Neural network e ncha ea Google e nepahetse haholo ebile e potlakile ho feta li-analogue tse tsebahalang

Convolutional neural networks (CNNs), e susumetsoang ke ts'ebetso ea likokoana-hloko ka har'a cortex ea pono ea motho, e loketse hantle mesebetsi e kang ntho le ho lemoha sefahleho, empa ho ntlafatsa ho nepahala ha tsona ho hloka ho khathatsa le ho lokisoa hantle. Ke ka lebaka leo bo-rasaense ba Google AI Research ba ntseng ba hlahloba mefuta e mecha e lekanyang li-CNN ka tsela e "hlophisitsoeng haholoanyane". Ba phatlalalitse liphetho tsa mosebetsi oa bona ka sehlooho "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," e ngotsoe ho portal ea saense ea Arxiv.org, hammoho le ho lingoliloeng ho blog ea hau. Bangoli-'moho ba bolela hore lelapa la litsamaiso tsa bohlale ba maiketsetso, tse bitsoang EfficientNets, li feta ho nepahala ha li-CNN tse tloaelehileng 'me li eketsa katleho ea marang-rang a methapo ka makhetlo a 10.

"Tloaelo e tloaelehileng ea ho phahamisa mefuta ke ho eketsa botebo kapa bophara ba CNN, le ho sebelisa qeto e phahameng ea setšoantšo se kentsoeng bakeng sa koetliso le tlhahlobo," ho ngola moenjiniere oa software ea basebetsi Mingxing Tan le rasaense ea ka sehloohong oa Google AI Quoc V .Le). "Ho fapana le mekhoa e tloaelehileng e lekanyang liparamente tsa marang-rang tse kang bophara, botebo, le tharollo ea ho kenya letsoho, mokhoa oa rona o lekanya boemo bo bong le bo bong ka mokhoa o tsitsitseng oa lintlha."

Ho tsoela pele ho ntlafatsa ts'ebetso, bafuputsi ba buella ho sebelisa marang-rang a macha a mokokotlo, mobile inverted bottleneck convolution (MBConv), e sebetsang e le motheo oa lelapa la EfficientNets la mehlala.

Litekong, EfficientNets e bonts'itse ho nepahala le ho sebetsa hantle ho feta li-CNN tse seng li ntse li le teng, ho fokotsa boholo ba paramethara le litlhoko tsa lisebelisoa tsa computational ka tatellano ea boholo. E 'ngoe ea mehlala, EfficientNet-B7, e bonts'itse boholo bo nyane ka makhetlo a 8,4 le ts'ebetso e ntle ka makhetlo a 6,1 ho feta CNN Gpipe e tsebahalang, hape e fihletse ho nepahala ha 84,4% le 97,1% (Top-1 le Top-5). Liphetho tse 50) tlhahlobong setšoantšo sa ImageNet. Ha ho bapisoa le CNN ResNet-4 e tsebahalang, mofuta o mong oa EfficientNet, EfficientNet-B82,6, o sebelisa lisebelisoa tse tšoanang, o fihletse ho nepahala ha 76,3% khahlano le 50% bakeng sa ResNet-XNUMX.

Mefuta ea EfficientNets e sebelitse hantle ho li-dataset tse ling, e fihletse ho nepahala ho holimo ho li-benchmarks tse hlano ho tse robeli, ho kenyeletsoa le CIFAR-100 dataset (91,7% ho nepahala) le lipalesa (98,8%).

"Ka ho fana ka lintlafatso tse kholo ts'ebetsong ea mekhoa ea li-neural, re lebeletse hore EfficientNets e na le monyetla oa ho sebetsa e le moralo o mocha oa mesebetsi ea nakong e tlang ea pono ea k'homphieutha," Tan le Li ba ngola.

Mohloli oa khoutu le mangolo a koetliso bakeng sa Google's cloud Tensor Processing Units (TPUs) li fumaneha mahala ho Github.

Source: 3dnews.ru