE yana biyan kuɗi kaɗan kaɗan kawai.
Wace gasa SENet ta yi nasara a cikin 2017?
SENet ta sami nasarar ƙalubalen rarraba hoto na ILSVRC 2017, yana tura kuskuren saman-5 zuwa kusan 2.25%.