Neural network yatsopano ya Google ndiyolondola kwambiri komanso yachangu kuposa ma analogue otchuka

Ma Convolutional neural network (CNNs), motsogozedwa ndi njira zachilengedwe mu kotekisi yamunthu, ndioyenera kugwira ntchito monga kuzindikira kwa chinthu ndi nkhope, koma kuwongolera kulondola kwawo kumafuna zotopetsa komanso zokonzekera bwino. Ichi ndichifukwa chake asayansi ku Google AI Research akuwunika mitundu yatsopano yomwe imakulitsa ma CNN mwanjira "yokhazikika". Iwo adasindikiza zotsatira za ntchito yawo mu nkhani "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," yolembedwa pa tsamba lasayansi la Arxiv.org, komanso mu zofalitsa pa blog yanu. Olemba anzawowo akuti banja la ma intelligence system, otchedwa EfficientNets, limaposa kulondola kwa CNNs wamba ndikuwonjezera mphamvu ya neural network mpaka ka 10.

"Mchitidwe wamba wa makulitsidwe amitundu ndikuwonjezera mopanda kuzama kapena m'lifupi mwa CNN, ndikugwiritsa ntchito kusamvana kwapamwamba kwa chithunzi chothandizira pakuphunzitsidwa ndikuwunika," alemba olemba mapulogalamu a antchito Mingxing Tan ndi wasayansi wamkulu wa Google AI Quoc V .Le). "Mosiyana ndi njira zachikale zomwe zimakulitsa mopanda malire magawo a netiweki monga m'lifupi, kuya, ndi kusintha kolowera, njira yathu imayang'ana mbali iliyonse ndi zinthu zokhazikika."

Kuti apititse patsogolo magwiridwe antchito, ofufuzawo amalimbikitsa kugwiritsa ntchito netiweki yatsopano yam'mbuyo, mobile inverted bottleneck convolution (MBConv), yomwe imakhala ngati maziko a banja la EfficientNets.

M'mayeso, EfficientNets yawonetsa kulondola kwapamwamba komanso kuchita bwino kuposa ma CNN omwe alipo, kuchepetsa kukula kwa magawo ndi zofunikira zowerengera motengera kukula kwake. Imodzi mwamitunduyi, EfficientNet-B7, idawonetsa kucheperako ka 8,4 ndikuchita bwinoko ka 6,1 kuposa CNN Gpipe yotchuka, komanso idakwanitsa 84,4% ndi 97,1% kulondola (Pamwamba-1 ndi Pamwamba-5). 50 zotsatira) pakuyesa pa chithunzi cha ImageNet. Poyerekeza ndi CNN ResNet-4 yotchuka, mtundu wina wa EfficientNet, EfficientNet-B82,6, wogwiritsa ntchito zofananira, adapeza zolondola za 76,3% motsutsana ndi 50% za ResNet-XNUMX.

Mitundu ya EfficientNets idachita bwino pamaseti ena, kukwaniritsa zolondola kwambiri pamabenchmark asanu mwa asanu ndi atatu, kuphatikiza CIFAR-100 dataset (91,7% kulondola) ndi maluwa (98,8%).

"Popereka kusintha kwakukulu pakuchita bwino kwa ma neural models, tikuyembekeza kuti EfficientNets ili ndi kuthekera kokhala ngati chimango chatsopano cha ntchito zamtsogolo zamakompyuta," Tan ndi Li alemba.

Khodi yochokera ndi zolemba zophunzitsira za Google's cloud Tensor Processing Units (TPUs) zimapezeka kwaulere pa Github.

Source: 3dnews.ru