Kwenzekeni
Iphepha elithi “Width-Independent Compressibility of Deep Neural Networks,” elithunyelwe ku-arXiv ngo-Agasti 22, 2026, lethula ithiyomu mayelana nokucindezela amanethiwekhi ajulile we-neural. Kunethiwekhi yothisha engaguquki, ebanzi ngokwanele enemisebenzi yokwenza kusebenze ukuhlaziya, ababhali bathi kunenethiwekhi encane yokujula okufanayo cishe emele umsebenzi ofanayo.
Ababhali, u-Hong-Yi Wang, uMingze Wang, kanye no-Liu Ziyin, bathi bafakazela i-theoremu yokucindezelwa okufanayo kwemibono ejulile ye-multilayer enemisebenzi yokuhlaziya yokuhlaziya. Ukusethwa kwabo kubheka inethiwekhi yothisha engaguquki, ejulile futhi ebanzi. Umphumela ofunwayo uwukuba khona kwenethiwekhi ewumngcingo enobubanzi obufanayo cishe obumele umsebenzi wokuphumayo wenethiwekhi yasekuqaleni.
Isimangalo esimaphakathi sephepha ukuthi ububanzi obucindezelwe obufinyelelekayo buzimele kububanzi bangempela benethiwekhi yothisha. Esikhundleni salokho, i-abstract inika isibopho se-oda sokuthi O((log(1/epsilon))^d_in), lapho i-epsilon iyiphutha lesilinganiso elivunyelwe futhi u-d_in ingubukhulu bokufaka obusebenzayo. Lokhu kusho ukuthi isibopho sobubanzi esishiwo sibuswa ukunemba okufiswayo kanye nobukhulu bokufakwayo, kunokuba ngokuqondile ukuthi ububanzi benethiwekhi yoqobo bungakanani.
Ukwakhiwa kusebenzisa amasu amabili achazwe emthonjeni. Eyokuqala iyindlela yokuphuma kokunye edizayinelwe ukulandisa ngokufakwa kwe-low-dimensional. Okwesibili ukukala kabusha okuhlakaniphile kohlaka, ababhali abathi kugcina imephu yokufaka-yokukhiphayo. Umthombo wethula lezi njengezingxenye zobufakazi, kodwa i-abstract enikeziwe ayichazi ukwakhiwa okugcwele noma icacise izinto ezingaguquki ezifihliwe nge-oda notation.
Lona umphumela wasetiyetha esikhundleni sokukhululwa komkhiqizo noma isistimu yokuminyanisa ebonisiwe. Irekhodi le-arXiv likhomba umsebenzi njengephepha elikhulu elinamakhasi ayi-11, amakhasi angama-28 esewonke, elinezibalo ezine. Umthombo awusho ukuthi indlela ihloliwe kumanethiwekhi e-convolutional, ama-transformer, amamodeli esisekelo, noma amasistimu e-AI asetshenzisiwe, futhi ayinikezi izilinganiso zokuminyanisa ezingokoqobo noma izinguquko ezilinganiselwe kusikhathi sokusebenza, inkumbulo, noma ukusetshenziswa kwamandla. Isiqinisekiso esishiwo siphathelene nokulinganiselwa komsebenzi omelwe ngaphansi kokuqagela kwephepha; emthonjeni ohlinzekiwe, ayikubeki ukuncipha okungokoqobo kwawo wonke amanethiwekhi othisha.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umphumela unikeza incazelo yethiyori yokuthi kungani amanye amanethiwekhi e-neural aqeqeshiwe angase aqukathe ukuphindaphindeka okususwayo okukhulu. Uma ukwakhiwa kunganwetshwa ngaphezu kokucatshangelwa kwephepha, kungase kwazise imizamo yokunciphisa usayizi wemodeli nokubala ngaphandle kokuphatha ububanzi benethiwekhi yasekuqaleni njengesithiyo esiyinhloko.
Ukucindezelwa kwemodeli kubalulekile ngoba ukwehlisa usayizi wenethiwekhi eqeqeshiwe kungase kwehlise isitoreji, inkumbulo, nezidingo zokuthintwayo. Iphepha likhuluma ngombuzo oyisisekelo osekelwe kulowo mgomo: noma ngabe ukuziphatha komsebenzi wenethiwekhi kudinga ngempela ububanzi obuqathaniswa nobubanzi benethiwekhi obufundile. Ithiyori yayo ithi, ngaphakathi kwesilungiselelo esicacisiwe, impendulo ingaba cha.
Ingxenye yomphumela encike ngobubanzi iyisimangalo esiwumphumela kakhulu. Uma uthisha obanzi angalinganiswa inethiwekhi encane ububanzi bayo obudingekayo buncike kakhulu kubukhulu bokufakwayo nokubekezelela iphutha, khona-ke ububanzi benethiwekhi bungase bube isilinganiso esinolwazi oluncane lobunkimbinkimbi obungaphakathi bomsebenzi kunalokho obuvela kusakhiwo sangempela. Lokho kungase kube nomthelela endleleni abacwaningi abacabanga ngayo mayelana nokungafuneki kanye nokusebenza kahle kokumelela.
Umphumela futhi unomkhawulo owusizo: ufakwe ngokucacile kufreyimu eduze kwemibono ejulile yezendlalelo eziningi ngokusebenza kokuhlaziya kanye nenethiwekhi yothisha engashintshi. Umthombo awuqinisekisi ukuthi isibopho esifanayo sezakhiwo ezilawula i-AI ekhiqizayo yamanje, noma ukuthi inethiwekhi ecindezelweyo ingatholwa kahle ngemodeli eqeqeshiwe. Ngakho-ke ithiyori ikhulisa ukuqonda kwethiyori ngaphandle, ngokwayo, ukukhombisa ipayipi elilungele ukusetshenziswa.
Iphepha lingasiza ekuhlukaniseni imibuzo emibili evame ukuhlanganiswa: ukuthi ingabe inethiwekhi ehlangene ikhona ngokomgomo kanye nokuthi onjiniyela bangakwazi yini ukuyakha eyodwa ngemali eshibhile kuyilapho begcina ukuziphatha okubalulekile. Umthombo usekela isimangalo sokuqala esimweni saso esishiwo. Ayiphenduli okwesibili, futhi abafundi akufanele bahumushe ithiyori njengobufakazi bokuthi amamodeli amakhulu e-AI akhona angancishiswa ngokushesha abe usayizi othile omncane.
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
Imibuzo esheshayo iwukuthi ingabe i-theoem iyasebenza ekwakhiweni kwezakhiwo ezisetshenziswa ezinhlelweni zamanje ze-AI, ukuthi makhulu kangakanani amanethiwekhi angumphumela kuzilungiselelo ezingokoqobo, nokuthi ukwakhiwa kungagcina ukunemba ngaphansi kokucindezelwa okungokoqobo kanye nemingcele yokusebenzisa. Umthombo onikeziwe awukubiki ukuhlolwa, ukuqhathanisa kwebhentshimakhi, imiphumela yokusetshenziswa ekhishiwe, noma ubufakazi bokusebenzisa.
Inkinga yokuqala okufanele ibukwe ububanzi. Umsebenzi owengeziwe uzodinga ukuhlola ukuthi ingabe umphumela uyadlulela kweminye imisebenzi yokusebenzisa, izakhiwo, ubukhulu bokufakwayo, nemisebenzi. Ikakhulukazi, umthombo onikeziwe awubheki amanethiwekhi e-convolutional, amamodeli asuselwe ukunaka, amasistimu avelayo, noma amamodeli ezinto eziningi.
Udaba lwesibili olokwakha kanye nezindleko. Nakuba iphepha lichaza ukwakhiwa okuhambisana nokuphuma kokunye kanye nokukala kabusha okuhlakaniphile kohlaka, umthombo awusho ukuthi kungakanani ukubala, idatha, noma ukufinyelela kumodeli yasekuqaleni okudingekayo ukuze kwakhiwe inethiwekhi encane. Ukuba wusizo okungokoqobo kuzoncika ekutheni inqubo ingenzeka yini ezinhlelweni eziqeqeshwe ngempela kunokuba kuqinisekiswe ngokwezibalo kuphela ukuthi zikhona.
Indaba yesithathu ukukalwa kwekhwalithi. I-theoem isebenzisa ibhajethi yephutha, kodwa umthombo awucacisi ukuthi iphutha lokulinganisa lihlobana kanjani nokunemba komsebenzi, ukuqina, ukulinganisa, ukuziphatha kokuphepha, noma amandla angajwayelekile. Imodeli ecindezelwe ingase ilinganisele umsebenzi ngaphansi kwesilinganiso esisodwa sezibalo kuyilapho ishintsha ukusebenza kokokufaka okubalulekile ekusebenzeni.
Okokugcina, ukukhiqiza kabusha okuzimele kanye nokuqinisekiswa kokuqina kuzocacisa ukubaluleka komphumela. Ubufakazi obuwusizo buzobandakanya ukusetshenziswa, ukuhlolwa kubo bonke ububanzi benethiwekhi nobukhulu bokufakwayo, ukuqhathanisa nezindlela zokuminyanisa ezimisiwe, nezilinganiso zenkumbulo, ukubambezeleka, namandla. Kuze kube yilapho kuvela lobo bufakazi, isiphetho esiqine kakhulu esisekelwe ukuthi iphepha linikeza isiqinisekiso esisha sokucindezelwa kwethiyori ngaphansi kokuqagela okuchaziwe, hhayi ukuthi selivele lenze amasistimu e-AI asetshenzisiwe abe mancane noma ashibhe.