Chii chaitika
Vatsvagiri Yicheng Mao naHongru Du vanokurudzira kubata kugoverwa kwedhata rekudzidzisa munzvimbo dzese seyedzidzo yemusanganiswa. Yavo dhizaini inoenzanisira kuti madomasi anokanganisa sei kurasikirwa kwekusimbisa uye anoshandisa nhamba yekuyedza-dhizaini nzira kusarudza kuti ndeipi proxy-kudzidziswa inomhanya kuita.
Iyo preprint, yakatumirwa kuarXiv musi waAug. 24, inotsanangura kusanganiswa kwedata sedambudziko rekugadzira: kana huwandu hwese hwema tokens hwekudzidziswa hwakagadziriswa, vashandi vanofanirwa kusarudza kuti ndeipi mugove unofanirwa kubva kune imwe neimwe domain. Vanyori vanopokana kuti iripo proxy-based workflows yatove nechimiro chekuyedza musanganiswa. Mukududzirwa ikoko, madomasi ndiwo mativi, migove yechiratidzo ndiyo yakaenzana, madiki-modhi yekudzidzira anomhanya mapoinzi ekuedza, uye kurasikirwa kwekusimbisa ndiyo mhinduro yakayerwa. Chikumbiro chepakati chebepa ndechekusarudza iwo mapoinzi ekuedza nemaune pane kubata musanganiswa weproxy semuunganidzwa wemabikirwo evamiriri akasarudzwa kunyanya kufanotaura. Mafuremu anochengeta tarisiro yekuti ungadzidza sei kubva kune iripo yekuedza proxy uku iyo tokeni yese yese inoramba yakagadziriswa. Saka inobata kusarudzwa kwezvidzidzo zvekuedza uye kududzirwa kwekusimbisa-kurasikirwa kwemhinduro, kwete kuchinja kwehuwandu hwemashoko aripo pakudzidziswa.
Vanyori vanogadzira sparse yechipiri-odha Scheffé mhinduro-yepamusoro modhi. Mukutaura kwakajeka, modhi inofungidzira zvese mupiro wega wega wedhata data uye mhedzisiro yekuisanganisa neimwe dura. Iyo abstract inoti iyo ongororo inoona kuti domain kukosha kune hukama hwakasimba: mamwe madomasi anoita seasina kusimba kana akatariswa kuburikidza neawedzero maitiro anove akanaka mune mamwe masanganiswa, kunyanya kana akabatanidzwa newebhu-inotorwa zvinyorwa. Ichi chirevo chekudyidzana mukuwongorora kwechidzidzo, kwete kuwanikwa kwese kwese data sosi inovandudza LLM kana yakasanganiswa newebhu zvinyorwa.
RegMix inoshandiswa seye bepa empirical kesi yekudzidza. Vanyori vanoti iyo sparse statistical modhi inochengetedza musanganiswa masikero pazvikero zvemuenzaniso uye inoramba ichikwikwidza neinoshanduka-shanduka muchina-yekudzidza kufanotaura, uku ichipawo kuparara kwakajeka kwekuwedzera uye kupindirana mhedzisiro. Mukuenzanisa kwakarongedzerwa kucherechedzwa mhinduro dzeproxy-kudzidzisa, yavo modhi-robust I-optimal dhizaini inodzoreredza kurongeka kwakakodzera kwemusanganiswa mushure mekunge 25% yepakutanga proxy run yabviswa. Iyo sosi haipe huwandu hwekumhanya, saizi yemhando, dhataseti, yekusimbisa-kurasikirwa kukosha kana kuenzanisa kweiyo chaiyo compute mutengo mune abstract.
Nei zvichikosha
Iyo nzira inogona kubatsira vaongorori kudzidza kudzidziswa-dhata kuumbwa nedikidiki-modhi yekuedza uku vachiita kuti kudyidzana kwedomasi kuonekwe. Mhedzisiro yakashumwa yekubudirira inobva pakuenzanisa kwakarimwa kuti icherechedzwe mhinduro dzekudzidziswa kweproxy, saka kukosha kwayo kunoenderana nekuti inoendesa kune mamwe ma dataset, modhi uye marongero ekudzidziswa.
Kudzidzira-dhata kuumbwa isarudzo yepakati mukuvaka mhando dzemitauro mikuru, asi kuyedza misanganiswa yakawanda inogona kuda kudzokororwa kudzidziswa kwevamiriri. Iyo yakarongwa sisitimu inogadzirisa maitiro ekuyedza pachayo: inoedza kuona kuti ndeapi masanganiswa anodzidzisa zvakanyanya pamberi pese mutevedzeri wemumiriri aitwa. Kana iyo yakashumwa simulation mhedzisiro inobata mukuita, vaongorori vanogona kushandisa zvishoma zviedzo vachidzidza kuti ndezvipi zvikamu zvinoita zvirinani, kana kushandisa imwechete yekuyedza bhajeti kuongorora dzimwe nzira.
Kusimbisa kwebepa pakudyidzana kwepaviri kunobatsirawo. Domain inoratidzika isingafadzi pakuzviparadzanisa nevamwe inogona kupa kukosha kana yasanganiswa neimwe dura, uye musanganiswa unoratidzika uchivimbisa kubva kune akasiyana domain zvibodzwa unogona kuita zvakasiyana kana zvikamu zvabatanidzwa. Nzira inofumura hukama ihwohwo inogona kuita kuti data-kusanganisa sarudzo zvive nyore kuongorora nekutsanangura pane kufanotaura kunongoburitsa chinzvimbo chekupedzisira. Kwainobva kunopa kududzira uku semukana weiyo sparse Scheffé modhi.
Mhedzisiro yacho inonyanya kukosha semupiro wemitoo, kwete sehumbowo hwekuti musanganiswa wekudzidziswa wave kusimbiswa zvakanyanya. Bepa harizivise mhando yemutauro mutsva, dataset kana chigadzirwa. Inopa nzira yekuronga zviedzo zvakatenderedza bhajeti rechiratidzo uye yekusimbisa-kurasikirwa mhinduro. Kukosha kwaro kunoshanda nekudaro kunoenderana nekunaka kwemaproxy modhi, kumiririrwa kweyakayerwa kurasika kwekusimbisa uye dhigirii rekuti zvimiro zvinoramba zvakatsiga kana kudzidziswa kuchikwira.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibvunzo yakakosha ndeyekuti nzira yacho inonatsiridza sarudzo muhukuru-hukuru kudzidzisa, kuvimbika kwayo masanjiro kunze kweRegMix kesi yekudzidza, uye kuti mari inochengeterwa inoenderera mberi kana proxy ichimhanya yakasiyana zvakanyanya kubva kune yekupedzisira modhi. Kwakabva hakurevi kutumirwa kukuru, kudzokorora kwakazvimiririra kana kudzidziswa-kuchengetwa kwemutengo wakakwana.
Nyaya yekutanga kuona ndeyekusimbisa kwekunze. Iyo abstract inoshuma empirical RegMix nyaya yekudzidza uye simulation yakakwenenzverwa kuti ione mhinduro dzeproxy-yekudzidzisa, asi haitauri kuti dhizaini yakarongwa yakaedzwa mukumhanya kutsva kukuru kwekutanga. Basa remangwana raizoda kuratidza kana kusarudza mashoma maproxy musanganiswa achitungamira kune imwecheteyo sarudzo kana yekupedzisira modhi, tokeni bhajeti, pombi yekugadzira data kana evaluation suite yachinja.
Yechipiri nyaya ndeye chiyero chekuda kuderedzwa kwe25%. Mazwi anoratidza kuti mhedzisiro inobva pakubvisa ingangoita chikamu chimwe muzvina cheiyo proxy yepakutanga inomhanya mukuenzanisa uku uchidzoreredza kurongeka kwakakodzera musanganiswa. Izvo hazvigadzirise kuderedzwa kwepasirese 25% mumutengo wekudzidziswa, nguva yewachi-wachi kana simba. Kuchengetwa kwacho kunogona kusiyana nehuwandu hwemadomasi, chimiro chenzvimbo yekupindura uye huwandu hweruzha muzviyero zvekusimbisa.
Kwakabva zvakare kunosiya zvakakosha zvekushandisa zvisingazivikanwi. Iyo abstract haitauri iyo yakazara yekuyedza dhizaini, iyo domains inosanganisirwa muRegMix, saizi yemaproxy modhi, iyo yakakwana yekusimbisa-kurasikirwa misiyano kana kangani iyo nzira inosarudza yakaderera musanganiswa. Haitsananguri kudzokorora kwakazvimiririra kana nguva yekusagadzikana. Iwo mameseji ndiwo achaona kana iyo dhizaini yakasimba zvakakwana kune yakakwira-mutengo sarudzo kana inonyanya kukosha yekuongorora lens yeongororo yeongororo.