Autoencoder yu néew ngir mëna tekki
Sparse autoencoders (SAEs) jumtukaay la buy xaajale mbir yu jaxasoo yi ci biir reso neuronal ci benn xeetu mbir yu gëna set, yu nit ñi mëna tekki.
Résumé
They are one of the leading techniques for opening the 'black box' and seeing what concepts a model actually represents.
Plongeur bu xóot
Ci biir transformateur bi, benn vecteur de activation dafay jaxase ay junni konsept ci benn yoon, te loolu tax na ñu koy jàng lu jafe. Autoencoder bu bariwul ay reso yu ndaw la yu am ñaari couche yuñ tàggat ngir defaraat activation yooyu jaaraleko ci couche bu nëbbu bu yaatu, waaye ak penalti bu bariwul buy forse yenn neuron yu bari yi ñuy fexe benn yoon. Ndax fitna googu, bépp yunit bu nëbbu defay yam ci benn konsept, lu mel ni ' wax ci Golden Gate Bridge' mbaa 'kodu Python'. Ci atum 2024 dafa yokk lii ba 3 Sonnet, génne ci lu tollu ci 34 milioŋ ciy màndarga, ba noppi DeepMind siiwal liggéey SAE bu méngoo. Suko defee gëstukat yi mën nañu tëye wala wàcce benn màndarga ngir saytu limuy def.
Gis-gis xarala
Benn SAE dafay xayma ab aktivasioŋ d-dimensionel ci ab couche bu nëbbu bu gëna yaatu (8x ba 100x lu gëna rëy), ba noppi defaraat original bi. Taggat yaram dafay wàññi njuumti tabaxaat boole ci L1 penalti ci activation yu nëbbu, luy ñaax sparsity suko defee unité yu bari nekk ci wetu zero. Variante yu melni TopK SAEs dañuy doolel sparsity ci saasi ci tëye K bi gëna mag ci liggéey, ak SAEs yu am buntu dañuy tàqale dogal biy fett ak magnitude, wàññi benn njuumte bu L1 dugal.
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 Autoencoder yu néew ngir mëna tekki
Xaarandi SAEs ñu joge ci curiosité gëstu dem ci audit pratique ak jumtukaayi kaaraange, boole ci dashboard yiy etikete man-man yi ak gis sircuit yu gëmloo wala yu wóorul. Jafe-jafe yi ubbeeku ñooy 'xaajale màndarga' (benn konsept buy xaaj lu bari), màndarga yu ñàkk, ak njëgu tàggat SAE ci bépp xeetu modelu frontiere. Yoon yu bees yu melni crosscoder, transcoder, ak matryoshka SAEs seen mébet mooy jàpp xayma ci diisaay yu bari ak ci granularite yu bari benn yoon.
Doxal ci àdduna dëgg
Anthropic 'Golden Gate Claude' demo, fu yokk benn man-mani SAE def model bi di royuwaale pont bi ci tontu bu nekk
Di génne ak etikete lu tollu ci 34 milioŋ ciy màndarga ci Claude 3 Sonet ngir xayma konsept yu melni sycophancy, njuumti kode, ak doxalin bu wóorul
Gis man-mani kaaraange yu am solo yu melni njuuj njaaj, tëddee, wala ëmbiit bu mëna lore buñu mëna saytu wala doxal ci diiru dugal
Debugging li waral benn model di juum ci xaaj ay dugal ci saytu ban man-man yuñ mëna tekki yuñ aktive ci benn laaj buñ jox
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 Sparse Autoencoders for Interpretability 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
Autoencoder yu néew ngir dindi màndarga
Laaj yi ñuy faral di laaj
What is Sparse Autoencoders for Interpretability?
Sparse autoencoders (SAEs) jumtukaay la buy xaajale mbir yu jaxasoo yi ci biir reso neuronal ci benn xeetu mbir yu gëna set, yu nit ñi mëna tekki. Benn la ci pexe yi gëna am solo ngir ubbi 'kees bu ñuul bi' ngir gis ban konsept la model bi di fësal.
Lan mooy jubluwaay bi gëna mag ci tàggat ab autoencoder bu néew ci ab model buy tàmbali?
SAEs tabaxaat activations jaaraleko ci ab couche bu yaatu, bu nëbbu, suko defee benn-benn unité mën méngoo ak benn konsept bu nit ñi mëna xam.
naka lay def ba SAE di ñaax bépp yunit bu nëbbu mu wane benn konsept?
Penalty sparsity (lu melni terme L1 wala constraint TopK) dafay forse unité yu nëbbu yu bari ñu des ci wetu zero, di push unité bu nekk ci lu am solo.
Roughly how many features did Anthropic extract when scaling SAEs to Claude 3 Sonnet in 2024?
Anthropic's 2024 'Eskalaasioŋ Monosemanticite' liggéey génne ci 34 milioŋ ci man-man ci benn xeetu liggéey.
What did the 'Golden Gate Claude' demonstration show?
Bi gëstukat yi yokke màndarga Golden Gate Bridge, dañu def model bi mu nekk ci kaw pont bi, di wane ni màndarga yi dañuy nekk levier causal, duñu etiket kese.
Lan moo waral layer bu nëbbu ci SAE di gëna yaatu ak activation bi muy defaraat?
Modèle yi dañuy boole konsept yu bari ci dimension yu gàtt (superposition); SAE bu mat te yaatu dafay jox konsept bu nekk palaas ngir jël boppam.