Chii chaitika
Chikwata chekutsvagisa chakurudzira DSSG, yekupedzisira-kusvika-kumagumo sisitimu yeChitubu-Yakazara-Yemahara Domain Adaptation yemamodheru emutauro wechiratidzo. Iyo nzira inoshandisa hova mbiri dzekutungamira-imwe yakavakirwa pamashoko akagadzirwa uye imwe yakavakirwa pamakirasi anchors-kuenzanisa kuchinjika kune imwe dura nekuchengetedza yepasi rose zvinoreva. Iyo yakabatana vhezheni, DSSG-PAC, nguva nenguva inodzokorora prototype anchors uye inoavharira pakati pekudzokorora. Pepa racho rinoti DSSG yakapfuura nzira dzemazuva ano-ye-the-art pane akawanda mabhenji, nepo DSSG-PAC yakachengetedza zvakanyanya kuita ikoko ne18.9% yakaderera yakazara nguva yekuchinja.
Preprint, yakatumirwa kuarXiv musi waNyamavhuvhu 28, 2026, inotarisa pa Source-Fully-Free Domain Adaptation, kana SFF-DA. Bepa rinotsanangura izvi sekugadzirisa mamodheru emutauro wechiratidzo pasina kuwana dhata kana basa-rakanangana nemodhiyo. Chirevo chayo chepakati ndechekuti kuseta uku kunogadzira dambudziko re "dual semantic drift". Static drift inouya kubva kune yakamisikidzwa kirasi embeddings iyo inogona kusaenderana kune itsva domain, nepo dynamic drift ichibva kune akagadzirwa mazwi anogona kutsauka sezvo modhi inochinja. Sekureva kwevanyori, matambudziko aya anowedzera kukakavara pakati pepurasitiki - kugona kudzidza ruzivo rwakanangana nedura - uye kugadzikana - kugona kuchengetedza zvinowirirana zveboka.
DSSG, iyo yakarongwa chimiro, inosanganisa maviri mafomu ekutungamira semantic. Rukova rwemazwi anoitirwa kutora ruzivo rwedomeini, nepo kirasi-anchor rukova rwakaitirwa kuchengetedza pasi rose categorical kuenderana. Iro bepa rinodaidza iyi nzira yakasanganiswa Dual Semantic Guidance, kana DSG. Inosumawo Dynamic Cross-Modal , iyo inoshandisa kugoverwa kwevadzidzisi kunoshanduka kuenzanisa kuenderana pakati pevadzidzisi nemhando dzevadzidzi. Sosi inotsanangura zvikamu izvi sechikamu chekupedzisira-kusvika-kumagumo chigadziriso chimiro, asi haitsanangure magadzirirwo emhando, dhatasethi, masheji ekudzidzisa, kana computational hardware inoshandiswa mukuyedza.
Vanyori vanobva vawedzera DSSG muDSSG-PAC nekudzokorora nguva nenguva prototype anchors uye nekuachengeta kusvika pakuenzanisa kunotevera. Chinangwa chavo chakataurwa ndechekuderedza kudzokororwa kwemavara-parutivi komputa uchichengeta kugona kwekutungamira kwekirasi kuchinjika sezvo nzvimbo yemavara inoshanduka. Iro bepa zvakare rinoti rinomisa SFF-DA miganho yenjodzi inobatanidza njodzi yemudzidzi kune semantic-yemudzidzisi mhando uye mutsauko wemudzidzisi-mudzidzi. Mukuedza kwayo kwakashumwa, vanyori vanoti DSSG inogara ichipfuura nzira dzemazuva ano dzepamusoro pamabhenji akawanda. Vanoenderera mberi vachiti DSSG-PAC yakachengetedza mashandiro ekugadzirisa uku ichidzikisa nguva yakazara yekugadziriswa ne18.9%. Kwakabva hakuratidze mabhenji, nzira dzekuenzanisa, zvibodzwa zvakakwana, kana kusava nechokwadi kwenhamba.
Nei zvichikosha
Mamodheru emutauro wekuona angangoda kushanda munzvimbo umo data ravo rekutanga rekudzidziswa kana kuti basa-rakanangana nemodhiyo risingawanikwe. Iyo yakarongwa nzira inogadzirisa izvo zvinomanikidza nekuyedza kugadzirisa iyo modhi uchishandisa chinangwa-domain ruzivo uchidzikamisa semantic kudonha. Kana ikagadzirwazve yakazvimirira, kuderedzwa kwenguva kwakataurwa kunogona kuita kuti kugadziridzwa kwemahara kusadhure zvakanyanya, kunyangwe kwakabva kusingapi ruzivo rwakakwana rwekuyedza kuongorora ukuru kana kuwanda kwezvinowanikwa.
Dambudziko rinoshanda rinogadziriswa nebepa rakatetepa uye rakanyanya kujeka kupfuura general modhi yakanaka-tuning. Muchigadziro chakatsanangurwa, sangano rinogona kunge riine modhi yemutauro wechiratidzo asi richishaya mvumo, chengetedzo, kana kuwana data uye basa-rakananga modhi rinoshandiswa panzvimbo inotangira. Nzira inogona kuchinjika pasi pezvirambidzo izvi inogona kuve yakakosha uko data rekutanga risingakwanise kutamiswa kana kuchengetwa. Iyo nzira yakarongwa yakagadzirirwa kutenderedza icho chinomanikidza kwete kubata iyo data data serinowanikwa nekusingaperi.
Mupiro wehunyanzvi webepa kuyedza kwayo kubata maviri anokwikwidza zvinodiwa kamwechete. Zvinyorwa zvinogadzirwa zvinogona kupa ruzivo nezve iyo inotarirwa dura, asi vanyori vanopokana kuti kuvimba navo chete kunogona kukonzera semantic kudonha. Yakamisikidzwa kirasi embeddings inogona kuchengetedza yakagadzikana chikamu chimiro, asi vanyori vanopokana kuti ivo vanogona kuomesesa kune inochinja dura. Kubatanidza caption uye class-anchor streams saka zvinoratidzwa senzira yekuwedzera chinangwa-chaicho kuchinjika pasina kusiya yakagadzikana categorical reference. Iyo miganho yenjodzi inoitirwa kugadzirisa kuti kunaka kwesemantic mudzidzisi uye gaka riri pakati pemudzidzisi nemudzidzi rinokanganisa sei kugadzirisa njodzi.
Iyo yakashumwa 18.9% kuderedzwa kweiyo yakazara kugadzirisa nguva ndiyo yakajeka inoshanda mhedzisiro mune kwakabva. Kana mhedzisiro yacho inobata mumhando dzakasiyana siyana, madomasi, uye magadzirirwo ehardware, kuderedza kudzokororwa kwemavara-parutivi komputa kunogona kudzikisa mutengo wekugadzirisa masisitimu emutauro wechiratidzo. Zvisineyi, kwakabva hakutauri kana nguva yekuchengeta ichiratidzira nguva yewachi, kushandisa komputa, kana chimwe chiyero, uye haitauri nheyo yakaverengerwa kudzikiswa. Iyo yainzi mukana wekuita ingoriwo mushumo wevanyori mune preprint abstract, saka inofanirwa kubatwa semhedzisiro yekutsvagisa yakamirira kubereka kwete kugona kwakagadzwa indasitiri.
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 yakakosha inotevera ndeipi mabenchmarks uye nheyo dzakashandiswa, yakawanda sei yakashandurwa pasi peDSSG-PAC, ndeapi hardware uye zvigadziriso zvekugadzirisa zvakaburitsa 18.9% nguva yekudzikisa, uye kana nzira yacho inoshanda kupfuura iyo yakaongororwa mabasa. Iro bepa ndeye arXiv preprint, uye sosi yacho haipe yakazvimiririra kusimbiswa, humbowo hwekutumira, huwandu hwekuita hwakakwana, kana ruzivo nezve kodhi yakabatana. Mikaha iyoyo inoganhurira zvingagumiswa pamusoro pokugadzirira kunoshanda.
Chekutanga chechokwadi chekutarisa izere rekodhi rekodhi. Tsime rinoti nzira iyi yakaedzwa pamabhenji akawanda, asi haazvitumidze mazita kana kutsanangura madomasi, dhatabhesi, evaluation metrics, nzira dzekutanga, kana magadzirirwo emhando. Iwo mameseji anodiwa kuti uone kana zvakashumwa zviwanikwa zvakakura kana kuti zvakanangana mune mamwe mabasa. Mutsara wekuti "izvino mamiriro-ye-iyo-maitiro maitiro" zvakare ndeye chirevo nebepa, kwete yakazvimiririra yakamisikidzwa kuenzanisa mune yakapihwa sosi.
DSSG-PAC inobvumidza kuongororwa kwakapatsanurwa nekuti iko kushanda kwayo kunosanganisira tradeoff iyo abstract inotsanangura chete zvemhando. Vanyori vanoti nzira iyi inochengetedza dhizaini yekuchinjisa panguva yekucheka nguva yakazara yekuchinja ne18.9%, asi ivo havapi hunoenderana kana nhamba dzengozi. Vaongorori nevashandisi vangazoda kuziva kuti kangani ma prototype anchors anodzokororwa, kuti yakawanda sei caching inokanganisa kushandiswa kwendangariro, uye kana yakaderera computation inoshandura mashandiro pane yakaoma kana kukurumidza kuchinja yakananga madomasi.
Kodhi yebepa inotsanangurwa seinowanikwa kuburikidza nearXiv-yakabatana URL, asi iyo yakapihwa hairatidze iyo repository kana kumisa kuzara kwayo, rezinesi, kuberekazve, kana mamiriro ekugadzirisa. Rimwe basa rinofanirwa kuyedzawo kana miganho yenjodzi inofanotaura maitiro akacherechedzwa uye kuti nzira yacho inoramba ichishanda here kana macaptions akagadzirwa ari emhando yakaderera kana mapoka ekirasi asina kujeka. Kusvika iyo mibvunzo yapindurwa, preprint inotsigira kushuma hurongwa hutsva hunogadziriswa uye vanyori vayo zvakayerwa zvichemo, asi kwete mhedziso nezve kugadzirira kugadzira kana hukuru hwepasirese.