Chii chaitika
Iyo arXiv preprint inogadzira inogumira-yakafara geometric chimiro chekudzidza yakasarudzika spectral alignment mune yakadzika neural network. Inoyera kudyidzana pakati pemagedhi, huremu-inogadzirwa covariance, kumashure sensitivities, avhareji ekunze zvigadzirwa uye neural feature matrices anoshandisa commutators. Bepa rinoshuma mienzaniso yekuongorora uye zviedzo zvenhamba umo kutakura, kusaenzana uye kukanzura chimiro kurongeka pazvikamu nezvikero.
Iro bepa, rakatumirwa kuarXiv musi waNyamavhuvhu 24, 2026, rinoongorora iyo yemukati geometry yeyakadzika neural network. Chinhu chayo chepakati ndicho commutator: chiyero chesvomhu chekusapindirana pakati pezvimiro zvingasaenderana kana kushanduka pamwechete. Vanyori vanoshandisa iyi pfungwa kuhukama hutatu: magedhi ane covariance, kumashure sensitivities ane covariance, uye avhareji ekunze zvigadzirwa zvine neural feature matrices. Chinangwa chakataurwa ndechekutsanangura maitiro akadzidzwa ejometri akarongwa, kutakurwa kuburikidza nematanho uye akasarudzika anoenderana panguva yekudzidziswa.
Kuzivikanwa kwakasarudzika mupepa kunoparadza senitivity-covariance commutator kuita zvitubu zvina: kutakura kwepasi, kusaenzana pakati pematanho ari padyo, kushanduka kwakajeka kwekunzwa uye kupindirana pakati pemasuwo asina mitsetse uye covariance. Kuora uku kunoitirwa kupatsanura nzira dzinogona kuoneka pamwe chete muzvikamu zvakaunganidzwa. Iro bepa rinotsanangurawo iyo AGOP-NFM commutator seyehumwe-value-yakaremerwa yekutakura yemukati commutator, iyo vanyori vanoti inotsanangura kuti sei kurongeka kunoonekwa padivi rechimiro kusingazviratidzi wega geometry yemukati yakaigadzira.
Iro bepa rinowedzera kusuma akavharirwa masimba epanzvimbo kugadzirisa musanganiswa pakati peakaparadzana covariance subspaces uye inopa fungidziro inosanganisira spectral mapundu, projekita shanduko uye kudzikamisa. Iyo inogadzira inomisikidzwa Lyapunov misimboti inogona kuburitsa kuora kana yakajeka geometric kukanganisa miganhu kana mukati damping fungidziro inobata. Iyo abstract inotaura kuti aya mamiriro haateveri kubva kune kuyerera chete. Mumienzaniso yekuongorora uye zviedzo zvenhamba, vanyori vanorondedzera factorization pakati pe spectral uye activation geometry, kukura kwenguva pfupi uye kudzima pakati penonzero masosi. Mumamiriro akaedzwa enguva-yenguva, vanoti kusafambiswa kwakashata-kusaenzana kwakadzora kudzima pakadzika, hupamhi uye mabhenji maviri ekudzoreredza.
Nei zvichikosha
Basa racho rinodenha fungidziro yekuti kudzidziswa kwakabudirira kana kuderera kwenjodzi yekufanotaura kunoburitsa ratidziro yakapfava, yemukati. Kana iyo framework ikabata kupfuura yakaedzwa marongero, inogona kupa vaongorori nzira yakanyatsojeka yekuongorora kuti neural features inorongwa uye kutakurwa sei panguva yekudzidziswa, ukuwo yambiro pamusoro pekurapa kwakacherechedzwa kurongeka sehumbowo hweimwe nzira yemukati.
Mupiro unoshanda ndiwo hurongwa hwekubvunza mubvunzo wakanyanya kupfuura kuti network iri kudzidza: ndezvipi zvimiro zvemukati zviri kuenderana, kupi, uye kuburikidza neicho nzira? Musiyano iwoyo une basa nekuti mamodheru maviri anogona kusvika panjodzi yakafanana uchigadzira akasiyana emukati geometries. Iro bepa rinopokana zvakajeka kuti kudzikisira njodzi hakufanire kureva kudonha kwevatambi, saka chikanganiso chekufungidzira chakaderera hachifanirwe kubatwa sehumbowo hwekuti zvinhu zvemukati zvetiweki zvanyatsoenderana.
Iyo dhizaini inogona kubatsira kune vaongorori vanodzidza optimization, kumiririra kuumbwa uye kududzira. Kupatsanura zvifambiso, padhuze-layer kusaenzana, sensitivity kusiyanisa uye nonlinear gedhi-covariance kupindirana kunogona kubatsira kuona kuti nei kurongeka kunooneka kuchikura, matanda kana kudzoka kumashure. Hurukuro yebepa yemapeji ekuona uye shanduko yeprojekita inonongedzawo kune mamiriro ayo nzvimbo dzepasi dzinogona kuramba dzichisiyanisa kana kudzikama. Aya ndiwo mamethodological mukana anotsanangurwa nekwakabva, kwete kuratidzwa kugona kugadzira kana maturusi akasimbiswa.
Mhedzisiro inoisawo miganho pazvichemo zvakakura nezve neural-network kudzidziswa. Iyo tsime haina kupa mutemo wepasirese iyo yese yakadzama network inoteedzera, uye mhedziso yaro inoenderana nemamiriro ezvinhu: kurongeka kunotsanangurwa seyero- uye chiyero-inotsamira kuenderana chiitiko chinotongwa nekutakura, kudyidzana, kudzima uye kunogoneka kunyorova. Iyo qualification yakakosha pakutsvaga kweAI nekuti zviyero zvemukati zvinogona kuve nehanya nekuvaka, chiyero, nhanho yekudzidzira uye sarudzo yekumiririra. Kucherechedzwa-kudivi rekutarisa kunogona kuve kunodzidzisa pasina kuve yakazara account yetiweki yemukati dynamics.
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
Mubvunzo mukuru wakavhurika ndewekuti dhizaini uye yakashumwa kudzima mhedzisiro inowanda kupfuura mienzaniso yekuongorora yebepa, inogumira-nguva mareji uye maviri ekudzoreredza mabhenji. Kudzokororwa kwakazvimirira, kuenzanisa nemaitiro aripo ekumiririra-yekuongorora, uye zviedzo pamamodheru akakura kana akasiyana-siyana angazodiwa kumisikidza kusvika kunoshanda. Sosi yacho ndeye arXiv vhezheni 1 preprint uye haimisi wongororo yevezera, kuwanikwa kwekodhi, chiyero chebhenji, kana masaizi ekuita.
Chirevo chakasimba chekuyedza ndeye kushingirira kunonzi kukanzura kunodzorwa nekutakurisa kwakashata- kusaenzana kwekudyidzana pakadzika, upamhi uye mabhenji maviri ekudzoreredza. Kwakabva hakupi mazita ebhenji, saizi yedataset, masisitimu emodhiyo, zvigadziriso zvekudzidzira kana saizi yemaitiro enhamba mune zvinyorwa zvinopihwa. Iwo madonhwe ndiwo achaona kuti zvakawanikwa zvingadudzirwa zvakadii uye kuti kupindirana kwakataurwa kwakasimba here kana kuti kwakanangana nehurongwa hwakaedzwa.
Basa rakazvimirira rinofanirwa kuongorora kana hwaro hunoramba huine ruzivo rwekuronga, mamodheru emitauro, mamodheru echiono, kutarisisa-kwakavakirwa zvivakwa uye nzira dzekudzidzisa kunze kwezvirongwa zvinoshandiswa mubepa. Inofanirawo kuenzanisa matanho evafambi nearipo diagnostics yekumiririra kufanana, chimiro chekufambisa uye optimization dynamics. Iyo abstract inotsanangura fungidziro yekufungidzira uye zviedzo asi hairatidze kuti huwandu hwataurwa hunovandudza kufanotaura, kugadzirisa kana dhizaini yemhando muhurongwa hwekushanda.
Nzvimbo yacho inozivisa pepa searXiv: 2608.22910, shanduro 1, yakatumirwa nemunyori mumwe chete akanyorwa muna Aug. 24. Hazvirevi kuti basa rakaitwa nevezera rekuongorora, uyewo peji yakapiwa haigadziri kuwanikwa kwekodhi kana zvinhu zvakakwana zvekuedza. Vaverengi vanofanira naizvozvo kubata mhedziso sezvikumbiro zvekutsvaga zvakamirira kusimbiswa kwakawanda. Iko kukurumidza kusimudzira ndiko kuburitswa kwechimiro uye bvunzo dzayo dzekutanga; kukosha kwayo kunoshanda kuchaenderana nekudzokorora, kushuma kwakajeka kwefungidziro uye humbowo hwekuti zviyero zvinowanda kupfuura zvakashumwa zviedzo zvenguva yekupedzisira.