Google tus tshiab neural network yog qhov tseeb thiab nrawm dua li cov analogues nrov

Convolutional neural networks (CNNs), tau txais kev tshoov siab los ntawm cov txheej txheem lom neeg hauv tib neeg lub ntsej muag cortex, zoo haum rau kev ua haujlwm xws li khoom thiab lub ntsej muag lees paub, tab sis kev txhim kho lawv qhov tseeb yuav tsum tau tedious thiab zoo-tuning. Tias yog vim li cas cov kws tshawb fawb ntawm Google AI Research tab tom tshawb nrhiav cov qauv tshiab uas ntsuas CNNs hauv "kev tsim qauv ntau dua". Lawv luam tawm cov txiaj ntsig ntawm lawv txoj haujlwm hauv Tshooj "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," tshaj tawm rau ntawm lub vev xaib tshawb fawb Arxiv.org, nrog rau hauv cov ntawv luam tawm ntawm koj qhov blog. Cov kws sau ntawv tau lees paub tias tsev neeg ntawm kev txawj ntse txawj ntse, hu ua EfficientNets, tshaj qhov tseeb ntawm CNNs tus qauv thiab ua rau kom muaj txiaj ntsig ntawm neural network nce txog 10 npaug.

"Qhov kev coj ua ntawm kev ntsuas cov qauv yog ua kom qhov tob lossis dav ntawm CNN, thiab siv cov kev daws teeb meem siab dua ntawm cov duab nkag rau kev cob qhia thiab kev ntsuas," sau cov neeg ua haujlwm software engineer Mingxing Tan thiab Google AI tus kws tshawb fawb Quoc V .Le). "Txawm li cas los xij, tsis zoo li cov txheej txheem ib txwm muaj uas tau txiav txim siab ntsuas qhov tsis sib xws xws li qhov dav, qhov tob, thiab kev daws teeb meem, peb cov txheej txheem sib npaug ntawm txhua qhov loj me nrog cov txheej txheem scaling.

Txhawm rau txhim kho kev ua tau zoo ntxiv, cov kws tshawb fawb tau tawm tswv yim siv lub pob txha caj qaum tshiab, mobile inverted bottleneck convolution (MBConv), uas yog lub hauv paus rau EfficientNets tsev neeg ntawm cov qauv.

Hauv kev ntsuam xyuas, EfficientNets tau pom tias muaj qhov tseeb dua thiab ua haujlwm zoo dua li CNNs uas twb muaj lawm, txo qis qhov loj me thiab cov kev xav tau ntawm kev suav los ntawm qhov kev txiav txim loj. Ib qho ntawm cov qauv, EfficientNet-B7, tau pom 8,4 npaug me me thiab 6,1 npaug ntawm kev ua tau zoo dua li lub npe nrov CNN Gpipe, thiab tseem ua tiav 84,4% thiab 97,1% qhov tseeb (Top-1 thiab Top-5). 50 qhov tshwm sim) hauv kev sim ntawm ImageNet teeb tsa. Piv rau CNN nrov ResNet-4, lwm tus qauv EfficientNet, EfficientNet-B82,6, siv cov peev txheej zoo sib xws, ua tiav qhov tseeb ntawm 76,3% piv rau 50% rau ResNet-XNUMX.

EfficientNets cov qauv ua tau zoo ntawm lwm cov ntaub ntawv, ua tiav qhov raug siab ntawm tsib ntawm yim qhov ntsuas, suav nrog CIFAR-100 dataset (91,7% raug) thiab paj (98,8%).

"Los ntawm kev muab kev txhim kho tseem ceeb hauv kev ua tau zoo ntawm cov qauv neural, peb cia siab tias EfficientNets muaj peev xwm los ua lub hauv paus tshiab rau kev ua haujlwm hauv computer pom yav tom ntej," Tan thiab Li sau.

Cov lej cim thiab cov ntawv qhia kev cob qhia rau Google's huab Tensor Processing Units (TPUs) muaj pub dawb rau ntawm github.

Tau qhov twg los: 3d xov.ru