Amanethiwekhi e-Convolutional neural network (CNNs), agqugquzelwe izinqubo zebhayoloji ku-cortex ebonakalayo yomuntu, ayifanele kahle imisebenzi efana nokubonwa kwento nobuso, kodwa ukuthuthukisa ukunemba kwawo kudinga ukucushwa okuyisicefe. Kungakho ososayensi ku-Google AI Research behlola amamodeli amasha akala ama-CNN ngendlela "eyakheke kakhulu". Bashicilela imiphumela yomsebenzi wabo ku "I-EfficientNet: Imodeli Yokucabanga Kabusha Yenethiwekhi Ye-Convolutional Neural," okuthunyelwe kungosi yesayensi i-Arxiv.org, kanye naku- kubhulogi yakho. Ababhali ababambisene bathi umndeni wezinhlelo zobuhlakani bokwenziwa, ezibizwa nge-EfficientNets, weqa ukunemba kwama-CNN ajwayelekile futhi ukhuphula ukusebenza kahle kwenethiwekhi ye-neural izikhathi ezingafika kwezi-10.
"Umkhuba ojwayelekile wamamodeli wokukala uwukukhuphula ngokungafanele ukujula noma ububanzi be-CNN, nokusebenzisa ukulungiswa okuphezulu kwesithombe sokufaka ukuze siqeqeshwe futhi sihlolwe," bhala unjiniyela wesofthiwe yabasebenzi uMingxing Tan kanye nososayensi oholayo we-Google AI u-Quoc V .Le). “Ngokungafani nezindlela ezivamile ezikala amapharamitha enethiwekhi ngokungafanele njengobubanzi, ukujula, kanye nokulungiswa kokufaka, indlela yethu ikala ngokulinganayo ubukhulu bonke ngesethi engashintshi yezinto zokukala.”
Ukuze kuthuthukiswe ukusebenza kahle, abacwaningi bakhuthaza ukusebenzisa inethiwekhi entsha yomgogodla, i-mobile inverted bottleneck convolution (MBConv), esebenza njengesisekelo somndeni we-EfficientNets wamamodeli.
Ekuhlolweni, i-EfficientNets ibonise kokubili ukunemba okuphezulu nokusebenza kangcono kunama-CNN akhona, yehlisa usayizi wepharamitha nezimfuneko zensiza yokubala nge-oda lobukhulu. Enye yamamodeli, i-EfficientNet-B7, ibonise usayizi omncane ngokuphindwe ka-8,4 nokusebenza okungcono ngokuphindwe izikhathi ezingu-6,1 kune-CNN Gpipe edumile, futhi yazuza nokunemba okungu-84,4% no-97,1% (Okuphezulu koku-1 nokuPhezulu-5). Imiphumela emi-50) ekuhlolweni Isethi ye-ImageNet. Uma kuqhathaniswa ne-CNN ResNet-4 edumile, enye imodeli ye-EfficientNet, i-EfficientNet-B82,6, esebenzisa izinsiza ezifanayo, ithole ukunemba okungu-76,3% uma kuqhathaniswa nokungu-50% kwe-ResNet-XNUMX.
Amamodeli we-EfficientNets asebenze kahle kwamanye amasethi edatha, afinyelela ukunemba okuphezulu kumabhentshimakhi amahlanu kwayisishiyagalombili, okuhlanganisa nesethi yedatha ye-CIFAR-100 (ukunemba okungu-91,7%) kanye (98,8%).
"Ngokuhlinzeka ngokuthuthukiswa okuphawulekayo ekusebenzeni kahle kwamamodeli we-neural, silindele ukuthi i-EfficientNets inamandla okusebenza njengohlaka olusha lwemisebenzi yombono wekhompyutha yesikhathi esizayo," u-Tan no-Li babhala.
Ikhodi yomthombo nezikripthi zokuqeqesha ze-Google Tensor Processing Units (TPUs) zitholakala mahhala .
Source: 3dnews.ru