Sabuwar hanyar sadarwa ta Google ta fi dacewa da sauri fiye da sanannun analogues

Cibiyoyin sadarwa na jijiyoyi (CNNs), waɗanda aka yi wahayi ta hanyar hanyoyin nazarin halittu a cikin baƙon gani na ɗan adam, sun dace da ayyuka kamar su gane abu da fuskar fuska, amma haɓaka daidaitonsu yana buƙatar mai daɗaɗawa da daidaitawa. Shi ya sa masana kimiyya a Google AI Research ke binciko sabbin samfura waɗanda ke auna CNN ta hanyar "mafi tsari". Sun buga sakamakon aikinsu a ciki labarin "EfficientNet: Sake Tunanin Model Scaling for Convolutional Neural Networks," wanda aka buga akan tashar kimiyya ta Arxiv.org, da kuma a cikin wallafe a kan blog. Marubutan sun yi iƙirarin cewa dangin tsarin bayanan ɗan adam, da ake kira EfficientNets, sun zarce daidaiton daidaitattun CNNs kuma suna ƙara haɓakar hanyar sadarwa ta jijiyoyi har sau 10.

"Aiki na yau da kullun na ƙirar ƙira shine ƙara girman zurfin ko nisa na CNN ba da gangan ba, da amfani da ƙuduri mafi girma na hoton shigarwa don horo da kimantawa," rubuta injiniyan software na ma'aikata Mingxing Tan da masanin kimiyyar Google AI Quoc V .Le). "Ba kamar hanyoyin gargajiya waɗanda ke daidaita ma'auni na cibiyar sadarwa ba bisa ka'ida ba kamar nisa, zurfin, da ƙudurin shigarwa, hanyarmu tana daidaita kowane girma tare da ƙayyadaddun abubuwan ƙima."

Don ƙara haɓaka aiki, masu binciken suna ba da shawarar yin amfani da sabuwar hanyar sadarwa ta kashin baya, juyin juya halin ƙwarya ta wayar hannu (MBConv), wanda ke zama tushen tushen EfficientNets na samfura.

A cikin gwaje-gwaje, EfficientNets ya nuna duka daidaito mafi girma da inganci fiye da CNNs na yanzu, rage girman sigina da buƙatun albarkatun lissafi ta tsari mai girma. Ɗaya daga cikin samfuran, EfficientNet-B7, ya nuna ƙarami sau 8,4 da mafi kyawun aiki sau 6,1 fiye da sanannen CNN Gpipe, kuma ya sami daidaiton 84,4% da 97,1% (Top-1 da Top-5). 50 sakamako) a gwaji akan saitin ImageNet. Idan aka kwatanta da sanannen CNN ResNet-4, wani samfurin EfficientNet, EfficientNet-B82,6, ta amfani da albarkatu iri ɗaya, ya sami daidaiton 76,3% da 50% na ResNet-XNUMX.

Samfuran EfficientNets sun yi kyau akan sauran bayanan bayanai, suna samun daidaito mai girma akan biyar daga cikin maƙasudai takwas, gami da saitin bayanan CIFAR-100 (daidaicin 91,7%) da Flowers (98,8%).

"Ta hanyar samar da ci gaba mai mahimmanci a cikin ingantaccen tsarin jijiyoyi, muna sa ran cewa EfficientNets yana da damar yin aiki a matsayin sabon tsari don ayyukan hangen nesa na kwamfuta na gaba," Tan da Li sun rubuta.

Lambar tushe da rubutun horarwa na Google's girgije Tensor Processing Units (TPUs) ana samun su kyauta Github.

source: 3dnews.ru