Chii chaitika
Apple Machine Kudzidza Tsvagiridzo yakatumira bepa, rakanyorwa pasi peChakavanzika uye Dhata Sainzi uye nzvimbo, ichikurudzira kuti muchina usingadzidzisi skip mapoinzi ekudzidzira ayo akayerwa pesvedzero pane zvakabuda zvemodhiyo haana basa, izvo vanyori vanorondedzera kucheka unlearning compute inosvika makumi mashanu muzana mumienzaniso yavo.
Apple Machine Kudzidza Tsvagiridzo yakaburitsa bepa rine musoro unoti "Kana Kusadzidza Kwakasununguka: Kushandisa Yakaderera Kufurira Mapoinzi Kuderedza Mari dzeComputational," yakatenderwa Udi Wieder, Vitaly Feldman, Robert Fisher, uye Anat Kleiman. Peji yekushambadzira inoisa basa pasi penzvimbo mbiri dzekutsvagisa, Yakavanzika uye Data Sainzi uye , uye inoisa iyo bepa. Chidzidzo chacho ndechekusadzidza kwemichina: dambudziko rekubvisa maitiro emuenzaniso wekudzidzisa kubva kumuenzaniso wakatodzidziswa, kazhinji nokuti munhu akumbira kuti data yavo ibviswe kana kuti chisungo chekuvanzika chinochida. Iro peji rinoratidza kuti Kleiman anodyidzana neHarvard uye kuti basa rakaitwa paApple.
Nhaurirano yebepa, sekupfupikiswa muchidimbu pane saiti yeApple, ndeyekuti nzira dziripo dzekusadzidzira dzinobata pese pese mu "kukanganwa set" - muunganidzwa wemienzaniso inobviswa - sezvakakosha zvakaenzana, uye kuti fungidziro iyi inopambadza. Vanyori vanoti vakafananidza pesvedzero mabasa mumutauro uye mabasa ekuona. Kufurira mabasa inzira yakadzikwa yekufungidzira kuti yakawanda sei muenzaniso wekudzidzisa wakashandura maparamendi emuenzaniso kana kufanotaura. Pahwaro hwekuenzanisa ikoko, vanyori vanoshuma vachizivisa zvidimbu zve data rekudzidzisa izvo zvinokanganisa pazvinobuda zvemuenzaniso, uye vanokurudzira chimiro chinoderedza dhatabheti isati yaitwa nzira yekusadzidza.
Nhamba yemusoro ndeye bepa pacharo: "yakakosha computational savings (kusvika 50 muzana) pamienzaniso chaiyo yepasirese." Chirevo ichocho, uye chirevo chekuti zvidiki-zvinokanganisa zvinogona kuzivikanwa nekuvimbika, zvinobva kune vanyori sekupfupikiswa paApple peji. Izvo hazvina kusimbiswa zvakazvimirira pano. Iyo abstract haitauri mamodheru, dhatabhesi, kana yekutanga unlearning algorithms inoshandiswa; haitauri kuti "zvisina basa" maitiro anotsanangurwa sei kana pachikumbaridzo; uye haitsananguri metrics anoshandiswa kutonga kana kusadzidza kwakabudirira. Izwi rekuti "yemahara" riri mumusoro rinoreva computational mutengo, kwete kune chero chirevo pamusoro pekuvanzika garandi.
Ruzivo rwakati wandei rwunongozivikanwa kubva kwarwo. Iro peji rinotakura mwedzi wekushambadza waNyamavhuvhu 2026, nepo inoenderana-yekuverenga yekupinda ine musoro wakafanana pane imwechete peji yaChikunguru 17, 2026; peji haritsananguri misi miviri. Iko hakuna chiratidzo chinooneka chenzvimbo yekuongorora-vezera kana kugamuchirwa kwemusangano, hapana chinongedzo chekodhi yakaburitswa, uremu hwemuenzaniso, kana ongororo yezvigadzirwa pane peji, uye hapana chirevo chekuti Apple inoshandisa nzira iyi mune chero chigadzirwa chekutumira kana pombi yemukati yedata. Iyo abstract zvakare haigadzirise mutengo wekufungidzira wekufungira pesvedzero pakutanga, iyo yemamodheru makuru pachayo mutengo wakakura uye unozoda kuverengerwa maererano nechero mari yakachengetwa.
Kwakabva mashoko: machinelearning.apple.com ↗
Nei zvichikosha
Kukudza zvikumbiro zvekudzima data pamamodeli akadzidziswa kunodhura, uye bepa rinogadziridza kusadzidza kubva pamutengo weunifomu pa-point kupinda mudambudziko retriage - asi maitiro-based triage igakava rinoshanda, kwete vimbiso yakasimbiswa yemutemo kana yekuchengetedza.
Kudzimwa kwedata rave dambudziko reinjiniya rekushanda pane kungori pamutemo. Kana mushandisi akabvisa mvumo kana kushandisa kodzero yekudzima, kubvisa mitsara kubva mudhatabhesi kwakatwasuka; kubvisa izvo zvakadzidziswa modhi yakatorwa kubva mumitsara iyo isiri. Sarudzo inovimbika ndeyekudzidzirazve pasina iyo data, iyo inodhura zvakakwana zvekuti masangano anowanzo kuita batch iyo mu periodic retraining cycles. Inenge nzira dzekusadzidza dzinovimbisa chimwe chinhu nekukurumidza, asi dzakaoma kuratidza. Basa rinodzikisa mutengo wekuvhiya rinotakura zvakananga kuti zvikumbiro zvekudzima zvinogona kukudzwa nguva nenguva uye nekukasira, kana kuti vanomira kusvika pakudzidziswa kunotevera.
Mafambiro epfungwa ari mubepa rino zvine nharo zvakanyanya kukonzeresa pane kugona kwenhamba. Kana muenzaniso chaiwo ukaratidza kusasiya kurongeka kunoyerwa mumaitiro emuenzaniso, ipapo panogona kunge pasina chekugadzirisa kumuenzaniso iwoyo. Izvo zvinoshandura kusadzidza kubva kune yunifomu pa-poindi mutengo kuita triage kurovedza: kuyera kutanga, wobva washandisa compute chete painokwanisa kushandura maitiro emuenzaniso. Kune zvikwata zvinobata zvikumbiro zvekudzima pahuwandu, mutengo unosiyana unoenderana nekuti data rakawanda sei idambudziko rekuronga rakasiyana pane mutengo wakatarwa unoyera kuti zvingani zvikumbiro zvinosvika.
Mupimo wakakosha ndewekuti zvisungo zvekuvanzika zvinowanzonyorwa zvakatenderedza chikumbiro, kwete kutenderedza maitiro akayerwa. Mutongi kana munhu akanganiswa anogona kusagamuchira "muenzaniso wacho haana kuzvidzidza" sehumbowo hwekuti chisungo chekudzima chaitwa, kunyanya nekuti fungidziro yefungidziro ifungidziro ine huchokwadi hunozivikanwa kuti hunosiyana nemuenzaniso, maitiro ekudzidzisa, uye nzira yekufungidzira yakashandiswa. Pepa racho, paumboo hwechinyorwa charo, rinoronga mupiro waro semari yekombuta; haritauri kuti kusvetuka-tsika-pesvedzero mapoinzi kunogutsa chero mutemo wakati, uye ichi chinyorwa hachifanirwe kuverengwa sekutaura kwazvinoita.
Kune zvakare chekuchengetedza dimension iyo abstract mashizha akavhurika. Kana poindi yakasarudzika seyakaderera-pesvedzero uye yasiiwa panzvimbo, mubvunzo unoshanda ndewekuti iyo poindi inoramba ichidzoserwa - kuburikidza nenhengo yekufungidzira, iyo inoedza kana rekodhi chaiyo yaive muchikamu chekudzidziswa, kana kuburikidza nekutorwa zvakananga kwezvinyorwa zvakadzidzwa nemusoro. Kubata nemusoro mumamodheru makuru kunozivikanwa kuve asina kuenzana, nemarekodhi asingawanzo kana ekunze anowanzo kufumurwa kupfuura akajairwa, zvinova ndizvo chaizvo huwandu hwevanhu uko fungidziro yepesvedzero ingangove isina chokwadi. Iyo abstract haitauri kana vanyori vakaongorora chimiro chavo kurwisa kurwiswa kwakadaro, saka kuramba kwavo kwavari hakuzivikanwe kubva kune ino sosi.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
crm_get_transaction(id='4092').In AI training, what is an "epoch"?
Zvekutarisa zvinotevera
Kunyangwe bepa rakazara richitaura mamodheru, dhatabheti, zvikumbaridzo, uye kuvanzika-kurwiswa kuongororwa kuseri kwehuwandu hwemakumi mashanu muzana, ingave mutengo wekomputa pesvedzero pachayo wakaburitswa, uye kana maregulator kana mamwe maLab anobvuma kana kupikisa pfungwa yekuti yakaderera-pesvedzero mapoinzi anogona kusiiwa munzvimbo.
Chinhu chekutanga kutsvaga ipepa rakazara kuseri kweabstract. Mibvunzo yakakosha inosanganisira kuti ndeapi mamodheru nemaseti akashandiswa, nzira dzipi dzekutanga dzekusadzidzira mari dzakayerwa nadzo, ndedzipi pesvedzero-yekuita fungidziro yakasarudzwa, kuti chikumbaridzo che "zvisingatarisike" chakaiswa sei, uye kuti kusvika-kusvika-50-muzana nhamba inoratidza yakajairika nyaya kana nyaya yakanaka pakati pemienzaniso yakaedzwa. Ingave basa rakagamuchirwa panzvimbo inotariswa nevezera, uye kana kodhi kana evaluation zvinyorwa zvinoburitswa, zvichaona kuti vamwe vangakurumidza sei kutarisa mhedzisiro.
Chechipiri, tarisa kuti kukanganwa kunaka kunoyerwa sei. Kubuda-kufanana metrics - inoratidza kuti modhi inoita sekunge haina kumbobvira yaona data - humbowo husina kusimba pane kuongororwa kweanopikisa uchishandisa nhengo-yekufungidzira kana kurwisa kwekubvisa. Zvakare zvakakosha kutariswa ndeye accounting: ingave iyo yakachengetwa yakachengetwa iri mambure emutengo wekombuta pesvedzero fungidziro, uye kuti iyo inodhura zvikero semamodhi uye kukanganwa seti inokura sei. Nzira inochengetedza hafu yekombuta yekusadzidza asi inoda pesvedzero inodhura inogona kutaridzika zvakasiyana pachiyero chemuganhu pane pamienzaniso midiki.
Chechitatu, tarisa regulatory reception. Effect-based triage inokoka mubvunzo wekuti zviremera zvekudzivirira data hazvina kugadzikana mugwara reveruzhinji: kana mutongi achigona kuramba kushandura modhi nechikonzero chekuti rekodhi rekodhi raive diki zvisingayereki. Chero nhungamiro, kuita kwekuita, kana zviyero basa rinogadzirisa zvisungo zvekudzima kwemamodhi akadzidziswa zvinogadzirisa kana nzira iyi ichishandiswa mukuteedzera workflows kana kuti inoramba iri yemukati inoshanda maitiro.
Pakupedzisira, tarisa kugamuchirwa uye kuramba. Kudzokorodza kwakazvimirira pazvivakwa zvakasiyana uye kugovera data kwaizoratidza kuti yakaderera-pesvedzero subsets yakajairika uye ino zivikanwa zvakajairika, kana artifact yemamwe maseti. Kuedza kutyora nzira - kuvaka makesi uko poindi inotaridzika yakaderera-pesvedzero ichiri kutorwa - ingave inodzidzisa zvakafanana. Uye zvinoramba zvisingazivikanwe kana Apple inoda kushandisa izvi mukugadzira; peji rekuburitsa haritauri zvakadaro, uye hapana chinofanira kufungidzirwa.