I-Convolutional neural networks (CNNs), ephefumlelwe yinkqubo yebhayoloji kwi-cortex ebonakalayo yomntu, ifaneleke kakuhle imisebenzi efana nento kunye nokuqatshelwa kobuso, kodwa ukuphucula ukuchaneka kwazo kufuna ukudinwa kunye nokulungiswa kakuhle. Yiyo loo nto izazinzulu kuphando lweGoogle AI ziphonononga iimodeli ezintsha ezikala ii-CNN ngendlela "eyakhiwe ngakumbi". Bapapashe iziphumo zomsebenzi wabo kwi "I-EfficientNet: UkuCinga kwakhona iModeli yokuKhawulwa kweeNethiwekhi zeNeural," ifakwe kwi-portal yesayensi i-Arxiv.org, kunye nakwi- kwiblogi yakho. Ababhali ababambeneyo babanga ukuba usapho lweenkqubo zobukrelekrele bokwenziwa, ezibizwa ngokuba yi-EfficientNets, idlula ukuchaneka kwee-CNNs eziqhelekileyo kwaye yandisa ukusebenza kakuhle kwenethiwekhi ye-neural ukuya kuthi ga kwi-10 amaxesha.
"Umkhuba oqhelekileyo weemodeli zokulinganisa kukunyusa ngokungenasizathu ubunzulu okanye ububanzi be-CNN, kwaye usebenzise isisombululo esiphezulu somfanekiso wegalelo loqeqesho kunye nokuvavanya," bhala injineli yesoftware yabasebenzi uMingxing Tan kunye nososayensi okhokelayo weGoogle AI uQuoc V.Le). "Ngokungafaniyo neendlela eziqhelekileyo ezilinganisa ngokungenamkhethe iiparamitha zenethiwekhi ezinje ngobubanzi, ubunzulu, kunye nesisombululo segalelo, indlela yethu ilinganisa ngokulinganayo umlinganiso ngamnye ngeseti emiselweyo yezinto zokulinganisa."
Ukuqhubela phambili ukuphucula ukusebenza, abaphandi bakhuthaza ukusebenzisa inethiwekhi entsha yomqolo, i-mobile inverted bottleneck convolution (MBConv), esebenza njengesiseko sentsapho ye-EfficientNets yeemodeli.
Kwiimvavanyo, i-EfficientNets ibonise ukuchaneka okuphezulu kunye nokusebenza okungcono kunee-CNN ezikhoyo, ukunciphisa ubungakanani beparameter kunye neemfuno zezixhobo zokubala ngomyalelo wobukhulu. Enye yeemodeli, i-EfficientNet-B7, ibonise amaxesha angama-8,4 amancinci kunye namaxesha angama-6,1 ukusebenza ngcono kune-CNN Gpipe eyaziwayo, kwaye iphumelele i-84,4% kunye ne-97,1% ngokuchanekileyo (i-Top-1 kunye ne-Top-5) iziphumo ezi-50) kuvavanyo iseti ye-ImageNet. Xa kuthelekiswa ne-CNN ResNet-4 eyaziwayo, enye imodeli ye-EfficientNet, i-EfficientNet-B82,6, esebenzisa izixhobo ezifanayo, ifumene ukuchaneka kwe-76,3% ngokubhekiselele kwi-50% ye-ResNet-XNUMX.
Iimodeli ze-EfficientNets ziqhube kakuhle kwezinye iiseti zedatha, ziphumeza ukuchaneka okuphezulu kwiibenchmarks ezihlanu kwezisibhozo, kubandakanywa nedathasethi ye-CIFAR-100 (ukuchaneka okungama-91,7%) kunye (98,8%).
"Ngokubonelela ngophuculo olubalulekileyo ekusebenzeni kweemodeli ze-neural, silindele ukuba i-EfficientNets inamandla okusebenza njengesakhelo esitsha semisebenzi yombono wekhompyutheni yexesha elizayo," uTan noLi babhala.
Ikhowudi yomthombo kunye nezikripthi zoqeqesho zeTensor Processing Units zikaGoogle (TPUs) zifumaneka simahla kwi .
umthombo: 3dnews.ru