Convolutional neural networks (CNNs), inofemerwa nebiological process mune yemunhu inoona cortex, inonyatsokodzera mabasa akadai sechinhu uye kucherechedzwa kwechiso, asi kunatsiridza kurongeka kwavo kunoda kunetesa uye kunyatsogadzirisa. Ndosaka masayendisiti kuGoogle AI Research ari kuongorora mhando nyowani dzinoyera CNNs nenzira "yakarongeka". Vakaburitsa mhedzisiro yebasa ravo mukati "EfficientNet: Rethinking Model Scalling for Convolutional Neural Networks," yakatumirwa pasainzi portal Arxiv.org, pamwe nemu pa blog yako. Vanyori-vanyori vanoti mhuri yeartificial intelligence system, inonzi EfficientNets, inodarika huroyi hwemhando dzeCNNs uye inowedzera kushanda kweneural network nekusvika kagumi.
"Tsika yakajairika yekuyera mamodheru ndeyekuwedzera kudzika kana hupamhi hweCNN, uye kushandisa resolution yepamusoro yemufananidzo wekupinza pakudzidziswa uye kuongororwa," nyora yevashandi software injiniya Mingxing Tan uye Google AI anotungamira sainzi Quoc V .Le). "Kusiyana nemaitiro echinyakare ayo anoyera zvisina tsarukano maparamita etiweki sehupamhi, kudzika, uye kugadzirisa kwekuisa, nzira yedu inoyera saizi imwe neimwe seti yakatarwa yezviyero zvinhu."
Kuti uenderere mberi nekuvandudza mashandiro, vaongorori vanotsigira kushandisa nyowani yemusana network, mobile inverted bottleneck convolution (MBConv), iyo inoshanda sehwaro hweEfficientNets mhuri yemamodheru.
Mumiedzo, EfficientNets yakaratidza zvese zviri zviviri huchokwadi hwepamusoro uye kushanda zvirinani pane dziripo CNNs, kudzikisa paramende saizi uye computational zviwanikwa zvinodiwa nekuraira kwehukuru. Imwe yemamodheru, EfficientNet-B7, yakaratidza diki zvakapetwa ka8,4 uye ka6,1 kuita zvirinani pane ine mukurumbira CNN Gpipe, uye yakawana 84,4% uye 97,1% chokwadi (Pamusoro-1 uye Pamusoro-5). 50 mhedzisiro) mukuyedza pa iyo ImageNet set. Zvichienzaniswa neyakakurumbira CNN ResNet-4, imwe EfficientNet modhi, EfficientNet-B82,6, ichishandisa zviwanikwa zvakafanana, yakawana huchokwadi hwe76,3% maringe ne50% yeResNet-XNUMX.
EfficientNets modhi dzakaita nemazvo pane mamwe madataset, ikawana huchokwadi hwepamusoro pamabenchmarks mashanu pamasere, kusanganisira iyo CIFAR-100 dataset (91,7% chokwadi) uye (98,8%).
"Nekupa kuvandudzwa kwakakosha mukushanda kwemaitiro emagetsi, tinotarisira kuti EfficientNets ine mukana wekushanda sehurongwa hutsva hwemabasa ekuona komputa," Tan naLi vanonyora.
Kwakabva kodhi uye zvinyorwa zvekudzidzisa zvegore reGoogle Tensor Processing Units (TPUs) zvinowanikwa pachena pa .
Source: 3dnews.ru