Fenomen ñaari wàcci
Ñaari wàcci mooy seetlu bu yéeme bi ñu gis ni lu model bi di gëna rëy, njuumte test bi dafay njëkka gëna bon ci wetu 'interpolation threshold' waaye ginaaw ga mu gëna baax - di ñaawlu tradeoff bu téere jàngat bu yàgg bi.
Résumé
It matters because it helps explain why enormous, overparameterized neural networks generalize well instead of overfitting.
Plongeur bu xóot
Statistique classique dafay jàngale courbe bu am formu U: lu complexité model bi di gëna yokk, njuumte test bi dafay wàññeeku, di wàcci, ginaaw ga mu yokk lu model bi di gëna méngoo. Ñaari wàcci, Belkin, Hsu, Ma, ak Mandal ñoo ko siiwal ci 2019 te AIU_PROTECTED_10_ jàngat ko ci escalier, dafay wane ni courbe amna ñaareelu wàcci. Jumtukaayu test bi dafay yegg ba ci buntu interpolation bi - barab bi model bi am parametre yu doy ngir mëna méngoo ak barabu tàggat bu nekk (njuumteg tàggat amul). Push romb loolu ci regime overparameterized ak njuumte test daanuwaat, lu bari ci suufu barab bu neex classique. Benn effet bi dafay feeñ ci dayo model bi, diiru tàggat bi ('epoch-wise' ñaari wàcci), ak dayo dataset bi. Mungi defaraat tiitaange bu yagg bi di wax ni 'paramet yu bari defay tektale lu ëppu.'
Gis-gis xarala
Ci buntu interpolation amna benn solution bu méngoo ndànk ak done yi, te dañu ko forse mu nekk jagged ak norm bu kawe, moo tax du generalise bu baax. Ci regime biñ paramétree lu ëpp, barina pexe yu amul benn njuumte, te wàccinu gradient bi dafay jubal ci bi gëna nooy, bi gëna suufe. Boobu tànneef ngir interpolatër yu woyof - du lim parametre bi ci boppam - mooy liy dawal ñaareelu wàcci ngir wàññi njuumti test bi.
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 fenomenu ñaari wàcci
Gëstukat yi dañuy jëfandikoo ñaari wàcci ngir setal yooni scaling ak tànn kañ lañu wara bàyyi tàggat yaram, ndax 'tàggat lu gëna yàgg, gëna bon, gëna baax' amna njeexital yu am solo. Xaarandil theorie bu gëna dëgër bu koy boole ak yamale bu nëbbu, kernel tangente neuronal, ak grokking. Ci jëf, njàngale mi - gëna rëy ak gëna gudd mën na jàppale romb zone danger - jota jël dogal ngir tàggat xeetu fondation yu gëna mag, du ñu gëna mag.
Doxal ci àdduna dëgg
Faram-fàcce li waral benn xeetu làkk bu am 175 milyaar ciy parametre di gëna mëna yamale lu ëpp benn bu digg buñ defar bu baax doonte amna kàttan gu gëna bari
Tann tàggat lu weesu barab bi ñàkka am validasioŋ di gëna bon ci diir bu gàtt, ndax ñaari wàcci epoch-wise dafay wax luy waaja am ci ginaaw
Saytu ab xeetu gis-gis bu njubteem wàññeeku ndànk su lim parametre bi méngoo ak dayo tàggat-set bi, ba noppi yóbbu ko ci gëna xóot ci parametrisation
Yegle dogal ci dayo model ci AutoML suko defee praktisien yi moytu zone interpolation bu yomba dagg
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.
Bindal fi Fenomenu Ñaari Wàcci di jàppale ak fi 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 Double Descent Phenomenon 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
Wàccig degrade
Laaj yi ñuy faral di laaj
What is Double Descent Phenomenon?
Ñaari wàcci mooy seetlu bu yéeme bi ñu gis ni lu model bi di gëna rëy, njuumte test bi dafay njëkka gëna bon ci wetu 'interpolation threshold' waaye ginaaw ga mu gëna baax - di ñaawlu tradeoff bu téere jàngat bu yàgg bi. Dafa am solo ndax dafay jàppale ci leeral lu tax reso neuronal yu rëy, yu bari ay parametre di generalise bu baax ci barabu overfitting.
Fan la njuumti test bi di gëna bari ci courbe descendant double bi?
Jumtukaay bi dafay am ci buntu interpolation bi, foofu la model bi am kàttan gu doy ngir dawal njuumte tàggat yaram ba zero ak benn pexe bu dëgër.
Lan mooy xew ci njuumti test bi model bi di gëna bari ay parametre?
Ci regime overparameterized njuumte test wàcci ñaareelu yoon, mooy màndarga biy joxe ñaari wàcci turam.
Lan moo waral wàccinu gradient di jàppale ci regime bu ëpp parametre?
Ak pexe yu bari yu amul benn njuumte, wàccinu gradient bi dafay jubal ci bi gëna nooy, bi gëna suufe, mooy gëna mëna yamale.
'Epoch-wise' ñaari wàcciku dafay wax ci efekt bi feeñ ci fonction lan?
Ñaari wàcci mën na am ci axis bi ci diiru tàggat-yaram: njuumti test mën na gëna tar ba noppi gëna baax su tàggat-yaram wéyee ci yeneen jamono.
Ban xalaat bu yàgg la double descendant di xeex?
Dafay werante ak téere biy wax ni courbe bu am formu U bi dafa wax ni complexité bu gëna bari lu weesu benn point dafay yokk njuumti test yi ci overfitting.