Asarar Hankali don Gano Rashin Daidaito
Asarar mai da hankali shine aikin asara da aka gyara wanda ke saukar da misalan masu sauƙi don haka mai ganowa zai iya mai da hankali kan masu wuya, masu wuya.
Dubawa
It solved the extreme background-versus-object imbalance that crippled one-stage object detectors.
Zurfafa nutsewa
A cikin gano abu, hoto na iya ƙunsar wasu abubuwa na gaske kawai amma dubun-dubatar wuraren ƴan takara, kusan dukkansu suna da sauƙi. Tare da daidaitaccen giciye-entropy, wannan ambaliya mai sauƙi na rashin ƙarfi ya mamaye gradient kuma yana nutsar da abubuwan da ba kasafai ba. Asarar mai da hankali, wanda Lin ya gabatar a cikin takarda RetinaNet na 2017 da abokan aiki a Facebook AI, yana gyara wannan ta hanyar ninka giciye-entropy ta hanyar (1 - p_t) ^gamma. Lokacin da aka rarraba samfurin gaba ɗaya kuma daidai, p_t yana kusa da 1, don haka yanayin yana raguwa zuwa sifili kuma ƙayyadadden misali da kyar ke ba da gudummawa. Masu wuya, misalan misalan da ba a tantance su ba suna kiyaye cikakken nauyi. Tare da gamma a kusa da 2, RetinaNet ya dace ko ya doke masu gano matakai biyu a hankali kamar Faster R-CNN yayin da yake zama cibiyar sadarwa mai sauƙi ta hanyar wucewa.
Fahimtar Fasaha
Ma'aunin gamma mai mai da hankali yana sarrafa yadda ake murkushe misalan masu sauƙi: a gamma 0 hasara mai mahimmanci daidai da giciye-entropy na yau da kullun, kuma gamma mafi girma yana haɓaka mai da hankali kan lokuta masu wahala. Ma'auni alpha mai nauyi (sau da yawa 0.25 don ajin da ba kasafai ba) yawanci ana haɗa shi da shi. Mahimmanci ma'aunin daidaitawa yana sake fasalin gradients, ba kawai ƙimar asara ba, don haka yaɗa baya a zahiri yana ƙarfafa samfuran da ba su da tabbas ba tare da haƙar ma'adinan misalan hannu ko sake yin samfuri ba.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar Asarar Hankali don Gane Rashin Daidaito
Asarar mai da hankali ya zama sinadari na asali fiye da RetinaNet, yana bayyana a cikin masu gano abubuwa kamar FCOS, a cikin yanki, kuma a cikin rarrabuwa mai tsayi. Bambance-bambancen kamar hasara mai inganci, asarar hangen nesa na rarrabawa, da asarar varifocal suna tace shi don na'urori marasa anka na zamani da na'urori masu auna canji. Yayin da ganowa ke motsawa zuwa ƙirar saiti kamar DETR waɗanda ke amfani da daidaitawa biyu, salon mai da hankali ya kasance kayan aiki mai amfani a duk inda mitocin aji suka karkata sosai.
Aiwatar da Gaskiyar Duniya
Gano ƙananan alamun hanya ko masu tafiya mai nisa a cikin firam ɗin tuƙi masu cin gashin kansu inda mafi yawan pixels suke bango.
Nemo ciwace-ciwacen ciwace-ciwace ko raunuka a cikin binciken likita wanda nama mai lafiya ya mamaye.
Lalacewar tsinkaya akan layin masana'anta inda galibin sassan da aka bincika sun kasance na al'ada.
Gano ƙananan jiragen ruwa ko motoci a cikin manyan tauraron dan adam da hotunan iska.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
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 Focal Loss for Imbalanced Detection 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
Jagora na gaba
Siamese Networks da Asarar Sau Uku
Tambayoyin da ake yawan yi
What is Focal Loss for Imbalanced Detection?
Asarar mai da hankali shine aikin asara da aka gyara wanda ke saukar da misalan masu sauƙi don haka mai ganowa zai iya mai da hankali kan masu wuya, masu wuya. Ya warware matsananciyar rashin daidaituwa ta bango-da-abu wanda ya gurgunta abubuwan gano abubuwa mataki-daya.
Wace babbar matsala a gano abu mataki ɗaya aka ƙera hasara mai mahimmanci don magancewa?
A cikin ganowa mai yawa, misalan bayanan baya masu sauƙi sun zarce abubuwa na gaske kuma sun mamaye madaidaicin giciye-entropy gradient. Asarar mai da hankali yana sake yin nauyi sosai, misalai masu wuya suna ƙidayawa.
Menene ma'anar daidaitawa wanda hasara mai hankali ke ninkawa kan giciye-entropy?
Asara mai hankali shine -(1 - p_t)^gamma * log(p_t). Kalmar (1 - p_t)^gamma tana raguwa zuwa sifili don ingantaccen tsinkaya.
Me zai faru da asarar hankali lokacin da aka saita gamma mai mai da hankali zuwa 0?
A gamma = 0 factor (1 - p_t) ^ 0 daidai da 1, don haka asarar hankali yana rage daidai zuwa daidaitattun giciye-entropy.
Menene haɓaka gamma ke yi don sauƙi, misalan da aka tsara?
Gamma mafi girma yana sa kalmar gamma (1 - p_t)^gamma ta ragu da sauri don babban p_t, don haka misalai masu sauƙi suna ba da gudummawa ko da ƙasa ga asara da gradient.
Wanne gine-ginen injin ganowa ya gabatar da asarar hankali?
An gabatar da asarar mai da hankali a cikin takarda na RetinaNet na 2017, yana ba da damar gano matakin mataki ɗaya don yin hamayya da ingantattun hanyoyin matakai biyu.