Fatte bu metti
Fatte bu metti mooy su reso neuronal jàngee liggéey bu bees ba noppi ci saasi mu ñàkka mëna def liggéey yi mu jota xam.
Résumé
It is a central obstacle to building AI that learns continually without retraining from scratch.
Plongeur bu xóot
Reseau neuronal yi dañuy denc xam-xam ci poid yuñ bokk. Soo tàggatee benn model ci liggéey bu bees, yeesali gradient yi dañuy soppi parametre yi kodewoon mën-mën yu njëkk ya, suko defee performance yu yàgg yi mën nañu daanu. Lii mooy fàtte bu metti, ñu koy woowe interference catastrophique, McCloskey ak Cohen ñoo ko njëkka bind ci 1989. Dafay tar ci jàng bu toppalante wala bu wéy, fu done yi di ñëw ci ay fase, duñu boole lépp. Ci misaal, defar chatbot bu bari ci mbindu yoon mën na wàññi kàttanam ci waxtaan. Li gëna am solo ci brute-force fix mooy ñu tàggataat ci liggéey yépp ñu bokk, waaye loolu seer na lool te dafay laaj nga am done yu yàgg yi. Lu moy loolu, gëstukat yi dañuy jëfandikoo pexe yuy aar poid yu am solo, baamtu misaal yu njëkk, wala yokk ay parametre yuñ jagleel liggéey bi, lépp ngir may model yi ñu dajale xam-xam ci anam wi nit ñi di defee.
Gis-gis xarala
Fatte dafay am ndax dañuy jëfandikoowaat benn poid ci liggéey yépp, ba noppi wàccinu gradient bu amul benn tënk ci done yu bees daf leen di toxal ci seen bopp. Mitigations yi bokkuna ci Elastic Weight Consolidation, luy yokk penalti buy yeexal coppite yi ci parametre yi ñu jàpp ni dañu am solo ci liggéey yu yàgg yi (ñu xayma ko ci xibaaru Fisher). Leneen luñu mëna def mooy repetition wala replay experience (interleaving misaal yu yàgg yuñ denc wala yuñ defar), ak pexe isolation parametre yu melni adaptateur wala LoRA yuy gelé modelu base bi ba noppi yokk ay module yu bees yu ndaw.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu fàtte bu metti
Ginaaw model yi dañu joge ci tàggat benn yoon dem ci sistem yuy wéy di yeesal, di saytu fàtte nekk na lu am solo. Pexe yu am njariñ ci parametre yu melni adaptatëri LoRA may ekip yi ñu yokk xam-xam te duñu yàq modelu basis bi, ba noppi sistem yu yokk-retrieval yi dañuy moytu jafe-jafe bi ci denc xam-xam bu bees ci benn butig bu nekk ci biti moo gën poid yi. Xaarandil benchmark yuy wéy di jàng, architectures modulaires, ak pexe consolidation yu yuur-inspiré ñu màgg, yóbbu nu ci model yuy yeesal ak xibaar yu bees boole ci tëye ci anam wu wóor li ñu xamoon ba pare.
Doxal ci àdduna dëgg
Benn chatbot buñ yamale bu baax ci mbindu pajum dafay ñàkka mëna waxtaan bu baax.
Dajale poid elastik dafay may ndawu jeu yi ñu jàng jeu Atari yu bees te duñu fàtte yu yàgg yi.
Ekip yi dañuy jëfandikoo adaptëru LoRA ngir yokk xam-xam domen bu bees ci noonu lañuy bàyyi mën-mënu modelu base bu liw.
Expérience replay dafay denc misaal yu njëkk yi ba noppi boole leen ci diiru tàggat bu bees ngir baña yàq performance yu yàgg yi.
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
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 Catastrophic Forgetting 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
Wéyal jàng ak fàtte musiba
Laaj yi ñuy faral di laaj
What is Catastrophic Forgetting?
Fatte bu metti mooy su reso neuronal jàngee liggéey bu bees ba noppi ci saasi mu ñàkka mëna def liggéey yi mu jota xam. Gallankoor bu mag la ci tabax IA buy wéy di jàng te doo tàggataat ci noonu rek.
Luy fàtte musiba?
Fatte bu metti mooy ñàkka am kàttan yu yàgg yi ci saasi suñu tàggatee reso bi ci liggéey bu bees, ndax poid yiñ bokk dañuy gëna bari.
Lan moo waral fàtte bu metti di am ci reso neuronal yi?
Xam-xam mingi dundu ci ay parametre yuñ bokk, kon tàggat ci done yu bees dafay soppi poid yooyu ba noppi efaase njàngale yu njëkk ya.
Lan lay def ngir wàññi fàtte?
EWC dafay yokk benn daan buy yeexal coppite yi ci parametre yiñ jàpp ni dañu am solo ci liggéey yu yàgg yi, ñu xayma ko ci xibaaru Fisher.
naka la ko jaar-jaar (repetition) di xeexe fàtte?
Repetition dafay jaxasoo ak misaal yu njëkk (denc wala defar) ci di jàng liggéey yu bees suko defee performance yu yàgg yi gëna am doole.
Lan moo waral adaptëru LoRA yi am njariñ ngir moytu fàtte musiba?
LoRA dafay tëye modelu base bi ba noppi jàng ay parametre yu ndaw, suko defee xam-xam bu bees du yàq mën-mën yi fi nekk.