Uburyo bwo gusobanura
Uburyo bwo gusobanura ni imbaraga zo guhindura-injeniyeri yo kubara imbere yimiyoboro yimitsi muri algorithms yumuntu.
Incamake
Rather than asking 'which input mattered,' it asks 'what is this network actually computing, circuit by circuit?'
Kwibira cyane
Aho uburyo nka SHAP busobanura ibyinjira nibisohoka, gusobanura imashini bifungura agasanduku kandi biga uburemere nibikorwa ubwabo. Abashakashatsi (cyane cyane kuri Anthropic, OpenAI, na academiya) bafata transformateur nka porogaramu igomba kubora, ikamenya 'imizunguruko': ibisobanuro bya neuron hamwe n'abayobozi bitondera gushyira mu bikorwa umurimo wihariye. Ibyagaragaye byingenzi birimo 'induction head,' imitwe yitondera ikoporora uburyo bwo kwiga murwego rwo kwiga, hamwe no kuvumbura ko neuron imwe ikunze kuba 'polysemantic,' irasa kubintu byinshi bidafitanye isano kuko icyitegererezo gipakira ibintu byinshi kuruta ibipimo (superposition). Imodoka zidasanzwe zikoreshwa ubu kugirango zivemo ibintu bisukuye, monosemantike 'ibintu,' nk'icyerekezo gikora ku kiraro cya Zahabu.
Ubushishozi
Inzitizi yibanze ni superpression: umuyoboro ufite ibipimo bya d urashobora kugereranya ibirenze d ibiranga ubibitse nkicyerekezo cya orthogonal, bityo neuron kugiti cye ikarasa kubintu bidafitanye isano. Imodoka zidasanzwe zikemura iki kibazo wiga inkoranyamagambo ituzuye yubaka ibikorwa ukoresheje ibice bike bikora icyarimwe, bikagaragara kubintu bisobanurwa. Abashakashatsi noneho bemeza imirongo hamwe nimpamvu zitera, gukuraho cyangwa 'gutobora' ibikorwa kugirango bemeze ko ikintu gikora ibarwa.
Ingaruka z'Ingamba
Igiciro na bije
Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.
Ibyemezo bisobanutse
Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.
Kugenzura ubuziranenge
Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.
Ejo hazaza ha Mechanistic Interpretability
Ubusobanuro bwibikoresho nibyingenzi mumutekano wa AI: gusobanukirwa imbere bishobora kutwemerera kugenzura imiterere yuburiganya, kumenya ubushobozi buteye akaga, hamwe nimyitwarire yo guhindura ibintu muburyo butaziguye. Ibikorwa byigihe gito byibanda ku gupima gake autoencoders kugera kumipaka yerekana imipaka, kuvumbura imizunguruko, no kubaka 'inkoranyamagambo ziranga.' Intego yifuzwa ni 'MRI kumiyoboro yimitsi', inzira yo gusoma ibitekerezo byikitegererezo mbere yo koherezwa, nubwo gusobanura mu budahemuka sisitemu ya miliyari ibipimo bikomeje kuba ikibazo gikomeye.
Gushyira mu bikorwa Isi
.
Abashakashatsi bagaragaje 'imitwe ya induction' muri transformateur ikopera kandi ikomeza ibimenyetso byerekana inshuro nyinshi, basobanura uburyo bwingenzi bwihishe inyuma yo kwiga.
Gukora patching ikoreshwa mugutandukanya aho icyitegererezo kibika ukuri (urugero, umurwa mukuru wigihugu), kigaragaza ibice byihariye nibice bishinzwe.
Amatsinda yumutekano akora iperereza kumiterere yimbere kugirango amenye niba icyitegererezo cyerekana ibitekerezo nkuburiganya cyangwa amabwiriza adafite umutekano, bigafasha gukurikirana cyangwa gutabara.
Ingaruka & Kurinda
Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.
Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.
Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.
Igishushanyo mbonera
Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.
Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.
Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.
Tegura inzira yo gusubiza ibyabaye mbere yo gupima.
Komeza Ubushakashatsi
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 Mechanistic 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
Ubuyobozi bukurikira
Imashini zidasanzwe zo gusobanura
Ibibazo bikunze kubazwa
What is Mechanistic Interpretability?
Uburyo bwo gusobanura ni imbaraga zo guhindura-injeniyeri yo kubara imbere yimiyoboro yimitsi muri algorithms yumuntu. Aho kubaza 'iyinjiza ryari rifite akamaro,' irabaza iti 'uyu muyoboro urimo kubara iki, umuzenguruko ukoresheje umuzunguruko?'
Niyihe ntego nyamukuru yo gusobanura imashini?
Uburyo bwo gusobanura bugamije gusobanukirwa algorithms nyirizina urusobe rushyira mubikorwa imbere, ntabwo rwinjiza-ibisohoka gusa.
Ni iki 'superposition' yerekeza mu muyoboro w'imitsi?
Superposition ireka umuyoboro uhuza ibitekerezo byinshi kuruta uko ufite neuron / ibipimo ubibika nkibyerekezo hafi ya orthogonal byuzuzanya, bitera neuron polysemantic.
'Imitwe yo kwinjiza' ni iki?
Imitwe ya induction itahura ibyabaye mbere yikimenyetso kigezweho hanyuma ikoporora ibyayikurikiranye, uburyo bwingenzi butuma imyigire yimiterere.
Nigute abashakashatsi bemeza ko umuzunguruko wavumbuwe ukora koko hypothesis?
Uburyo bwimpamvu nko gukora patching cyangwa gukuramo ibizamini niba guhindura ibice mubyukuri bihindura imyitwarire ijyanye, kwemeza uruhare rwayo.
Kuki gusobanura imashini bifatwa nkibyingenzi kumutekano wa AI?
Gusobanukirwa ibintu byimbere hamwe nizunguruka bishobora gutuma igenzura ryuburiganya cyangwa ubushobozi buteye akaga kandi ryemerera gutabara kugamije, gushyigikira kohereza neza.