Chii chaitika
Vatsvakurudzi Richard Cornelius Suwandi naFeng Yin vakaunza GRAPE, ipfupi yeGradient Refinement uye Progress-Aware Exploitation, muarXiv preprint yakatumirwa Aug. 25. Nzira iyi yakagadzirirwa kudhura, yepamusoro-dimensional dema-bhokisi optimization, apo optimizer inofanira kuvandudza chigumisiro kuburikidza nekuita mivhunzo yakaganhurirwa pane kunyatsoona.
Iyo preprint inotaridza dambudziko mune yakakwira-dimensional dema-bhokisi optimization: iripo yenzvimbo yeBayesian-optimization nzira dzinogona kufarira nzira dzinogona kuvandudza chinangwa asi dzichingoburitsa zvidiki kwazvo. Mukugadzira bepa, izvi zvinogona kutambisa mivhunzo mishoma pane inochengetedza mafambiro. Mibvunzo ine basa nekuti basa riri kugadziriswa rinotsanangurwa sekudhura, zvichireva kuti kuongorora kwega kwega kunogona kutakura mutengo wakakura wekombuta kana unoshanda.
GRAPE inogadzirisa dambudziko mumatanho maviri. Chekutanga, inonatsa gradient yeko kumashure ichishandisa zvinotsanangurwa nevanyori seyakavharwa-fomu yekutora basa. A posterior pano inomiririra kusavimbika kwenzira pamusoro pegwara renzvimbo yekuvandudza. Chechipiri, GRAPE inosarudza gwara rekuvandudza nekuwedzera kuderera kunotarisirwa kudzika pasi. Izvi zvinoitirwa kutsigira nhanho dzisiri nyore kungobatsira chete, asi dzinogona kuunza kufambira mberi kune musoro.
Bepa rinoshuma maitiro maviri edzidziso. Inoti iyo gradient-yekunatsiridza nhanho monotonically inoderedza kusagadzikana kwenzvimbo, uye kuti kufambira mberi-kuziva gwara rinoshanduka richienda kumawere echokwadi sezvo iyo yekumashure inopenya. Izvi zvirevo kubva kuvanyori 'theoretical analysis. Nzvimbo yakapihwa haipe humbowo, fungidziro, kana mamiriro ekuti chirevo che convergence chinoshanda.
Izvo zvakashumwa zviedzo zvinovhara anokwana maviri AI-anoenderana marongero. Mukurwiswa kwevatema-bhokisi, vanyori vanoti GRAPE yakawana avhareji 5.4-nguva yekumhanyisa pamusoro penzira dzekutanga. Mumutauro wakakura modhi yekukurumidza optimization, vanoshuma kuderedzwa kwe3.8-log-unit mukudemba kweavhareji kwekupedzisira zvichienzaniswa neyechipiri-yakanakisa nzira. Iyo abstract hairatidze mamodheru, zvibodzwa zvekurwisa, mabasa ekukurumidza, bhajeti remibvunzo, dhatasethi, Hardware, kana nzira chaidzo dzekukwikwidza dzinoshandiswa mukuenzanisa ikoko.
Nei zvichikosha
Kana zvakashumwa zvakawanikwa zvikabata kupfuura kuedza kwevanyori, GRAPE inogona kuderedza mutengo wekugadzirisa maAI masisitimu uye kuongorora kusasimba kwavo. Bepa rinoshuma avhareji ye5.4-nguva yekumhanyisa pamusoro pezvigadziro zvevatema-bhokisi vavengi kurwiswa uye 3.8-log-uniti kuderedzwa kwekupedzisira avhareji yekuzvidemba pamitauro mikuru yemhando yekukurumidza-optimization mabasa.
Chinhu chinoshanda ndechekushandisa zviwanikwa. Mazhinji AI kusimudzira uye ekuongorora mabasa anoda inodzokororwa miedzo: inogadzirisa inopa shanduko, inobvunza opaque system, inoona mhedzisiro, uye inopa imwe shanduko. Nzira inosvika kune inofananidzwa kana iri nani chinangwa ine ongororo shoma inogona kuderedza nguva uye mutengo weiyo mafambiro ebasa. Kutariswa kwebepa saka kwakabatana zvakananga nemagadzirirwo eAI masisitimu uye kuyedzwa, pane kubata AI sechiitiko chekushandisa.
Mhedzisiro yeadversarial-attack ingangove yakakosha kune zvese chengetedzo kuyedzwa uye yekudzivirira tsvagiridzo. Nekukurumidza kurwiswa kwebhokisi dema kunogona kubatsira vatsvagiri kuwana kukundikana mumasisitimu asingaburitse magiraidhi emukati kana modhi mamiriro. Panguva imwecheteyo, kugona kumwe chete kunogona kuita kuti kuferefeta kuwanikwe kumapato ari kutsvaga hunhu hunoshandisika. Iyo sosi inomisikidza yakashumwa kukurumidza asi hairatidze kuti nzira yacho inoshandura sei-chaiyo-yepasirese njodzi njodzi kana kuti inobudirira kupikisa masisitimu akaiswa.
Mhedzisiro yekukurumidza-yekugadzirisa inonongedza kushandiswa kwakasiyana: kutsvaga otomatiki zvirevo zvinovandudza mashandiro emhando yemutauro pachinangwa chakatsanangurwa. Kuzvidemba kwakaderera kunoratidza kuvandudzwa kuri nani pasi pekumisikidzwa kwebepa, asi chinyorwa hachitsanangure chinangwa chakashandiswa kana kuti mhedzisiro yacho yakavandudza chokwadi, chengetedzo, kuvimbika, kubatsira, kana chimwe chinhu. Kugonesa kupikisa chibodzwa chakatetepa kunogona kuburitsa zviwanikwa izvo zvisingaendere kune zvakakura zvevashandisi zvinodiwa.
Iro basa rinogonawo kubatsira semupiro wekuwedzera nekuti mashandisirwo aro anoratidzwa seanoshanda kune akakwira-dimensional dema-bhokisi mabasa, kwete emitauro chete. Nekudaro, sosi yacho haipe humbowo nezveminda inodarika nzvimbo mbiri dzakataurwa dzekunyorera. Izvo zvakare hazviratidze kuti iyo nzira iri nani pasi rose, kuti mibairo inoenderera pasi yakaenzana compute pane yakaenzana mibvunzo, kana kuti ayo ekuwedzera posterior-kunatsiridza maverengero haadhure maererano nemabasa ari kuvandudzwa.
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 mikuru yakavhurika ine chekuita neruzivo rwekuyedza kuseri kwenhamba idzo, nheyo dzakashandiswa, huwandu hwemhando uye mabasa akaedzwa, uye kuti nzira yacho inoramba ichishanda kunze kwezvirongwa zvakashumwa. Iyo preprint haina kusimbiswa yakazvimiririra mune yakapihwa sosi, uye zvichemo zvaro zvinofanirwa kubatwa semhedzisiro yakataurwa nevanyori pane kusimbiswa kwekuita vimbiso.
Chekutanga chechokwadi chekutarisa ndiyo yakazara yekuyedza protocol. Vaverengi vanofanira kutarisa nhamba nemhando yemabasa, chiyero chedambudziko rega rega, bhajeti remibvunzo, nzira yekumisa, tsananguro yekumhanyisa, uye kuti maavhareji anovanza musiyano mukuru. Mutsara wekuti "avhareji 5.4-nguva kumhanya" unoreva chete padivi pezviya, nekuti kumhanya kunogona kuyerwa mumibvunzo, nguva yapfuura, kana humwe huwandu.
Kuenzanisa kwacho kunokosha zvakare. Kwakabva kunoti GRAPE yakapfuura yekutanga uye yechipiri-yakanakisa nzira, asi haina kudoma nzira idzo mune abstract. Ongororo yekutevera inofanira kutarisa kuti nheyo dzakanyatsorongedzwa here, kana nzira dzese dzakawana zviwanikwa zvemakomputa zvakaenzana, uye kuti mhedzisiro yacho yakasimba here pambeu dzisina kurongeka uye matambudziko. Kuitwa kwakazvimirira kana kudzokorora kwaizopa humbowo hwakasimba kupfuura preprint chete.
Pakuedza kwemhando yemutauro, zvakakosha zvisingazivikanwe identity identity, nzvimbo yekutsvaga nekukurumidza, sarudzo yebasa, evaluation metric, uye kuendesa kune zvisingaonekwe. Imwe nzira inogona kuderedza kuzvidemba pabhenji uku uchigadzira zvinokatyamadza, zvakakurisa kune seti yebvunzo, kana isina kukodzera kushandiswa kwekuchengetedza. Zvichave zvakakoshawo kudzidza kuti mhedzisiro yakashumwa inoshanda kumhuri dzese dzemhando kana chete kune mamwe masisitimu akadzidzwa.
Zvichemo zvedzidziso zvinofanira kuverengwa pamwe chete nemafungiro avo. Kwakabva hakutauri kuti giredhidhendi yepasi inofanira kutsanangurwa nemazvo sei, kuti nzira yacho inoita sei nezvinangwa zvine ruzha kana zvisingaendereki, kana kuti mashandiro anochinja sei pakukura. Mavhezheni eramangwana ebasa, kodhi kana kuburitswa kwedata, wongororo yevezera, uye bvunzo dzakazvimiririra pakuongorora kweAI uye kuchengetedzeka kwemabasa zvingabatsira kuona kana GRAPE iri nzira inobatsira zvakanyanya kana mhedzisiro inovimbisa inongogumira kune zvakashumwa zviedzo.