Kwenzekeni
I-arXiv preprint ithuthukisa uhlaka lwejiyomethri olunobubanzi obunomkhawulo ukuze kufundwe ukuqondanisa okukhethiwe kwe-spectral kumanethiwekhi ajulile we-neural. Ikala ukusebenzisana phakathi kwamasango, i-covariance ekhiqizwa isisindo, ukuzwela emuva, imikhiqizo yangaphandle ye- emaphakathi kanye namatrices wesici se-neural usebenzisa abagibeli. Iphepha libika izibonelo zohlaziyo kanye nokuhlolwa kwezinombolo lapho ezokuthutha, ukungalingani nokukhansela kumomotheka ukuqondanisa kuzo zonke izendlalelo nezikali.
Iphepha, elithunyelwe ku-arXiv ngomhla ka-Aug. 24, 2026, lifunda ijometri yangaphakathi yamanethiwekhi ajulile we-neural. Into yayo emaphakathi i-commutator: isilinganiso sezibalo sokungahambisani phakathi kwezakhiwo okungenzeka zingaqondani noma ziguquke ndawonye. Ababhali basebenzisa lo mbono ebudlelwaneni obuthathu: amasango ane-covariance, ukuzwela okubuyela emuva nge-covariance, kanye nemikhiqizo yangaphandle ye- emaphakathi enomatikuletsheni wesici se-neural. Umgomo oshiwo uwukuchaza ukuthi izici ezifundiwe zejiyometri zihlelwa kanjani, zithuthwe ngezendlalelo futhi ziqondaniswe ngokukhetha phakathi nokuqeqeshwa.
Ukuhlonza okune- layerwise ephepheni kuhlukanisa ukuzwela-i-covariance commutator ibe yimithombo emine: ezokuthutha ezansi nomfula, ukungalingani phakathi kwezendlalelo eziseduze, ukushintshashintsha okuqondile kokuzwela nokusebenzisana phakathi kwamasango angaqondile kanye ne-covariance. Lokhu kubola kuhloselwe ukuhlukanisa izindlela ezingase zivele ndawonye ezilinganisweni ezihlanganisiwe. Leli phepha liphinde lichaze umgibeli we-AGOP-NFM njengokuthutha okunesisindo esisodwa senani lomgibeli wangaphakathi, ababhali abathi kuchaza ukuthi kungani ukuqondanisa okubonakalayo ohlangothini lwesici kungakhombisi ngokwako ijiyomethri yangaphakathi eyikhiqizile.
Leli phepha liphinde lethule amandla asendaweni avikelwe ukuze kubhekwane nokuxubana phakathi kwezindawo ezihlukene ze-covariance futhi lethula izilinganiso ezibandakanya izikhala ze-spectral, ukuvela kwe-projektha nokuzinza. Yakha izimiso ezinemibandela ze-Lyapunov ezingaveza ukubola lapho kubekwe obala imingcele yamaphutha ejometri noma kubambe ukuqagela kokudambisa kwangaphakathi. I-abstract ithi lezi zimo azilandeli ukusuka ekugelezeni kwe- kuphela. Ezibonelweni zokuhlaziya nokuhlolwa kwezinombolo, ababhali babika ukwenziwa kwe-factorization phakathi kwe-spectral ne-activation geometry, ukukhula kwesikhashana nokukhansela phakathi kwemithombo ye-nonzero. Kuzilungiselelo zesikhathi esilinganiselwe ezihloliwe, bathi ukusebenzisana okungalungile kwezokuthutha–ukungalingani kulawule ukukhanselwa phakathi kokushona, ububanzi kanye namabhentshimakhi amabili okuhlehla.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umsebenzi ubekela inselele umcabango wokuthi ukuqeqeshwa okuphumelelayo noma ukwehla kwengozi yokubikezela kuveza ukumelwa okulula, okuhambisana ngaphakathi. Uma uhlaka lubambelela ngale kwezilungiselelo ezihloliwe, lunganikeza abacwaningi indlela enembe kakhudlwana yokuxilonga ukuthi izici ze-neural zihlelwa futhi zithuthwe kanjani ngesikhathi sokuqeqeshwa, kuyilapho isexwayiso ngokwelapha ukuqondanisa kwesici esiboniwe njengobufakazi bendlela ethile yangaphakathi.
Igalelo elibonakalayo liwuhlaka lokubuza umbuzo othize kunokuthi ngabe inethiwekhi iyafunda: yiziphi izinhlaka zangaphakathi ezihambisanayo, kuphi, futhi ngayiphi indlela? Lowo mehluko ubalulekile ngoba amamodeli amabili angafinyelela ubungozi obufanayo ngenkathi ethuthukisa amajiyometri angaphakathi ahlukile. Iphepha liphikisa ngokusobala ukuthi ukwehlisa ubungozi akudingekile ukuthi kusikisele ukuwa komgibeli, ngakho-ke iphutha eliphansi lokuqagela akufanele lithathwe njengobufakazi bokuthi izingxenye zangaphakathi zenethiwekhi ziqondaniswe ngokufanayo.
Uhlaka lungaba usizo kubacwaningi abacwaninga ngokusebenza kahle, ukumelwa nokutolika. Ukuhlukanisa ezokuthutha, ukungalingani kwesendlalelo esiseduze, ukuhluka kokuzwela kanye nokusebenzisana okungaqondile kwesango-covariance kungase kusize ukukhomba ukuthi kungani ukuqondanisa okubonakalayo kukhula, izitebele noma kuhlehla. Ingxoxo yephepha yezikhala ze-spectral kanye nokuvela kwe-projektha nakho kukhomba ezimeni lapho izindawo ezingaphansi zingahlala zihlukaniseka noma zizinzile. Lawa amathuba e-methodological achazwe umthombo, hhayi amakhono okukhiqiza abonisiwe noma amathuluzi aqinisekisiwe.
Umphumela uphinde ubeke imikhawulo ezimangalweni ezibanzi mayelana nokuqeqeshwa kwe-neural-network. Umthombo awethuli umthetho wendawo yonke olandelwa yiwo wonke amanethiwekhi ajulile, futhi isiphetho sawo sinemibandela ecacile: ukuqondanisa kuchazwa njengento ehambisanayo encike kungqimba nesikali ebuswa ezokuthutha, ukusebenzisana, ukukhansela kanye nokuncibilika okungenzeka. Lezo ziqu zibalulekile ocwaningweni lwe-AI ngoba izilinganiso zangaphakathi zingazwela ekwakhiweni kwezakhiwo, esikalini, esiteji sokuqeqesha kanye nokukhetha kokumelela. Ngakho-ke ukubonwa kohlangothi lwesici kungase kufundise ngaphandle kokuba i-akhawunti ephelele yokuguquguquka kwangaphakathi kwenethiwekhi.
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 oyinhloko ovulekile owokuthi ingabe uhlaka kanye nemiphumela yalo yokukhansela ebikiwe ivelela ngale kwezibonelo zezibalo zephepha, imithetho yesikhathi esinqunyelwe kanye namabhentshimakhi amabili wokuhlehla. Ukuphindaphinda okuzimele, ukuqhathanisa nezindlela ezikhona zokuhlaziya ukumela, kanye nokuhlolwa kwamamodeli amakhulu noma ahlukahlukene omsebenzi kuyodingeka ukuze kutholakale ukufinyelela okungokoqobo. Umthombo uyinguqulo engu-1 ye-arXiv futhi awusunguli ukubuyekezwa kontanga, ukutholakala kwekhodi, isikali sebhentshimakhi, noma osayizi bomphumela.
Isimangalo esinamandla okufanele sihlolwe ukuphikelela okubikiwe kokukhansela okubuswa ukungahambi kahle kwezokuthutha–ukungalingani phakathi kokujula, ububanzi namabhentshimakhi amabili okuhlehla. Umthombo awunikezi amagama ebhentshimakhi, osayizi bedathasethi, ukulungiselelwa kwamamodeli, izilungiselelo zokuqeqesha noma osayizi bomphumela wezinombolo embhalweni onikeziwe. Leyo mininingwane izonquma ukuthi okutholiwe kungahunyushwa kabanzi nokuthi ingabe ukusebenzelana okubikiwe kunamandla noma kuqondile kuhlelo oluhloliwe.
Umsebenzi ozimele kufanele uhlole ukuthi ingabe uhlaka luhlala lunolwazi lokuhlukanisa, izibonelo zolimi, izibonelo zemibono, izakhiwo ezigxile ekunakekelweni kanye nezinqubo zokuqeqesha ngaphandle kwezilungiselelo ezisetshenziswe ephepheni. Kufanele futhi iqhathanise izinyathelo zabahambisi nokuxilonga okukhona kokufana kokumelela, izici zokuthutha kanye namandla okuthuthukisa. I-abstract ichaza izilinganiso zokuhlaziya nokuhlola kodwa ayiqinisekisi ukuthi amanani ahlongozwayo athuthukisa ukuqagela, ukulungisa iphutha noma ukwakheka kwemodeli kusistimu yokusebenza.
Umthombo ukhomba iphepha njenge-arXiv:2608.22910, inguqulo 1, ehanjiswe umbhali oyedwa osohlwini ngo-Aug. 24. Awusho ukuthi umsebenzi ubuyekezwe ontanga, futhi ikhasi elinikeziwe alikuqinisekisi ukutholakala kwekhodi noma izinto zokuhlola ezigcwele. Ngakho-ke abafundi kufanele baphathe iziphetho njengezimangalo zocwaningo ezilindele ukuqinisekiswa okubanzi. Intuthuko esheshayo ukushicilelwa kohlaka kanye nokuhlolwa kwalo kokuqala; ukubaluleka kwayo okungokoqobo kuzoncika ekuphindaphindeni, ukubika okucacile kokucatshangelwa kanye nobufakazi bokuthi izinyathelo zenzakalela ngale kokuhlolwa kwesikhathi esinqunyiwe okubikiwe.