Ubuyobozi bwa tekiniki

Imashini zidasanzwe zo gusobanura

Imodoka zidasanzwe (SAEs) nigikoresho gikurura ibikorwa byimbere byimbere byurusobe rwimitsi mumurongo munini cyane wibintu bisukuye, bisobanurwa nabantu.

Incamake

Imodoka zidasanzwe (SAEs) nigikoresho gikurura ibikorwa byimbere byimbere byurusobe rwimitsi mumurongo munini cyane wibintu bisukuye, bisobanurwa nabantu. Nimwe mubuhanga buyobora mugukingura 'agasanduku kirabura' no kureba icyo icyitegererezo cyerekana.

Sparse Autoencoders for Interpretability ni tekinoroji yubaka tekinike igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo.

Kwibira cyane

Imbere ya transformateur, icyerekezo kimwe cyo gukora ivanga hamwe nibihumbi nibihumbi icyarimwe, bigatuma bigorana gusoma. Autoencoder ni gake ni urusobe ruto rwibice bibiri byatojwe kongera kubaka ibyo bikorwa binyuze murwego rwagutse rwihishe, ariko hamwe nigihano gike gihatira bake gusa muri neuron nyinshi zacyo kurasa icyarimwe. Kubera uwo muvuduko, buri gice cyihishe gikunda kuba umuhanga mubitekerezo bimwe, nka 'kuvuga ikiraro cya Zahabu Zahabu' cyangwa 'code ya Python'. Muri 2024 Anthropic yapimye ibi kuri Claude 3 Sonnet, ikuramo ibintu bigera kuri miriyoni 34, kandi OpenAI na DeepMind byasohoye imirimo ibangikanye SAE. Abashakashatsi barashobora noneho gufunga ikintu hejuru cyangwa hepfo kugirango bagerageze neza icyo ikora.

Ubushishozi

SAE ishushanya ibikorwa bya d-bipima muburyo bwagutse cyane (akenshi 8x kugeza 100x binini), hanyuma ikubaka umwimerere. Amahugurwa agabanya ikosa ryo kwiyubaka hiyongereyeho L1 ibihano kubikorwa byihishe, bitera imbaraga nke kuburyo ibice byinshi biguma hafi ya zeru. Ibihinduka nka TopK SAEs byubahiriza muburyo butaziguye mugukomeza ibikorwa bya K binini gusa, kandi amarembo SAE atandukanya icyemezo cyo kurasa nubunini, kugabanya kubogama kuri gahunda L1 itangiza.

Kumenya gake Autoencoders yo gusobanura

Imodoka zidasanzwe (SAEs) nigikoresho gikurura ibikorwa byimbere byimbere byurusobe rwimitsi mumurongo munini cyane wibintu bisukuye, bisobanurwa nabantu. Nimwe mubuhanga buyobora mugukingura 'agasanduku kirabura' no kureba icyo icyitegererezo cyerekana. Sparse Autoencoders for Interpretability ni tekinoroji yubaka tekinike igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo. Kugirango wubake byimbitse, fata Sparse Autoencoders yo gusobanura nkicyitegererezo gikora, ntabwo ari ikintu kimwe: gusobanura ibyagezweho, gusobanura ibitekerezo, no gutandukanya ibyo sisitemu ishobora gukora byizewe nibisaba guca imanza zinzobere.

Mubimenyerezo, amakipe akomeye akoresha Sparse Autoencoders kugirango asobanure atezimbere ubwubatsi, amakuru, hamwe nibikorwa remezo birwanya kwizerwa nigiciro. Bandika ibipimo ngenderwaho byerekana intsinzi, bagerageza kurwanya amakuru afatika hamwe nakazi keza, kandi bagasubiramo bashingiye kubikorwa byagaragaye ko batsinzwe aho gutsinda inshuro imwe. Aha niho imyumvire yubumenyi ihinduka mubushobozi burambye kubicuruzwa, politiki, nibikorwa.

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka. Mugihe kimwe, Guhindura igipimo kimwe gishobora guhisha intege nke za sisitemu. Uburyo bukomeye cyane ni uguhuza umuvuduko wikigereranyo hamwe na disipuline yimiyoborere: kuyobora abaderevu, gufata ibimenyetso, gutangaza ibyemezo byicyemezo, no gukomeza kuvugurura uburyo bwo kwirinda nkimyitwarire yicyitegererezo, ibyo abakoresha bategereje, nibisabwa n'amategeko bigenda bihinduka.

Ingaruka z'Ingamba

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.

Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.

Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Kazoza ka Autoencoders idasanzwe yo gusobanura

Tegereza SAEs kuva mumatsiko yubushakashatsi yerekeza kubigenzuzi bifatika no gukoresha ibikoresho byumutekano, harimo imbaho ​​zerekana ibimenyetso biranga ibimenyetso byerekana uburiganya cyangwa umutekano muke. Ibibazo byafunguye birimo 'gutandukanya ibice' (igitekerezo kimwe kigabanyamo byinshi), kubura ibintu, hamwe nigiciro cyamahugurwa SAEs kuri buri cyiciro cyurugero rwimipaka. Icyerekezo gishya nka crosscoders, transcoders, na matryoshka SAEs bigamije gufata ibarwa mubice kandi kuri granularies icyarimwe.

Gushyira mu bikorwa Isi

__AIU_PROTECTED_4 __ 'Irembo rya Zahabu Claude' demo, aho kongera imbaraga imwe ya SAE byatumye icyitegererezo cyerekana ikiraro mubisubizo byose

Gukuramo no gushyira akamenyetso hafi miriyoni 34 ziranga Claude 3 Sonnet kugirango ushushanye amakarita nka sycophancy, amakosa ya code, nimyitwarire idakwiye.

Gushakisha ibintu bijyanye n'umutekano nko kubeshya, kubogama, cyangwa ibintu bishobora guteza akaga bishobora gukurikiranwa cyangwa kuyobora mugihe cyoherejwe

Gukemura impanvu icyitegererezo cyerekana nabi inyongeramusaruro mugusuzuma ibintu bisobanurwa byakorewe mugihe runaka

Uburyo bwo Gushyira mu bikorwa

Sparse Autoencoders yo gusobanura mubikorwa

__AIU_PROTECTED_4 __ 'Irembo rya Zahabu Claude' demo, aho kongera imbaraga imwe ya SAE byatumye icyitegererezo cyerekana ikiraro mubisubizo byose.

__AIU_PROTECTED_4 __ 'Irembo rya Zahabu Claude' demo, aho kongera imbaraga imwe ya SAE byatumye icyitegererezo cyerekeza ku kiraro muri buri gisubizo Amakipe akunze kubona ibisubizo byiza iyo asobanuye ibipimo ngenderwaho byimbere imbere, akanagaragaza inyungu zumusaruro hamwe nibiciro byigihe.

Sparse Autoencoders yo gusobanura mubikorwa

Gukuramo no gushyira akamenyetso hafi miriyoni 34 ziranga Claude 3 Sonnet kugirango ushushanye ikarita yibitekerezo nka sycophancy, amakosa ya code, nimyitwarire idakwiye.

Gukuramo no gushyira akamenyetso kuri miriyoni 34 ziranga kuva Claude 3 Sonnet kugirango ikarita yerekana ikarita nka sycophancy, amakosa ya code, hamwe nimyitwarire idahwitse Amakipe akunze kubona ibisubizo byiza iyo asobanuye ibipimo ngenderwaho byimbere, agakomeza inzira yo kuzamuka kwabantu kubibazo bikurikirana, kandi agakurikirana inyungu zumusaruro hamwe nibiciro byamakosa mugihe runaka.

Sparse Autoencoders yo gusobanura mubikorwa

Gushakisha ibintu bijyanye n'umutekano nko kubeshya, kubogama, cyangwa ibintu bishobora guteza akaga bishobora gukurikiranwa cyangwa kuyobora mugihe cyoherejwe.

Gushakisha ibintu bijyanye n'umutekano nko kubeshya, kubogama, cyangwa ibintu bishobora guteza akaga bishobora gukurikiranwa cyangwa kuyoborwa mugihe cyoherejwe Amakipe ubusanzwe abona ibisubizo byiza iyo asobanuye ibipimo byujuje ubuziranenge imbere, agakomeza inzira yo kuzamura abantu kubibazo, kandi akurikirana inyungu zibyara umusaruro hamwe nibiciro byamakosa mugihe runaka.

Sparse Autoencoders yo gusobanura mubikorwa

Gukemura impanvu icyitegererezo cyerekana nabi inyongeramusaruro mugusuzuma ibintu bisobanurwa byakorewe mugihe runaka.

Gukemura impanvu icyitegererezo cyerekana nabi inyongeramusaruro mugusuzuma ibintu bisobanurwa bikoreshwa mugihe cyihuse Amakipe akunze kubona ibisubizo byiza mugihe asobanuye ibipimo ngenderwaho byimbere, agakomeza inzira yo kuzamura abantu kubibazo byimbitse, kandi agakurikirana inyungu zibyara umusaruro nibiciro byigihe.

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

1

Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.

Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

2

Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.

Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

3

Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.

Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

4

Tegura inzira yo gusubiza ibyabaye mbere yo gupima.

Tegura inzira yo gusubiza ibyabaye mbere yo gupima. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

Komeza Ubushakashatsi