I-Online and Hard Negative Mining
Izimayini ezinzima ezingezinhle zikhetha izibonelo ezifundisa kakhulu, okunzima ukuzihlukanisa ukuziqeqesha esikhundleni sokumosha umzamo ezintweni ezilula imodeli esezilungile.
Uhlolojikelele
It is the trick that makes metric learning and object detection converge fast and accurately.
I-Deep Dive
Uma uziqeqesha ngokulahleka okuthathu noma okuphambene, ama-negative amaningi athathwe ngokungahleliwe asekude ne-ancho, ngakho akhiqiza ukulahlekelwa okuyiziro futhi akukho gradient, izitebele zokuqeqesha. Ukumbiwa kwezimayini okungekuhle kulungisa lokhu ngokukhetha ama-negative aqinile: izibonelo eziseduze nehange ngokungalungile. Ezimayini ezingaxhunyiwe ku-inthanethi, ngezikhathi ezithile uskena idathasethi ukuze uthole lezi, ezihamba kancane futhi ezidala. Ukumbiwa kwezimayini ku-inthanethi kubala ngokuhamba kwesikhathi phakathi kwenqwaba encane ngayinye: ngemva kokudlula phambili, ubheka wonke amabanga ahamba ngamabili eqeqebeni bese ukhetha abephula umthetho oqine kakhulu. I-FaceNet yethule ukumba izimayini eziqinile, ikhetha ukungalungi okude kakhulu kunokuhle kodwa okusengaphakathi kwe-marji, igwema ukungazinzi okungabangelwa ama-negative aqine kakhulu ekuqaleni kokuqeqeshwa.
I-Technical Insight
Ukumba izimayini ku-inthanethi kusebenzisa inqwaba osuvele uyenze ikhompuyutha. Ngokushumekwa kuka-B uthola i-matrix yebanga le-B-by-B ngokuyisisekelo mahhala, ukuze ukwazi ukuhlola izinombolo ezinkulu zamaphalethi amathathu ekhandidethi isinyathelo ngasinye. Ukumbiwa kwezimayini ngenqwaba kuyakhetha, kuhange ngalinye, okukude kakhulu okuphozithivu kanye negethivu eseduze eqeqebeni. Ukumbiwa kwezimayini kancane kancane kucindezela ama-negative ukuthi alale phakathi kwebanga eliphozithivu kanye nebanga eliphozithivu kanye nemajini, okukhiqiza ama-gradient angeyona eyero kodwa azinzile. Amaqoqo amakhulu anikeza iqoqo elicebile lamakhandidethi aqinile, yingakho usayizi weqoqo uthinta kakhulu ikhwalithi yokufunda imethrikhi.
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 Lezimayini Eziku-inthanethi Nezimbi Kanzima
Umgomo, ukuqeqesha kulokho okunzima, manje ushayela ukufunda okuzimele okuhlukile, lapho amachibi amakhulu e-batch negative (namabhange enkumbulo afana ne-MoCo) ahlinzeka ngokuqhathanisa okunzima ngaphandle kwamalebula. Abacwaningi balungisa ukuthi i-negative kufanele ibe lukhuni kangakanani, njengoba ama-negative aqine kakhulu avame ukuvela abhalwe ngokungeyikho noma acishe abe yimpinda ebonisa ukuqeqeshwa okonakele. Lindela izimayini ezihlakaniphile, eziqaphela ukungaqiniseki kanye nama-negative aqinile okwenziwa akhiqizwe imodeli ngokwayo, kanye nokuhlanganiswa okuqinile namasistimu okubuyisa ama-negatives aqinile emibuzweni yabasebenzisi bangempela.
Ukuqaliswa Komhlaba Wangempela
Ukuqeqeshwa kokuqaphela ubuso: I-FaceNet isebenzisa izimayini eziku-inthanethi ezingaqinile ukuze ifunde ukushumeka okwehlukanisa abantu abafanayo.
Ukutholwa kwento: I-SSD kanye nezitholi ezifanayo zisebenzisa izimayini ezinegethivu kanzima ukuze zilinganisele isikhukhula samabhokisi angemuva alula ngokumelene namabhokisi ezinto ezingavamile.
Ukubuyisa isiqephu esiminyene: ukusesha kanye nezinhlelo ze-RAG zemba imibhalo eqinile engemihle ebukeka ibalulekile kodwa ingafanele, ilola isitholi.
Amasistimu wokuncoma: amamodeli ezimayini izinto umsebenzisi angazichofanga kodwa ezifana nezinto ezichofoziwe, ezifundisa ukuhlukanisa okuhle kokunambitha.
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 Online and Hard Negative Mining 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
Isici Esiku-inthanethi Nesingaxhunyiwe Ku-inthanethi Sokusebenzela I-skew
Imibuzo evame ukubuzwa
What is Online and Hard Negative Mining?
Izimayini ezinzima ezingezinhle zikhetha izibonelo ezifundisa kakhulu, okunzima ukuzihlukanisa ukuziqeqesha esikhundleni sokumosha umzamo ezintweni ezilula imodeli esezilungile. Iqhinga elenza ukufunda imethrikhi nokutholwa kwento kuhlangane ngokushesha nangokunembile.
Kungani ama-negative amaningi athathwe ngokungahleliwe enikeza isignali encane yokuqeqeshwa ekulahlekeni kwe-triplet?
Ama-negative alula ahlala ngaphezu komkhawulo, anelisa ukulahlekelwa kakade, futhi anikele ngokuyisisekelo akukho gradient, ngakho-ke izitebhisi zokuqeqesha.
Yini ehlukanisa izimayini ze-inthanethi nezimayini ezingaxhunyiwe ku-inthanethi?
Ukumba izimayini ku-inthanethi kubala amabanga alandelanayo phakathi kwenqwaba yamanje isinyathelo ngasinye, kuyilapho izimayini ezingaxhunyiwe ku-inthanethi ngezikhathi ezithile ziskena yonke idathasethi futhi ingase iphelelwe yisikhathi.
Yini 'i-semi-hard' engalungile njengoba kuchazwe ku-FaceNet?
I-semi-hard negative iqhelelene ne-ancho kunokuphozithivu kodwa iwela ngaphakathi kwemajini, enikeza ama-gradient awusizo, azinzile ngaphandle kokungaqini kwezimo ezinzima kakhulu.
Emayini ye-batch-hard, iyiphi i-negative ekhethwa ihange ngalinye?
I-batch-hard mining ikhetha izimo ezinzima kakhulu: kuhange ngalinye lithatha elikude kakhulu eliphozithivu kanye negethivu eseduze (edida kakhulu) eqeqebeni.
Kungani usayizi weqoqwana omkhulu evame ukuthuthukisa imiphumela yokufunda yemethrikhi?
Iqoqo elikhulu likhiqiza i-matrix yebanga elilandelana ngokubili elikhudlwana, ngakho-ke kukhona ama-negative amaningi okumba anzima kakhulu esinyathelweni ngasinye.