Chii chaitika
Iro bepa "Width-Yakazvimirira Compressibility yeDeep Neural Networks," rakatumirwa kuarXiv musi waNyamavhuvhu 22, 2026, rinopa dzidziso yekumanikidza yakadzika neural network. Kune yakagadziriswa, yakafara zvakakwana network yevadzidzisi ine analytic activation mabasa, vanyori vanoti kune yakamanikana network yehudzamu hwakafanana iyo inenge inomiririra basa rimwechete.
Vanyori, Hong-Yi Wang, Mingze Wang, naLiu Ziyin, vanotaura kuti vanoratidza yunifomu yekumisikidza theorem yezvakadzama multilayer perceptrons ine analytic activation mabasa. Kuseta kwavo kunofunga yakagadziriswa, yakadzika uye yakafara mudzidzisi network. Mhedzisiro yavanoti ndeye kuvapo kwetiweki nhete ine hudzamu hwakafanana hunenge hunomiririra yekutanga network yekupinza-kubuda basa.
Chirevo chepakati chebepa ndechekuti iyo inosvikika yakamanikidzwa hupamhi yakazvimirira pahupamhi hwemudzidzisi network. Pane kudaro, iyo abstract inopa kurongeka kweO((log(1/epsilon))^d_in), uko epsilon ndiyo inotenderwa fungidziro kukanganisa uye d_in ndiyo inobudirira yekuisa dimension. Izvi zvinoreva kuti hupamhi hwakataurwa hunodzorwa nehuroyi hunodiwa uye dimension yekupinda, kwete zvakananga nekufara kwakaita network yepakutanga.
Kuvaka kunoshandisa maitiro maviri anotsanangurwa mune sosi. Yekutanga iderivative-matching nzira yakagadzirirwa kuzvidavirira kune yakaderera-dimensional yekupinda. Yechipiri ndeye-law-wise reweighting, iyo vanyori vanoti inochengetedza yekuisa-yekubuda mepu. Kwakabva kunopa izvi sezvikamu zvehumbowo, asi iyo yakapihwa abstract haitsananguri kuvakwa kuzere kana kutsanangura zvinoramba zvakavanzwa neodha notation.
Uyu ndiwo mhedzisiro yedzidziso kwete kuburitswa kwechigadzirwa kana yakaratidza compression system. Iyo arXiv rekodhi inozivisa basa sebepa guru remapeji gumi nerimwe, mapeji makumi maviri nemasere pamwe chete, aine manhamba mana. Kwainobva hakutauri kuti nzira yacho yakaedzwa pa convolutional network, transformers, foundation model, kana deployed AI systems, uye haipe anoshanda compression ratios kana kuyerwa shanduko munguva yekumhanya, memory, kana simba rekushandisa. Iyo vimbiso yakataurwa ine chekuita nekufungidzira kweiyo inomiririrwa basa pasi pekufungidzira kwebepa; hazviiti, mune yakapihwa sosi, kudzikisira kunoshanda kwese kune network yega yega yevadzidzisi.
Nei zvichikosha
Mhedzisiro yacho inopa tsananguro ye theoretical yekuti sei mamwe akadzidziswa neural network angave ane yakakura inobvisika redundancy. Kana kuvakwa kuchigona kukwidziridzwa kupfuura fungidziro yebepa, inogona kuzivisa kuedza kudzikisa modhi saizi uye computation pasina kubata iyo yekutanga network hupamhi sechinhu chikuru chinomanikidza.
Kutsikirira kwemuenzaniso kwakakosha nekuti kudzikisa saizi yetiweki yakadzidziswa kunogona kudzikisa kuchengetedza, ndangariro, uye inference zvinodiwa. Iro bepa rinotarisa mubvunzo wekutanga uri pasi pechinangwa ichocho: kana mashandiro etiweki achitoda hupamhi hunofananidzwa nehupamhi hwetiweki hwakadzidza. Theorem yayo inotaura kuti, mukati meiyo yakatarwa, mhinduro inogona kuva kwete.
Iyo yakafara-yakazvimiririra chikamu chemhedzisiro ndiyo inonyanya kukosha. Kana mudzidzisi akakura achigona kuenzaniswa netiweki diki ine hupamhi hunodiwa hunoenderana zvakanyanya nechiyero chekuisa uye kukanganisa kushivirira, ipapo upamhi hwetiweki hunogona kunge husina ruzivo rune ruzivo rwekuoma kwemukati kwebasa pane zvarinoratidzika kubva kune yekutanga dhizaini. Izvi zvinogona kukanganisa kuti vaongorori vanofunga sei nezve redundancy uye kumiririra kunyatsoshanda.
Mhedzisiro yacho zvakare ine chidziviso chinobatsira: yakanyatso dhizainiwa yakatenderedza yakadzika multilayer perceptrons ine analytic activation uye yakagadziriswa mudzidzisi network. Kwacho hakuratidze kuti chisungo chimwechete chezvivakwa zvinotonga zvazvino zvinogadzira AI, kana kuti network yakamanikidzwa inogona kuwanikwa nemazvo kune yakasarudzika yakadzidziswa modhi. Iyo theorem saka inowedzera kunzwisisa kwedzidziso pasina, pachayo, kuratidza yakagadzirira-kushandisa-kumanikidza pombi.
Iro bepa rinogona kubatsira kupatsanura mibvunzo miviri inowanzo sanganiswa: ingave compact network iripo musimboti uye kana mainjiniya achigona kuvaka imwe zvakachipa achichengeta maitiro akakosha. Kwakabva kunotsigira chirevo chekutanga mumagadzirirwo ayo akataurwa. Iyo haipindure yechipiri, uye vaverengi havafanire kududzira theorem sehumbowo hwekuti hombe dzeAI modhi dzinogona kuderedzwa nekukurumidza kune imwe saizi diki.
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 iripo ndeyekuti iyo theorem inoshanda kune zvivakwa zvinoshandiswa mune zvazvino ma AI masisitimu, akakura zvakadii ma network ari mumamiriro echokwadi, uye kuti kuvaka kunogona kuchengetedza kurongeka pasi pekumanikidza kunoshanda uye zvipingaidzo zvekuendesa. Nzvimbo yakapihwa haitauri zviedzo, kuenzanisa kwekuenzanisa, zvakaburitswa zvekushandisa, kana humbowo hwekutumira.
Nyaya yekutanga kutarisa ndeye scope. Rimwe basa raizoda kuyedza kuti mhedzisiro yacho inosvika kune mamwe mabasa ekuita, zvivakwa, zviyero zvekuisa, uye mabasa. Kunyanya, iyo yakapihwa sosi haigadzirise convolutional network, yekutarisisa-yakavakirwa modhi, inodzokororwa masisitimu, kana mamodhi akawanda.
Nyaya yechipiri ndeyekuvaka uye mari. Kunyangwe bepa richitsanangura chivakwa-chinofananidza kuvaka uye dhizaini-huchenjeri reweighting, kwainobva haritauri kuti yakawanda sei computation, data, kana kuwana kune yekutanga modhi inodiwa kuvaka yakatetepa network. Kubatsira kunoenderana nekuti maitiro acho anogoneka kune chaiwo akadzidziswa masisitimu pane kungovimbiswa nemasvomhu kuti aripo.
Nyaya yechitatu ndeyekuyera kwehutano. Iyo theorem inoshandisa bhajeti yekukanganisa, asi sosi haitaure kuti chikanganiso chekufungidzira chinobatana sei nekurongeka kwebasa, kusimba, kuenzanisa, maitiro ekuchengetedza, kana kugona kushomeka. Modhi yakamanikidzwa inogona kuenzanisa basa pasi peyero imwe yemasvomhu uku ichichinja maitiro pane zvinopinza zvine basa.
Pakupedzisira, kubereka kwakazvimirira uye kusimbiswa kwesimba kunojekesa kukosha kwemhedzisiro. Humbowo hunoshanda hungasanganisira mashandisirwo, zviedzo pahupamhi hwetiweki uye zviyero zvekupinza, kuenzanisa nenzira dzakatarwa dzekumanikidza, uye kuyerwa kwendangariro, latency, uye simba. Kusvikira humbowo hwakadaro hwaonekwa, mhedziso yakasimba yakatsigirwa ndeyekuti bepa rinopa vimbiso nyowani yekumanikidza pasi pekufungidzira kwakatsanangurwa, kwete kuti yakatoita kuti maAI masisitimu ave madiki kana kudhura.