Chii chaitika
Vatsvagiri vanokurudzira Segment-to-Vhidhiyo Supervision, kana S2V, yemultimodal AI masisitimu anopindura mibvunzo nezve mavhidhiyo marefu. Nzira yacho inogadzira mienzaniso yemibvunzo-nemhinduro kubva muzvikamu zvipfupi, zvemukati zvevhidhiyo, zvino inoshandisa iyo mienzaniso kudzidzisa modhi pamavhidhiyo azere anoenderana. Vanyori vanorondedzera kuvandudzwa kune akawanda-akareba-vhidhiyo-kunzwisisa mabhenji vachishandisa zviuru gumi VQA samples, ine imwechete yekupfuura yekupfuura uye mashoma ekuburitsa tokens pakufungidzira.
Iro bepa, rakatumirwa kuarXiv musi waNyamavhuvhu 21, 2026, rinogadzirisa kunzwisiswa kwevhidhiyo yakareba nemamodhi akawanda emitauro mikuru. Kutanga kwayo ndechekuti mavhidhiyo akareba uye akaomarara ane zvinhu zvinotsausa zvinogona kuvanza ruzivo rwenzvimbo. Maererano nevanyori, izvi zvinogona kutungamirira muenzaniso kuti uise pfungwa pauchapupu husina kunaka uye kuunza mhinduro isina kururama. Tsvagiridzo saka yakanangana neiyo chaiyo AI kugona: kubatanidza mubvunzo nezve vhidhiyo kune yakakodzera nguva kana chikamu mukati meiyo vhidhiyo.
Nzira yakarongwa inonzi Segment-to-Video Supervision, kana S2V. Vatsvakurudzi vanotanga vabudisa mienzaniso yekuona-mhinduro yemibvunzo kubva munzvimbo, zvikamu zvipfupi. Iyo mienzaniso inozodzoserwa kumashure kune yakazara-vhidhiyo kuseta panguva yekudzidziswa. Chirevo chakataurwa ndechekuti zvikamu zvipfupi zvinoita kuti zvinyatso-gwara zvive nyore kuona, ukuwo kudzidziswa pavhidhiyo yakazara kunodzidzisa modhi kubatanidza izvo nemibvunzo kunyangwe kuvepo kwezvinhu zvisingaenderane. Iyo nzira inoshandisa kusimbisa kudzidza nezvinotsanangurwa neabstract semubairo wakapusa-wakavakirwa mubairo uye zviuru gumi zveVQA samples.
Panguva yekufungidzira, iyo inokonzeresa S2V modhi yakagadzirirwa kupindura nekamwe chete kumberi uye nenhamba shoma yematokeni ekubuda. Vanyori vanoti zviedzo zvavo zvinoratidza kuvandudzwa kunoenderana-refu-vhidhiyo-kunzwisisa mabhenji kana ichienzaniswa neakajairika mamodhi emhando dzakawanda uye kufunga-kwakavakirwa nzira. Ivo zvakare vanoti mabhenefiti mukudzidziswa uye inference kunyatsoita. Kwakabva hakuzivise mabenchmarks kana kupa saizi yezvakaziviswa zvakawanikwa mune abstract, saka izvo zvichemo zvinofanirwa kubatwa semibairo yakataurwa nebepa kwete chokwadi chakazvimiririra.
Nei zvichikosha
Mavhidhiyo marefu ane huwandu hwakakura hwezvinhu zvisina basa, zvichiita kuti zviome kuAI masisitimu kuwana humbowo hunodiwa kupindura mubvunzo. Nzira yebepa inonangana nedambudziko irori uchitsvaga kudzivirira mitengo yakakwira yekuzivisa, dhizaini yemubairo yakaoma uye latency yakabatana nedzimwe nzira dzakadzama dzekufunga. Kana yakasimbiswa yakazvimiririra, hunyanzvi hwacho hunogona kuita kuti ongororo yevhidhiyo yakareba inyatsoshanda mumasetimu uko komputa, nguva yekupindura kana zviwanikwa zvekunyora zvinomanikidzwa.
Kuongorora kwenguva refu kwevhidhiyo kwakaoma nechikonzero chakatwasuka: ruzivo rwunodiwa kupindura mubvunzo runogona kungotora chikamu chidiki chechimiro chakakura. Iyo AI sisitimu inobata vhidhiyo yese inofanirwa kusiyanisa humbowo hwakakodzera kubva kumashure chiitiko, inodzokororwa zviono uye zviitiko zvisina hukama. Iyo S2V maitiro anotarisa kudzidziswa pane humbowo hwenzvimbo kutanga, yobva yashandisa vhidhiyo yakazara semagadzirirwo ayo humbowo hunofanirwa kudzoserwa. Iyo imhinduro yakanangwa kune yepakati muganho muvhidhiyo-inokwanisa AI masisitimu.
Chirevo chekushanda chinogona kukosha nekuti kuvandudza kufunga kazhinji kunouya nemutengo wakawedzerwa. Bepa rinoti maitiro ekutanga anogona kuda kusimbaradza-kunatsa-tuning pamusoro, inodhura dhizaini uye yakaoma mibairo dhizaini. Inotiwo mamwe ega ega kana iterative-maonero masisitimu anoburitsa mhinduro refu uye anowedzera inference latency. S2V ine chinangwa chekudzikisa mitoro iyo kuburikidza nediki VQA yekudzidziswa seti, mubairo wakapusa wechokwadi uye imwe-pass mhinduro maitiro. Kana iyo yakashumwa kuita ikadzokororwa, nzira yacho inogona kuve yakakosha kune vanogadzira mavhidhiyo makuru kana kushanda pasi pekupindura-nguva uye compute zvipingaidzo.
Kukosha kunoshanda kunoramba kune zvimiso. Kuita kurinani kwebhenji hakuzogadzirise kunzwisiswa kwechokwadi muvhidhiyo yepasirese, apo mibvunzo ingave isina kujeka, humbowo hwakakodzera hunogona kuparadzirwa nguva dziri kure, uye kukanganisa kunogona kuve nemhedzisiro yakasiyana zvichienderana nekushandisa. Kunobva hakutauri mhedzisiro yekutumirwa, kuongororwa kwevanhu, kuyedzwa-kwakanangana nedomeni kana mitengo yekukundikana. Izvo zvakare hazviratidze kuti iyo nzira inoderedza mutengo wakazara mune yega yega, sezvo kugadzira emunharaunda kudzidziswa mienzaniso uye kugadzirira-yakazara-mavhidhiyo ekuisa kunogona kumanikidza ivo pachavo basa.
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
Iro bepa ndeye arXiv preprint, uye chirevo chayo hachipi chaizvo zvakawanikwa, mazita ebhenji, compute zvinodiwa kana kuenzanisa mune zvenhamba. Kumwe kuongorora kunofanirwa kuongorora kana kuvandudzwa kwakataurwa kunobata kureba kwevhidhiyo, madomasi uye mhando dzemibvunzo, kana kusarudzwa kwechikamu kuchiunza mapofu, uye kuti nzira yacho inoramba ichishanda here kana ichishandiswa nemhando dzakakura kana dzakasiyana dzemultimodal.
Humbowo hunotevera hwakakosha ndehwenhamba uye nzira kubva pabepa rakazara. Vaverengi vanofanirwa kutsvaga kuzivikanwa uye hukuru hwemabhenji emavhidhiyo marefu, mamodheru ekutanga, shanduko chaiyo uye kudzidziswa uye kuyerwa kwekufungidzira. Izvo zvakare zvine basa kana kuenzanisa kunoshandisa yakafanana modhi mhuri, mavhidhiyo ekuisa uye compute bhajeti. Pasina iwo ruzivo, iyo abstract inotsigira kuvepo kweiyo nzira yakatsanangurwa uye vanyori vakashuma gwara remhedzisiro, asi kwete fungidziro chaiyo yemukana wayo.
Segment-based supervision inogona kuunza tradeoff kana zvikamu zvakasarudzwa zvisina zvakakwana. Ongororo inofanirwa kuyedza mibvunzo ine mhinduro dzinoenderana nezviitiko zvisati zvaitika uye mushure mechikamu chenzvimbo, kusangana munzvimbo dziri kure dzevhidhiyo, kana hukama hwenyama husina kujeka. Inofanirawo kuongorora kana chizvarwa chechikamu chichifarira zvinoonekwa zviri pachena asi chisipo odhiyo, nguva, kuzivikanwa kwemutauri kana nyaya yakafararira. Sosi inotsanangura mhinduro yemubvunzo wevhidhiyo asi haitsanangure kuti idzi modali kana nyaya dzakaoma dzakabatwa sei.
Kudzokorora kwakazvimirira kunozoona kana S2V iri nzira yekudzidzisa inoshanda zvakanyanya kana mhedzisiro inosungirirwa kune imwe data uye sarudzo dzemhando. Basa rekutevera rinobatsira raizofananidza nzira yacho nedzimwe nzira dzakasimba dzevhidhiyo-refu, dziedze pane zvisingaonekwi uye kuyera maitiro ekukanganisa pamwe chete nekururama. Iro bepa rakangotumirwa uye harina yakashumwa yekunze yekusimbisa mune yakapihwa sosi. Kusvika iwo macheki awanikwa, mupiro wayo wakasimba unotsigirwa izano rekongiri rine munyori-akashumwa bhenji uye kuvandudzwa kwehunyanzvi, kwete humbowo hwekuti-refu-vhidhiyo AI yakagadzirisa kufunga kwenzvimbo.