Wéyal jàng ak fàtte musiba
Luy wéyal jàng mooy tàggat IA ci liggéey yu bees yu bari ci diir bu gàtt te baña efaase limu xamoon bu njëkk.
Résumé
Its central obstacle is catastrophic forgetting: when a neural network learns a new task, gradient updates overwrite the weights that encoded earlier tasks, and old skills collapse.
Plongeur bu xóot
Reseau neuronal yiñ miin dañu jàpp ni done yépp am nañu ci benn yoon. Ci àdduna dëgg, done yi dañuy ñëw ci anam wu toppalante, te naïvely fine-tuning ci liggéey yu bees yi waral ñu fàtte catastrophique - performance ci liggéey yu njëkk yi dafay wàññeeku ndax diisaay yi ñu bokk dañuy binndaat. Jàngat bu wéy dafay fexe ngir yemale stabilite (teg xam-xam bu yàgg) ak plastisite (jël xam-xam bu bees), dilemma stabilite-plastisite bu yàgg bi. Ñatti famiy yu mag yu am ci pexe yi am: pexe yu ñuy yamale lu melni Elastic Weight Consolidation biy daanel coppite yi ci poid yi ñu jàpp ni dañu am solo ci liggéey yu yàgg yi; replay pexe yiy denc wala defar misaali liggéey yu njëkk ba noppi boole leen ci diiru tàggat; ak pexe architecture yiy xaaj paramet wala module yu bees ci liggéey bu nekk. Amul benn njuréef bu ko mëna saafara, te jàngat bi dafay laal jekkal liggéey, domen, ak jekkal yokkute ci klaas.
Gis-gis xarala
Fatte musiba dafay am ndax wàccinu gradient ci liggéey bu bees dafay toxal diisaay yiñ bokk ci optimum bu bees te amul benn tere ngir des ci wetu gox yi baax ci liggéey yu yàgg yi. Consolidation poid elastik dafay xayma solo bu poid bu nekk (jaaraleko ci matrix Fisher) ba noppi yokk ci penalti quadratic buy ancre poid yu am solo yi ci seen valeur yu yàgg yi. Replay mingi jegeel distribution bi njëkk ci jaxase misaal yu yàgg yiñ denc wala yuñ defar ci ay lots yu bees, suko defee gradient yi di wane liggéey yu yàgg yi ak yu bees yi, wàññi bind buy yàq.
njeextalu pexe
dogal yu gëna leer
Daf lay jàppale nga tàqale kàddu yu leer ci wàllu xarala ak làkku fësal njaay.
Njëgg ak budget
Mën nga laaj laaj yu gëna baax ci samp gi balaa ngay dugal xaalis wala sa jotu liggéey.
Ekip ak def liggéey
Ekip yi bokk xam-xam ñoo gëna mëna jël yenn dogal ci wàllu produit, politik ak jàng.
Ëlëgu jàng bu wéy ak fàtte bu metti
Jàngat bu wéy dafay gëna am solo ngir mëna wéy di jëfandikoo model yu mag yi te duñu soxla tàggataat bu mat sëkk. Gëstu dafay push ci yeesali parametre-efficace yuy wéy (adaptatër, module LoRA yokk ci liggéey bu nekk), replay bu gëna baax jëfandikoo model generative, ak pexe yuy yeesal xam-xam ci model fondation yi di moytu fàtte ak drift bu bëggul. Xaarandi lëkkalekaay yu gëna seere ak ndawu liggéey yiy jàng ci aparey bi, replay buy baña denc ay done yu ñor, ak ay benchmark yu gëna wane dëgg, ay done yu amul benn taxawaay, du ay pexe liggéey yu rafet.
Doxal ci àdduna dëgg
Benn xeetu nataal buñ dugal bu wara jàng kategori produit yu bees weer wu be nekk te baña fàtte yi njëkka am.
Personalisasioŋ bi am ci aparey bi (klaweer wala assistant vocal) buy méngoo ak jëfandikukat bi ci diir bu gàtt te du ñàkk njubte gu mat sëkk.
Robot yuy jàng xam-xam bu bees ci wàllu jëfandikoo ay mbir, di tëye xam-xam bi ñu njëkka xam.
Yeesal xeetu làkk ak ay mbir yu bees wala ay domen yu jëfandikoo ay adaptatër suko defee ñu baña yàq kàttan yu njëkk ya.
Risk yi ak balustrade yi
Ekip yu bari mën nañu jëfandikoo benn baat ci anam wu wuute, kon teela leeral yaatuwaayam.
Benchmark yi mën nañu nuru lu am doole waaye performance yi ci àdduna bi duñu tolloo.
Bëgg kalite done ak palaŋu jàngat dafay faral di jur njariñ yu yomba dagg.
Roadmap ngir samp gi
Tàmbaleel ci joxe leeral ci làkk wu leer ci njariñ li nga soxla.
Tannal benn metric bu baax ak benn anam bu baaxul balaa ngay saytu.
Doxal ab pilote bu ndaw ak ay done yu representatif, du ab demo bu leer.
Dokumenteer fu jàng bu wéy ak fàtte Catastrophic di jàppale ak fu pexe yu gëna yomba gëna baax.
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 Continual Learning and 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
Fatte bu metti
Laaj yi ñuy faral di laaj
What is Continual Learning and Catastrophic Forgetting?
Luy wéyal jàng mooy tàggat IA ci liggéey yu bees yu bari ci diir bu gàtt te baña efaase limu xamoon bu njëkk. Gallankoor bi gëna mag mooy fàtte: su reso neuronal jàngee liggéey bu bees, yeesali gradient yi dañuy soppi poids yi encode liggéey yu njëkk ya, ba noppi xam-xam yu yàgg yi dañuy daanu.
Luy 'fatte bu metti'?
Fatte bu metti mooy ñàkka mëna def liggéey bu yàgg sudee dañuy soppi poids yiñ bokk ci reso bi, fekk ñu ngi jàng lu bees.
Luy boole poid elastik (EWC)?
EWC yokk na penalti (jëfandikoo xibaaru Fisher) buy ancre poid yi ñu jàpp ni dañu am solo ci liggéey yu yàgg yi, wàññi fàtte.
naka la pexe yu lalu ci replay di xeexe fàtte?
Replay dafay jaxase misaal yu njëkk yi ci ay lote yu bees suko defee gradient yi wane liggéey yu yàgg yi ak yu bees yi, di jegesi done yiñ boole.
Lan moo nekkul benn ci ñatti famiy yu mag yi ci njuréefi njàng mu wéy?
Ñatti famiy yu mag yi ñooy yamale, replay, ak pexe architecture; kompresioŋ bokkul ci ñoom.
Lan moo waral ñu fàtte liggéey bu bees bi nga def ci liggéey bu bees?
Coppite gradient yi amul ay tënk dañuy binndaat poids yiñ bokk ci optimum bu bees bi, bàyyi gox yi baaxoon ci liggéey yu njëkk ya.