Jàng liggéey yu bari
Jàng liggéey yu bari dafay tàggat benn model mu def liggéey yu bari yu jëm ci benn yoon, di séddoo seeni xalaat ci biir.
Résumé
By learning shared structure, each task helps the others, often improving accuracy and data efficiency over training separate models.
Plongeur bu xóot
Duñu tabax benn model bu wuute ci liggéey bu nekk, jàng liggéey yu bari (MTL) dafay jëfandikoo yaxu ndigg buñ bokk buy xaajaloo ci boppu liggéey bu nekk. Ci misaal, reso buy dawal boppam mën na bokk benn encodeur vision ba noppi ñu xaaj ko ñaari pàcc ngir gis oto yi, xaaj tali bi, ak xayma xóotaayu yoon wi. Layer yiñ bokk dañuy jàng man-mani yu am njariñ ci liggéey yépp, fekk njiit bu nekk amna lu muy spesialise. Loolu dafay nuru benn xeetu njuumte ak yamale: siñaal yi bawoo ci benn liggéey dañuy tënk représentation buñ bokk, wàññi overfitting ak gëna suqali generalisation, rawatina sudee yenn liggéey yi amul ay done yu néew. Jafe-jafe bi gëna mag mooy yemale liggéey yi - sudee seeni balansu ñàkk wala seeni gradient dañuy xëccoo, benn liggéey mën na ëpp doole ñeneen ñi di dundu lu metti, jafe-jafe bu ñuy woowe toxal bu baaxul. Pexe yu melni pondération perte, pondération bu sukkandiko ci ñàkka wóor, ak operation gradient dañuy fexe ba liggéey yi di wéy di jëflante, du ñuy xëccoo.
Gis-gis xarala
Mébet bi dafay faral di nekk limu pondéree ci perte yi ci liggéey bu nekk, L = Σ w L, ba noppi tànn pondération w lu jafe la ndax liggéey yi dañu wuute ci yaatuwaay ak jafe-jafe. Séddoo parametre yu dëgër (bañ bu ñu bokk, bopp yu wuute) mooy anam wi gëna yomba te gëna yamale; séddoo bu woyof dafay tax model yu wuute di booloo bu baax. Gradient yiy xëccoo ci liggéey yi mën nañu ko fomm, kon pexe yu melni pondération incertitude (jàng w ci saasi) wala PCGrad (projection fu sori composants gradient yiy xëccoo) jàppale liggéey yi ñu tàggatoo ci anam wu dëgër.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu jàng liggéey yu bari
Jàngat liggéey yu bari dafay jàppale tendaas bi jëm ci xeetu generalist yi. Modèlu làkk yu mag yi dañuy liggéey lu bari — benn reso mooy yonnee tekki làkk, tënk, kodage, ak Q&A — te sistem multimodal yi dañuy yokk lii ci bind, nataal ak audio. Xaarandi jëfandikoo bu gëna bari ci architecture yuñ boole ak tuning instruction yuy boole liggéey yu bari ci benn model, boole ci balancing ak routing bu gëna baax ci liggéey (ni ci njaxasu-ekspert) kon yokk liggéey mënul tekki yokk model yu wuute.
Doxal ci àdduna dëgg
Gis-gis biy dawal boppam dafay bokk benn enkodeer gis-gis ngir gis mbir, xaaj yoon wi ak xayma xóotaayu mbir mi.
Royuwaayi làkk yu yaatu yuy jëfandikoo tekki làkk, tënk, yëg-yëg ak tontu ci laaj ak tontu ci benn reso buñ bokk.
Sistem yiy xelal nit ñi dañuy wax luy waaja am ci klike yi, waxtu seetaan yi ak jënd yi ngir gëna mëna jëflante ak jëfandikukat yi.
Modèlu nataalu pajum yiy gis tumër ci benn yoon, xaaj ay pexem, ba noppi tànnal xeetam ci benn scanner bi.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
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 Multi-Task 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
Gis bi ci topp
Séddoo ay parametre yu jafe ci reso yu bari liggéey
Laaj yi ñuy faral di laaj
What is Multi-Task Learning?
Jàng liggéey yu bari dafay tàggat benn model mu def liggéey yu bari yu jëm ci benn yoon, di séddoo seeni xalaat ci biir. Bu ñu jàngee jumtukaay buñ bokk, liggéey bu nekk dafay jàppale ñeneen ñi, lu ci gëna bari mooy gëna dëppoo ak njariñu done yi ci tàggat ay xeetu liggéey yu wuute.
Lan mooy xalaatu jàng liggéey yu bari?
MTL dafay tàggat benn model ci liggéey yu bari suko defee ay couche yuñ bokk jàpp structure bu am njariñ ci ñoom ñépp.
Ci reso MTL bu am paramet yu dëgër yuñ séddoo, lan lañuy séddoo ak lan lañuy tàqale?
Séddoo parametre yu dëgër yi dañuy jëfandikoo benn bagaas buñ bokk ngir man-mani yu mag yi ak bopp yu wuute yuñ jagleel liggéey bu nekk.
Lan moo waral jàng liggéey yu bari mën na yokk yamale gi?
Jàng liggéey yu bari dafay tënk man-man yiñ bokk, di def ni regularisation buy faral di jàppale liggéey yu néew done yi.
Luy 'toxal bu baaxul' ci jàng liggéey yu bari?
Gradient yuy xëccoo wala perte yu ëpp doole mën na waral benn liggéey di naxasal ñeneen ñi, lu wuute ak toxal bu am njariñ.
naka lañuy faral di defaree perte générale ci jàng liggéey yu bari?
Mébet bi mooy Σ w L, te tànn diisaay yi lu jafe la ndax liggéey yi dañu wuute ci yaatuwaayu ak jafe-jafe.