Kwenzekeni
Ukuphrinta kuqala okusha kwe-arXiv kuthuthukisa ithiyori yezibalo yokujwayelekile, ukufaka ngokweqile kanye nekhanda kumamodeli okusabalalisa aqeqeshelwe ukulungiselelwa okuphezulu, okunepharamitha ngaphezulu. Ababhali bathola ama-trajectories engcuphe ngqo ngaphansi kokuqeqeshwa kokugeleza kwe-gradient futhi bakhombe izilinganiso ezintathu ezihlukene ngokwekhwalithi ezingavela ngesikhathi sokuqeqeshwa.
Iphepha, elithunyelwe ku-arXiv ngo-Aug. 25, licwaninga amamodeli akhiqizayo asuselwa kumaphuzu, okuhlanganisa amamodeli okusabalalisa asetshenziselwa ukuhlanganiswa kobukhulu obuphezulu. Isiqalo sayo siwukushuba kwesimo ekufanisweni kwe-denoising score: uma izinkinga zokuhlehla ezisetshenziswe phakathi nokuqeqeshwa bezixazululwa ngokuqondile ngedatha enomkhawulo, inqubo yokukhiqiza ewumphumela izogcina ikhiqize kabusha amasampula okuqeqesha. Ngakho-ke ababhali bagxila ekumisweni okujwayelekile-okusobala noma okusobala-okuvumela izinhlelo ezinjalo ukuthi zikhiqize amasampula angaphezu kwamakhophi angokoqobo edatha yabo yokuqeqeshwa. Lokho kusetha kuvumela ababhali ukuthi balandelele ukuthi ukuziphatha okufanele kushintsha kanjani njengoba kuqhubeka ukuqeqeshwa, kuyilapho begcina umbuzo ugxile emthonjeni wokwenza okuvamile.
Ukuhlaziywa kwenziwa ngombuso we-proportional high-dimensional lapho inani lamasampuli nobukhulu bedatha bukhula ngamanani aqhathanisekayo. Imodeli yababhali yokufanisa amaphuzu endaweni ye-kernel enenani le-vector ekhiqiza kabusha i-Hilbert ene-kernel yomkhiqizo wangaphakathi futhi ithola imikhondo yobungozi yokuqeqeshwa kokugeleza kwe-gradient. Lokhu ukwakhiwa kwetiyori, hhayi umbiko wemodeli entsha yezohwebo, umphumela wokulinganiswa noma ukuthunyelwa. Ukukala okulinganayo kuyingxenye yelensi yokuhlaziya yephepha, futhi ukumelwa kwekernel kunikeza isilungiselelo lapho lawo ma-trajectories angabalwa khona.
Leli phepha lichaza izigaba ezintathu ezibuswa izilinganiso ezahlukene. I-spectral estimator ijwayelekile; amaphuzu anomsindo omsulwa akhiqiza iziqongo zasendaweni ezihlanganisa inhloso yokuqeqeshwa; futhi i-empirical Bayes estimator ibamba ngekhanda idatha. Ababhali babe sebehlaziya ukuthi lezo zilinganiso zihlangana kanjani ne-reverse-time stochastic differential equation esetshenziselwa ukukhiqiza amasampula. Bathi isithombe esiba umphumela sihlanganisa izindlela ezijwayelekile zokufunda okugadiwe, ezifana ne-kernel linearization kanye nokuzenzela okujwayelekile, kuyilapho kukhombisa indlela yokuziphatha eqondene ngqo nokumodela okuzikhiqizayo. Ngakho-ke umehluko umayelana nesilinganiso esilawula isikolo esifundiwe, hhayi isimangalo sokuthi wonke amasistimu asetshenzisiwe adlula ngokulandelana okufanayo.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umsebenzi unikeza uhlaka lokuqonda ukuthi kungani amamodeli okusabalalisa ekwazi ukukhiqiza izibonelo zokuqeqeshwa ngokomgomo kodwa ngokuvamile akhiqize amasampula anoveli ekusebenzeni. Iphinde ihlukanise ukuziphatha okuningana okuvame ukuhlanganiswa ndawonye njengokuthi "ukugcwalisa ngokweqile," okungasiza abacwaningi bacabange ngokunembe kakhulu ngekhwalithi yamamodeli, ukuzinza kokuqeqeshwa kanye nezingozi zokukhumbula ngekhanda.
Inzuzo engokoqobo yomsebenzi isilulumagama sawo esibukhali sokuxoxa ngezindlela zokuhluleka zemodeli yokusabalalisa. Imodeli ingafanelana nenhloso yayo yokuqeqeshwa ngaphandle kokuziphatha njengetafula lokubheka elilula, futhi ingaqukatha izingxenye ezibanjwe ngekhanda kuyilapho ifunda isakhiwo esibanzi. Ukwehlukanisa ukwenziwa okuvamile, ukuhumusha okunomsindo nokubamba ngekhanda abacwaningi bagweme ukuphatha lonke iphutha eliphansi lokuqeqeshwa noma okukhiphayo okuphindaphindiwe njengento efanayo. Lowo mehluko ubalulekile ngoba imiphumela efanayo ingabonisa izindlela ezihlukene eziyisisekelo, ezinemithelela ehlukene yokuhumusha.
Uhlaka lungazisa umsebenzi wesikhathi esizayo mayelana nokuxilongwa kokuqeqeshwa. Uma ukwakheka kwesigaba sephepha kusinda kuzilungiselelo ezingokoqobo, abacwaningi bangase basebenzise ubudlelwano phakathi kokuguquguquka kokuqeqeshwa kanye nokuziphatha kwesampula okwenziwe ukuze baphenye lapho imodeli ifunda isakhiwo esidlulisekayo, lapho iwumsindo ofanele, nalapho ikhiqiza izibonelo. Lokho kungase kubaluleke ekuhlolweni kwemodeli, ukukhethwa kwedathasethi nezinqumo mayelana nokuthi kungakanani ukuqeqeshwa okufanele idatha ebucayi. Ukusetshenziswa okuhlongozwayo kuwukuhlola nokuqhathanisa, kunenqubo yokuqapha eseyenziwe kakade yezinhlelo zokukhiqiza.
Okutholakele kuhlobene ikakhulukazi ezingxoxweni zomphakathi mayelana nokuthi amasistimu okukhiqiza ayayibamba ngekhanda idatha yawo. Umthombo awusho ukuthi amasistimu asetshenzisiwe avamise ukudalula izibonelo eziyimfihlo, futhi awunikezi izilinganiso zokuvuza kobumfihlo, ukukhiqizwa okuyimpinda noma ukulimala komhlaba wangempela. Igalelo layo esikhundleni salokho kuyitiyori yokuhlaziya izimo lapho ukukhumbula kungavela ngaphansi kwemodeli eyenziwe lula yokuqeqeshwa nokuthatha amasampula. Noma iyiphi inqubomgomo noma isiphetho somkhiqizo singadinga ubufakazi obunamandla ngale kwalokhu kunyatheliswa ngaphambili. Ngakho-ke umngcele phakathi kwethiyori kanye naleyo mibuzo yomhlaba wangempela uyingxenye yokubaluleka kwephepha.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Umbuzo obalulekile ukuthi le thiyori ifinyelela kude kangakanani ngaphandle kwesilungiselelo sephepha. Umthombo awubiki ukuhlolwa kumasistimu ezithombe ezisetshenzisiwe, ezilalelwayo noma zevidiyo, futhi awusunguli imingcele yokukhumbula ngekhanda okuyingozi noma ubungozi bobumfihlo obungokoqobo. Umsebenzi wokulandelela uzodinga ukuhlola ukuthi izigaba ezintathu ziyavela yini ekwakhiweni okungokoqobo, amasethi edatha kanye nezinqubo zokuqeqesha.
Ukuhlolwa okubaluleke kakhulu ukuphindaphinda okunamandla. Umthombo awusho ukuthi izigaba zawo ezintathu ziye zabonwa yini kuwo wonke ama-architecture amanje okusabalalisa, amashejuli omsindo, izindlela zokuthuthukisa noma amasethi edatha. Futhi ayibiki ukuhlolwa kwesithombe, umsindo noma ividiyo, naphezu kokuhlonza lezo njengezindawo zohlelo lokusebenza zamamodeli okukhiqiza asuselwa kumaphuzu. Izifundo zokulandelela zizodinga ukuqhathanisa ama-trajectories engozi abikezelwe namajika okuqeqesha angempela namasampuli akhiqiziwe. Ukuqhathanisa okunjalo kuzobonisa ukuthi ukulandelana kwethiyori kuyincazelo ewusizo yokuziphatha okubhekiwe noma kuhlala kuqondile kumodeli ehlaziyiweyo.
Abafundi kufanele babheke ukuthi ithiyori ishintsha kanjani lapho ukuqagela kuxegisiwe. Iphepha ligxile ekuqeqesheni okuvilaphayo, ukubusa okuphezulu, ukwakheka kwe-kernel ekhiqizayo kanye nokuguquguquka kokugeleza kwe-gradient. I-abstract ayiqinisekisi ukuthi amanethiwekhi obubanzi obunomkhawulo we-neural, izinyathelo zokuqeqesha ezinomkhawulo, ukwenziwa ngcono kwe-minibatch, ukukhetha kwezakhiwo noma ukusatshalaliswa kwedatha okuhlukile kuyishintsha kanjani imiphumela. Lokho okungaziwa kukhawulela ukuhumusha okuqondile ezinhlelweni zokukhiqiza. Igebe libalulekile ngoba ukuqagela ngakunye kunciphisa izimo lapho imikhondo etholiwe ingafundwa njengesibikezelo esisebenzayo.
Olunye udaba oluvulekile umngcele phakathi kwethiyori ngekhanda nengcuphe engenzeka. Umthombo ukhomba isilinganisi esibamba idatha ngekhanda, kodwa awunikezi umkhawulo walapho ukukhumbula kutholakala, kukhipheka noma kube yingozi. Umsebenzi wesikhathi esizayo kufanele uhlole ukuthi ingabe izigaba ezibikezelwe ziyahambisana yini nokuphumayo okuyimpinda, ukuvuza kobumfihlo noma okunye ukuziphatha okungalinganiseka, nokuthi ingabe ukwenza njalo noma izilawuli zokuqeqesha zingagudluza isistimu iye ekwenzeni okuvamile ngaphandle kokululaza ukukhiqizwa okuwusizo. Lokhu kuzoxhumanisa izigaba zezibalo zephepha nemiphumela engahlolwa ngaphandle kokwakhiwa kwethiyori.