Amanethiwekhi Okucindezela Nokukhuthaza
Amabhulokhi we-Squee-and-Excitation (SE) avumela inethiwekhi ye-convolutional ukuthi ifunde ukuthi isiteshi ngasinye sesici singasikala kangakanani, sisilinganise kabusha ngokusekelwe kumongo womhlaba.
Uhlolojikelele
This cheap attention-like mechanism won the 2017 ImageNet competition and became a standard CNN building block.
I-Deep Dive
Yethulwe ngu-Hu, Shen, kanye neLanga ngo-2017, i-SE block yengeza ukunakwa kwesiteshi okusobala ku-CNN. Isebenza ngezinyathelo ezimbili. 'Ukuminyanisa' kusebenzisa ukuhlanganisa okumaphakathi komhlaba wonke ukuze kugoqwe imephu yesici ngasinye (ubude x ububanzi) ibe inombolo eyodwa, okukhiqiza incazelo eyodwa yesiteshi ngasinye efingqa ukwenziwa kwayo kusebenze emhlabeni jikelele. 'I-excitation' iphakela leyo ivekhtha ngezandlalelo ezimbili ezincane ezixhunywe ngokugcwele nge-bottleneck (i-ReLU bese i-sigmoid) ukuze kukhiqizwe isisindo somzila ngamunye phakathi kuka-0 no-1. Lezo zisindo ziphindaphinda amamephu wesici sangempela, zikhulise iziteshi eziwusizo futhi zidambise ezingabalulekile. I-SENet iwine inselele yokuhlukaniswa kwe-ILSVRC 2017, yehlisa iphutha eli-5 eliphezulu laya cishe ku-2.25%. Ibhulokhi ingeza amaphesenti ambalwa kuphela amapharamitha engeziwe kanye nokubala, futhi ingena ku-ResNet, Inception, noma i-MobileNet enoshintsho oluncane.
I-Technical Insight
Ukuminyanisa kukhiqiza ivekhtha engu-C yobude lapho i-z_c iyisilinganiso sendawo sesiteshi c. I-excitation computes s = sigmoid(W2 * ReLU(W1 * z)), lapho i-W1 yehlisa ubukhulu ngenani lokunciphisa u-r (ngokuvamile u-16) futhi i-W2 iyibuyisela, igcina izindleko ezingeziwe zizincane. Okukhiphayo kuyisici semephu yesici esikalwa ngobuhlakani besiteshi ngo-s. Kuwuhlobo lokuzigaxa wena ngokwakho: inethiwekhi iyanquma, kusukela kuzibalo zomhlaba, ukuthi yiziphi iziteshi ezibalulekile kulokhu kufaka phakathi.
I-Strategic Impact
Izindleko kanye nesabelomali
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka.
Izinqumo ezicacile
Imfundo yobuchwepheshe isiza amaqembu ukuthi akhethe isitaki esifanele, hhayi nje esisha.
Ukulawulwa kwekhwalithi
Izinketho ezingcono zobunjiniyela zinciphisa izehlakalo ezinokwethenjelwa ekukhiqizeni.
Ikusasa Lamanethiwekhi Okuminyanisa Nokuvusa Injabulo
Amabhulokhi e-SE aphila ngaphakathi kwezakhiwo ezisebenza kahle: I-EfficientNet ne-MobileNetV3 ziwashumeka kumabhulokhi wawo wokwakha. Umbono ukhulise umndeni wamamojula wokunaka, i-CBAM yengeza ukunakwa kwendawo, i-ECA-Net ithatha indawo yebhodlela nge-convolution eshibhile ye-1D, futhi lawa maqhinga okulungisa kabusha angasindi manje avela ekutholeni, ekuhlukaniseni, ngisho nakwamanye amahybrids okuguqula umbono. Lindela ukunakwa kwesiteshi ukuthi kuhlale kuyisilinganisi sokunemba esinezindleko eziphansi nomaphi lapho ukungqubuzana kuqhubeka khona.
Ukuqaliswa Komhlaba Wangempela
I-SENet iwine inselele yokuhlukaniswa kwe-ImageNet ILSVRC 2017 ngokwengeza amabhlogo we-SE kumgogodla we-ResNeXt
I-EfficientNet ne-MobileNetV3 zishumeka amamojula e-SE kuwo wonke amabhulokhi ukuze kukhuliswe ukunemba kumadivayisi eselula
Izitholi zezinto namamodeli okuhlukanisa afaka amabhulokhi e-SE ukuze agcizelele iziteshi zezici ezifundisayo
I-ECA-Net kanye ne-CBAM banweba umbono we-SE ngokulungiswa kabusha kwesiteshi okushibhile noma okuqaphela indawo
Izingozi & Guardrails
Ukuthuthukisa ibhentshimakhi eyodwa kungafihla ubuthakathaka obubanzi besistimu.
Izindleko zengqalasizinda nezokulungisa zivame ukubukelwa phansi.
Izikhala zokuphepha nokubonakala zingakhula njengoba izinhlelo ziba nzima kakhulu.
Ukuqalisa Umhlahlandlela
Chaza ukubambezeleka, ikhwalithi, nezindleko ezihlosiwe ngaphambi kokuqaliswa.
Ibhentshimakhi ngaphansi komthwalo wangempela nezimo zedatha.
Ukuqapha amathuluzi amaphutha, ukukhukhuleka, nomthelela wabasebenzisi.
Lungiselela izindlela zokuhlehlisa nezigameko ngaphambi kokukala.
Qhubeka Uhlole
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Squeeze-and-Excitation Networks quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Umhlahlandlela olandelayo
Amanethiwekhi weCapsule
Imibuzo evame ukubuzwa
What is Squeeze-and-Excitation Networks?
Amabhulokhi we-Squee-and-Excitation (SE) avumela inethiwekhi ye-convolutional ukuthi ifunde ukuthi isiteshi ngasinye sesici singasikala kangakanani, sisilinganise kabusha ngokusekelwe kumongo womhlaba. Le nqubo eshibhile efana nokunakwa iwine umncintiswano we-ImageNet ka-2017 futhi yaba ibhulokhi yokwakha evamile ye-CNN.
Senzani isinyathelo 'sokucindezela' ku-SE block?
Ukuminyanisa kusebenza ukuhlanganisa okumaphakathi komhlaba wonke, kugoqe imephu yesici ngasinye sibe isichazi esisodwa sesiteshi ngasinye.
Isinyathelo 'sokuthakasela' senzeni?
Injabulo idlula ivekhtha ekhanyiwe ezendlaleloni ezimbili ze-FC igcine ku-sigmoid, ikhiqiza izisindo zokukala zesiteshi ngasinye.
Zisetshenziswa kanjani izisindo zenjabulo kumamephu wesici sokuqala?
Imephu yesici yesiteshi ngasinye iphindaphindwa ngesisindo sayo esifundiwe, ikhulisa noma icindezela leso siteshi.
Kungani isasasa lisebenzisa ibhodlela elinesilinganiso sokunciphisa (ngokuvamile 16)?
Ibhodlela liyashwabana bese libuyisela ubukhulu besiteshi, ngakho i-SE block ibiza amaphesenti ambalwa kuphela engeziwe.
Yimuphi umncintiswano i-SENet eyawunqoba ngo-2017?
I-SENet iwine inselele ye-ILSVRC 2017 yokuhlukaniswa kwezithombe, iphusha iphutha eliphezulu-5 ku-2.25%.