Shabakadda neerfaha cusub ee Google aad ayey uga saxsan tahay ugana dhaqso badan tahay analoogyada caanka ah

Shabakadaha neerfaha ee Convolutional (CNNs), oo ay dhiiri galiyeen hababka bayooloji ee kortexka aragga ee bini'aadamka, ayaa si fiican ugu habboon hawlaha sida shay iyo aqoonsiga wejiga, laakiin hagaajinta saxnimadooda waxay u baahan tahay daal iyo hagaajin. Taasi waa sababta saynisyahannada Google AI Research ay u sahamiyaan moodallo cusub oo cabbiraya CNN-yada qaab "ka habaysan". Waxay daabaceen natiijada shaqadooda maqaal "EfficientNet: Dib-u-Fikirka Qaabka Qaabaynta Isku-xidhka Neural-ka," ayaa lagu dhajiyay bogga sayniska ee Arxiv.org, iyo sidoo kale in daabacadaha on your blog. Qorayaashu waxay sheeganayaan in qoyska nidaamyada sirdoonka macmalka ah, ee loo yaqaan EfficientNets, ay ka sarreeyaan saxnaanta CNN-yada caadiga ah waxayna kordhiyaan waxtarka shabakada neerfaha ilaa 10 jeer.

"Dhaqanka caadiga ah ee moodooyinka la isku qurxiyo waa in si aan macquul ahayn loo kordhiyo qoto dheer ama ballaca CNN, oo loo isticmaalo xallinta sare ee sawirka gelinta ee tababarka iyo qiimeynta," qor injineer software Mingxing Tan iyo Google AI ee saynisyahanka Quoc V .Le). "Si ka duwan hababka dhaqameed ee si aan kala sooc lahayn u cabbira xuduudaha shabakada sida ballaca, qoto dheer, iyo xallinta wax gelinta, habkayagu wuxuu si isku mid ah u cabbiraa cabbir kasta oo leh qaybo go'an oo cabbiraya."

Si loo sii wanaajiyo waxqabadka, cilmi-baarayaashu waxay u ololeeyaan adeegsiga shabakad cusub oo laf-dhabarta, is-bedbeddelka dhalo roga ee mobilada (MBConv), kaas oo u adeega sida aasaaska qoyska EfficientNets moodooyinka.

Tijaabooyin, EfficientNets waxay soo bandhigeen saxnaan sare iyo hufnaan ka wanaagsan CNN-yada jira, iyadoo hoos u dhigaysa cabbirka cabbirka iyo shuruudaha agabka xisaabinta iyadoo loo eegayo siday u kala sarreeyaan. Mid ka mid ah moodooyinka, EfficientNet-B7, ayaa muujiyay 8,4 jeer ka yar cabbirka iyo 6,1 jeer waxqabad ka wanaagsan kan caanka ah ee CNN Gpipe, sidoo kale wuxuu ku guuleystay 84,4% iyo 97,1% saxnaanta (Top-1 iyo Top-5). 50 natiijadii) ImageNet set. Marka la barbar dhigo CNN-ka caanka ah ee ResNet-4, nooc kale oo EfficientNet ah, EfficientNet-B82,6, oo isticmaalaya ilo la mid ah, ayaa gaadhay saxnaanta 76,3% iyo 50% ee ResNet-XNUMX.

Moodooyinka EfficientNets waxay si fiican u sameeyeen xog-ururinta kale, iyagoo gaaray saxnaansho sare shan ka mid ah sideed bartilmaameed, oo ay ku jiraan xogta CIFAR-100 (saxnimada 91,7%) iyo ubax (98,8%).

"Iyadoo la siinayo horumarro la taaban karo oo ku saabsan hufnaanta moodooyinka neerfaha, waxaan fileynaa in EfficientNets ay awood u leedahay inay u adeegto qaab cusub oo loogu talagalay hawlaha aragga kombiyuutarka mustaqbalka," Tan iyo Li ayaa qoray.

Koodhka isha iyo qoraalada tababarka Google's Cloud Tensor Processing Units (TPUs) ayaa si xor ah loo heli karaa Github.

Source: 3dnews.ru