ʻOi aku ka pololei a me ka wikiwiki o ka pūnaewele neural hou a Google ma mua o nā analogues kaulana

ʻO nā pūnaewele neural Convolutional (CNNs), i hoʻoulu ʻia e nā kaʻina olaola i loko o ka cortex ʻike maka kanaka, ua kūpono i nā hana e like me ka mea a me ka ʻike maka, akā ʻo ka hoʻomaikaʻi ʻana i kā lākou pololei e pono ai ke hoʻoluhi a hoʻoponopono maikaʻi. ʻO ia ke kumu e ʻimi nei nā ʻepekema ma Google AI Research i nā hiʻohiʻona hou e hoʻohālikelike i nā CNN ma kahi ʻano "ʻoi aʻe". Ua hoʻopuka lākou i nā hopena o kā lākou hana ma 'ōlelo "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," i kau ʻia ma ka puka ʻepekema ʻo Arxiv.org, a me nā palapala ma kāu blog. Ua ʻōlelo nā hoa kākau ʻo ka ʻohana o nā ʻōnaehana naʻauao artificial, i kapa ʻia ʻo EfficientNets, ʻoi aku ka pololei o nā CNN maʻamau a hoʻonui i ka pono o kahi pūnaewele neural a hiki i 10 mau manawa.

"ʻO ka hana maʻamau o ka scaling kumu hoʻohālike e hoʻonui arbitrarily i ka hohonu a laula o ka CNN, a hoʻohana kiʻekiʻe hoʻonā o ke kiʻi hoʻokomo no ka hoʻonaʻauao a me ka loiloi 'ana," kakau lako polokalamu 'enekinia Mingxing Tan a me Google AI alakai 'epekema Quoc V .Le). "ʻAʻole e like me nā ala kuʻuna e hoʻohālikelike i nā ʻāpana pūnaewele e like me ka laula, ka hohonu, a me ka hoʻonā hoʻokomo ʻana, ʻo kā mākou ala e hoʻohālikelike i kēlā me kēia ʻāpana me kahi hoʻonohonoho paʻa o nā mea scaling."

No ka hoʻomaikaʻi hou ʻana i ka hana, ʻōlelo ka poʻe noiʻi i ka hoʻohana ʻana i kahi pūnaewele iwi hope hou, mobile inverted bottleneck convolution (MBConv), kahi kumu no ka ʻohana EfficientNets o nā hiʻohiʻona.

I nā hoʻāʻo, ua hōʻike ʻo EfficientNets i ka pololei kiʻekiʻe a me ka maikaʻi ʻoi aku ka maikaʻi ma mua o nā CNN i loaʻa, e hōʻemi ana i ka nui o nā ʻāpana a me nā koi waiwai helu helu ma kahi o ka nui. ʻO kekahi o nā hiʻohiʻona, ʻo EfficientNet-B7, hōʻike i ka 8,4 mau manawa ʻoi aku ka liʻiliʻi a me 6,1 mau manawa ʻoi aku ka maikaʻi ma mua o ka CNN Gpipe kaulana, a loaʻa pū i ka 84,4% a me 97,1% pololei (Top-1 a me Top-5). 50 hopena) i ka hoʻāʻo ʻana ma ka hoʻonohonoho ImageNet. Hoʻohālikelike ʻia me ka CNN ResNet-4 kaulana, ʻo kekahi kumu hoʻohālike EfficientNet, ʻo EfficientNet-B82,6, me ka hoʻohana ʻana i nā kumuwaiwai like, loaʻa ka pololei o 76,3% versus 50% no ResNet-XNUMX.

Hana maikaʻi ʻia nā hiʻohiʻona EfficientNets ma nā ʻikepili ʻē aʻe, e loaʻa ana ka pololei kiʻekiʻe ma nā ʻelima o ʻewalu mau hiʻohiʻona, me ka CIFAR-100 dataset (91,7% pololei) a Flowers (98,8%).

"Ma ka hoʻolako ʻana i nā hoʻomaikaʻi koʻikoʻi i ka maikaʻi o nā hiʻohiʻona neural, manaʻo mākou e hiki i ka EfficientNets ke lawelawe ma ke ʻano he hoʻolālā hou no nā hana ʻike kamepiula e hiki mai ana," kākau ʻo Tan lāua ʻo Li.

Loaʻa wale ʻia nā code kumu a me nā palapala hoʻomaʻamaʻa no ka Google Cloud Tensor Processing Units (TPUs). Github.

Source: 3dnews.ru