Ukulahlekelwa Okuthathu kanye Nokufunda Kwemethrikhi
Ukulahleka kwe-triplet kufundisa inethiwekhi ye-neural ukubeka izinto ezifanayo eduze nezinto ezingafani ziqhelelene endaweni yokushumeka.
Uhlolojikelele
It is the foundation behind face recognition, image search, and recommendation systems that need to compare things rather than just classify them.
I-Deep Dive
Ukufunda kwemethrikhi kuqeqesha imodeli ukuthi ikhiqize okushumekiwe, ama-vector lapho ibanga libonisa ukufana. Ukulahleka kukathathu kwenza lokhu kusetshenziswa okokufaka okuthathu ngesikhathi: ihange, iphozithivu (isigaba esifanayo nesikhonkwane), kanye nenegethivu (isigaba esihlukile). Injongo iphusha ihange eduze kwephozithivu kunenegethivu okungenani ngemajini egxilile. Ngokusemthethweni, ukulahlekelwa ngu-max(0, d(a,p) - d(a,n) + margin), lapho u-d ngokuvamile eyibanga le-Euclidean. I-FaceNet ka-Google ka-2015 yenze le ndlela yaduma, ifunda ukushumeka kobuso obungu-128-dimensional ngqo. Uma usuqeqeshiwe, uqhathanisa noma yiziphi izinto ezimbili ngokubala ibanga, akukho ukuqeqeshwa kabusha okudingekayo kobunikazi obusha. Leli khono lesethi evulekile yingakho amandla okufunda emethrikhi eqinisekisa futhi ukuhlukaniswa kwemisebenzi yokubuyisa kungakwazi ukuphatha kalula.
I-Technical Insight
Imajini yiyona eyenza ukulahlekelwa kathathu kusebenze. Ngaphandle kwayo, imodeli ingabhidliza kancane konke okushumekiwe kuye endaweni eyodwa, kwenze lonke ibanga libe uziro noku-oda kungasho lutho. Imajini iphoqelela isilondolozi: inegethivu kufanele okungenani ibe imajini ibe kude kunephozithivu ngaphambi kokuba ukulahlekelwa kufinyelele kuziro. Ukushumeka kuvamise ukwenziwa kwe-L2 ku-hypersphere yeyunithi, ngakho amabanga ahlala elinganiselwe futhi eqhathaniswa. Ukukhetha imajini (imvamisa cishe ku-0.2) kuhwebelana ngokuthi izigaba zihlangana kanjani ngokumelene nokuhlukana phakathi kwazo.
I-Strategic Impact
Izinqumo ezicacile
Kukusiza ukuthi uhlukanise izimangalo ezicacile zobuchwepheshe kusukela olimini lokumaketha.
Izindleko kanye nesabelomali
Ungabuza imibuzo yokusebenzisa kangcono ngaphambi kokusebenzisa imali noma isikhathi.
Ithimba kanye nokusebenza komsebenzi
Amaqembu anokuqonda okwabiwe enza izinqumo ezingcono zomkhiqizo, inqubomgomo, nokufunda.
Ikusasa Lokulahlekelwa Okuthathu kanye Nokufunda Kwemethrikhi
Ukulahlekelwa kwe-triplet ehlanzekile kuya ngokuya kuthathelwe indawo yizinjongo ze-batch-wide ezifana nokufana okuningi, i-proxy-anchor, nokulahlekelwa okuphambene (InfoNCE) okuqhathanisa amapheya amaningi ngesinyathelo ngasinye futhi ahlangane ngokushesha. Izindlela zokuziqondisa ezifana ne-SimCLR zibonisa ukufunda kwemethrikhi kungasebenza ngaphandle kwamalebula ngokuphatha imibono ethuthukisiwe njengokuhle. Njengoba imininingwane egciniwe yevekhtha kanye nokutholwa kwesizukulwane esithuthukisiwe, ukushumeka okufundiwe kusekela ukusesha kwe-semantic esikalini sezinto eziyibhiliyoni, ngakho-ke umqondo oyinhloko webanga lokufana uya umaphakathi kakhulu, njengoba ukwakheka kwe-triplet ethile kuncipha.
Ukuqaliswa Komhlaba Wangempela
Ukuqinisekiswa kobuso besitayela se-FaceNet: amafoni namasango ephasipoti aqinisekisa ubuwena ngokubheka ukuthi ukushumeka kobuso okubili kuwela ngaphakathi komkhawulo webanga yini.
Ukusesha komkhiqizo obonakalayo: amasayithi e-e-commerce avumela abathengi ukuthi balayishe isithombe futhi bathole izinto ezifanayo ngokubheka okushumekiwe komakhelwane.
Ukuqinisekiswa kwesipikha: abasizi bezwi bashumeka isampuli yezwi futhi bayiqhathanise nephrofayela ebhalisiwe ukuze kuqinisekiswe ukuthi ubani okhulumayo.
Ukuqinisekiswa kwesignesha nombhalo wesandla: amabhange ashumeka ireferensi namasiginesha emibuzo kanye namafulegi afojiwe lapho ibanga lidlula umkhawulo ofundiwe.
Izingozi & Guardrails
Amaqembu ahlukene angasebenzisa igama elifanayo ngokuhlukile, ngakho chaza ububanzi kusenesikhathi.
Amabhentshimakhi angabukeka eqinile kuyilapho ukusebenza komhlaba wangempela kungalingani.
Ukuziba ikhwalithi yedatha nezinhlelo zokuhlaziya kuvame ukudala imiphumela entekenteke.
Ukuqalisa Umhlahlandlela
Qala ngencazelo yolimi olulula yomphumela oyidingayo.
Khetha imethrikhi eyodwa yempumelelo nesimo esisodwa sokuhluleka ngaphambi kokuhlolwa.
Qalisa umshayeli omncane onedatha emele, hhayi isethi yedemo ephucuziwe.
Idokhumenti lapho Ukulahlekelwa Okuthathu kanye Nokufunda Ngemethrikhi kusiza nalapho izindlela ezilula zingcono khona.
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 Triplet Loss and Metric Learning 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 we-Siamese kanye Nokulahlekelwa Okuthathu
Imibuzo evame ukubuzwa
What is Triplet Loss and Metric Learning?
Ukulahleka kwe-triplet kufundisa inethiwekhi ye-neural ukubeka izinto ezifanayo eduze nezinto ezingafani ziqhelelene endaweni yokushumeka. Kuyisisekelo esingemuva kokuqashelwa kobuso, ukusesha izithombe, namasistimu wokuncoma adinga ukuqhathanisa izinto kunokumane azihlukanise.
Yiziphi izinto ezintathu ezifakiwe ezakha i-triplet ekulahlekeni kwe-triplet?
I-triplet iqukethe ihange, iphozithivu ehlanganyela isigaba se-ancho, kanye nenegethivu evela esigabeni esihlukile.
Kungani ukulahlekelwa kathathu kufaka phakathi igama le-margin?
Ngaphandle kwemajini, imodeli ingakwazi ukubeka yonke into endaweni efanayo, yenze wonke amabanga abe iqanda futhi anelise ukulahlekelwa kancane. Imajini iphoqa ukwehlukana kwangempela.
Iluphi uhlelo oludumile lwango-2015 olwaduma ukulahlekelwa kathathu ukuze kubonwe ubuso?
I-FaceNet ye-Google isebenzise ukulahlekelwa kathathu ukuze ifunde ukushumeka kobuso obuhlangene futhi yamisa ibhentshimakhi ekuqinisekiseni ubuso.
Ithini imodeli eqeqeshiwe ye-metric-learning yokokufaka ngakunye?
Imodeli iveza okokufaka kokushumeka ama-vector; bese uqhathanisa izinto ngokulinganisa ibanga phakathi kokushumeka kwazo.
Kungani ukufunda kwemethrikhi kufaneleka kahle ezinkingeni ezivulekile njengokwengeza ubuso obusha?
Ngenxa yokuthi ukuqashelwa kusekelwe ebangeni lokushumeka, ungabhalisa ubuwena obusha ngokugcina ukushumeka kwakho, akukho ukuqeqeshwa kabusha kwemodeli okudingekile.